Model Predictive Control Thesesgood PDF

PH.D.
DISSERTATION

MODEL
-
BASED MULTIVARIATE CONTROL OF
CONDITIONING SYSTEMS FOR OFFICE BUILDINGS

Yuebin Yu
Center for Building Performance and Diagnostics
School of Architecture
Carnegie Mellon University
Pittsburgh, Pennsylvania
August, 2012

Advisory Committee
Dr. Volker Hartkopf (Chair), Professor, Architecture
Vivian Loftness, University Professor, Architecture
John Wiss, Adjunct Professor, ME
Dr. Bruce Krogh, Professor, CEC
Dr. Yan Lu, Siemens

MODEL
-
BASED MULTIVARIATE CONTROL OF
CONDITIONING SYSTEMS FOR OFFICE BUILDINGS

A Thesis Submitted
in Partial Fulfillment of the Requirements
for the Degree of Doctor of Philosophy
in Building Performance and Diagnostics
in the School of Architecture
Carnegie Mellon University
Pittsburgh, PA

by

Yuebin Yu
August 2012
Copyright Yuebin Yu, 2012. All rights reserved.

Copyright Declaration

I hereby declare that I am the sole author of this thesis.
I authorize Carnegie Mellon University, Pittsburgh, Pennsylvania to lend this thesis to
other institutions or individuals for the purpose of scholarly research.
I authorize Carnegie Mellon University, Pittsburgh, Pennsylvania to reproduce this thesis
by photocopying or by other means, in total or in part, at the request of other institutions
or individuals for the purpose of scholarly research.
Copyright Yuebin Yu, 2012. All rights reserved.

Acknowledgements
I would like to take this opportunity to memorize my previous adviser, Dr. David H.
Archer, for his invaluable instructions on my study during my first two years at CMU.
His dedication to work, love to life, and care to students will never be forgotten.
I express my most sincere gratitude to my thesis adviser, Prof. Vivian Loftness. Her
support, guidance and comprehension of the research topic during the entire project made
it possible to structure and finish the thesis.
I am grateful to my committee chair and the director of Center for Building Performance
and Diagnostics, Prof. Volker Hartkopf, for his continuous and energetic support during
my research.
For their comments and help during the thesis process, I would like to extend my
appreciation to the members of my committee: Dr. Yan Lu, Prof. Bruce Krogh, and Prof.
John Wiss. Their help both in terms of research and personal matters made my life easier.
I am very thankful to the project team members and my friends at CBPD and from
Siemens Corporate Research, Bertrand Lasternas, Flore Marion, Dr. Zhen Song, Dr. Kun
Ji and Dr. Dong Wei, for their insightful discussions.
The financial support of Department of Energy for my research is gratefully
acknowledged. The research presented in this thesis is part of the project under award
DE-EE0003843: Advanced, Integrated Control for Building Operations to Achieve 40%
Energy Savings.

Dedication

To
My mom, Xiling Huang, and my wife, Daihong Yu, for
their love, support and understanding.

Abstract
In the U.S., 55% of the energy used in office buildings is consumed by heating,
ventilation, and air-conditioning (HVAC) systems. Although the efficiency of individual
equipment has increased considerably with better design and manufacturing in general,
the energy use has continued to increase because of the greater density of occupants and
increased use of energy consuming devices. One way to reduce the energy consumption
of mechanical conditioning is to improve the overall energy efficiency through integrated
advanced controls. It is undoubted that incorporating a model and utilizing future
information in building operation offers great energy saving potential.
Building systems have unique features that make multivariate model-based predictive
control (MPC) technique as a promising option in the future. Thermal comfort and energy
conservation are the two basic reasons for having a building automation system in
operation in buildings. However, neither of the two objectives can be explicitly included
in conventional single-input-single-output (SISO) control laws. In addition, a building
system is normally a nonlinear multiple-input-multiple-output (MIMO) system which
could have coupling issues among the controlled objects. Conventional control laws, such
as an proportional-integral-derivative (PID) controller, are linear time-invariant SISO-
based. Finally, there are hard and soft constraints present in building systems that cannot
be explicitly handled in PIDs.
A literature review carried out in this research reveals that there are plenty of studies of
model assisted control conducted in the building industry in the past decades. However,
the majority focused on utilizing nonlinear models and applied a simple "simulator plus
optimization solver" approach. The advantage of this approach is that it simplifies the
construction efforts since a building system model can be encapsulated in simulation
tools, and it does not require much understanding on the optimization solver. However,
since the simulator becomes a black-box to the solver, this method is usually
computationally costly and not suitable for dynamic optimization and online

implementation. The approach and benefits of applying linear classical MPC in, which is
much faster and has mature analyzing theory, buildings have not been not fully explored.
This thesis research concentrates on three enabling techniques for utilizing linear
classical MPC in building systems. The MPC is designed for the implementation in a
dual-loop predictive structure to explore the full potential. More specifically, in this
thesis, multivariate control-oriented dynamic models for diversified building energy
systems are developed based on first principles. This includes constructing simulators for
the various components included in building systems. Second, linear MPC controllers for
the mechanical components in the diversified systems are investigated. Third, fast
nonlinear MPC controllers utilizing linear classical MPC as the core for the space thermal
conditioning for both thermal comfort and energy conservation are designed. The
performances of the controllers are evaluated in the Intelligent Workplace north (IWn)
test bed that includes a hydronic system.
The MPCs for complex primary equipment, including an enthalpy recovery unit, a vapor
compression unit, and an active desiccant unit, are designed and evaluated in a
decentralized architecture. It is found that, by applying MPCs, the multiple variables are
coordinated properly toward the objective with minimum efforts. In an intensive swing
season test, the MPC on the enthalpy recovery unit successfully manages the outdoor air
intake and enthalpy recovery wheel across the three different operation conditions: full
outdoor air cooling without recovering the enthalpy, economizer, and minimum outdoor
with enthalpy recovering. The MPC for the vapor compression unit modulates the
compressor speed, the condenser fan, and the supply fan toward the steady state optimal
values with minimum movement when the step disturbances on the set point and outdoor
air temperature are introduced.
Wiener nonlinear MPCs and Hammerstein-Wiener nonlinear MPCs are studied for the
space thermal conditioning. The approach remains the advantages of linear classical MPC
and solves the nonlinearity issues involved in thermal comfort oriented control. With a
linear time-invariant first-principle model for the building and linear MPC as the core, the

controllers achieve the optimal solution for IWn in less than five minutes for five days
simulation in a full look-ahead scenario. The fastest scenario takes only thirty more
seconds to accomplish, which makes the approach practical for online implementation.
The techniques of using MPC for achieving smooth day-night switch, band control,
dynamic constraints, and dynamic weighting are discussed. In addition, the energy saving
potential of applying the proposed MPCs is found to be between six to thirty six percent
compared to a 24/7 temperature oriented control and seven to forty two percent compared
to a temperature oriented conventional control with office schedule. The techniques are
applicable to other space conditioning systems other than a hydronic system.
In addition, this thesis also addresses the under-sensing and modeling issues in hydronic
system control. A "node-branch-state" based approach is proposed to simplify modeling
efforts and facilitate thermal-hydronic coupled simulation. Unlike conventional "node-
loop" approach, the modeling method can handle the discontinuity caused by pump
on/off and valves close/open operation in a large hydronic network simulation. With the
approach, an advanced control with the integration of thermal and hydronic network
becomes possible. Preliminary case studies, including two for water distribution and one
for coupled thermal and hydronic network, are presented.
In conclusion, the work and findings presented in this thesis make important
contributions to the research of model predictive control of air-conditioning systems in
office buildings. The results can be partially utilized or as a whole for other dynamic
system oriented research. Future work, such as automated model generation, model
parameter acquirement, is briefly discussed.

III
TABLE OF CONTENTS

TABLE OF CONTENTS .................................................................................................. III
LIST OF FIGURES ....................................................................................................... VIII
Chapter 1 Introduction ................................................................................................. 15
1.1 Motivation .......................................................................................................... 15
1.2 Background ........................................................................................................ 19
1.3 Objective ............................................................................................................ 21
1.4 Methodology ...................................................................................................... 22
1.5 Outline of the thesis ............................................................................................ 23
Chapter 2 Literature Review........................................................................................ 26
2.1 General survey .................................................................................................... 26
2.2 Model/simulator assisted control ....................................................................... 30
2.3 Model predictive control .................................................................................... 35
2.4 Thermal comfort control .................................................................................... 42
2.5 Summary ............................................................................................................ 45
Chapter 3 Model-based Predictive Control ................................................................. 47
3.1 Introduction ........................................................................................................ 47
3.2 Linear quadratic MPC ........................................................................................ 48
3.2.1 Controller formulation....................................................................... 49
3.2.2 Solving methods ................................................................................ 51
3.3 MPC in broad definition ..................................................................................... 54
3.3.1 Controller formulation....................................................................... 54
3.3.2 Solving methods ................................................................................ 55
3.4 Summary ............................................................................................................ 56

IV
Chapter 4 Model Development .................................................................................... 57
4.1 Introduction ........................................................................................................ 57
4.2 Building thermal system ..................................................................................... 58
4.2.1 Building air and envelop model ........................................................ 58
4.2.2 Thermal comfort model ..................................................................... 64
4.3 Hydronic thermal system ................................................................................... 66
4.4 Hybrid ventilation unit ....................................................................................... 68
4.4.1 Enthalpy recovery unit and desiccant dehumidifier .......................... 70
4.4.1.1 Full dynamic model ................................................................................. 71
4.4.1.2 Simplified fast dynamic model ............................................................... 77
4.4.2 Vapor compression unit .................................................................... 84
4.4.2.1 Compressor.............................................................................................. 87
4.4.2.2 Expansion valve ...................................................................................... 89
4.4.2.3 Full dynamic model ................................................................................. 91
4.4.2.4 Simplified fast dynamic model ............................................................... 96
4.4.2.5 Refrigerant states conversion ................................................................ 105
4.4.3 Model of miscellaneous components .............................................. 106
4.4.3.1 Wheel power consumption .................................................................... 106
4.4.3.2 Fans ....................................................................................................... 107
4.4.3.3 Bypass damper ...................................................................................... 107
4.4.3.4 Gas burner and heat exchanger ............................................................. 108
4.4.3.5 Heat transfer coefficients ...................................................................... 109
4.5 Summary .......................................................................................................... 109
Chapter 5 Control of Hybrid Ventilation Unit ........................................................... 112

V
5.1 Introduction ...................................................................................................... 112
5.2 Base structure and control of the hybrid ventilation unit ................................. 114
5.3 Model linearization and state space realization ................................................ 118
5.4 MPC architecture and performance evaluation ................................................ 120
5.4.1 Enthalpy recovery wheel ................................................................. 125
5.4.2 Vapor compression unit .................................................................. 142
5.4.3 Active desiccant unit ....................................................................... 153
5.5 Conclusion and discussion ............................................................................... 161
Chapter 6 Control of Space Thermal Conditioning ................................................... 164
6.1 System description ........................................................................................... 164
6.2 Base control of the space condition .................................................................. 167
6.2.1 Base control ..................................................................................... 167
6.2.2 Space layout .................................................................................... 167
6.3 MPC architecture and performance evaluation ................................................ 168
6.3.1 Thermal comfort and system model ................................................ 169
6.3.1.1 Thermal comfort evaluation .................................................................. 169
6.3.1.2 Building dynamic model ....................................................................... 171
6.3.2 Hammerstein-Wiener structure based fast nonlinear MPC ............. 172
6.3.2.1 Formulation ........................................................................................... 176
6.3.2.2 Results ................................................................................................... 179
Chapter 7 Control Related Issue on Hydronic System .............................................. 196
7.1 Introduction ...................................................................................................... 196
7.2 Hypotheses ....................................................................................................... 199

VI
7.3 Modeling methodology .................................................................................... 200
7.4 Model and simulating structure ........................................................................ 205
7.5 Evaluation cases ............................................................................................... 207
7.5.1 Case 1- Hydraulic distribution ........................................................ 208
7.5.2 Case 2- Thermal- hydraulic co-simulation ...................................... 212
Chapter 8 Conclusions And Discussion .................................................................... 216
Reference ........................................................................................................................ 221
Publication ...................................................................................................................... 230
Appendix A: .................................................................................................................... 232
Appendix B: .................................................................................................................... 242
Appendix C: .................................................................................................................... 247

VII
LIST OF TABLES
Table 5-1: Design parameters of the desiccant wheels ................................................... 115
Table 5-2: Design efficiency of the wheels .................................................................... 116
Table 5-3: Facts of the hybrid desiccant ventilation unit ................................................ 117
Table 5-4: The typical control modes of the hybrid unit ................................................ 117
Table 5-5: Manipulated variables of the hybrid ventilation unit .................................... 121
Table 5-6: Controlled variables of the hybrid ventilation unit ....................................... 122
Table 5-7: MPC control scenarios for summer and swing weather conditions .............. 133
Table 5-8: MPC control scenarios for the DX unit ......................................................... 148
Table 5-9: MPC control scenarios for the active desiccant unit ..................................... 155
Table 6-1: Control scenarios for classical MPC and HW nonlinear MPC ..................... 178
Table 6-2: Comparison of energy consumption, thermal comfort, and simulation duration
......................................................................................................................................... 192
Table A-1: Construction and thickness ........................................................................... 232
Table A-2: Fenestration parameters ................................................................................ 232
Table A-3: Enthalpy recovery unit efficiency ................................................................ 240

VIII
LIST OF FIGURES

Figure 1-1: Illustration of two-level air-handling unit control system ............................. 16
Figure 1-2: Illustration of the Intelligent Workplace conditioning system ....................... 18
Figure 1-3: Illustration of simulation-assited predictive control ...................................... 20
Figure 1-4: Illustration of dual-loop MPC ........................................................................ 21
Figure 2-1: Number of articles about model-based supervisory control in building
industry ............................................................................................................................. 26
Figure 2-2: Distribution of modeling methods in building industry ................................. 27
Figure 2-3: Target system or components and the modeling method ............................... 28
Figure 2-4: Optimization techniques applied in the reviewed papers ............................... 30
Figure 2-5: Illustration of basic MPC structure for HVAC system control ...................... 37
Figure 2-6: Acceptable range of operative temperature and humidity for thermal comfort
........................................................................................................................................... 42
Figure 3-1: Illustration of model predictive control scheme ............................................ 48
Figure 4-1: Outdoor view of the IW building ................................................................... 58
Figure 4-2: Geometry model of the IWn .......................................................................... 58
Figure 4-3: Heat transfer through a building construction ................................................ 60
Figure 4-4: 3R2C analogue heat transfer model ............................................................... 60
Figure 4-5: Thermal network in a multizone building ...................................................... 61
Figure 4-6: Picture of a cooled beam ................................................................................ 66
Figure 4-7: Picture of water mullions ............................................................................... 66
Figure 4-8: Picture of a radiant ceiling ............................................................................. 67
Figure 4-9: Picture of the hybrid ventilation unit ............................................................. 69
Figure 4-10: Illustration of the hybrid unit ....................................................................... 69
Figure 4-11: Illustration of an enthalpy recovery wheel ................................................... 71

IX
Figure 4-12: Picture of the active desiccant wheel ........................................................... 71
Figure 4-13: Illustration of the rotary wheel and sections ................................................ 76
Figure 4-14: Discretization of the wheel in thickness and angle ...................................... 77
Figure 4-15: Illustration of enthalpy recovery wheel work process ................................. 79
Figure 4-16: Illustration of active desiccant wheel dehumidification process ................. 81
Figure 4-17: Cycle of refrigerant in a vapor compression unit ......................................... 85
Figure 4-18: Illustration of full model .............................................................................. 87
Figure 4-19: Scroll type vapor compressor ....................................................................... 88
Figure 4-20: Thermal expansion valve ............................................................................. 90
Figure 4-21: Electronic expansion valve .......................................................................... 90
Figure 4-22: Picture of the condenser ............................................................................... 93
Figure 4-23: Picture of the evaporator .............................................................................. 95
Figure 4-24: Illustration of fast model dynamics .............................................................. 97
Figure 4-25: EES and Matlab co-simulation .................................................................. 106
Figure 4-26: Bypass damper and mixing chamber ......................................................... 108
Figure 5-1: Illustration of the hybrid ventilation unit with denotes ................................ 115
Figure 5-2: MPC architecture- example 1 ...................................................................... 123
Figure 5-3: MPC architecture- example 2 ...................................................................... 123
Figure 5-4: Trended outdoor air temperature in 2006 .................................................... 124
Figure 5-5: Room air and supply air temperature ........................................................... 125
Figure 5-6: Abstracted enthalpy recovery unit MPC structure ....................................... 125
Figure 5-7: Main S-function for the enthalpy recovery wheel ....................................... 126
Figure 5-8: Subfunction for the enthalpy recovery ......................................................... 127
Figure 5-9: ERV model wrapped as a block ................................................................... 127

X
Figure 5-10: ERV model with PID controllers ............................................................... 128
Figure 5-11: MPC controller for the ERV in Matlab ...................................................... 130
Figure 5-12: Comparison of PIDs and MPC for ERV control- winter 1 hour ................ 131
Figure 5-13: Comparison zoom in- winter 1 hour .......................................................... 131
Figure 5-14: Outdoor air intake and ERV rotation speed of MPC ................................. 132
Figure 5-15: Outdoor air intake and ERV rotation speed in summer, ERV- 1 ............... 133
Figure 5-16: ERV supply air humidity control in summer, ERV- 1 ............................... 134
Figure 5-17: ERV air temperature control in summer, ERV- 1 ...................................... 134
Figure 5-18: ERV air humidity control in summer, ERV- 2 .......................................... 136
Figure 5-19: ERV air temperature control in summer, ERV- 2 ...................................... 136
Figure 5-20: ERV manipulated variables in summer, ERV- 2 ....................................... 137
Figure 5-21: ERV manipulated variables in summer, ERV- 3 ....................................... 138
Figure 5-22: Disturbance conditions, ERV- 4 ................................................................ 139
Figure 5-23: ERV manipulated variables in swing season, ERV- 4 ............................... 140
Figure 5-24: ERV air temperature control in swing season, ERV- 4 ............................. 140
Figure 5-25: ERV air temperature tracking zoom in, ERV- 4 ........................................ 141
Figure 5-26: ERV air humidity control in swing season, ERV- 4 .................................. 141
Figure 5-27: Abstracted DX unit MPC structure ............................................................ 143
Figure 5-28: DX unit model in Simulink ........................................................................ 144
Figure 5-29: Condenser model in Simulink .................................................................... 144
Figure 5-30: Evaporator model in Simulink ................................................................... 145
Figure 5-31: Compressor model in Simulink.................................................................. 145
Figure 5-32: Expansion valve model in Simulink .......................................................... 146
Figure 5-33: Air temperature, DX- 1 .............................................................................. 149

XI
Figure 5-34: Control variables 1, DX- 1 ......................................................................... 150
Figure 5-36: Air temperature conditions, DX- 1 ............................................................ 151
Figure 5-37: Air temperature 1, DX- 2 ........................................................................... 152
Figure 5-40: Active desiccant dehumidification model in Simulink .............................. 154
Figure 5-41: Active desiccant dehumidification model in Simulink .............................. 154
Figure 5-42: Manipulated variables, ADU- 1 ................................................................. 156
Figure 5-43: Air humidity, ADU- 1 ................................................................................ 157
Figure 5-44: Air temperature, ADU- 1 ........................................................................... 157
Figure 5-47: Air temperature, ADU- 2 ........................................................................... 159
Figure 5-49: Air Temperature, ADU- 3 .......................................................................... 160
Figure 6-1: External view of the building system ........................................................... 164
Figure 6-2: Illustration of the integrated building energy system ................................... 166
Figure 6-3: Illustration of the space layout ..................................................................... 168
Figure 6-4: Pure linear classical MPC for building systems ........................................... 172
Figure 6-5: Hammerstein-Wiener nonlinear model ........................................................ 173
Figure 6-6: Control structure for Wiener model plant .................................................... 174
Figure 6-7: Control structure for Hammerstein model plant .......................................... 175

XII
Figure 6-8: IW water based system in Hammerstein-Wiener structure .......................... 176
Figure 6-9: Illustration of the variables in HW nonlinear MPC ..................................... 177
Figure 6-10: IWn model without nonlinearity ................................................................ 179
Figure 6-11: IWn model with nonlinear hydronic system .............................................. 180
Figure 6-12: Outdoor air condition and solar radiation .................................................. 181
Figure 6-13: Simulation results, SC- 1 ........................................................................... 182
Figure 6-14: Simulation results, SC- 1' ........................................................................... 183
Figure 6-19: Simulation results, SC- 6-1 ........................................................................ 188
Figure 6-20: Simulation results, SC- 6-2 ........................................................................ 189
Figure 6-21: Illustration of desired transition from unoccupied condition to occupied
condition in winter .......................................................................................................... 190
Figure 7-1: Liquid flow network with one joint nodes ................................................... 200
Figure 7-2: Illustration of a secondary thermal hydronic network ................................. 202
Figure 7-3: Node-branch-state hydraulic model process flowchart ............................ 206
Figure 7-4: Example integration of the hydraulic network simulator ............................. 207
Figure 7-5: Pump curves for the branches (U) and main pump (D) ............................... 208
Figure 7-6: Water flow rate through each branch (U) and node balance error (D) ........ 209
Figure 7-7: Water flow rate through each branch (U) and node balance error (D) ........ 211
Figure 7-8: Simulink model for thermal-hydraulic co-simulation .................................. 212
Figure 7-9: Simulation results for 7 zones in the space .................................................. 213

XIII
Figure A-1: Mote sensor installed in the north-east bay ................................................. 233
Figure A-2: Underfloor surface temperature measurement ............................................ 234
Figure A-3: Mote sensor array in north-east bay ............................................................ 234
Figure A-4: Mote sensor array in the meeting area ........................................................ 235
Figure A-5: Illustration of space temperature distribution- initial .................................. 235
Figure A-6: Illustration of space temperature distribution- final .................................... 236
Figure A-7: Simulated air temperature vs real temperautre in the plenum ..................... 236
Figure A-8: Simulated air temperature vs real temperautre in the space ........................ 237
Figure A-9: Hybrid ventilation unit control algorithm air temperature set point to
compressor to percentage ................................................................................................ 238
Figure A-10: Hybrid ventilation unit control algorithm compressor output percentage to
VFD................................................................................................................................. 239
Figure A-11: Spreadsheet based manufacturer model of the hybrid ventilation unit ..... 239
Figure A-12: Transfer efficiency vs wheel face velocity ................................................ 240
Figure A-13: Efficiency Comparison .............................................................................. 241
Figure B-1: Thermal properties of R22 .......................................................................... 242
Figure B-2: Temperature vs enthalpy of saturated liquid ............................................... 243
Figure B-3: Temperature vs enthalpy of saturated vapor ............................................... 243
Figure B-4: Temperature vs pressure of two-phase mixture .......................................... 244
Figure B-5: Temperature vs density of saturated liquid ................................................. 244
Figure B-6: Temperature vs density of saturated vapor .................................................. 245
Figure B-7: Temperature vs specific heat of superheated vapor .................................... 245
Figure B-8: Temperature vs humidity of air ................................................................... 246
Figure B-9: Air density vs air density ............................................................................. 246

XIV
Glossary
Abbreviations
ADU Active desiccant unit
ARMAX Auto-regressive moving average with exogenous input
BAS Building automation system
BIM Building information modeling
CARIMA Controlled auto regressive integrated moving average
DDC Direct digital control
DX Direct expansion
ERV Enthalpy recovery wheel
FIR Finite impulse response
HMPC Hammerstein nonlinear model predictive control
HVAC Heating ventilation and air-conditioning
HW Hammerstein-Wiener
IWn Intelligent Workplace north
LTI Linear time invariant
MIMO Multiple input multiple output
MPC Model predictive control
PID Proportional integral and derivative
PMV Predicted mean vote
PPD Predicted percentage of dissatisfied
RC Resistance capacitance
SISO Single input single output
SR Step response
VFD Variable frequency drive
WMPC Wiener nonlinear model predictive control

15
CHAPTER 1 INTRODUCTION
1.1 Motivation 1.2 Background 1.3 Objective 1.4 Methodology 1.5 Outline of the
thesis
1.1 Motivation
Buildings account for about 40% of the total energy consumption in the U.S. (DOE
2008), and most of the energy is used to heat, ventilate, and air-conditioning the space.
Although the efficiency of individual HVAC equipment has been increased considerably
with better design and manufacturing in general, the energy use has continued to increase
because of the greater use of energy consuming devices (CBECS 2003). To improve the
management effectiveness and efficiency, a building automation system (BAS) is
typically deployed in modern buildings to interconnect and automate the various systems
and devices. One of the important goals of applying BAS is to use the control functions
to satisfy occupants comfort requirements with minimal energy use (Mathew 2001).
Considering the scale of energy use by buildings, improving the performance of controls
is thus critical for both energy and indoor environmental sustainability.
Control functions of BAS are typically fulfilled using a two-level control structure: a
supervisory level and a local level. The supervisory-level controller coordinates the
systems and components by specifying the set-points and directing the mode switch of
operation. The local controllers modulate the individual device (e.g. a relay, an actuator, a
VFD, etc.) to track the set-points provided by the supervisory level controller. The two
levels are networked together via direct digital control technology (DDC) in modern
BAS. The overall energy efficiency and control effectiveness of a building system is
jointly determined by the performance and coordination of the two levels control. The
layout of a control system is illustrated below in Fig.1-1 for an air handling unit.

16
schedules & set points
space return air
H
C
C
C
T H
exhaust
air
F
T F H
outdoor
air
supply air
pressure controller
VFD
VFD
return fan
supply fan
heating
coil
cooling
coil
P
actuator
outside air
controller
return fan
controller
supply air
temperature
controller
Local controllers
Supersivory controller
supply
air
weather, occupancy, schedules, etc

Figure 1-1: Illustration of two-level air-handling unit control system
In a conventional control system, the supervisory-level controller comprises a group of
logics and typically resides in a central BAS station. It takes actions based on preset
conditions and/or rules, or the commands from the operators. The basic functions include
heating/cooling switch over, and components sequencing, event scheduling, etc. The
goal of a supervisory level control is to optimize the operations of building systems for
human comfort with minimal energy or monetary consumption. For example, to avoid an
extra electricity demand charge, a load shedding function can be implemented in the
supervisory level control. Alternatively, to use natural cooling sources, a night cooling
algorithm may be considered; and so forth.
Unlike the supervisory-level control, the local-level control focuses only on maintaining
the set-points by manipulating typically a single device. A specific example would be the
regulation of a supply air temperature in an air handling unit where a hot water valve is

17
modulated to track the set-point. Depending on the applications, the local-level control
can be a discrete or continuous, open or closed loop. A closed loop adds a feedback with
output sensing on the open loop to improve the set-point tracking accuracy and eliminate
the offset. So far, the most prevalent feedback control law used for modulating devices in
the building automation industry is proportional-integral-derivative (PID) control (Geng
and Geary 1993). Well-tuned local-loop controllers can enhance the space thermal
comfort, reduce energy use, and/or extend the component life.
Modern building mechanical/energy systems have a wide range of subsystems and
components to service the needs of occupants. HVAC systems and artificial lighting
systems are the two that have a great impact on building energy consumption. A
conventional HVAC system includes primary heating and/or cooling sources (e.g. chiller,
cooling tower, boiler, heat exchanger, etc.), movers (e.g. pump, fan, etc.), and secondary
distribution devices (e.g. air handling unit, terminal box, etc.). In addition to conventional
HVAC systems, passive energy systems, which mainly utilize natural sources, such as
day lighting, natural ventilation, and so forth are also becoming more common in
buildings for energy conservation. Passive energy sources are gradually being integrated
into BAS control to exploit the potentials and avoid conflicts due to improper operations,
which may inversely lead to more energy consumption. For example, opening the
window for the natural ventilation while the mechanical cooling is on can lead to
increased cooling energy use while the intent was to reduce cooling demands through
natural conditioning. The diversified energy systems in a modern building present a
challenge and significant opportunity for control engineers.
Fig. 1-2 illustrates the augmented building system of the Intelligent Workplace North
(IWn). The IWn is conditioned predominantly by multiple water-based sub-systems. The
water mullions are deployed along the external walls to provide heating and cooling to
the perimeter area. Ten additional chilled beams and four sets of radiant ceilings are
installed in the office area for additional thermal control capability. The space is
ventilated by a direct expansion refrigerant based rooftop unit. To ensure that the space
humidity is under control with no condensation on the internal surfaces in summer, the

18
rooftop unit is equipped with an enthalpy recovery unit and an active desiccant
dehumidifier. The conditioned air is evenly distributed into the space via the floor
diffusers. Passive systems, mainly day lighting and natural ventilation, are also installed.
All the west and east facing walls have motor-driven louvers and blinds to control the
indoor illuminance and solar heat gain. Eight windows are equipped with motors to
control the inlet of outside air when the outside air condition is favorable for natural
cooling. The integrated systems and combined heat and power in the IW are of
significance for existing and new office buildings for energy conservation (Betz 2009).
1
campus
thermal grid
integrated
RTU
water system
1. radiant panels
and cool beams
2. water mullions
air system
3. air diffusers
lighting system
4. dimmable
artificial lights
2
3
4
5. operable windows
6. motorized blinds
7. motorized louvers
5
6
7
Active system
Passive system
natural
resources
nature gas
electricity

Figure 1-2: Illustration of the Intelligent Workplace conditioning system
The multiple systems in IWn bring in more energy saving potentials and present a
challenge to the control at the same time. An advanced control system that can coordinate
the range of devices over time and occupancy variations efficiently is needed to
maximize the benefits. Due to the heterogeneous nature and varying requirements in
buildings, today's supervisory control lacks the standardization of control logic and is
more-or-less implemented as rules in the form of IF-THEN statements often in propriety
language. When diverse building systems are to be coordinated, the control logic
becomes complicated, non-standard, and difficult to troubleshoot (Timothy 2005). Rules

19
are typically constructed by the onsite control engineers based on expert knowledge,
dated information on the system themselves, and the templates used for other projects.
The multivariate system is often decomposed into multiple single-input-single-output
(SISO) loops and modulated with conventional responsive control locally.
Conventional ON/OFF and PID controllers are expensive in the long run since they
operate at a non-optimal efficiency. They are simple and normally work with one single
objective, e.g. reference tracking, offset prevention, etc. To fulfill the function of an
optimal control for other objectives (e.g. energy efficiency, thermal comfort, etc.), a
building control system needs to possess a predictive capability with building dynamics
and deterministic and stochastic disturbances considered. While predefined rules and
single-loop PID controls are interpretable to and preferred by onsite engineers, obtaining
optimal setting for a complex system is not an easy job. Meanwhile, neither of the two
control objectives, thermal comfort and energy savings, can be explicitly expressed in
conventional control laws. Utilization of models in control, the basis of model predictive
control, serves as a potential measure in the future to systematically improve the
performance in the future.
1.2 Background
Model-based optimal control technology provides opportunities for improving the
building energy system performance by incorporating multiple system models. From a
general point of view, model-based predictive control (MPC) is not a specific control
law, but a strategy which uses a combination of control methods linked by common ideas
(Clark 1994, Camacho 1995). The utilization of building system models has been widely
found in the design and evaluation period. As computers decrease in cost while
increasing in speed and performance, the utilization of models for real-time building
system control is becoming more and more practical. The integration of computer and
information technology into the building system operation has been gaining continuous
interest recently. The model-based control approach also has a natural tie to the broad
application of building information modeling (BIM), where a model is built and
maintained through the whole building life cycle.

20
A building system and/or equipment model can be integrated into a control system to
compare the various options, in a defined time interval, with given inside and outside
conditions. The results can be further fed back to the control system, utilized to generate
rules, or presented to operators to make proper actions in advance. Utilization of model
and simulation in real-time system control can possibly compensate the shortcomings of
predefined rules. It is very useful in load shedding, thermal storage, and passive cooling
control to minimize the operation cost where the interaction is too complex to be
evaluated in a simple intuition. The layout of utilizing a model/simulation in the
supervisory level controller in a receding horizon manner is illustrated in Fig. 1-3.
Building and plant systems
simulator
suggested action,
setpoints, etc
weather condition/prediction
occupancy/prediction
optimization
current system states
( energy consumption,
illuminance, temperautre, humidity, etc)
database
EMCS and
local controllers
building
building mechanical system
control actions
e
n
e
r
g
y

f
l
o
w

Figure 1-3: Illustration of simulation-assited predictive control
Models can also be integrated into the control for optimization. With the decreasing cost
of sensors or virtual sensors (Li et al. 2011), modern DDC has the capability to provide a
vast quantities of system states information. With the increasing computation speed of
processors, the information can be filtered and utilized to develop the models for the
building system MPC. For linear control, the original time-variant nonlinear system is
linearized around the operation points. With the model and dynamic system information,
MPC can make an optimal action in the chosen prediction horizon to minimize the cost
function. The optimization may be implemented in both the supervisory level and the
local level, termed dual-loop MPC, as illustrated in Fig. 1-4.

21
cost function +
constraints
Linearized
system model in
state space form
High level dynamic optimizer
System model
Low level dynamic
optimizer
u (manipulated
variables)
Building plant
State
estimator
MPC
(centralized or distributed)
x
y r (setpoints)
(system states)
cost function +
constraints
Weather condition
/ prediciton
Occupancy
/ prediciton
Occupants' feedback
Supervisory controller
unmodeled disturbances
modeled/measured disturbances

Figure 1-4: Illustration of dual-loop MPC
The research on utilizing a model to guide the control and operation has been widely
studied in process industry since early 1970s. However, it has just started to enter the
building industry. Roughly, the research can be categorized into two approaches, either a
simulation-assisted predictive control, or a model based predictive control. The detailed
review on the research from both building mechanical domain and control domain is
provided in Chp. 2.
1.3 Objective
Improving design is far from making buildings energy efficient. Compared to
conventional control, utilizing models in building system control and/or operation in a
predictive manner offers many more benefits. The flexible operation criteria, enhanced
handling of constraints, and nature multiple-input-multiple-output (MIMO) features in
model predictive control (MPC) provide the potential to further improve the building
system performance. However, there is no literature so far systematically studied the
utilization of various models in the optimal predictive control of both the fast dynamic
equipment and slow dynamic space system. The application and performance of linear
classical MPC in building systems are not fully explored.
The hypothesis of this thesis is:

22
Advanced model predictive control that integrates sensor and controller technologies
with dynamic predictive information will save energy and improve thermal comfort.
Beyond conventional model predictive control, the following enabling techniques are
investigated as the objectives:
- Developing multivariate control-oriented dynamic models for diversified energy
systems. This includes constructing simulators for the various components
included in building systems;
- Designing linear MPC controllers for mechanical components in the diversified
systems and conducting the control performance evaluation;
- Designing fast nonlinear MPC controllers for the space thermal conditioning,
which incorporates thermal comfort and energy conservation, and evaluating the
performance.
The research uses the IWn as the test bed that includes a hydronic system.
1.4 Methodology
The following aspects are conducted to accomplish the objectives of this research:
- Comprehensive literature review: The building automation, model/simulator
assisted control, thermal comfort control, and MPC theory are surveyed.
- Mathematical deduction: The dynamic models of the diversified building system,
including the building space, hydronic system, hybrid active desiccant ventilation
unit (enthalpy recovery wheel, vapor compression unit, and active desiccant unit)
are deduced. The models are in different time-scales and serve different operation
purpose in the ideal dual loop MPC structure.
- Computational simulation: MPC with the various models for equipment control
and space conditioning are evaluated by a set of simulations. The reference
tracking performance and energy saving potentials of using MPC for the hybrid
ventilation unit is conducted in the level of seconds. The MPC of space thermal
conditioning and system optimization is conducted in the level of minutes.

23
- Experiments: Experiments and historical data, which are conducted and collected
for the space, as part of the DoE funded project within or beyond the thesis, are
utilized in the thesis when proper.
1.5 Outline of the thesis
The remaining chapters are as follows:
Chapter 2: Literature Review
This chapter provides a critical literature survey on the existing study in terms of model
assisted control, model predictive control, and thermal comfort oriented control. The
literatures in the past decades in the aforementioned areas are reviewed. Firstly, the
modeling methodology, approaches, optimization methods, etc., are summarized and
categorized. Then, the concept of general model assisted control in building industry is
presented and the recent key publications are reviewed. Furthermore, model based
predictive control with mathematical and linear model and the studies in the field are
surveyed. Research on utilizing thermal comfort in advanced control is briefly
introduced. A number of issues deserving more study as systematically conducted in the
thesis are identified.
Chapter 3: Model Predictive Control
Model predictive control is a control methodology rather than a fixed set of control law.
It consists of three basic elements: an internal model, an optimization problem, and a
moving horizon mechanism. This chapter mainly focuses on the theory of classical linear
MPC, problem formulation, and solution method. The internal dynamic model of
classical MPC is linear for one or a set of given operation points. The cost function is
defined as a weighted norm of reference tracking errors, control actions, and control
action rates. The constraints on the inputs, outputs, and moving rates are naturally
included. MPC with broad definition, which converts a control problem into an
optimization programming problem, and the solving methods for identifying solutions are
introduced.

24
Chapter 4: Model Development
The chapter deduces the dynamic models for the diversified building systems in IWn
through first principles. For the purpose of dynamic control, a lumped-parameter
approach is generally applied to realistically simplify the models. A resistance-
capacitance (RC) analogue model is constructed for the IWn space. The various aspects,
including the heat transfer through the building enclosure, solar heat gain through the
fenestrations, natural infiltration, internal heat gain, etc., are considered. The hydronic
thermal heating and/or cooling system in the space is modeled based on the dominant
heat transfer mechanism. The dynamic models for the enthalpy recovery unit, vapor
compression unit, and active desiccant unit are deduced based on energy and mass
balance with minimum geometry parameters involved. The corresponding assumptions
for the aforementioned models are stated accordingly. The dynamic models serve as the
base for the simulators that can be used for various investigation purposes.
Chapter 5: Control of Hybrid Ventilation Unit
This chapter applies classical linear MPC algorithm to the devices in the hybrid
ventilation unit. The nonlinear models are linearized around the operation points to serve
as the internal model. It is found that, with the flexible formulation of MPC, the control
guides the device toward the goal with minimum efforts. For example, in swing season,
when the space is in need of free cooling outdoor air, the MPC for the enthalpy recovery
wheel deactivates the wheel and speeds up the exhaust fan. MPC for the vapor
compression unit coordinates the condenser fan and compressor to track the supply air
temperature set point or the humidity set point. Control of the active desiccant wheel
imposes a challenge since the dehumidification performance is highly dependent on the
regeneration air temperature with less manipulation allowed. The results of MPC
performance are presented and discussed as well.
Chapter 6: Control of Space Thermal Conditioning
To take the advantage of linear MPC for online implementation, based on the slow
dynamic feature of building thermal system, a Wiener MPC and a Hammerstein-Wiener

25
(HW) nonlinear MPC are proposed and studied in this chapter. A tool that can be used to
automatically generate the state space model is developed. The original nonlinearities
included in a dynamic building mechanical system are handled through the combination
of a linear dynamic model and one or two nonlinear static mappings. The performance
with pure linear MPC, Wiener nonlinear MPC with thermal comfort nonlinear mapping,
and Hammerstein-Wiener nonlinear MPC with both input and output mapping is
evaluated. Cases with different formulation of the optimization problem are studied. The
results based on the simulations show the benefits of using the proposed control for both
thermal comfort and energy conservation with minimum computational cost.
Chapter 7: Control Related Issue on Hydronic Systems
This chapter presents how a hydronic system can be properly modeled for balancing,
knowledge of flow distribution, coupled simulation, and evaluation of control, etc. It
focuses on water-based heating and cooling systems, which generally have high energy
efficiency in design but can perform poorly in reality due to the under-sensing and strong
thermal-water coupling. The study introduces the motivation, presents the simplified
modeling methodology, and illustrates the model and simulating structure. The approach
can handle the discontinuity involving valve open/close and pump on/off control. A
preliminary evaluation of the method is conducted with two simple simulations. The
node-branch-state modeling approach can be easily modified, expanded and integrated
into a detailed thermal model. The chapter concludes that the method can serve as virtual
sensing for potential use in control, operation and fault detection and diagnostics.
Chapter 8: Conclusions and Discussion
The conclusions and discussion on the thesis and future work are stated in this chapter.

26
CHAPTER 2 LITERATURE REVIEW
2.1 General survey 2.2 Model assisted
control
2.3 Model predictive
control
2.4 Thermal comfort
control
- The existing study on
modeling and control
for building systems are
surveyed.
- This part reviews the
study on conventional
model assisted control.
- Study of control with
explicit inclusion of
models in a moving
horizon manner is
reviewed.
- Thermal comfort
evaluation and its use
for building system
control is briefly
surveyed.
2.1 General survey
Model based predictive control has been utilized in the manufacturing process industry
since the early 1970s. The utilization of models for controlling integrated building
systems has just begun to thrive. Most of the efforts conducted in the direction of
utilizing models are in design and evaluation stages. After reviewing the literature in the
building industry, about 100 papers till the time of this survey are identified as the studies
focusing on the model/simulation assisted control of building systems (real time energy
management related) for higher energy efficiency. The research of utilizing model based
control can be categorized into the following four charts from Fig. 2-1 to Fig. 2-4.

Figure 2-1: Number of articles about model-based supervisory control in building
industry
0
5
10
15
20
25
30
Before1990 19911995 19962000 20012005 2006tonow
P
a
p
e
r
n
u
m
b
e
r
s

27

Figure 2-2: Distribution of modeling methods in building industry
Depending on the purpose and application, the model can be constructed in many
different ways. Generally, they are categorized into three types: physical model, gray-box
model, and black-box model. A physical model is deduced from the physical
understanding of the system and based on fundamental laws of energy, mass, heat
transfer, momentum, and flow balance, etc. Such a model requires physical meaning
parameters. A detailed physical model demands high computational power during the
simulation, while it generally has very high accuracy. A black-box model discards the
physical meaning of the system and establishes a direct relationship between the selected
inputs and outputs. While the model is usually simple, it needs a large number of trended
data to derive the relationship and no extrapolation is allowed outside of the training data.
A gray-box model has less complexity than a detailed physical model but still remains
some meaningful structure. The advantage of using a grey-box model is the fast speed
and relaxed extrapolation.
Most papers were published from 1991 to 2005 alongside the application of direct digital
control. As shown in Fig. 2-1, in the past decades, the research on utilizing building
energy system models or the rules generated from models for energy conservation has
been continuously active. These papers represent four types of modeling methods. About
35.90%
15.38%
42.31%
6.41%
Physicalmodel Blackbox GrayboxandHybrid Modelfree

28
36% is built upon the first-principle physical model and 15% upon data-driven black-box
model. The largest portion, about 42% of the reviewed research, utilizes simplified gray-
box and hybrid model methods.

Figure 2-3: Target system or components and the modeling method
These model developments were typically allied to specific building systems, aimed at
improving building system operations. For multiple systems, building wide controls, the
31.71%
3.66%
12.20%
39.02%
9.76%
3.66%
Cooling Heating Storage Buildingwide Airside Pump
0
2
4
6
8
10
12
14
16
Airside Building
wide
Cooling Heating Pump Storage
P
a
p
e
r
n
u
m
b
e
r
Blackbox GrayboxandHybrid Modelfree Physicalmodel

29
most widely used modeling method is physical model, followed by gray-box and hybrid
model (Fig. 2-3). Typically a global search engine is applied on the models to identify the
optimal set points. After building wide development of model-based controls, the next
most prevalent development is for cooling systems, followed by cooling storage for peak
load management. Gray-box and hybrid model dominates the model assisted cooling
operation optimization. Only modest efforts have been paid to the operation of heating
systems and water loop movers. Building systems with thermal storage have been widely
studied to either shave the load or switch among the cooling sources.
Utilization of a building system model for control in real applications involves the
optimization techniques. Depending on the modeled system and control challenge, the
optimization problem could be roughly categorized as linear or nonlinear, local or global,
direct search or gradient based. Finding the optimal solution to a problem in a given
circumstance is a key issue for the utilization of model assisted supervisory control
application (Wang 2008) as well as for the model based local regulation control. The two
in practice have some level of overlaps, although the definition differs. The local control
action in a given prediction horizon can be parameterized and the time depending control
problem is converted into an optimal problem (Yu 2009). Selection of the optimization
methods are highly problem specific. Thanks to the development from the mathematics
and computing area, most problems can be formulated in an advance language and be
solved with existing program packages. The main focus from the domain of building
industry is defining the control problem statements, formulating and encapsulating them
with a proper solver. (Google page rank finds the eigenvector of a 10
10
by 10
10
matrix,
which is one of the largest computations in the world.)
Depending on the problem statement, for a given set of specifications, there always exist
several optimization techniques which can be applied but differ in one or several aspects.
In other words, the approach is not unique. Since research generally has mixed utilization
of optimization techniques and the categorization of optimization techniques is broad,
Fig. 2-4 is not to be exhaustive but to qualitatively reflect the status and the preference. In
the reviewed HVAC and building industry research papers, about 90% use nonlinear

30
optimization techniques since the problems in HVAC systems are nonlinear with
constraints. Genetic algorithm (GA) is the most widely used global optimization
technique followed by sequential quadratic programming. The other 10% of papers
utilize linear optimization methodology for their problems. It shall be pointed out that
artificial intelligence is also considered as a measure for optimization in many other
research papers. They are not reviewed here.

Figure 2-4: Optimization techniques applied in the reviewed papers
2.2 Model/simulator assisted control
Even if models are incorrect, they can be useful. The use of system models to improve
the performance of supervisory control systems is not a new idea. In 1988, a review study
on control system simulation was conducted (Kelly 1988). While a model can be utilized
in any part of system operation, model assisted control in this thesis mainly refers to
the supervisory level control. It relies on model simulation to provide optimal
recommendations. It is important to mention that the regulating, sequencing, and
0%
2%
4%
6%
8%
10%
12%
14%
16%
18%
b
r
a
n
c
h
a
n
d
b
o
u
n
d
e
v
o
l
u
t
i
o
n
a
r
y
a
l
g
o
r
i
t
h
m
s
a
n
d
g
e
n
e
t
i
c
a
l
g
o
r
i
t
h
m
e
x
t
r
e
m
e
v
a
l
u
e
g
e
n
e
t
i
c
a
l
g
o
r
i
t
h
m
s
i
m
u
l
a
t
e
d
a
n
n
e
a
l
i
n
g
c
o
n
j
u
g
a
t
e
g
r
a
d
i
e
n
t
m
e
t
h
o
d
d
i
r
e
c
t
s
e
a
r
c
h
l
a
g
r
a
n
g
e
m
e
t
h
o
d
s
e
q
u
e
n
t
i
a
l
q
u
a
d
r
a
t
i
c
p
r
o
g
r
a
m
m
i
n
g
s
t
e
e
p
e
s
t
d
e
s
c
e
n
t
u
n
i
v
a
r
i
a
t
e
s
e
a
r
c
h
m
a
x
i
m
u
m
p
r
i
n
c
i
p
l
e
r
e
c
u
r
s
i
v
e
l
e
a
s
t
s
q
u
a
r
e
s
i
n
g
u
l
a
r
p
e
r
t
u
r
b
a
t
i
o
n
global local NA
nonlinear other

31
supervisory control inevitably overlap one another in a complicated building system. The
utilization of a model in the building control can be either offline and advise the system
operators with the best operating strategy, or online and determine the operating
decisions for optimal performance.
A model is an abstraction of a physical system, expressed in a set of equations. When the
model is well-tuned to capture the systems thermal dynamics, both quantitatively and
qualitatively, it becomes useful to be included in the control decision making. The role of
the model is to help identify a set of energy or cost efficient settings for the local
controllers, taking into account the capacity of and interactions among the systems, and
the given inside and outside conditions. A fairly accurate short term weather prediction
can be acquired via the internet. With the increasing computational power of general
purpose computers, a model can evaluate thousands of different settings in a relatively
short time. In reality, such an evaluation is usually performed with a smaller scale of
potential settings picked from a candidate pool.
Henze et al. (1997) investigated the predictive control of building thermal storage
systems using a physical model-based approach. They pointed out that, while a thermal
storage system can utilize the energy price difference and do load shifting, the
conventional control has very poor performance lacking efficient strategies. Heuristic
rules usually prefer direct cooling during off-peak hours and only use a storage system to
complement the shortage of a chiller. In their study, hourly steady-state models were used
for the cooling plant with ice storage systems ignoring the dynamics with
charging/discharging changes. Dynamic programming method was utilized to find the
operation strategy with minimum cost. They claimed that the strategy can bring in
significant cost savings even with load prediction uncertainty.
Zhang and Hanby (2006) studied a model assisted control for a system with different
energy sources and active and passive thermal storage. Daylighting, natural ventilation,
solar water and air heating, and a water tank were the elements that presented in the
space. The question to be solved with a model was to find, when a heat source was

32
available, whether it should be allowed into the building directly, stored for later use, or
simply rejected. A simplified zone and system model was developed to avoid a high
computational cost. Space dynamics were represented by the temperature evolution of
mean air and mean wall. The heat gain, including air terminals heat, internal gains and
solar gain, was lumped as one into the space air. The model of the building and plant
were built in a Matlab/Simulink environment. An evolutionary method and a micro-
genetic algorithm are used to devise the control trajectory for the next day. They found
that while significant improvements in system operation are possible, the simulation time
with 13 hourly commands makes the global search approach prohibitive for online
implementation.
Chen (2001) formulated an optimal model with a number of practical operation
constraints to supervise the operation of building systems with big thermal mass. In the
operation structure, a supervisor was built upon a weather predictor and a weather
monitor. The weather predictor generates ambient temperature and solar radiation
profiles for the next 12 to 24 hours based on the local daily weather forecast and recent
measurements. The model parameters in a z-transfer function form were estimated using
recursive least squares for efficient real time implementation. It was recommended that
the constraints should be utilized when adopting dynamic programming in searching for
the optimal set points to reduce the searching efforts. The simulation results indicate that
the predictive supervisory control can save around 10 to 27% for the case building. The
proposed online implementation method was also fast that determined a 24-hour
operation strategy in about 15 seconds.
Miyajima et al. (2007) applied a supervisory optimal controller for an air-conditioning
system in a field test. The control system was established based on the physical models of
each component with optimization engine software to find the optimums. Five control
variables, including the chilled-water temperature and flow rate, condenser water
temperature and flow rate, and supply air temperature were selected as the controller
outputs. The simulator first calculated the heat load using trended data and then
postulated initial values for the target control values. The simulation further deduced the

33
possible energy consumption of the devices in the water and air loop. According to their
study, the simulation assisted controller optimized the set points in a given range for
different conditions. It could bring about 24% energy savings over the ordinary manual
operation. The greatest energy-saving effect was achieved in the chiller, followed by the
condenser-water pump and the chilled-water pump.
Fong et al. (2006) used the simulation-optimization approach to obtain the optimums of a
subway station HVAC system operation. They argued that, due to the individual features
of HVAC system, conventional operation methods and set points might not be
necessarily energy-efficient in daily operations. Evolutionary program was adopted to
handle the discrete nonlinear and highly constrained optimization problems. The focus
was to identify a monthly optimum reset scheme for the chilled water and supply air
temperature. The model followed a component-based approach and simplified the multi-
AHUs into one. It was found that the scheme could achieve 7% saving potentials
compared to the existing operational settings.
Cascia (2000) implemented a near-optimal global set point control in an energy
management system. The chilled water, hot water, and coil discharge air set points were
the main focuses in his study. A simplified linear chiller component model and a more
general quadratic chiller model were considered. All power consumptions were expressed
as a function of temperature difference between chilled-water supply and return
temperatures. The model parameters were obtained from onsite measurements via the
direct digital system. A third-party energy accounting program showed that a monthly
energy reduction in kWh could be obtained from 3% to 14%. The method assumed that
the condenser water flow rate was constant.
Wang and Jin (2000) utilized the genetic algorithms with a detailed model to optimally
control a VAV air-conditioning system by identifying the proper supply air temperature,
outdoor ventilation flow rate, and chilled water temperature. The GA optimizer started
with a group of random set-points as the first generation and computed each generation
until the optimums were found within the given range. The findings were then further

34
checked by a rule-based supervisor to avoid the action of the optimums that might only
bring in trivial savings. The model adopted in their study was a simplified but adaptive
physical model that could give accurate simulation in finite time step. The parameters and
the slow-varying variables were considered as constant that could be estimated by using
recursive least-squares with exponential forgetting technique. Eight to 37% of total cost
savings were identified in a four-day test. The indoor air quality was improved
significantly with the simulation assisted control.
Keblawi et al. (2011) developed an optimized online supervisory controller for the chilled
ceiling displacement ventilation. The energy consumption was minimized without
compromising the indoor air quality and thermal comfort. The control set points
including the chilled water flow rates, the reheat amount, and the fan speed were adjusted
with genetic algorithm. The physical-model-based supervisory control was built upon the
component models, the load and environment predictive space thermal model, and an
optimizer. Simplification with lumped parameters was conducted to the models to avoid
the nonlinear radiative heat transfer process of the radiant ceiling and the heat storage
effect in the wall. A test case was performed in an office space and proved the feasibility
of this approach.
Clarke et al. (2001) described the prototype of model assisted control embedding a
simulation program in a building energy management system. A physical mode was
considered better than a black-box model in their research. The clear cause-effect
relationship, good adaption to the changes, and better scenario flexibility were regarded
as the good features of physical models. They claimed that model assisted control could
be very useful in the circumstances when significant prediction time was allowed,
alternatives and the implications were evaluated, high interaction exists, or large
variations were known in advance. A dynamic simulation program ESP-r was used
together with LabVIEW to illustrate the performance in predicting the optimum heating
start time for a two-room facility. 5% temporal error was observed between the measured
and simulated results, and the maximum daily temperature error in one day was around 1
Celsius degree.

35
Carina et al. (2011) studied a run-time co-simulation with a middleware between the
simulation and the advance language to develop a building controller. The model was
used to assess the possibility of utilizing the simulations to improve the performance of
building controls. The low-level controllers in EnergyPlus were remained while only the
high-level decisions were made in Matlab. The temperature set points and plant and zone
equipment availability signals, as well as some heat gain settings, were provided from the
outside algorithms. They pointed out that this approach was flexible with clear structure
to controller development, even for complicated problems. However, the direct
involvement of physical simulation program also significantly increased the debugging
efforts and simulation time.
Mahdavi (1997) argued that since a virtual model could move both backward and
forward in time, it could be calibrated first to predict the buildings response to
alternative control scenarios. He argued that a critical task to realize a model simulation
assisted control was to create the set of control options. Two approaches, a generate-and-
test method and a bi-directional inference method, were suggested for the goal. The first
method had a rule-based option generator to produce a finite number of discrete control
options. The second method used an explicit definition of control and performance
variables. A single one-zone example was provided to illustrate the proposed
methodology.
As a summary, the basic idea behind the model/simulator assisted control is to minimize
the overall system energy consumption and transfer the finding in a model into the real
implementation. The model is usually implicit to the optimization solvers. The selection
of the optimization technique and formulating the cost and objective function comes to
the next after the development of models.
2.3 Model predictive control
"The objective of optimal control is to determine the control signals that will cause a
process to satisfy the physical constraints and at the same time minimize (or maximize)
some performance criterion." (Kirk 1998). Model predictive control (MPC) is one type of

36
optimal control among many others. The term MPC refers to a suite of control strategies
that use an explicit model of a physical system to derive the set of controller actions that
minimize a cost function, and subject to a set of constraints. MPC was originally
developed for petroleum process control, and then gradually adopted in a wide variety of
application areas including chemicals, food processing, automotive, and aerospace
applications (Qin et al. 2003).
The inner model used for linear classical MPC is typically a linearized system expressed
explicitly in state space form. This allows the utilization of powerful tools to study the
system, the linear quadratic regulator theory, Kalman filtering theory, internal model
theory, etc (Rawlings 2000). In the case of linear MPC, the solution of the optimal
control problem can be casted as the solution of a quadratic program and solved
efficiently even on-line. Thus linear MPC is widely used in industry. Another trait of
using MPC is it can explicitly consider state and input constraints into the solution
seeking process, which is never a simple task for conventional controllers. In contrast to
classical control, which uses a pre-computed state or output feedback control law, MPC
estimates the system future behavior with a plant model in an open-loop manner. The
control inputs are obtained from all feasible sequences with the assistance of optimization
solvers. Only the first element of the computed sequence of optimal controls is then
applied to the system, and the process is repeated in a receding horizon manner. As
indicated above, models are not perfect forecasters; therefore a state observer is included
with the measurements feedback on the system outputs to overcome some effects of poor
models. The structure and operation mechanism of a basic MPC can be illustrated as in
Fig. 2-5.

37
Time
k k+Nc k+Np
reference
current
time stamp
past output
measurements
past control action
(manipulated variable)
control horizon
prediction horizon
predicted output
) ( k Nc k y +
control action
) ( k Nc k u +
y

Figure 2-5: Illustration of basic MPC structure for HVAC system control
From a mathematic point of view, MPC is actually a constrained optimization problem,
where the solution is numerically obtained. For classical reference tracking control, MPC
has a cost function in a quadratic form, with weighting matrices on the tracking errors
and control efforts, subjected to a set of hard and soft constraints. The cost function can
also be reformed differently with the terms to represent the control needs, such as energy
cost.
While MPC can be used in the control of SISO system, the main application is for MIMO
systems, where the coupling among the inputs and outputs is strong. In conventional
controls, sequential responsive proportional, integral, and derivative control is widely
applied in HVAC system for local control, with the coupled multivariate system treated
as decoupled multiple SISO loops. By doing so, the cross-coupling effect among the
loops is usually ignored or compensated afterwards (Krakow 1995). However, there exist
many cases that the coupling affects among the inputs and outputs cannot be ignored in
order to achieve the energy efficiency and occupants comfort. Living space climate

38
regulation is naturally a multivariate problem having no unique solution. The optimal
control law targets to satisfy thermal comfort, visual comfort, and indoor air quality with
minimum energy consumption or cost. MPC is suitable for the situation since it can
easily handle multivariable control problems naturally. The optimization features allow it
to operate closer to the constraints compared to responsive control that can have higher
energy efficiency.
Compared to SISO, model based MIMO control is more effective in simultaneously
achieving multiple control objectives in an integrated building energy system. The
performance of conventional SISO control strategies is inherently inferior to that of
MIMO control strategies (He 1998). Study of utilizing the good features of MPC on
multivariate HVAC systems is gradually gaining more interest recently.
Kolokotsa et al. (2009) used a simplified bilinear-model based predictive controller,
combined with a building energy management system, to predict and control the indoor
environmental conditions of a 125m
2
singe zone laboratory. The standard least square
estimation was utilized to identify the parameters in the model. The control space was
discretized to simplify the minimization procedure. The set points for the room air
environment were selected as conventional static values with dead bands. Trial-and-error
method was utilized in the process of selecting the weighting matrix and tuning the model
and the prediction horizon.
Morosan et al. (2010a, 2010b) studied a distributed predictive control structure for
thermal regulation in buildings to avoid the exponentially growing computational
demand. MPC was considered in their study good for intermittent operated space where a
set point during unoccupied time was not even needed. Future occupation profile was
incorporated into the problems as the error weighting term. The control in terms of the set
point tracking only got penalized during the occupied hours. The prediction horizon was
chosen as five hours for the control system to be able to see the indoor temperature up to
the desired set point in the worse situation. A three-zone simplified case was studied with
the proposed structure, decentralized structure and centralized structure. The centralized

39
and distributed MPC showed an improvement about 36% on thermal comfort meeting
hours with 13% of energy savings.
Ma et al. (2010) used a simplified two-mass nonlinear system to model thermal zone
dynamics and a distributed MPC to regulate the indoor air temperature. The controller
was implemented by using combined sequential quadratic program and dual
decomposition for outer iteration and inner iteration to reduce the time of solving the
optimization problem. In the simplified model, the system pressure and the individual
component dynamics were not considered. A nominal prediction of disturbance load was
assumed available as a function of time and weather. The cost function was formed for
temperature tracking, comfort range violation, and control efforts minimization. An
artificial load profile was applied to a group of identical room models with one day
simulation. Compared to a rule based controller, the proposed MPC control had better
pre-cooling performance for load shifting when the system consumed less energy for
cooling in the morning.
Yuan and Perez (2006a, b) applied MPC to a variable air volume system to achieve an
acceptable indoor air quality. They pointed out that the conventional ventilation control
applied a ventilation rate to the total supply air flow while the individual zone air flow
rate was determined by the load. Poor ventilation could happen to those rooms with low
load demand while over-ventilation could occur when a building was under-occupied.
The MPC was to coordinate the total outside air intake, the zone supply air quantity and
reheat for both temperature and ventilation controls. The PI sequencing controller in the
outer loop is replaced by the MPC and the inner loop PI controller remains. The set
points of supply air temperature and air flow rate from MPC were tracked by the inner PI
controller. The ventilation condition at the six zones with MPC was found improved
compared to the traditional control.
Qi and Deng (2009) investigated the application of a MPC for the operation of a direct
expansion air conditioning unit in a single room. The temperature and humidity coupling
problem was solved with the control on the compressor speed and the fan speed. A good

40
disturbance rejection capability was observed when the system was imposed with an
unmeasured heat load disturbance. It took the MIMO around 20 minutes to achieve their
respective new set points. Compared to SISO strategy, the MIMO control strategy was
regarded giving better simultaneous control on the indoor air temperature and relative
humidity with adequate control sensitivity and accuracy. The controller was designed
based on a linearized dynamic model at a fixed operating point and worked around the
operating point with a relative small range. The control only concerned about the air side
without considering the vapor side.
Oldewurtel et al. (2010) developed a stochastic MPC strategy for building climate control
based on a simplified bilinear model. The building-wide energy use was treated as one
single room in their model while the illuminance and CO
2
concentration were modeled
by instantaneous responses. An uncertainty driven by Gaussian noise was considered to
improve the estimation based on the weather prediction. Sequential linear programming
was applied to iteratively solve the non-convex optimization with non-linearities. The
constraints were not required to be satisfied at all times but with a predefined probability.
The controller performance was evaluated for different system types in different climates
by calculating the performance bound. The controllers determine the delivered heating or
cooling power, the used air change rates as well as lighting and blind positioning. MPC
showed less energy use than rule-based control while four out of six cases had smaller
thermal comfort violation.
Gondhalekar et al. (2010) proposed a method to achieve least-restrictive robust MPC
with the open-loop prediction in periodic MPC formulation for building climate control.
They argued that the building dynamics are time-invariant and the control problem was
periodic, such as, the fluctuation of activities of office workers, the time-dependent
outside air temperature. It was pointed that these stochastic distribution were bounded
with periodically time-dependent bounds depending on the time of day. The control
objective (e.g. the energy price) is periodically time-dependent. The environmental inputs
(outside air temperature, solar radiation, and internal heat gain) were treated as two parts:
periodically time-dependent deterministic portion and stochastic but bounded portion.

41
The four tailed control laws demonstrated that the methodology was applicable for a
large variety of control problems with nearly standard methods.
Privara et al. (2011) and Cigler et al. (2010) used MPC for the temperature control of a
water-based heating system. They pointed out that conventional On/off and PID control
could hardly work with MIMO systems in practice, only if the systems were specially
decoupled. Subspace identification was applied in their study to obtain the system model
and a classical quadratic optimization problem was concerned in the MPC. The penalty of
temperature tracking was only imposed when the room air temperature is lower than the
set points to avoid energy waste due to precise reference tracking. It was found that the
MPC could save around 17-24% energy compared to the conventional weather-
compensated strategy. The main reason was that MPC was able to employ the building
thermal capacitance and the optimal inputs were not so aggressive in comparison to
conventional control strategy.
Gwerder and Todtli (2005) studied the performance of MPC in a room air temperature
control by applying linear programming algorithms. They argued that predictive control
in building application had a potential to be more energy and cost effective than
conventional responsive control because of the thermal capacitance, wide rang controlled
variables, time-dependent cost, etc. A simple lumped-parameter model that could
produce the buildings essential static and dynamic thermal properties was considered in
their high-level controller. A constant weighting matrix was applied in their study and a
negative weighting factor was assigned to the blind position to encourage day lighting
and avoid higher cost on artificial lighting. The simulation result supported that
predictive control solution was superior to conventional responsive control. The saving
feature was considered more obvious when low cost heating or cooling can be used to
cool or heat the building at an early time.
MPC and model-assisted control have the similarities. By utilizing a model, the different
actions can be evaluated to improve the performance of a system for future. The
evaluation can be performed periodically. Unlike model-assisted control, MPC requires

42
more explicit involvement of model and usually has more strict format to obtain the
optimal solution in a receding horizon manner.
2.4 Thermal comfort control
Humans are adaptive to a wide range of indoor air condition without feeling discomfort.
A conventional air conditioning terminal control in building system is typically set to
provide a constant indoor air temperature. Research has revealed, however, that a human
indoor comfort is neither static, nor depending on air temperature settings alone.
Environmental parameters including air temperature, mean radiant temperature, air
velocity, and relative humidity, along with the personal parameters, are the main factors.
As a result, a wider range of room air temperature control could lead to more energy
savings without sacrificing the occupants thermal comfort. To achieve the benefit, the
thermal comfort concept shall be included into the control law of building regulation.

Figure 2-6: Acceptable range of operative temperature and humidity for thermal
comfort
(ASHRAE 2004)

43
Thermal comfort is generally defined as that condition of mind which expresses
satisfaction with the thermal environment (ASHRAE 2004). It is a combination of
subjective and objective factors, and varies from person to person. An absolute comfort
status for all occupants in an air conditioned building is never reachable, but it is possible
to estimate the thermal comfort for the majority of the occupants and impose
corresponding control actions on the secondary air conditioning system. There are many
ways to define such an indoor environment. The most well known and widely accepted
methods are Fangers comfort model with two indices: predicted mean vote (PMV) and
predicted percentage of dissatisfied (PPD) (Fanger 1972). The index uses -3 for cold, -2
for cool, -1 for slightly cool, 0 for natural, +1 for slightly warm, +2 for warm, and +3 for
hot. An indoor environment is considered generally acceptable with PMV between -0.5
and +0.5 and PPD less than 10.
The theoretical evaluation of thermal condition is made based on the heat exchange
between a human body and the surrounding environment. When human adaptation is
taken into account, the comfort range can even be wider than the condition prescribed by
applying PMV index (de Dear and Brager, 1998). Behavior adjustments, relaxation of
expectations, and acclimatization to the conditions are the means to obtain the adaptation.
In a literature survey by Frontczak (2011), it was suggested that the type of building and
outdoor climate including seasons, should be considered to develop system for indoor
environment control. Room decoration and light color are found as negligible factors that
do not have impact on thermal comfort sensation. The occupants in natural ventilated
buildings have a more forgiving attitude on the indoor thermal conditions and allow
higher indoor temperatures in summer and lower temperatures in winter. Similar results
were also presented by Brager and de Dear (1998) in a review about thermal adaptation
in the built environment.
There are plenty of research conducted in terms of evaluation and improvement of
occupants thermal comfort for different buildings and layouts (Daum et al. 2011, Mazzei
et al. 2005, etc). Few research actually utilized advance control to achieve both thermal
comfort and energy conservation. Mui and Chan (2003) developed new notions about

44
adaptive comfort temperature in buildings in humid sub-tropical area. The adaptive
interface relationship of indoor comfort temperature with outdoor air temperature was
determined by a large-scale survey. The clothing value and air velocity were presented as
linear regression equations of the mean operative temperature. With the integration of the
adaptive control temperature model, it is possible to reduce temperature shock and
thermal comfort dissatisfaction rate while achieving about 7% energy savings.
Alcala et al. (2005) proposed the use of weighting linguistic fuzzy rules to control the
indoor thermal comfort, indoor air quality, and system stability with explicit energy
conservation as the main criteria. A genetic optimization combining human experience
was applied to select the linguistic rules and the weighting for the fuzzy logic controller.
Compared to the on-off control, 9% to 14% of energy savings were reported achievable
by using the intelligent control approach.
Yonezawa et al. (2000) applied the comfort index PMV into a room air temperature
regulation system. A neuro-PMV calculating system based on neural network was
applied to estimate the thermal comfort condition. The users were allowed to provide
feedback into the control system. The learning process took place at night and the
weighting coefficients were calculated by back propagation method. A fuzzy calculation
unit determines the room air temperature set point based on the deviation of the targeted
PMV value and calculated PMC. +0.3 was selected as the upper limit of the comfort
range in the cooling mode and -0.3 in heating. With the field experiment in the summer,
about 17% of energy savings were claimed compared to conventional control system for
an 11-sotry building.
Freire et al. (2008) investigated the performance of using MPC for thermal comfort
optimization and energy consumption minimization as two separated objectives. A
heating system with power input as the single manipulated variable and multiple
measured outputs in a single zone was considered in their study. A nonlinear static model
was applied to obtain the predicted mean vote (PMV) as thermal comfort indicator and
the nonlinear optimization was solved by using sequential quadratic programming

45
algorithm. A black box auto-regressive with exogenous input (ARMAX) model was
identified to represent a building coupled to the HVAC system. The simulation results
showed that MPC could simultaneously promote thermal comfort and energy
consumption reduction with no violation on occupants' thermal comfort.
Thermal comfort is not a directly measurable variable but represents the indoor air
condition that occupants preferred for. Two indexes, PMV and PPD, could be calculated
based on the measured room air temperature, air velocity, air relative humidity, and other
parameters. The high nonlinearity makes it difficult for the direct use in conventional PID
controller in buildings. More study is needed to investigate the measure and potential
effects of incorporating thermal comfort in advanced controls.
2.5 Summary
This chapter reviews the literature about the utilization of models in building system
controls. The modeling methods, scale of utilization in building systems, and
optimization solution approaches are surveyed. Models for operation and design
conventionally are not differentiated other than the availability and/or accuracy of the
physical parameters. When a model is wrapped in a structure (such as a simulation tool)
transparent to the optimization engine, a global heuristic search is resorted to. This
approach is typical for the evaluation of actions in building systems but can hardly be
used for online application due to the high computational cost. MPC, originated and
widely used in process industry after PID, is an optimal control method. It has relatively
strict structures, fixed solution approaches, and mature analysis tools. The problem is
defined as a convex quadratic programming problem and solved in a moving horizon
manner. Thermal comfort, which is not a directly measurable physical variable, can be
included in advanced controls to achieve the purpose of space conditioning. The research
on applying optimal control utilizing models is getting more and more attention recently.
However, the review reveals that there are few or no systematic studies of:

46
- the development of control-oriented models for an integrated system with radiant
heating/cooling and hybrid dehumidification ventilation;
- the utilization of various models in the optimal predictive control of both the fast
dynamic equipment and slow dynamic space system;
- the application and evaluation of the performance of linear MPC on a complex
hybrid ventilation unit, including a vapor compression unit, enthalpy recovery
unit, and active desiccant unit;
- the structure and application Wiener nonlinear MPC and Hammerstein-Wiener
nonlinear MPC with linear MPC as the core for multivariate building systems and
multiple operation criteria;
- the modeling, sensing, and control challenge of the involved discontinuity and
thermal coupling in a hydronic system.

47
CHAPTER 3 MODEL-BASED PREDICTIVE CONTROL
3.1 Introduction 3.2 Linear quadratic MPC 3.2 MPC in broad definition
- Introduction - Controller formulation
- Solving methods
- Controller formulation
- Solving methods
3.1 Introduction
Model based optimal control technology provides us opportunities for improving the
building energy system performance by incorporating the system models and utilizing
future information. From a general point of view, MPC is not a specific control law, but a
strategy which uses a combination of several control methods linked by means of some
common ideas (Clark 1994, Camacho 1995). With over 2000 industrial installations,
MPC is currently the most widely implemented advanced control technology for process
plants (Qin and Badgwell, 2003). A lot of literatures (Michael 2010, Qin 2003, Rawlings
and Mayne 2009, Wang 2009, etc) summarized the MPC theory and recent studies in
both industry and academia.
Although MPC paradigm includes many different variants with all kinds of features, all
MPC systems involve the basic elements to generate the control trajectory for a process: a
model, a solver, and a moving horizon manner. Michael (2010) provided a good
illustration showing how an MPC works as Fig. 3-1. Different MPCs have the varieties
on the model format, disturbance estimation method, objective, solver, etc. Many
techniques are utilized in addition to improve the performance for given circumstances,
for instance, moving block, reference look-ahead, time-variant weighting, time variant
constraints. In this chapter, only the basic knowledge of MPC is introduced and the focus
is on the classical quadratic MPC, which is one of the most mature optimal control
strategies and serves as the base to the understanding and formulation of other general
MPCs.

48
@ time k
take measurements
from the process
utilize the process model and future information
on inputs/outputs, weights, etc
current and future
: control actions
: disturbance
(measured or estimated)
future process
outputs
Objectives
Constraints
solver: solve the constructed optimization problem
best current and future control actions
implement best current control action
for control horizon n
time k+n

Figure 3-1: Illustration of model predictive control scheme
3.2 Linear quadratic MPC
Linear classical MPC refers to a group of optimization problem where:
- The internal model is linear and time-invariant (LTI);
- The objective function is a quadratic function in terms of the control inputs,
measureable outputs, and rates of control inputs;

49
- The constraints are on the control inputs, measureable outputs, and rates of
control inputs.
No direct feed-through from the inputs to the outputs is allowed in the classical MPC
formulation, since normally the output is measured at time k before the new input at time
k is computed and implemented.
3.2.1 Controller formulation
The classical MPC has the following quadratic programming formulation:

=
+ A +
k
i
R S
Q
y
u u e
1
2 2
2
min
(3-1)

U L
U L
U L
u L
u u u
u u u
y y y
x x x
Cx y
Bu Ax x
A s A s A
s s
s s
s s
=
+ =
(3-2)
where
y
e
is the outputs errors, u A is the rates of control inputs,
u
is the control inputs,
x is the derivative of state variables,
x
is the vector of system state variables, A is the
matrix of states, B is the matrix of inputs, C is the matrix of outputs. Subscripts
Q
, S
and R denote the weighting matrixes on the reference errors, rates of control inputs, and
control inputs respectively. Subscripts L and U mean the lower bound and upper bound
of the corresponding variables.
The internal model utilized in a classical MPC has linear and time-invariant features.
Depends on the applications, different types of models may be used. The common model
types are impulse and step response models, auto-regressive moving average with
exogenous input (ARMAX)/ controlled auto regressive integrated moving average
(CARIMA) models, transfer function models, and more general, state space models.

50
A finite impulse response (FIR) or step response (SR) model is widely used in system
control since it can fit arbitrarily complex stable linear dynamics based on experimental
data. No or less advance knowledge about the system is required to develop the model.
Methods, such as ordinary least-square or optimization based on minimization of
prediction and measured errors, can be used to process the data and obtain the model
parameters. Eq. 3-3 and Eq. 3-4 give the general expression of FIR and SR model
respectively.

( ) ( ) i k u b k y
n
i
i
=

=1
(3-3)

( ) ( ) ( ) 1
1
+ A =

=
n k u s i k u s k y
n
n
i
i

(3-4)
where
u
is the input signal,
y
is the output signal, b is the filter coefficient (tap
weights),
n
is the number of parameter terms,
s
is the model coefficient, u A is the unit
input change, and k is the current time stamp.
For a MIMO system, the coefficients become a matrix instead of a vector to represent the
mapping relationship between the multiple inputs and multiple outputs. As seen
obviously from the expression, a large number of data are generally needed to identify
the parameters used in an FIR or SR model for a complex system. The parameters can be
up to hundreds for a complex system. For a MIMO system, the disadvantage of great
number of parameters becomes troublesome. Since the model does not include any
structure information from the system, data over-fitting and high modeling uncertainty
can exist.
Eq. 3-5 and Eq. 3-6 show the structure of an ARMAX and a CARIMA model
respectively. In Eq. 3-5, from left to right, the terms are auto regressive part, moving
average part, and exogenous inputs. In Eq. 3-6, the moving average term includes an
integral part.

51

( ) ( ) ( ) k Bu k Ce k Ay + =
(3-5)

( ) ( ) ( ) k Bu k e
C
k Ay +
A
=

(3-6)
Both models can be deduced based on measurements and in an adaptive (predictive)
manner. With CARIMA model, good output predictions and future control sequence can
be obtained alternatively to minimize the cost function. The derivation of optimal
prediction of CARIMA model can be obtained by recursion of Diophantine equation
(Ydstie 2012). The MPC utilizing ARMAX and/or CARIMA model is also referred as
Generalized Predictive Control.
A transfer function model is applicable to both stable and unstable plants. It can be
deduced based on regression or from first-principles. As the term implied, the model
maps the direct relation between the inputs and outputs. The advantage of using a transfer
function model also includes its compact expression, which requires less parameter than
in FIR model. Eq. 3-7 gives the structure for a transfer function model. Like an FIR
model, a transfer function is considered less effective for MIMO plants.

( ) ( ) k Bu k Ay =

(3-7)
Any aforementioned linear dynamic model can be transformed to a state space model. A
state space model provides a nice structure for MPC, and it can easily handle MIMO
system. It has uniform treatment of stable, integrating, and unstable processes. Extra
measurement signals of non-measurable disturbances can be included as feed-forward to
improve the performance. Eq. 3-2 gives the simplified state-space model. On the right
side, the matrix can be augmented to include the noise and disturbances.
3.2.2 Solving methods
Since the model and the constraints are linear, and the cost function is formed in
quadratic, classical linear MPC can be eventually casted as a quadratic programming

52
problem. A quadratic programming problem is a convex optimization problem with a
unique optimal solution and can generally be solved in a short time, which makes it
suitable for online implementation. To offset nonzero values on the variables, an integral
term of manipulated variables based on augmented model is generally used in MPC.
Wang (2009) provided a detailed algebraic solution to unconstrained MPC and
introduced the solution methods to constrained MPC. The related information is briefly
reviewed here.
An algebraic method can be applied to find the optimal analytic solution for
unconstrained MPC. The system in Eq. 3-2 can be written into discrete format as:

( ) ( ) ( ) k u B k Ax k k x A + = +1

(3-8)
( ) ( ) ( )
( ) ( ) ( ) 1
1 1 2
2
+ A + A + =
+ A + + = +
k u B k u AB k x A
k u B k k Ax k k x

(3-9)

( ) ( ) ( ) ( )
( ) 1
...
2 1
+ A +
+ A + A + = +

c
N N
N N N
p
N k u B A
k u B A k u B A k x A k N k x
c p
p p p

(3-10)
where (k+1|k) denotes the future value at time k+1 based on time k information, p
N
is the
prediction horizon, and c
N
is the control horizon.
Similarly for the output variables, we have

( ) ( ) ( ) k u CB k CAx k k x A + = +1

(3-11)

( ) ( ) ( ) ( ) 1 2
2
+ A + A + = + k u CB k u CAB k x CA k k y

(3-12)

( ) ( ) ( ) ( )
( ) 1
... 1
2 1
+ A +
+ + A + A + = +

c
N N
N N N
p
N k u B CA
k u B CA k u B CA k x CA k N k y
c p
p p p

(3-13)
Define vectors:

53

( ) ( ) ( ) | |
T
p
k N k y k k y k k y Y + + + = ... 1 1

(3-14)

( ) ( ) ( ) | |
T
c
N k u k u k u U 1 ... 1 + A + A A = A

(3-15)
Therefore, the equations can be collected as:

( ) U k Fx Y uA + =

(3-16)
where
(
(
(
(
=
p
N
CA
CA
CA
F
...
2
, and
(
(
(
(
= u
B N N N N
c p p p
CA B CA B CA
CB CAB
CB
...
. ... . .
0 ...
0 ... 0
2 1
.

To find the optimal solution for the given set point and objective function:

| | ( ) k r R
T
s
1 ... 1 1 =

(3-17)

( ) ( ) U S U Y R Y R J
T
s
T
s
A A + =

(3-18)
where ( ) k r is the set point at sample time k, S is the weighting matrix on the control
movement.
Substitute Eq. 3-17 into Eq. 3-18, and take the derivative of Eq. 3-18 in terms of U A :

( ) ( ) ( ) U S k Fx R
U
J
T
s
T
A + u u + u =
A c
c
2 2

(3-19)
By setting it to zero, the solution can be found for the control signal as:

( ) ( ) ( ) k Fx R S U
s
T T
u + u u = A
1
(3-20)
For most plants in building systems, there are hard constraints and/or soft constraints on
the inputs, outputs, and/or changing rates. For instance, the total input to a fan cannot be

54
more than the maximum capacity. The supply air temperature needs to stay in an
acceptable range. Constraints can also be utilized to guide MPC toward the desired
operation direction. A constraint imposed on compressor speed changing rate can prevent
the compressor from frequent and large scale speed modulation. A wider constraints
during unoccupied hours and narrower constraints during occupied hours on the indoor
air temperature can lead to more energy conservative operation. In many circumstances
of space thermal conditioning, what we want is a band rather than a strict set point.
When constraints are considered in the controller, a quadratic programming problem can
be posed to obtain the solution. Techniques, such as Lagrange multipliers, Kuhn-Tucker
conditions, active set methods, primal-dual method, interior-points, etc., are the
candidates.
3.3 MPC in broad definition
The classic model-based predictive control is restricted to some certain degree. For
example, it requires the internal dynamic model to be linear and generally in state-space
format, the cost function to be quadratic in terms of the outputs and inputs, etc. (Frank et
al 2004). While the solution is fast relying on mature quadratic optimization techniques,
the application is not adequately general. In reality, most systems are nonlinear and may
need arbitrary expression on the cost function. A general model plus an numerical
optimization solver forms a more general MPC control. From a broad point of view, any
control strategy with the three elements can be considered as a form of MPC.
3.3.1 Controller formulation
The formulation can be put as:
( ) y x u f J , , min =

(3-21)

55
( )
( )
U L
U L
U L
u u u
y y y
x x x
u x h
w u x f x
s s
s s
s s
=
=
, 0
, ,

(3-22)
Compared to Eq. 3-1 and 3-2, the general MPC has much more flexible formulation in
terms of the objective function. It can be expressed as almost anything, as long as it has a
clear mathematic expression. The system model, LTI in the classical MPC, is replaced by
a concept of more general model, such as a set of differential algebraic equations, neural
network, or model bank. The modeling methods in classical MPC are also applicable in
the general MPC.
3.3.2 Solving methods
Very often, the dynamic control problem is casted as a numerical optimization
programming problem. The control inputs, or both the control inputs and the states in the
modeling equations are parameterized for the prediction horizon. With only control
inputs parameterized, or termed sequential approach, the system model is remained as a
simulator to obtain the cost and evaluate the constraints for different control actions. For
the second approach, termed collocation or simultaneous approach, both the control and
system model are converted into polynomial form that satisfy the system differential
algebraic equations. The models in general need to be expressed in a set of explicit
equations and follow some strict syntax in order to perform automatic numeric
processing. After that, many methods and optimization packages are then applied to solve
the optimization programming problems. The solution approaches vary a lot in terms of
involved mathematic techniques/principles and computational cost. The selection of a
proper solver can be difficult, which depends on the problem features and the required
performance of the control.
In general, a numerical solver is usually designed for a particular problem type. No
optimization solver is universally applicable, or it works at an extra cost. Meanwhile, the
solution of a general optimization problem (e.g. constrained, nonlinear, mixed integer

56
etc.) can be computationally expensive and no global optimization is guaranteed. The
satisfaction of conditions is only guaranteed at the selected points. Findeisen (2002)
categorized three principle approaches to solving optimal control problems: Hamilton-
Jacobi-Bellman equations, Euler-Lagrange differential equations, and finite
parameterization on the control inputs and constraints. A few literatures available
discussed the related issues involved in general MPC from different points of view (e.g.
Findeisen 2002, Allgower 2004, Michael 1998, Bequette 2007). For optimization
programming, a decision tree is collected providing a summary of the solver candidates
(http://plato.asu.edu/sub/hlores.html). A detailed discussion on the solving methods is out
of the scope of this thesis.
3.4 Summary
Model based predictive control is briefly reviewed in this chapter. The effort is by no
means to be comprehensive but to provide a basic overview of MPC categories,
formulation methods, and solving methods. Compared to MPC in general definition,
research and theory on a classical MPC with LTI model, linear constraints, and quadratic
objective function, is mature. While in reality, most systems are nonlinear and across a
wide range of operation conditions, a well-developed theory for MPC in broad definition,
especially nonlinear MPC, is not available. It is not clear whether a computationally
expensive nonlinear MPC will result in an additional benefit over linear classical MPC.
The stability, robustness, state estimation, disturbance estimation, fast optimization
algorithms, etc, remain as the future research topics.

57
CHAPTER 4 MODEL DEVELOPMENT
4.2 Building thermal system 4.3 Hydronic thermal system 4.4 Hybrid ventilation unit
- Building air and envelop model
- Thermal comfort model
- Models for the hydronic
terminals
- Enthalpy recovery unit and
desiccant wheel unit model
- Vapor compression unit model
- Model of miscellaneous
components
Models serve as the base for the integrated multivariate predictive control in order to
evaluate the evolution of the controlled system and determine the proper actions. The
format and time scale of a model is determined by the purpose of the use. In a building
thermal system, the building construction is usually of the longest time scale and the
equipment has the shortest time scale. This information is very useful in the
determination of the control system and its optimization. For the study of the dynamics of
a given system, generally the parts with the relatively long time constant shall remain in
dynamics, and the parts with the relatively short time constant can be simplified into
quasi-steady-state.
4.1 Introduction
In this chapter, models for diversified components in building systems are developed,
using first-principle method. The overall building system is divided into three sections
based on different time scales, relationship to the occupants, and the functionality. The
indoor air, building envelope, and thermal comfort are grouped as one section since they
directly link to the end-users of buildings and determine how a proactive control system
shall function. The building thermal system is served by the hydraulic system and hybrid
desiccant unit in the investigated test-bed. The modeling is described thereafter. A hybrid
desiccant unit, which provides the ventilation and latent load control to the space, is
additionally modeled. Fig. 4-1 is the picture of the test-bed, which sits on the top of a
four-story building. The geometry and space model is constructed in EcoTec, as depicted
in Fig. 4-2.

58

Figure 4-1: Outdoor view of the IW building

Figure 4-2: Geometry model of the IWn
4.2 Building thermal system
4.2.1 Building air and envelop model
The thermal dynamics of a building is determined mainly by the thermal mass of the
building materials and the inside air. The time constant of a heavy building can be up to
hours. To predict the future evolution of indoor condition for acceptable thermal comfort,
the dynamics need to be represented accurately in the model. Calculating instantaneous

59
space sensible load is critical in dynamic modeling of building thermal systems. Building
envelop includes the external walls, roof, and floor. They are the boundary between the
indoor and the outdoor environment. For energy consumption and thermal comfort
evaluation, the dynamics and interaction of the indoor air and the envelop are of great
importance since they determine the actual thermal load and evolution of indoor air
temperature and influence the thermal comfort. The time constant of a model tells how
fast the building system can response to the change of the mechanical system while the
model provides some insights on what corresponding actions shall be taken for a given
indoor temperature goal.
Among the variables in the model, the interior surface temperatures are unknowns, which
link the heat transfer through the enclosures and the indoor air. Additional models are
needed for the enclosures. They can be divided mainly into two different types: opaque
structure and transparent fenestration. An extensive literature exists on dynamic modeling
of the heat transfer mechanism. The most popular method is the heat balance method
proposed by ASHRAE (Pedersen et al. 2003, Chapter 30, ASHRAE Handbook-
Fundamentals 1997).
The deduction of the space model is based on the energy conservation, the first law of
thermodynamics. The basic assumptions include:
- The air in the space is well-mixed. The variation and temperature distribution in a
given zone is of very little effect on the energy consumption;
- The heat transfer within a construction is one-dimensional. Heat only transfers
perpendicularly to the surface;
- The temperature in the construction is layer-uniform. There is no temperature
difference existing along the length of one layer. Multi-layer structure can be
either lumped as one node or treated as layers lumped individually.
Fig. 4-3 illustrates the mechanism of heat transfer through a piece of building
construction. The thermal capacitance and resistance involved in the heat balance of a
building thermal system is analogue to the capacitance and resistance as in an electric

60
network. Similar to the electric network, the thermal network has the linearity and can be
put in the state space format and be solved easily. This approach is also used in the whole
building system simulation tools (EnergyPlus, Trnsys, etc). Fig. 4-4 depicts such an
analogue of a wall divided into two layers with two capacitances and three resistances. A
heat flux, such as radiation, can be regarded as a current source acting on the inside or
outside capacitance.

Figure 4-3: Heat transfer through a building construction

Figure 4-4: 3R2C analogue heat transfer model
Heat transfer through an opaque wall as shown in Fig. 4-4 can be modeled as:
Rw,e
Rw,m Rw,i
To Tr
Cw,e
Cw,i
Tw,e
Tw,i
qr
o
T
e w
T
,
i w
T
,
r
T
o
h
i
h
k
Q
r
q

61
( ) ( )
m w
i w e w
e w
e w o e w
e w
R
T T
R
T T
dt
dT
C
,
, ,
,
, ,
,

=
(4-1)
( ) ( )
i w
r i w
m w
i w e w
r
i w
i w
R
T T
R
T T
q
dt
dT
C
,
,
,
, , ,
,

+ =
(4-2)
where C is the total thermal capacitance of the given layer (kJ/K), w
T
is the
representative mass temperature of the given layer wall (K),
o
T
is the outdoor air
temperature (K),
w
R
is the thermal resistance between the nodes (K/kW),
r
q is the heat
flux due to the radiation through the fenestration on the internal surface (kW). Subscripts,
e, m, and i denote the external, middle, and internal layer of the construction respectively.
Fig. 4-5 illustrates the thermal network of one construction with an opaque wall and a
transparent window for a building with four rooms.
Rwin
To
room 1
room 2
room 3
room n
window
Rw,e
Rw,m
Rw,i
Cw,e
Cw,i
Cr,1
qr
Tr,1
wall
qplant+qgain

Figure 4-5: Thermal network in a multizone building

62
Convection is the main heat transfer mechanism between the air node, the internal heat
gains, and the enclosures. Based on the heat balance, a first-order equation can be
obtained to model the evolution of room air temperature:

j j gain j w j g j plant
j r
p j r
q q q q q
dt
dT
c V
inf, , , , ,
,
,
+ + + + =

(4-3)
where
is the air density (kg/m

3
), j r
V
, is the total air volume in the room (m
3
), p
c
is the
specific heat of room air (kJ/kg K), j r
T
, is the room air temperature (K), j plant
q
, is the
total heat flux from the mechanical system into the air node (kW), j g
q
, is the total heat
flux into the space from the fenestration (kW), j w
q
, is the total heat flux into the space
from the opaque construction (kW), j gain
q
, is the total internal heat gain from the
occupants, equipment, lighting, and so forth into the room air (kW), and j
q
inf, is the total
heat flux into the room air due to infiltration (kW). Subscript j denotes the j th room.
The thermal components in Eq. 4-3 can be further decomposed to the driven factors, e.g.
temperature differences. For simplicity, some of the subscripts for enumeration are
omitted in the following text.

( )
=
e j r g g j g
T T U A q
, ,

(4-4)

( )
=
w j r w w j w
T T U A q
, ,

(4-5)

( ) 3 /
, , inf, o j r j r j
T T V ACH q =

(4-6)
where g
A
is the surface area of the fenestration (m
2
), g
U
is the lumped average heat
transfer coefficient of the fenestration (kW/m
2
K),
e
T
is the outdoor air temperature (K),
w
A
is the surface area of the opaque construction (m
2
),
w
U
is the lumped average heat
transfer coefficient of the opaque construction (kW/m
2
K),
w
T
is the inside surface

63
temperature of the opaque construction (K), ACH is the air change rate due to
infiltration (/Hr).
The moisture balance in the air node can be similarly modeled. Unlike the air temperature
modeling, the overall space humidity balance is typically lumped as one node without
differentiating the spaces:

( ) ( )
g
l
r o r r plant plant
r
r
h
q
g g V ACH g g m
dt
dg
V + + = 3600 / 2 . 1

(4-7)
where
g
is the average humidity ratio (kg/kg), plant
m
is the total supply air flow rate into
the space from the plant (kg/s),
r
V
is total air volume in the space (m
3
), l
q
is the total
latent heat gain in the space (kW), and g
h
is the specific latent heat of water vaporization
(kJ/kg). Subscripts r and o denote room air and outdoor air respectively.
The heat gain from solar radiation through fenestrations is treated differently due to the
two-step heat transfer mechanism. The heat flux of solar radiation first penetrates
fenestrations, hits the floor, and then is reflected to all other surfaces. The surfaces are
warmed up by the heat flux then gradually transfer the heat into the air node mainly via
convection.
The complex process can be possibly modeled by making the following assumptions:
- The heat flux is evenly distributed among the internal surfaces;
- The surfaces are perfectly diffusive surfaces;
- Multiple reflections are ignored.
The algorithm proposed by Skartveit & Olseth (1987, 1998) based on the sky cleanness
index is used to process the hourly horizontal solar radiation. Normal solar radiation is
then corrected into the plane of building surface at any azimuth and tilt angle with the
method established by Liu & Jordan (1963). The model of the heat flux through a
fenestration has the following relationship:

64

surf r g g r
q A q

= t c
(4-8)
where
r
q is the solar radiation transmitted into the room (kW/m
2
),
c
is the average solar
absorptivity of the surfaces, g
t
is transmissivity of the fenestration, surf r
q
is the total
solar radiation corrected to the plane of the window (W/m
2
).
Natural ventilation is difficult to model accurately with first-principle approach. The
main reason is that the driven factors are complex including the building shape, proximity
of other buildings, wind direction, wind speed, etc. Local wind direction and speed have
high frequency of change and can largely impact the prediction results. A data regression
model is developed for the test-bed by Mahdavi (2008). The model is adopted here under
the condition when the space is running in natural ventilation mode.
The internal heat gains can be regarded as instant load and simplified as a function of
total load intensity:

( )
s s
I f q =

(4-9)

( )
l l
I f q =

(4-10)
where
q
is the total load from the occupants, equipment, and lighting, etc. (kW), and I
is the total load intensity (kW). The subscripts s and l denote sensible and latent part,
respectively.
f
maps the corresponding load based on the schedule.
4.2.2 Thermal comfort model
Thermal comfort is defined as the condition of mind which expresses satisfaction with
the thermal environment and is assessed by subjective evaluation (ASHRAE 55-2004).
As a subjective thing, an absolute satisfaction is almost impossible to achieve in a
conditioning space with more than one occupant. However, a relatively acceptable
thermal comfort to the occupants can be evaluated based on the existing findings.

65
Predictive mean vote (PMV) and predicted percentage of dissatisfaction (PPD) are the
indexes utilized to assess the thermal comfort condition of an indoor environment.
PMV gives the mean value of the votes of a large group of people. Values from -3 to 3
are utilized to express the conditions from very cold to very hot. A value, between -0.5 to
0.5, indicates a good thermal comfort indoor environment. PPD is another index in the
quantitative evaluation of thermal comfort. A PPD value higher than 10% indicates the
indoor environment may not be acceptable. Model of PMV and PPD are quasi-steady-
state and can be expressed as the following equations:
( )
( ) ( ) | |
( ) | | ( )
( ) ( ) | |
( )

+

+ =

a cl c
cl a
M
T T h fcl
T fcl T M
Pa M W M
Pa W M W M
e PMV
4 8
036 . 0
273 10 96 . 3 34 0014 . 0
867 . 5 0173 . 0 15 . 58 42 . 0
000699 . 0 733 . 5 05 . 3
3033 . 0 028 . 0
(4-11)
2 4
2179 . 0 03353 . 0
95 100
PMV PMV
e PPD

=

(4-12)
where M is metabolism (W/m
2
), W is external work (W/m
2
), Pa is partial water vapor
pressure (Pa),
fcl
is ratio of clothed body surface area to nude body surface area,
cl
T
is
surface temperature of clothing, cl
I
is thermal resistance of clothing (clo), and
c
h
is
convectional heat transfer coefficient (W/m
2
K).
The surface temperature of clothing is given by:

( )
( )
( )
( )

)
(
(
+
+

=

a cl c
mrt
cl
cl cl
T T h fcl
T
T
fcl
I W M T
4
4
8
273
273
10 96 . 3
155 . 0 028 . 0 7 . 35
(4-13)

( ) ( )
( )
s +
> +
=
air a cl air
air a cl a cl
c
V T T for V
V T T for T T
h
1 . 12 38 . 2 1 . 12
1 . 12 38 . 2 38 . 2
25 . 0
25 . 0 25 . 0
(4-14)
where
mrt
T
is the mean radiant temperature (K), and
air
V
is the air velocity (m/s).

66
4.3 Hydronic thermal system
The test bed space is mainly conditioned by a group of radiant pipes deployed along the
envelope. Additional radiant panels and cool beams are installed in the offices to provide
extra thermal conditioning capacity during extreme days. Figs. 4-6 to 4-8 show the
pictures of the different water based heating and cooling terminals installed in the space.

Figure 4-6: Picture of a cooled beam

Figure 4-7: Picture of water mullions

67

Figure 4-8: Picture of a radiant ceiling
The water based system of IWn was studied by Gong (2007). His study approached the
heat transfer performance of the water based heating and cooling system with first-
principle based finite difference method. While the approach was good to provide the
insight of the mechanism, it was not suitable for energy and control oriented simulation.
However, the study did tell that the water based terminals could be simplified with
lumped parameters with a very good accuracy.
Lumped parameter models are adopted in this study to simulate the heat transfer of the
water based terminals. The average representative temperature of the terminal mass can
be found by solving:

out in
t
t p t
q q
dt
dT
c M =
,
(4-15)
where
t
M
is lumped mass of the terminal (W/m
2
),
t
T
is the representative temperature of
the terminal (K), t p
c
, is the specific heat of the terminal (kJ/kg K),
in
q
is heat transferred
from the fluid to the mass (kW), and
out
q
is heat dissipated from the terminal to the space
(kW).
The total heat getting into the fluid residing in the terminals is simply:

68

( )
t s w p in
T T c m q =
,

(4-16)
where m is the fluid mass flow rate (kg/s), and s w
T
, is the supply water temperature (K).
The heat transferred from the terminals to the air node can be modeled by:

( )
n
r t t t out
T T A U q =
(4-17)
where t
U
is the overall heat transfer coefficient (kW/m
2
K),
t
A
is the representative
surface area (m
2
), and n is the adjusting coefficient.
The order of the model expressed with Eqs. 4-15 to 4-17 can be increased to achieve a
higher accuracy for the study of the terminals.
4.4 Hybrid ventilation unit
The ventilation of IWn is provided by a hybrid rooftop unit installed outside of the space.
This unit assumes the latent load of the space to avoid surface condensation in cooling
seasons (Yu et al. 2012a). Conventionally, dehumidification is achieved by cooling the
air down to a dew point. This practice couples the sensible load and latent load in the
cooling mode. A great amount of energy is wasted if this operation triggers on reheat to
offset the excessive sensible cooling when the space is in partial load condition (Yu et al.
2012b). Another drawback of this operation is that the humidity ratio of the supply air is
limited by the cooling coil surface temperature. When the space is sensitive to the air
humidity in humid weather, an improvement on the outdoor air processing is needed.
The hybrid rooftop air conditioner provides additional flexibility for the circumstance. It
consists of three essential parts that are, following the flow path of outside air, a total heat
recovery wheel, a revertible direct expansion coil, and a solid desiccant wheel. The
desiccant wheel is regenerated by a parallel air stream with a gas burner and a pulling
fan. Fig. 4-9 shows the picture of the hybrid unit and Fig. 4-10 illustrates the internal
structure.

69

Figure 4-9: Picture of the hybrid ventilation unit
e
n
t
h
a
l
p
y

r
e
c
o
v
e
r
y
w
h
e
e
l
exhaust fan
supply fan
outside air
exhaust air
D
X
c
o
o
l
i
n
g

c
o
i
l
d
e
s
i
c
c
a
n
t
d
e
h
u
m
i
d
i
f
i
e
r
mixing damper
m
i
x
i
n
g

c
h
a
m
b
e
r
supply air
return air
gas burner regeneration fan
outside air
1
2'
1'
2
3
4
5 5'
6
7 7'
8

Figure 4-10: Illustration of the hybrid unit
The outside air first passes through the enthalpy recovery wheel to exchange the heat and
moisture with the exhaust air. Then the air, with/without partially mixing with the return
air, is pulled into a vapor compression unit for further thermal processing. An active
desiccant dehumidifier is installed in one of the two air paths behind the outlet of the
direct expansion coil. It is used in humid weather to further remove the moisture from a
portion of the air. This active dehumidification process can bring the air to an extremely

70
low dew point (e.g. in the range of 10 to 15 grains, 0.65 to 0.97 g) (Fischer 2007). The
dehumidifier is regenerated with the air warmed up by the gas burner. Variable frequency
drivers are installed on all the fans in the unit to adjust the air flow rates when needed. A
pressure modulating valve is used on the natural gas pipe outlet to provide a good control
capability of the regeneration air temperature.
4.4.1 Enthalpy recovery unit and desiccant dehumidifier
Desiccant-coated enthalpy recovery and dehumidification devices have been widely used
due to the potential to enhancing the dehumidification performance of HVAC systems.
With proper maintenance, the device can reduce or eliminate the energy penalty. Zhai
(2008) proposed equation-based mathematic models to study the performance of both the
enthalpy recovery wheel and the active desiccant dehumidification wheel. In her study,
the effect of the practical issues on the operating performance was considered.
Experiments were conducted to evaluate the model and reasonable agreements between
the simulated and measured were identified.
A desiccant-coated enthalpy recovery wheel is a rotary device which links the process air
and the exhaust air to migrate both heat and moisture between the two air streams. While
there are many different structures, a honeycomb structure is widely used to increase the
heat and mass transfer areas for high efficiency. Such a device is usually installed for
outside air intake processing. In a summer cooling season, the outside air can be humid
and warm. Heat and moisture transfer from the outside air into the relatively dryer and
cooler exhaust air due to the gradients. In a winter heating season, the outside air is dry
and cold. The recovery wheel recycles the moisture and heat from the exhaust air into the
outside air stream. A designed efficiency for such a recovery wheel can be as high as
85% (Graham 2009). Fig. 4-11 and 4-12 illustrate the structure of the desiccant enthalpy
recovery wheel installed in the hybrid unit.

71
adsorption
section
n
process air
inlet
process air
outlet
desorption
section
regeneration air
inlet
regeneration air
outlet

Figure 4-11: Illustration of an enthalpy recovery wheel

Figure 4-12: Picture of the active desiccant wheel
4.4.1.1 Full dynamic model
The following assumptions are generally used for developing the model of enthalpy
recovery wheel and desiccant dehumidifier (e.g. Zhai 2006, Nia 2011, Ge. 2010, Harshe
2005):
- The axial heat and mass transfer in the air and the desiccant are considered
negligible;

72
- The mass and heat transfer coefficients are constant. There is no heat or mass
hysteresis in the coating;
- The convective heat and mass transfer rates are represented using the bulk mean
air temperature and humidity;
- Heat conduction in the desiccant is negligible;
- The inlet air conditions are uniform across the wheel surface;
- The heat and moisture transfer through the duct work is small;
- The two air streams are separated and there is no carry-over between the two air
streams.
Based on the assumptions, the following deduction can be made with the application of
energy and mass balance laws.
The moisture balance of the air stream is:

( ) 0 =
c
c
+
c
c
+
t
A
x
uA p h
vg vg
vm vg m
e e
e e

(4-18)
where
m
h
is the mass transfer coefficient (m/ s),
p
is the perimeter length of the air flow
channel (m), vg
e
is the concentration of water vapor in the air (kg/m
3
),
vm
e
is the
concentration of water vapor in the desiccant matrix (kg/m
3
),
u
is the air velocity (m/s),
A is the cross sectional area of the airflow channel (m
2
),
x
is the distance in axial
direction (m), and t is time (s).
It is assumed that the moisture load in the desiccant is linearly related to the moisture
concentration in the desiccant matrix. The moisture balance of the desiccant is described
as:

( ) 0 =
dt
dw
A p h
m
m m vm vg m
e e

(4-19)

73
where
m
A
is the cross sectional area of the desiccant matrix (m
2
) and
m
w
is the water
content of desiccant material (kg moisture/kg dry desiccant).
Additional equations are needed to link the water content of desiccant material to the
humidity ratio with the desiccant. The total derivative of
vm
w
in terms of the temperature
and water vapor concentration is:

m
m
m
vm
m
vm
vm
m
m
dT
T
w w
d
w
dw
|
|
.
|
\
|
c
c
+
c
c
c
c
+
c
c
c
c
=
e
e
e

(4-20)
where
is the relative humidity of desiccant matrix,

m
T
is the desiccant material
temperature (K).
The adsorption capacity of the desiccant material depends on both the temperature and
the humidity of the surrounding air. A correlation is needed so that the capacity becomes
a known parameter when the air temperature and humidity are given. Steich (1994)
studied the performance of rotary enthalpy exchangers. The following equations can be
used:

(
(
|
.
|
\
|
+
(
(
|
.
|
\
|
=
2 2
3140
exp 242 . 0
8590
exp 106 . 0
b b
w

(4-21)

ln RT b = (4-22)
Relative humidity and saturation pressure can be calculated by (Zhai 2006):

sat vm
vm
,
e
e
=

(4-23)

74
With the ideal gas law nRT PV = , Eq. 4-23 becomes
sat vm
v vm
p
T R
,
=
. sat vm
p
, is a function of
temperature only. With the same assumptions made by Zhai (2006), we can obtain the
saturation vapor pressure equation:

|
|
.
|
\
|
+
- =
273
5196
exp 10 12 . 1
1
11
, m sat vm
T p

(4-24)
Combining Eq. 4-21 to 4-24, the relative humidity can be achieved by:

( )
|
|
.
|
\
|
+
+ - =

273
5196
exp 273 10 09 . 4
9
m
vm m
T
T

(4-25)
Eventually, the terms in Eq. 4-20 can be replaced by taking the corresponding
derivatives.
The sensible energy balance of the air stream is described as:

( ) 0
, ,
=
c
c
c
c

t
T
c A
x
T
c uA T T hp
g
g p g
g
g p g g m

(4-26)
where h is the heat transfer coefficient (W/m
2
K),
m
T
is the desiccant matrix temperature
(K), g
T
is the air temperature (K), g
is the air density (kg/m

3
), and g p
c
, is the air specific
heat (J/kg K).
The energy balance of the desiccant composite is described as:

( )
( ) ( ) 0
, ,
2
2
=
c
c
+
+ +
c
c
t
T
c A c A T T hp
dt
dw
A A q
x
T
A k
m
sub p sub sub m p m m g m
m
sub sub m m st
m
sub sub

(4-27)

75
where k is the thermal conductivity (W/(m K)), and
st
q
is the adsorption heat (J/kg).
Subscript sub denotes the substrate of the desiccant material.
The first term in Eq. 4-27, which represents the heat conduction through the substrate of
the coating, is typically small and can be ignored in the analysis. Rearrange Eq. 4-19 for
dt
dw
m
and submit it into Eq. 4-27:

( )
( ) ( ) ( ) ( )
m m vm vg m sub sub m m st g m
m
sub p sub sub m p m m
A p h A A q T T hp
dt
dT
c A c A
e e

/
, ,
+ +
= +

(4-28)
The heat transfer coefficient is given by (Bejan 1984):

eff
g
D
Nu k
h =

(4-29)
where g
k
is the air thermal conductivity (W/ (m K)), Nu is Nusselt number, and eff
D
is
the hydraulic diameter of the channel (m).
The mass transfer coefficient is determined by applying the heat and mass transfer
analogy (Incropera and Dewitt 1996):

Le C
h
h
g g
m
=

(4-30)
where g
is the constant air density (kg/m

3
) and Le is the Lewis number.
The above model governs the mass and heat transfer of the individual channel. Since only
one dimensional heat and mass transfer is considered, the output of an individual channel
in a section can be regarded as the same of the wheel in the given section. To simulate the
performance of the enthalpy and desiccant wheel in a continuous rotating status, the
effect of rotation speed shall be included. The handling method is to discretize the

76
channel into small continuous channels and the outlet of the previous channel becomes
the inlet of the one behind.
The enthalpy and desiccant wheel is a rotation device A channel along the boundary for
adsorption at one time step can be the regeneration channel at the next time step with the
reversed flow direction. To simulate the dynamics due to the rotation, the states of the
moisture and temperature of the desiccant wheel shall be stored up and shifted between
the two modes. Time step of such a switch is picked based on the controlled rotary speed
since the discretization of the wheel in the direction of thickness and the angular surface
is needed to obtain the accurate results. Figs. 4-13 and 4-14 illustrate the mechanism
involved in the heat and mass transfer of a rotating desiccant wheel. The discritization as
shown in Fig 4-14 is partially adopted from (Nia 2006).
rotation direction
0
30
180
Inward
Outward

Figure 4-13: Illustration of the rotary wheel and sections

77
m
e
s
T
i
e
i
T i
e
i
T
o
e
o
T
m
e
s
T
m
e
s
T
m
e
s
T
m
e
s
T
m
e
s
T
e
l
e
m
e
n
t

1
m
e
s
T
i
e
i
T
o
e
o
T
m
e
s
T
m
e
s
T
m
e
s
T
m
e
s
T
m
e
s
T
e
l
e
m
e
n
t

1
o
e
o
T
i
e
i
T
o
e
o
T
m
e
s
T
i
e
i
T i
e
i
T
o
e
o
T
m
e
s
T
m
e
s
T
m
e
s
T
m
e
s
T
m
e
s
T
e
l
e
m
e
n
t

1
m
e
s
T
i
e
i
T
o
e
o
T
m
e
s
T
m
e
s
T
m
e
s
T
m
e
s
T
m
e
s
T
e
l
e
m
e
n
t

1
o
e
o
T
i
e
i
T o
e
o
T
Angel 1
e
l
e
m
e
n
t

n
e
l
e
m
e
n
t

n
e
l
e
m
e
n
t

n
e
l
e
m
e
n
t

n
Angel m1
Angel m1+1
Angel m1+m2
discretize in thickness
d
i
s
c
r
e
t
i
z
e

i
n

a
n
g
l
e

a
n
d

r
o
t
a
t
i
o
n

Figure 4-14: Discretization of the wheel in thickness and angle
4.4.1.2 Simplified fast dynamic model
The previous full model starts from the basic heat and mass transfer between the
desiccant matrix and the air. It can be utilized to obtain the detailed states of the
temperature and humidity of the desiccant wheel and the air along the flow and for the
different angles. Since the time step is determined by the discretization, the simulation
can take a significant amount of computational memory and time. Therefore, they are not
directly suitable for online control-oriented study.
For building energy simulation and equipment control, the physical states of the matrix
and the distribution within the matrix are not of the concern. Another approach, which is
based on lumped parameter method, can be utilized to obtain the bulk air temperature and
humidity from the outlet of the process and regeneration side.

78
In the moisture balance Eq. 4-19, setting
m m
m
v
A
p h
t =
, it can be seen this term is fixed as
the desiccant property is time invariant. The vapor concentration in the desiccant is a
first-order dynamic response to the difference between the vapor concentration in the air
stream and the desiccant wheel. The same time constant is applicable to the dynamic of
the bulk mean concentration of water vapor in the air where the moisture concentration
along the stream is not considered. So, the air stream moisture concentration shows first-
order dynamic behavior.
In Eq. 4-27, the second heat generation term is due to the moisture adsorption. Assuming
that the heat of adsorption is totally released into the desiccant composite, this term
disappears as well. Similarly, set
sub p sub sub m p m m
T
c A c A
hp
, ,
1

t
+
=
for Eq. 4-27 and
g p g
T
c A
hp
,
2
t =
for Eq. 4-26. As the expression shows, the two time constants in the first
order differential equations are time-invariant and can be considered as a design
parameter instead of control parameter. The deduction holds for both the process air and
the regeneration air stream.
From the physical model and previous deduction, it is known that the moisture and heat
transfer of the enthalpy recovery wheel and desiccant wheel are governed by the wheel
characteristics. The dynamic process can be expressed as a first-order system with fixed
time constants.

79
4.4.1.2.1 Enthalpyrecoverywheel

Figure 4-15: Illustration of enthalpy recovery wheel work process
An enthalpy recovery wheel aims to ensure the thorough moisture and heat exchange
between the two air streams. The process is illustrated in Fig. 4-15. The operation of an
enthalpy recovery wheel pulls the outdoor air and exhaust air toward each other. The
calculation of the sensible and latent heat transfer efficiency in the design and partial load
conditions is, as a set of look up table, provided in the manufacturers product category.
The process gets the base recovery effectiveness based on the wheel face velocity. The
efficiency is further adjusted for unbalanced air flows.
A more general model of the enthalpy heat recovery wheel is provided in the
documentation of EnergyPlus (2010) and ASHRAE (2008). To obtain the exchange
efficiency for any given air flow condition, the efficiency shall be provided for the
balanced air flow condition with air flow rates at 75% and 100% of the nominal rate. The
operating conditions for obtaining the exchanger performance are given by ARI (2003).
The efficiency dependency on the rotation speed is made based on European Standard for
laboratory testing of air-to-air heat recovery devices.
Temperature
H
u
m
i
d
i
t
y
exhaust air in
exhaust air out
outdoor air in
outdoor air out

80

( ) ( ) n wrf
m
m m
rate
s s s s

|
|
.
|
\
|
|
|
.
|
\
|
|
|
.
|
\
|
+
+ = 25 . 0 75 . 0
2
2 1
% 75 , % 100 , % 75 ,

c c c c

(4-31)

( ) ( ) n wrf
m
m m
rate
l l l l

|
|
.
|
\
|
|
|
.
|
\
|
|
|
.
|
\
|
+
+ = 25 . 0 75 . 0
2
2 1
% 75 , % 100 , % 75 ,

c c c c

(4-32)

( )
in in
p
p
s in out
T T
c m
c m
T T
, 1 , 2
1 ,
min ,
, 1 , 1

|
|
.
|
\
|
+ =
c (4-33)

( )
in in
p
p
l in out
c m
c m
, 1 , 2
1 ,
min ,
, 1 , 1
e e c e e
|
|
.
|
\
|
+ =
(4-34)
( )
2 , 2 1 , 1 min ,
, min
p p p
c m c m c m =

(4-35)
where
c
is the energy transfer effectiveness, m is the air mass flow rate (kg/s), T is the
air temperature (K),
rate
n
n
r =
as a function of the rotation speed is the wheel rotation
factor (rpm/rpm),
e
is the air humidity ratio (kg/kg), and p
c m
is the heat capacity rate
(W/K), Q
is the heat flow (W). Subscript s denotes sensible portion, l means latent
portion, 75% and 100% means the rating conditions according to ARI 2003, 1 denotes the
process air stream, 2 means the exhaust or regeneration air stream, and out denotes the
outlet, in means the inlet.
The process air outlet condition can be expressed as:

( )
out in in
p
p
s in
out
T
T T T
c m
c m
T
dt
dT
, 1 , 1 , 2
1 ,
min ,
, 1
, 1

|
|
.
|
\
|
+ =
c t
(4-36)

( )
out in in
p
p
l in
out
v
c m
c m
dt
d
, 1 , 1 , 2
1 ,
min ,
, 1
, 1
e e e c e
e
t
|
|
.
|
\
|
+ =
(4-37)

81
To this end, a simplified fast dynamic model for an enthalpy recovery wheel is obtained.
4.4.1.2.2 Activedesiccantwheel
Temperature
H
u
m
i
d
i
t
y
process air in
reg air out
reg air in
process air out

Figure 4-16: Illustration of active desiccant wheel dehumidification process
Unlike the enthalpy recovery wheel, the partition and the rotation speed of an active
dehumidification unit are designed to drive the moisture and heat move closer along the
isenthalpic line. The process is illustrated in Fig. 4-16. In dehumidification process, the
change of the process air and the regeneration air is along the isenthalpic line. The
process air passes through 3/4 of the desiccant wheel and the regeneration air passes
through 1/4 of the wheel. The operation and dynamic modeling is much more complex
than enthalpy recovery wheel.
An analogue method is adopted to obtain the fast dynamic model for the moisture and
heat transfer of an active desiccant wheel. The method utilizes potential efficiency factors
and mimics the modeling approach of a heat exchanger. Two artificial efficiency factors
are introduced as
1
F and
2
F . More details of the physical meaning are referred to (Banks
1972, 1985). Factor
1
F lines lie close to adiabatic saturation lines, and
2
F lines lie close
to the relative humidity line on the psychometric chart (Maclaine-Cross and Banks 1972,

82
Neti and Wolfe 2000, Jurinak 1982). The governing equations of an active desiccant
wheel based on effectiveness can be put as:

( )
8624 . 0
49 . 1
, 1
1000
344 . 4
273
2865
|
.
|
\
|
+
+
=
i
i
i
T
F
e

(4-38)

( )
07969 . 0
49 . 1
, 2
1000
127 . 1
6360
273
|
.
|
\
|
+
=
i i
i
T
F
e
(4-39)

( ) 2501 84 . 1
1000
006 . 1
, 1
+ + = =
i
i i
T h F
e
(4-40)

( )
( )
8 . 0
49 . 1
, 2
1 . 1
6360
273
i
i
i
T
F e +
+
=
(4-41)

in in
in out
F
F F
F F
1 , 1 2 , 1
1 , 1 1 , 1
1
= c
(4-42)

in in
in out
F
F F
F F
1 , 2 2 , 2
1 , 2 1 , 2
2
= c
(4-43)
where
c
is the artificial transfer effectiveness, i refers to the air node of process air inlet
(1,in), process air outlet (1,out), and regeneration air inlet (2,in).
The highly nonlinear expression for artificial potential factors can be linearized around
the operation points:

out
F F
out
T
T
b a b a
A b A b A b A b
dt
dT
, 1
1 5 5 1
4 1 2 5 3 2 1 1 1 5
, 1
+
=
c c
t
(4-44)

( )
out
F F
out
v
b a b a
A a A a A a A a
dt
d
, 1
1 5 5 1
4 1 2 5 3 2 1 1 1 5
, 1
e
c c
e
t

+
=
(4-45)

83

( ) ( )
2 , 1 4 , 1 4 , 2 3 , 2 3 1
c b T a b T a A
in in in in
+ + + = e e

(4-46)

( )
1 , 1 2 , 1 2 2
c b T a A
in in
+ = e

(4-47)

( ) ( )
4 , 1 8 , 1 8 , 2 7 , 2 7 3
c b T a b T a A
in in in in
+ + + = e e

(4-48)

( )
3 , 1 6 , 1 6 4
c b T a A
in in
+ = e

(4-49)
where:
ss
out
out
T
F
a
, 1
1 , 1
1
c
c
=
,
ss
out
out
F
b
, 1
1 , 1
1
e c
c
=
,
ss
in
in
T
F
a
, 1
1 , 1
2
c
c
=
,
ss
in
in
F
b
, 1
1 , 1
2
e c
c
=
,
ss
out
out
T
F
a
, 1
1 , 2
3
c
c
=
,
ss
in
in
F
b
, 2
2 , 1
3
e c
c
=
,
ss
in
in
T
F
a
, 1
1 , 1
4
c
c
=
,
ss
in
in
F
b
, 1
1 , 1
4
e c
c
=
,
ss
out
out
T
F
a
, 1
1 , 2
5
c
c
=
,
ss
out
out
F
b
, 1
1 , 2
5
e c
c
=
,
ss
in
in
T
F
a
, 1
1 , 2
6
c
c
=
,
ss
in
in
F
b
, 1
1 , 2
6
e c
c
=
,
ss
in
in
T
F
a
, 2
2 , 2
7
c
c
=
,
ss
in
in
F
b
, 2
2 , 2
7
e c
c
=
,
ss
in
in
T
F
a
, 1
1 , 2
8
c
c
=
,
ss
in
in
F
b
, 1
1 , 2
8
e c
c
=

ss
in
ss
out
F F c
1 , 1 1 , 1 1
c c =
,
ss
in
ss
in
F F c
1 , 1 2 , 1 2
c c =
,
ss
in
ss
out
F F c
1 , 2 1 , 2 3
c c =
,
ss
in
ss
in
F F c
1 , 2 2 , 2 4
c c =
.
The simplified desiccant model based on Eqs. 4-38 and 4-39 has been adopted by Trnsys
and, the model based on Eqs. 4-40 and 4-41 has been implemented in SPARK. Panaras et
al. (2010) experimentally investigated the artificial transfer effectiveness factors and
identified that the two factors can be regarded as constant pairs in a very wide range of
different air flow rates. The drawback of this approach is that the model needs to be
constantly linearized when the actual operation points move away from the vicinity of
system linearization point.
An alternative simplified modeling approach specifically for control purpose was
proposed by Zhao et al. (2005). The model lumped together the active desiccant wheel

84
and heat exchanger. The manipulated variable is reduced to be the power input of the heat
exchanger. The model is constructed based on the following facts:
- The heat exchanger is a heat source and the only root drive for dehumidification;
- For a given active desiccant wheel unit, the optimal functionality is based on fixed
active desiccant wheel structure and air flow ratio. Given a fixed heat input, the
higher the inlet process air humidity is, the higher the outlet process air humidity;
- Without turning on the heat exchanger, the process air temperature and humidity
do not change.

w k T T
in out
+ =
1 , 1 , 1

(4-50)

w k
in out
e Rh Rh

=
2
, 1 , 1

(4-51)
where Rh is the air relative humidity,
w
is the heat exchanger power input, and k is the
regression coefficients.
The dynamics is then established based on first-order system assumption.
4.4.2 Vapor compression unit
Vapor expansion systems are a type of thermal dynamic machinery, which utilize the
compressible refrigerant fluid to transfer the heat between the high and low temperature
sources. Since the units usually have a compact size and require less duct work
construction, they are widely utilized in commercial and public buildings to provide
inexpensive, efficient heating and cooling with minimal outdoor air (Westphalen and
Koszalinksi 2001).
Although the structure varies a little bit from one to one, the closed loop cycle of a vapor
compression unit includes the following main components: compressor, condenser,
expansion device, and evaporator. The cycle can be illustrated in a pressure-enthalpy
chart as shown in Fig. 4-17. This cycle moves the refrigerant from the vapor phase to two

85
mixed, and then liquid, and transfers the heat between the high temperature source and
low temperature source. When it works for heating purpose, it is termed heat pump. From
the outlet of the evaporator to the outlet of the compressor is generally regarded as an
isentropic or polytropic process where the superheated refrigerant in vapor phase is
compressed and then enters the condenser. In a condenser, the refrigerant vapor at a
higher temperature goes through the heat exchanger where it passed the heat to the fluid
outside of the condenser and cooled down into a liquid. As shown in Fig. 4-17, this ideal
process is isobaric, where the pressure of the vapor and liquid remains constant. The
liquid could be saturated or super-cooled before it goes through a throttling device to
reduce the pressure. An expansion device can be a thermal expansion valve, an electronic
expansion valve, or a long capillary pipe. The expansion process is considered isentropic
in an ideal case analysis. This results in a mixture of liquid and vapor at a lower
temperature and pressure. In the evaporator, the cold mixture gets vaporized again in a
constant pressure manner and removes the heat from the ambient.
P
H
Evaporator
Condenser
C
o
m
p
r
e
s
s
o
r
E
x
p

d
e
v
i
c
e
Liquid
Vapor
Two phase

Figure 4-17: Cycle of refrigerant in a vapor compression unit
The difficulty of modeling the transient of the closed loop vapor compression cycle is the
associated phase and system states change with the refrigerant. A couple of literatures are
identified providing the excellent review on the transient modeling. ASHRAE sponsored

86
a project in 2002 to review the dynamic models of vapor compression equipment
(Bendapudi and Braun 2002). The report summarized the common simplification
assumptions, modeling approaches, and provided a comprehensive review on the
modeling literatures up to 2002. It was found that almost invariably the heat exchangers
and expansion devices were modeled from first-principles. Apparently, modeling of the
thermal dynamic process of the refrigerant in the two heat exchangers (evaporator and
condenser) is the main challenge. The actuating components are generally considered
instant. The report summarized the modeling approaches of the heat exchangers into the
following three types: phase-dependent moving boundary method, phase independent
finite difference methods, and moving boundary with finite difference method. Ding
(2007) reviewed the recent developments in simulation techniques for vapor compression
refrigeration systems and further summarized the modeling algorithms.
The following assumptions are made for the moving-boundary-lumped-parameter
dynamic modeling of the vapor compression unit in the hybrid air-conditioner
(Bendapudi and Braun 2002):
- The heat exchangers are long and thin and the refrigerant flow remains one
dimensional and homogenous within the sections. The liquid and vapor phases are
considered to be in thermal equilibrium, and moving at the same velocity;
- Axial conduction in unidirectional pipe is negligible as compared to the heat
transfer through the heat exchangers;
- Thermal resistances of metallic elements in the system are negligible in
comparison with their capacitances;
- Effects of pressure wave dynamics and the pressure drop across the heat
exchangers are negligible;
- Expansion process is isenthalpic with no energy lost across the expansion device;
- Compression of the fluid is adiabatic with an isentropic efficiency.
The compression cycle with three fluid regions in the condenser, two fluid regions in the
evaporator, and the compression and expansion device is illustrated in Fig. 4-18. The full
nonlinear dynamic model is developed in the following sections for the direct expansion

87
unit. The manipulated variables for the direct expansion unit are: the compressor speed,
expansion valve opening, and condenser fan speed. To fully model the unit, the mass,
energy balance and the state equations are needed. Compared to the dynamics of the heat
exchangers, the time constant for the expansion valve and the compressor are much
shorter, and can be considered as instant (He et al. 1997, Rasmussen 2002). The same
holds true for the fan air flow control.
x = 1 x = 0 0 < x < 1
CV1
CV2a
CV2w-1 CV2w-2 CV2w-3
CV2-1 CV2-2 CV2-3
CV3
0 < x < 1 x = 1
CV4-1 CV4-2
air
air
CV4w-1
CV4w-2
CV34a
b
c d
e
f
a
3 4

Figure 4-18: Illustration of full model
4.4.2.1 Compressor
The compressor used in the unit is a variable speed scroll compressor. It houses two
interleaving scrolls to pump, compress or pressurize fluids with one of the scrolls fixed
and the other orbits eccentrically. It is well-known for its high efficiency and smooth
operation. A detailed model was studied by (Morishita et al. 1984).

88

Figure 4-19: Scroll type vapor compressor
(Carrier, 2004)

The compression process is considered a polytropic process. To fully model the scroll
compressor, the mass and energy balance equations are needed.

v co co a co
S V m q =

(4-52)
where co
m
is the refrigerant mass flow rate through the compressor (kg/s), d
is the fluid
density at the input of the compressor (kg/m
3
), co
V
is the volumetric displacement of
compressor (m
3
), co
S
is the compressor rotational speed (rps), and v
q
is the volumetric
efficiency of the compressor.
The swept volume of the compressor can be obtained from the manufacturer datasheet, or
using the following estimation (Chen 2005):

|
.
|
\
|
=
R
e
R
e
l r V
a co
2
2
t

(4-53)
where
l
is the stoke of cylinder (kg/s), a
r
is the radius of rotor (m),
e
is the rotor
eccentricity, and R is the radius of cylinder (m).

89
A regression model is adopted for the volumetric efficiency:

|
|
.
|
\
|
+ =
a
b
v
p
p
a a
2 1
q

(4-54)
where b
p
is the pressure of refrigerant at outlet (Pa), and a
p
is the pressure of refrigerant
at inlet (Pa).
1
a
and
2
a
are constants.
A linear model can be used to evaluate the power consumption of the compressor:

( )
1 2 1
b a p a w
b co
+ =

(4-55)
where co
w
is the compressor power input (kJ),
1
a
,
2
a
and
1
b
are constants.

co a b
w h h + =

(4-56)
where b
h
is the specific enthalpy of refrigerant at outlet (kJ/kg), and a
h
is the specific
enthalpy of refrigerant at inlet (kJ/kg).
4.4.2.2 Expansion valve
An expansion device is an electronic or hydraulic expansion valve where the mass flow
rate is controlled by adjusting the valve openness. The function is to hold a certain degree
of evaporator superheat. For a thermal expansion valve, as shown in Fig. 4-20, the
variation of the compressor suction line surface temperature leads to the deformation of
the diaphragm to change the valve openness. It is analogue to a pure proportional
controller. For an electronic expansion valve, as illustrated in Fig. 4-21, the operation is
directly controlled by the electronic signals determined by a control law (He 1985).

90

Figure 4-20: Thermal expansion valve
(CarsonDunlop.com)

Figure 4-21: Electronic expansion valve

Therefore the mass flow rate of the refrigerant through the expansion valve can be put as
(Li and Braun 2009):

( )
e d v v v v
p p C A m =

(4-57)

91
where v
m
is the mass flow rate of refrigerant through the expansion valve (kJ/kg), v
A
is
the effective valve area (m
2
), v
C
is the discharge coefficient, v
is the density of the

refrigerant through the valve (kg/m
3
), and d
p
and e
p
are the pressure of refrigerant at the
inlet and outlet of the valve (Pa).

( )
dsh sh r v
T T G A A A A + =

(4-58)
Where
r
A
is the set point of expansion valve opening area corresponding to dsh
T A
(m
2
),
G

is the gain from the control law, and dsh
T A
is the measured super heat degree between
'
f
T

and ef
T
.
Because a Reynolds number (Re) of single-phase working fluid flow is more than 104,
v
C
can be considered as a constant whose value is between 0.61 and 0.62 (Hu and Meng
2011). The enthalpy after the throttling process through the valve can be assumed to be
equal to that at the inlet:

e d
h h =

(4-59)
4.4.2.3 Full dynamic model
MacArthur and Grald (1989, 1992) deduced the detailed partial differential equations that
govern the conservation of mass, momentum, and energy of refrigerant within a heat
exchanger tube. The heat exchanger links the refrigerant side and the air side to form the
system. An additional energy equation could be established for the heat exchanger wall:

0 =
c
c
+
c
c
z
u
t

(4-60)

( ) ( )
2
2
2
1 4
u
D z
p u
t
u
i

=
c
+ c
+
c
c

(4-61)

92

( )
( )
r w i
i
T T k
D z
uh
t
p h
=
c
c
+
c
c 4

(4-62)

( ) ( ) ( )
w a o o w r i i
w
w
p
T T k D T T k D
t
T
A C + =
c
c
t t

(4-63)
where
is the density of the refrigerant (kg/m

3
),
u
is the axial velocity of the refrigerant
flow in the tube (m/s),
p
is the pressure of refrigerant (Pa), i
D
and o
D
are the inside and
outside diameter of the tube (m),
is the friction coefficient,

h
is the enthalpy of
refrigerant (kJ), T is the temperature (K) , i
k
is the heat transfer coefficient between the
tube wall and the refrigerant per unit area, o
k
is the heat transfer between the tube wall
and the air, p
C
the enthalpy of the tube wall (kJ/kg), and A is the cross section area of
unit length tube wall (m
2
). Subscript
w
is the wall,
r
is the refrigerant, and
a
is the air.
The set of equations can be solved numerically by discretizing the heat exchanger tube
into sections, and serves as the base for further analysis. However, as the deduction made
for the enthalpy wheel, the computational effort of this approach is high and does not suit
for control purpose. A moving interface lumped approach can be deduced based on the
engineering analysis of the refrigerant within the condenser and the evaporator without
losing the modeling accuracy. The partial differential equations can be turned into
ordinary differential equations by utilizing Leibnizs equation as below (He 1996,
Rasmussen 2002):

( )
( )
( )
( )
( )
( )
( ) ( )
( ) ( )
( ) ( )
( ) ( )
dt
t z d
t t z f
dt
t z d
t t z f dz t z f
dt
d
dz
t
t z f
t z
t z
t z
t z
1
1
2
2
, , ,
,
2
1
2
1
+
|
|
.
|
\
|
=
c
c
} }
(4-64)
Wedekind et al. (1978) first proposed a lumped-parameter method to obtain relatively
lower-order models for the purpose of control. As illustrated in Fig. 4-18, the refrigerant
in a condenser can be divided into three different regions and the evaporator can be
divided into two regions (He 1996, Rasmussen 2002). In the following two sections, the
models are presented for the condenser and evaporator in ordinary differential equation
form based on the literatures.

93
4.4.2.3.1 Condenser

Figure 4-22: Picture of the condenser

Fig. 4-22 gives the picture of the condenser. It includes the header, tubes and fins. A
superheated vapor section, a liquid vapor mixture section and a sub-cool liquid section
are considered adequate to describe the thermal dynamics of the refrigerant in a
condenser. The pressure is assumed uniform along the entire condenser tube. The states
of the condenser can be expressed as a vector:

| |
T
w w w d cd cd
T T T h P L L x
3 2 2 2 1 2 2 2 1 2
=

(4-65)
where
1 2
L
and
2 2
L
are the length of condenser tube where the refrigerant is in superheat
and two-phase, (m), respectively, cd
P
is the refrigerant pressure within the condenser
(kPa), d
h
is the enthalpy of the refrigerant at the outlet of the condenser (kJ), 1 2 w
T
, 2 2 w
T
, and
3 2 w
T
are the wall temperature of the condenser (K).
The overall model is seventh-order in state-space form:

94

) , (
cd cd cd cd cd
u x f x A =

(4-66)
Where:

(
(
(
(
(
(
(
(
(
=
77 72 71
66
55 51
43 42 41
34 33 32 31
23 22 21
13 11
a a a
a
a a
a a a
a a a a
a a a
a a
A
ed
(4-67)
( ) ( )
( ) ( )
( ) ( )
( ) ( )
( ) ( )
( ) ( )
(
(
(
(
(
(
(
(
(
+
+
+

+
+
+
=

3 2 2 3 2 3 2 , 3 2 ,
2 2 2 2 2 2 2 , 2 2 ,
1 2 2 1 2 1 2 , 1 2 ,
3 2 , 3 2 3 2 3 2 , 3 2
2 2 , 2 2 2 2 2 2 , 3 2 2 2
1 2 , 1 2 1 2 1 2 , 2 2
w a o o w r i i
w a o o w r i i
w a o o w r i i
d c
r w i i d d
r w i i d
r w i i c c
cd
T T D k T T D k
T T D k T T D k
T T D k T T D k
m m
T T L D k h h m
T T L D k h h m
T T L D k h h m
f
t t
t t
t t
t
t
t

(4-68)
where
k
is the heat transfer coefficient (kJ/m
2
K), and D is the tube diameter (m).
Subscripts i and o denote the internal and external of the tube. The numeric subscripts
denote the sections as shown in Fig. 4-19. The elements in the A matrix could be deduced
theoretically through linearization of first-principle model, or found in the literature.

95
4.4.2.3.2 Evaporator

Figure 4-23: Picture of the evaporator

Fig. 4-23 shows the picture of evaporator. The refrigerant in an evaporator goes through
an inverse thermodynamic process than in a condenser. The sub-cooled low pressure
liquid enters the evaporator, exchanges the heat through the wall with the external heat
source, and gets evaporated into vapor. A two-region lumped parameter model is
generally adopted for the simulation of the thermodynamics (He 1996). The pressure of
refrigerant is assumed uniform along the entire evaporator tube, therefore the momentum
equation drops out.
The states of the evaporator can be expressed as:

| |
T
w w f ef ef
T T h P L x
2 4 1 4 1 4
=

(4-69)

96
where
1
L
is the length of evaporator tube where the refrigerant is in liquid, (m), ef
P
is the
refrigerant pressure within the evaporator (bar), f
h
is the enthalpy of the refrigerant at
outlet (kJ), 1 4 w
T
and 2 4 w
T
are the wall temperature of the evaporator (K).
The fifth-order model of the evaporator in state-space form is:

) , (
ef ef ef ef ef
u x f x A =

(4-70)
Where:

(
(
(
(
(
(
=
55 51
44
33 32 31
23 22 21
12 11
a a
a
a a a
a a a
a a
A
ef
(4-71)
( )
( )
( ) ( )
( ) ( )
(
(
(
(
(
(
+
+

+
+
=

2 4 34 2 , 2 4 2 4 , 2 4 ,
1 4 34 1 , 1 4 1 4 , 1 4 ,
2 4 , 2 4 2 4 2 4 , 2 4
1 4 , 1 4 1 4 1 4 , 2 4
w a o o w r i i
w a o o w r i i
f e
r w i i f f f
r w i i e e e
ef
T T D k T T D k
T T D k T T D k
m m
T T L D k h m h m
T T L D k h m h m
f
t t
t t
t
t

(4-72)
The matrix elements can be deduced in the way similar to the condenser model. The
subscripts denote the sections as shown in Fig. 4-18.
4.4.2.4 Simplified fast dynamic model
The full nonlinear lumped dynamic model is widely adopted to evaluate the dynamics of
vapor compression units. However, it seems obvious that, compared to the air side
dynamics, the refrigerant side has much faster dynamics and can be considered instant.
The many state variables of the refrigerant are not of the concern of the overall unit
control in this study. Meanwhile, the detailed refrigerant side model requires constant
evaluation of the refrigerant properties and has a major impact on the model runtime.

97
For the control of a vapor compression unit, the main control variables are the
compressor speed, condenser fan speed, and electronic expansion valve openness. The
following deduction is made to obtain the dynamic model of the vapor compression unit.
Similar simplification for the transient model and control on the refrigerant side can be
found in (Vargas and Parise 1995, Deng 2000, Chi and Didion 1982). The deduction
divides the heat transfer around the condenser and the evaporator into the parts of
refrigerant, metal container, and the external air. Fig. 4-24 illustrates the physical loop on
a p-h chart.
e a f a f a
p p h h = = = , ,
'

cd b b b
p p h m = , ,
a
b
c
d
e
T
a cd
T
,
f' f
e
( )
cd cd cd
x T f h , =
w cd
Q
,
a cd
Q
,
a cd
T
,
b c b c
h h m m = = ,
cd cd
T p ,
( )
ef ef ef
x T f h , =
P
h
ef ef
T p ,
d
cd d l cd d v d
p p h h m m
= = = , ,
,

d d v
p m , ,
w
Q
4
a
T
34
3
T
4
T
ef e d e v e
p p h h m m = = = , ,
1 34 a
Q
2 34 a
Q
c
b
a
f'
d
e
f
' '
,
f f
T
4

-

4
0
b
a
r
160 - 450kJ/kg

Figure 4-24: Illustration of fast model dynamics

98
4.4.2.4.1 Condenser
The mass change rate of refrigerant in the condenser is the function of the mass flow rate
from the compressor and the mass flow rate to the expansion valve.

d c
cd
m m
dt
dm
=

(4-73)
where c
m
is the mass flow into the condenser (kg/s), d
m
is the mass flow out of the
condenser (kg/s), and cd
m
is the total refrigerant in the condenser (kg).
The enthalpy change of the refrigerant mixture in the condenser is the difference of the
inlet and outlet enthalpy and the heat dissipated through the metal wall:

w cd d d c c
cd
cd
Q h m h m
dt
dh
m
,
=

(4-74)
where cd
h
is the enthalpy of refrigerant mixture in the condenser (kJ/kg), c
h
is the
enthalpy of refrigerant out of the compressor (kJ/kg), and d
h
is the enthalpy of the liquid
refrigerant at the outlet of the condenser (kJ/kg).
The heat dissipation from the refrigerant to the condenser wall is:

( )
w cd cd cd i cd i w cd
T T A k Q
, , , ,
=

(4-75)
where w cd
Q
, is the heat dissipation (kJ/s), cd i
k
, is the inside heat transfer coefficient of tube
(kJ/m
2
K), cd i
A
, is the total inside heat transfer area of tube (m
2
), cd
T
is the temperature of
the refrigerant (K), and w cd
T
, is the average temperature of the condenser wall (K).
The metal wall of the condenser is considered as a thermal capacitance. The temperature
change of the metal is governed by:

99

a cd w cd
w cd
w cd w cd
Q Q
dt
dT
c M
, ,
,
, ,
=

(4-76)
where w cd
M
, is the total mass the condenser metal (kg), w cd
c
, is the average specific heat of
the metal (kJ/kg K), and a cd
Q
, is the heat dissipated into the outdoor air (kW).
The heat dissipation from the metal to the air is given by:

( )
a cd w cd cd o cd o a cd
T T A k Q
, , , , ,
=

(4-77)
where cd o
k
, is the external average heat transfer coefficient (kJ/m
2
K), w cd
c
, is the average
specific heat of the metal (kJ/kg K), and a cd
Q
, is the heat dissipated into the outdoor air
(kW).
The air temperature change of the air around the condenser is governed by:

( )
a cd a cd e p a cd
a cd
a cd a p
Q T T c m
dt
dT
V c
, , ,
,
,
+ =

(4-78)
where p
c
is the specific heat of air (kJ/kg K), a
is the density of air, a cd

V
, is the total air
volume of the control volume around the condenser (m
3
), a cd
m
,
is the mass flow rate of

condensing air (kg/s), e
T
is the outdoor air temperature (K), and a cd
T
, is the outlet air
temperature (K).
The refrigerant in the condenser is lumped as a mixture of both liquid and vapor. The
specific enthalpy of the refrigerant in the condenser can be given by:

( )
l cd cd v cd cd cd
h x h x h
, ,
1 + =
(4-79)
where cd
x
is the quality of the refrigerant, v cd
h
, and l cd
h
, are the specific enthalpy of the
saturated vapor and liquid refrigerant in the condenser (kJ/kg) respectively.

100
The enthalpy of saturated refrigerant vapor and liquid can be regressed as an empirical
function of the mixture temperature. Therefore, the temperature and the enthalpy of the
outlet refrigerant liquid can be determined by:

|
.
|
\
|
=
cd cd
cd cd
x T
dt
dh
f
dt
dT
, ,

(4-80)

|
.
|
\
|
~
cd
cd d
T
dt
dT
f
dt
dh
,

(4-81)
The temperature and the density of the refrigerant leaving the condenser can be regarded
as the same of the saturated liquid and identified by data regression:

( )
d d
h f T =

(4-82)

( )
d d
T f =

(4-83)
d
T
and d
are the liquid temperature (K) and density (kg/m

3
) of the saturated refrigerant
liquid respectively.
As an isobaric process in the condenser, the liquid pressure entering the expansion device
is regarded the same as the mixture in the condenser, and determined by the temperature
of the liquid refrigerant:

( )
cd d
T f p =

(4-84)
d
p
is the pressure of the refrigerant liquid at the outlet of the condenser (bar).
4.4.2.4.2 Evaporator
The heat transfer process in an evaporator is similar to that in a condenser, except:

101
- The phase change of the refrigerant in an evaporator is from two-phase to vapor
phase;
- The outlet vapor has some degree of superheat;
- The air side dynamics is more complex due to the possible moisture condensation.
The mass conservation of the refrigerant in the evaporator is given by:

f e
ef
m m
dt
dm
=
(4-85)
where e
m
is the mass flow into the evaporator (kg/s), f
m
is the mass flow out of the
evaporator (kg/s), and ef
m
is the total refrigerant in the evaporator (kg).
The mixture enthalpy in the evaporator is determined by the heat conservation in the
refrigerant:

w f f e v
ef
ef
Q h m h m
dt
dh
m
4
= (4-86)
where ef
h
is the enthalpy of refrigerant mixture in the evaporator (kJ/kg), v
m
is the mass
flow into the evaporator (kg/s), f
m
is the mass flow rate out to the compressor (kg/s), e
h

is the enthalpy of the refrigerant at the inlet of the evaporator (kJ/kg), f
h
is the enthalpy
of the refrigerant at the outlet of the evaporator (kJ/kg).
w
Q
4 is the heat transferred from the refrigerant side to the evaporator metal. It can be
determined by:

( )
w ef ef i ef i w
T T A k Q
4 , , 4
=

(4-87)
where ef i
k
, is the heat transfer coefficient of the inside of tube (kJ/m
2
K), ef i
A
, is the inside
total heat transfer area of tube (m
2
), ef
T
is the temperature of the refrigerant (K), and w
T
4
is the average temperature of the evaporator wall (K).

102
The metal wall of the evaporator is considered as another thermal capacitance as we
handle the condenser. The temperature change of the metal is governed by:

a w
w
w w
Q Q
dt
dT
c M
34 4
4
4 4
=

(4-88)
where w
M
4 is the total mass of the evaporator metal (kg), w
c
4 is the average specific heat
of the metal (kJ/kg K), and a
Q
34 is the heat absorbed by the evaporator metal from the air
(kW).
For the air side of an evaporator, a simple dry air-cooling in a dry-cooling region, or a
coupled air cooling and dehumidification in the wet-cooling region can happen. Analogue
to the two region analysis on refrigerant side, it is assumed in this study that there are two
regions of heat and mass transfer on the air side as well (Qi and Deng 2009). The process
air gets into the area of evaporator. It first passes the dry-cooling region, and then the
wet-cooling region.
Define the total heat dissipated from the evaporator wall to the surrounding air as:

|
.
|
\
| +
+ |
.
|
\
| +
=
2 2
4 34
4 2 , 2 ,
34 3
4 1 , 1 , 34
T T
T A k
T T
T A k Q
a
w ef o ef o
a
w ef o ef o a

(4-89)
where 1 , ef o
k
and 2 , ef o
k
are the heat transfer coefficients of the two regions (kJ/m
2
K),
1 , ef o
A
and 1 , ef o
A
are the contact area of the two regions (m
2
), 3
T
is the air temperature
before entering the cooling coil (K), a
T
34 is the average air temperature at the vicinity of
the cooling coil (K).
The governing equations for the two artificial areas are given as below:

( ) |
.
|
\
| +
+ =
2
34 3
4 1 , 1 , 34 3 34
34
34
a
w ef o ef o a p a
a
a a p
T T
T A k T T c m
dt
dT
V c

(4-90)

103

( ) |
.
|
\
| +
+ =
2
4 34
4 2 , 2 , 4 34 34
4
1 34
T T
T A k h h m
dt
dh
V
a
w ef o ef o a a a a

(4-91)
where 1 34 a
V
and 2 34 a
V
are the air volumes of the two regions (m
3
), a
m
34
is the air flow rate

through the coil (kg/s),
4
h
is the enthalpy of the outlet air (kJ/kg), a
h
34 is the specific
enthalpy of the air in the air close to the coil (kJ/kg).
With the relationship between air enthalpy, temperature, and moisture content, the
following equation holds:

w h T c h
g p
+ =

(4-92)
where
w
is the humidity ratio in the air (kg/kg), and g
h
is the latent vaporization heat of
water (kJ/kg). Since 34a is a dry-cooling region, 3 34
w w
a
=
.
The expression can be written as:

( )
( ) |
.
|
\
| +
+
+ = +
2
4 34
4 2 , 2 , 4 3 34
4 34 34
4
4
4
4
T T
T A k w w h m
T T m c
dt
dw
h V
dt
dT
V c
a
w ef o ef o g a
a a p g p

(4-93)
The air leaving the evaporator is assumed to be close to saturation and expressed by the
equation through curve fitting.

|
.
|
\
|
=
4
4 4
, T
dt
dT
f
dt
dw

(4-94)
The average enthalpy of the refrigerant mixture in the evaporator is a function of quality
and the specific enthalpy of the saturated liquid and vapor:

( )
l ef ef v ef ef ef
h x h x h
, ,
1 + =

(4-95)

104
ef
x
is the quality of the refrigerant, v ef
h
, and l ef
h
, are the enthalpy of the saturated vapor
and liquid refrigerant in the evaporator (kJ/kg) respectively.
The enthalpy of saturated refrigerant vapor and liquid can be regressed as an empirical
function of the mixture temperature. Therefore, the temperature and the enthalpy of the
outlet refrigerant liquid can be determined by:

|
|
.
|
\
|
=
ef ef
ef ef
x T
dt
dh
f
dt
dT
, ,

(4-96)

|
|
.
|
\
|
=
ef
ef f
T
dt
dT
f
dt
dh
,
'

(4-97)
where ' f
h
is the enthalpy of the superheated refrigerant (kJ/kg).
The control of an expansion valve is to maintain several degree of superheat from the
evaporator. To obtain the temperature of the superheated vapor, an ideal gas is considered
for the vapor and the following equation is utilized:

dt
dh
c dt
dT
f
r p
f
,
1
=

(4-98)
The specific heat, r p
c
, , can be determined by the pressure of the refrigerant mixture in the
evaporator:

( )
ef r p
p f c =
,
(4-99)
The pressure ef
p
in two phase stage can be determined by the temperature:

( )
ef ef
T f p =

(4-100)

105
The vapor density at the inlet of the compressor is needed to determine the mass flow rate
through the compressor. Utilizing the ideal gas law and the regression of saturated vapor
density against the temperature, we can obtain the superheated vapor density:

( )
ef
f g s
f ef f
T
T
T p f
,
,

= =

(4-101)

( )
ef g s
T f =
,

(4-102)
where g s,
is the vapor density entering the compressor (kg/m

3
).
4.4.2.5 Refrigerant states conversion
As deduced in the previous sections, the thermal dynamics of the refrigerant within the
close-loop of the vapor compression unit is also governed by a large number of state
relationships. A pressure-enthalpy diagram or a temperature-enthalpy diagram is
normally used to read out the unknown values from the known ones. For different types
of refrigerants, the relationships are different. One potential way to handle this in
simulation is to couple Matlab with EES, where very rich state functions are provided for
various refrigerants, through an EES-Matlab interface (EES). Fig. 4-25 illustrates the
mechanism where two txt-files are used as the exchange media. This method is
convenient when a system model is applied with many different refrigerants.

106

Figure 4-25: EES and Matlab co-simulation

The other way is to build up a set of look-up tables or regressed equations in Matlab. The
normal operation status of an air-conditioning cycle is in the range from -10 C to 85 C
(3.5 bar to 40 bar). The charts and equations collected in Appendix B are utilized to
determine the states for the refrigerant in the condenser or evaporator.
4.4.3 Model of miscellaneous components
4.4.3.1 Wheel power consumption
The two wheels in the hybrid unit are linked through a belt to two motors controlled by
VFDs. The control is considered as instant with no time delay. The power consumption
of the wheel in different rotational speed can be obtained by using the power law:

a
rate
rate
r
r
Q Q
|
|
.
|
\
|
=

(4-103)
where Q
rate
is the rated power consumption (kW), r
rate
is the rated rotation speed (rpm), r
is the rotation speed (rpm), a is the regression coefficient.

107
4.4.3.2 Fans
The fans in the unit are controlled by the VFDs, whose dynamics compared to the
thermal counterparts is much faster. The fan airflow control is regarded with good
accuracy and no time delay. A quasi-steady-state model is adopted here for all the three
fans in the hybrid unit. The relationship between the airflow rate and fan power
consumption is expressed as quadratic equations:

2
2 1 0
m a m a a Q + + =

(4-104)
where Q is the fan power consumption (kW),
m
is the air flow rate (kg/s), a
0
to a
2
are the
regression coefficients.
4.4.3.3 Bypass damper
A bypass damper is installed in the hybrid unit to increase the modulation flexibility on
the supply air temperature and the space relative humidity. Fig. 4-26 illustrates the air
path in the packaged unit. A portion of the process air from the cooling coil is bypassed
directly to the mixing chamber, while the other goes through the active desiccant wheel
for dehumidification. The control of the damper in this study is reduced to the fixed air
flow ratio of the main stream. The power consumption by the motor is ignored.

108

Figure 4-26: Bypass damper and mixing chamber
4.4.3.4 Gas burner and heat exchanger
Yu et al (2011) studied the performance of natural gas burner in a general purpose
rooftop air-conditioning unit. Their findings support that the heat exchanger heating
process is gas burning dominant. The model can be simplified and reduced to a minimum
number of geometric and operating parameters. The time constant for the dominant
dynamics of the gas regulating valve and the burning process is much shorter than the
heat exchanger. Therefore, a simple first-order model is adopted in this study for the gas
burner and heat exchanger:

( )
6 7
7
7
7
'
T T c m Q
dt
dT
c M
p g p
=

(4-105)
where 7
M
is the total mass of the air around the heat exchanger (kg), p
c
is the specific
heat of the air (kJ/kg K), 7
T
is the average temperature of the air (K), g
Q
is the heat flux
from the gas (kW), 7
m
is the mass flow rate of regeneration air (kg/s), and 7
T
and 6
T
are
the inlet and outlet air temperature (K).

109
The humidity ratio change of the regeneration air due to the heat input from the gas
burner cannot be ignored in this case. The moisture is burned out due to the air density
change, and is approximated by the following equation:

6
6
7
7
'

e
e
=

(4-106)
4.4.3.5 Heat transfer coefficients
The heat transfer coefficient on the refrigerant side in the closed loop for both evaporator
and condenser are regarded as constants in this study. The major heat transfer resistance
comes from the air side. The value of air side heat transfer coefficients can be obtained
from the datasheet provided by the manufacturer, or estimated from the onsite
measurement. The outside heat transfer coefficient is largely determined by the air
velocity and the external fin structure. The following equations are used after reviewing
the literature (Jang and Lai 1997, Chen et al. 2001, Kays and London 1964):

3
2
Pr
p
e o
c
v j k =

(4-107)

3
1
Pr Re
=
Nu
j
e

(4-108)
where Pr is Prandtl number,
v
is air velocity (m/s), e
j
is Colburn factors, Re is Reynold
number, and Nu Nusselt number.
4.5 Summary
In this chapter, the dynamic models for the diversified systems in the test bed are
analyzed and established. These models include:
- A building thermal system;
- A hydronic thermal system; and
- A hybrid ventilation unit system.

110
First-principle based physical modeling approach is mainly applied during the deduction
of the dynamic models. For the purpose of control study, the need of detailed geometry
information is reduced to the minimum by using a lumped-parameter method. The order
of the dynamic systems is kept adequately low to reduce the computational cost while
remaining the accuracy.
The building thermal system is modeled as a set of thermal network with capacitance,
resistance, and current components. The air in the space and the mass of the construction
are lumped as capacitances. The solar radiation, outdoor air infiltration, and internal heat
gains are treated as current heat sources acting on the wall nodes or the air nodes. The
thermal comfort model is expressed by PMV and PPD values as functions of internal air
physics and factors related to occupants. The hydronic terminals are modeled as first-
order nonlinear systems.
For the three components in the hybrid ventilation system, an enthalpy recovery unit, a
vapor compression unit, and an active desiccant unit, a detailed deduction by using
energy and mass conservation laws is conducted to obtain both the full dynamic models
and simplified fast dynamic models. The full dynamic models requires higher
computational cost and can produce more detailed information about the devices. The
simplified dynamic models keep the accuracy of the concerned variables with reduced
simulation cost. The latter one is suitable for online implementation of advanced control.
The control-oriented models obtained here can be utilized to build simulators for building
systems or the individual subsystem in order to evaluate the dynamic performance. They
can also be either used directly or converted into the internal models for MPC controllers.
For a linear classical MPC, a proper linearization of the original models around the
operation points is needed in order to construct the internal model.
Compared to the hybrid ventilation unit, the dynamics of the building thermal system and
hydronic thermal system is much slower. Since there is no large transient involved in the
space thermal system, the models can be utilized for short time, long time, and wide
range simulation, such as controller design, energy consumption evaluation, etc. This is

111
not directly applicable to the model for the hybrid ventilation unit, where the system has
fast dynamics and the models require short time-step for transient simulation. It is not just
because the equipment is more complex with three individual functional components and
coupled heat and mass transfer, but also because it involves large dynamic transients,
which are hard to be captured with physical models. The outputs of concern determined
by the inputs/outputs relationship in the model evolve fast. The start-stop, defrosting,
heating/cooling switch, etc., make it impossible to be covered fully in this thesis. There is
no literature so far providing a dynamic model that can deal with all the transients for a
complex hybrid ventilation unit.

112
CHAPTER 5 CONTROL OF HYBRID VENTILATION
UNIT
5.2 Base control of the hybrid
unit
5.3 Model linearization and state
space realization
5.4 MPC architecture and
performance evaluation
- Introduction to the base control
of the hybrid ventilation unit
- Introduction to the model
linearization method and the
state space realization
- Enthalpy recovery wheel
- Vapor compression unit
- Active desiccant unit
5.1 Introduction
The hybrid ventilation unit in the test bed includes three individually functional
components: the enthalpy recovery unit, vapor compression unit, and active desiccant
dehumidifier. The unit provides necessary ventilation air to the space and maintains the
space humidity level during summer to avoid condensation on the surface of the water-
based hydronic thermal system. The supply air flow rate and supply air temperature set
points are given by the central workstation. The space sensible load is mainly assumed by
the water-based hydronic system while the hybrid unit takes care of the latent load. To
avoid a large fluctuation of the indoor air conditions and as a matter of fact, the operation
of the hybrid unit, in terms of supply air flow rate and supply air temperature, is
constrained in a narrow range.
With a conventional cooling-based dehumidification method, the dehumidification
capacity is coupled with the sensible cooling process and limited by the evaporator
surface temperature. Dehumidification is mainly a result of cooling rather than a control
target. The supply air can only be dehumidified to the extent where the supply air flow
rate and supply air temperature matches the space sensible cooling load. With the
application of variable speed drivers, the coupling issue can be mitigated with improved
control performance. For example, reducing supply airflow rate can decrease the air side
sensible heat transfer coefficient and therefore increase the latent cooling capacity. The
cooling capacity delivered by a vapor compression unit is largely determined by the

113
control of the compressor. According to the vapor cycle chart and the model deduction
presented in Chp. 4, the refrigerant side mass flow and enthalpy change dominate the
cooling output. Increasing the compressor speed, with the condenser fan modulated
accordingly, leads to lower evaporator surface temperature and higher sensible and latent
cooling capacity.
To achieve a desired control objective, different control strategies and actions may be
taken. An optimal control may only be achieved when the action of the individual
components are coordinated. Traditional PID controlled multi-SISO approach does not
consider the strong coupling among and within the individual component in equipment as
the hybrid ventilation unit. Take the vapor compression unit as an example. Jakobsen
(2000) proposed a MIMO control strategy to identify the optimal settings for individual
components. They pointed out that adjusting the compressor speed, modulating the
condenser fan speed, or controlling the supply fan speed could achieve the same marginal
cooling capacity increment expectation, but only one optimum existed. Their study
showed the energy consumption penalty in a conventional multiple SISO control where
the set points for the components are fixed.
Denis et al. (2006) studied the application of nonlinear predictive control of a water
cooled vapor compression unit. The multi-objectives, including cooling capacity, energy
efficiency, smooth command, etc., were weighted to form the quadratic cost function.
The compressor speed, expansion valve openness, and pump speeds for evaporator and
condenser were manipulated variables. A 20s sampling period and four steps prediction
were adopted after trial and error. The simulation results showed a higher COP by about
20% could be achieved with optimal control. Since a nonlinear optimization algorithm is
utilized, the computational cost was reported higher in their study.
In reality, an optimal operation of an equipment is almost always subjected to a set of
constraints. For example, the compressor speed can only be lower than the maximum but
no less than a value (e.g. 30%) for motor overheat protection. The optimization is usually
also a compromise among the multiple objectives, e.g. additional dehumidification vs.

114
energy consumption. With a desired low energy consumption, the cooling or heating
capacity has to be satisfied first. In addition to the obvious cost, the changing rate of the
key components, such as the compressor, shall also be considered. A frequent and large
speed change results in an early burn-out. While the overall unit may not cost big money,
the fault diagnosis, installation, component replacement, and the discomfort during the
maintenance all count into the cost.
In this chapter, the classical MPC with linearized internal model is investigated for the
optimal control of the hybrid unit. The base structure and rule-based multi-SISO control
are first introduced. The controllable variables in the unit and their modulating range are
stated. Three model linearization methods are briefly described to obtain the internal
linear model for the MPC controllers. The performance of the proposed MPC for the
three components in the hybrid unit is presented. The simulation results and findings are
discussed.
5.2

Base structure and control of the hybrid ventilation unit
Fig. 5-1 illustrates the base structure of the hybrid ventilation unit. There are five variable
speed drives (VFDs) installed, four of them modulate fans' speed and one modulates the
compressor speed. The enthalpy recovery wheel and the active desiccant wheel are
equipped with speed-adjustable motors. A gas burner is fed with natural gas through a
modulating valve to control the regeneration air temperature. The binary points provided
by the manufacturer include: system on/off, supply fan on/off, compressor on/off,
compressor VFD status, condenser fan status, condenser fan VFD status, regeneration fan
on/off, and burner power on/off. The airflow, temperature, dew point, relative humidity
and enthalpy of the outdoor air, return air, air entering the coil, air leaving the coil, etc.,
are monitored through the manufacturer control panel.

115
e
n
t
h
a
l
p
y

r
e
c
o
v
e
r
y
w
h
e
e
l
exhaust fan
supply fan
outside air
exhaust air
D
X
c
o
o
l
i
n
g

c
o
i
l
d
e
s
i
c
c
a
n
t
d
e
h
u
m
i
d
i
f
i
e
r
mixing damper
m
i
x
i
n
g

c
h
a
m
b
e
r
supply air
return air
gas burner
regeneration fan
outside air
1
2'
1'
2
3
4
5 5'
6
7 7'
8
in in
T
, 1 , 1
,e
out out
T
, 1 , 1
,e
in in
T
, 2 , 2
,e
out out
T
, 2 , 2
,e
3 3
,e T ' 4
m
2 3 1
m m m =
2
m
3
m
4 4
,e T
7 7
,e T
34 34
,e T
' 7
m
8 8
,e T
' '
5 5
,e T
VFD-ex
VFD-erv
VFD-3
V
F
D
-
c
o
VFD-dsc
VFD-reg
modulating
gas valve
C
o
m
p
r
e
s
s
o
r
Condenser
VFD-cd
outside air
EEV
a
b
c d
e
f
4'
' 5
m
Qg
V4'

Figure 5-1: Illustration of the hybrid ventilation unit with denotes

Table 5-1 to 5-3 collect the parameters of the hybrid ventilation unit. The enthalpy wheel
and desiccant wheel are the same as studied by Zhai (2006). The efficiency of design
conditions are taken from the manufacture datasheet.
Table 5-1: Design parameters of the desiccant wheels
Enthalpy Recovery
Wheel
Active Desiccant Wheel
Nominal airflow rate (cfm) 2000 660
Wheel diameter (mm) 787 787
Wheel depth (mm) 152 102
Wheel split ratio 1/2 2/3
Rotary speed (rpm) 30 0.42
Channel size (mm) 1.8*4.2 1.5*3.4
Desiccant thickness (microns) 25 65

116
Substrate thickness 15 75
Desiccant material 3A modecular sieves Silica gel
Desiccant density (kg/m
3
) 760 700
Desiccant specific heat (J/kg K) 1000 1000
Desiccant conductivity (W/m K) 0 0
Separation factor 0.1 1
Maximum moisture loading (kg/kg) 0.2 0.36
Heat of adsorption (J/kg) 2,791,000 2,791,000
Substrate material Aluminum Glass fiber paper
Substrate density (kg/m
3
) 2,700 500
Substrate specific heat (J/kg K) 900 900
Substrate conductivity (W/m K) 237 0
Convective heat transfer coefficient
(W/m
2
K)
36.2 43.3
Convective mass transfer coefficient
(m/s)
0.03 0.036
Table 5-2: Design efficiency of the wheels
Enthalpy recovery wheel Desiccant wheel
75% 100% 75% 100%
Moisture transfer 0.735 0.686 0.74 0.68
Heat transfer 0.734 0.685 0.74 0.68

117
Table 5-3: Facts of the hybrid desiccant ventilation unit
DX unit + active desiccant unit
Single variable speed compressor, pressure: 144 - 278 PSI, total refrigerant charge: 15- 21 lbs
(6.8 - 9.53 kg), maximum cooling capacity: 12.5 tons (44 kW), maximum heating capacity:
150,000 BTU/Hr (44 kW), maximum burner capacity: 74,592 BTU/Hr (21 kW);
Supply fan motor: 3 HP (2.23 kW), supply air flow: 2250-3600 cfm (1.062 - 1.7 kg/s), condenser
fan motor: 1 HP (0.74 kW), condenser fan airflow rate: 9375 cfm (5.3 kg/s), Regen fan motor:
1/2 HP (0.37 kW), regeneration fan maximum airflow: 660 cfm (0.311 kg/s). Sensible heating
ratio: 0.38-0.6 (with 40% OA), 0.24- 0.5 (with 100% OA), Maximum compressor power 16.2
kW, Exhaust fan motor: 1.5 HP (1.12 kW).
The original control of the hybrid unit is provided by the manufacturer with an integrated
direct digital control system to enhance the control capabilities. The following four types
of control methods are obtained from the product catalog (Fischer, 2004).
Table 5-4: The typical control modes of the hybrid unit
Mode Description Notes
A Space temperature and
humidity control
The air temperature and relative humidity are monitored in
a space and used to determine the supply air temperature
and humidity set points. Temperature dominant control.
B Space temperature and
return humidity control
The air temperature and relative humidity are monitored in
the return air duct to determine the set point of supply
humidity to maintain humidity set point. Temperature
dominates.
C Supply temperature and
space (return) humidity
control
Used where other air conditioning units are deployed to
maintain the space temperature. This unit is controlled to
provide a predetermined moderate or neutral supply air
temperature. Humidity dominates.
D Building automation
variable integration with
supply temperature and
return humidity control
Used to condition multiple spaces or feeding VAV boxes.
The temperature and relative humidity are monitored in
the return air duct to determine what supply humidity is
needed to maintain humidity set points. Temperature
dominates.

118
As with a typical rule-based control, the hybrid unit is divided into multiple loops in the
original control algorithm: the outdoor air intake loop, the heat pump loop, the active
dehumidification loop, and the mixing air loop. The outdoor air intake loop modulates the
VFD on the module to maintain the outdoor air flow rate set point. The enthalpy recovery
wheel is controlled at a full rotation speed for most of the time to recycle the heat and
moisture from the exhaust air. The vapor compression unit loop modulates the
compressor speed and coordinates the VFD on the condenser to provide a predefined
supply air temperature. The active dehumidification loop controls the active desiccant
wheel and the natural gas burner to maintain the return air humidity ratio. Currently, with
control logic provided by the manufacturer, the supply air flow and the outside air flow
are constantly fixed at arbitrary values.
As discussed in the previous chapters and the introduction section, conventional rule-
based control can hardly deal with the strong coupling within and among the elements.
Meanwhile, the control law is usually set for a fixed reference point, without considering
the control efforts. In an MPC optimal control, the control law for a MIMO is naturally
included through the internal model. The cost function can be defined to represent the
control efforts and the tracking errors. The constraints on the inputs and outputs are
naturally embedded as well. It is suitable for an equipment control to reduce the energy
consumption for the given set points. By incorporating multiple criteria and/or
constraints, the hardware lifecycle can also be extended.
5.3 Model linearization and state space realization
The internal model for the low-level MPC requires a linear format. This section handles
the nonlinear model linearization for the direct expansion unit. A.J. Jordan (2006)
provided an excellent overall summary of linearization methods of nonlinear state
equations. The main methods widely used in applications are: Taylors series expansion,
optimal linearization method, and global linearization method.
A nonlinear system can be expressed in the following form:

119

( ) t u x f x , , =
,
( )
0
0 x x =

(5-1)
where
x
is the states vector,
u
is the control inputs,
t
is time, and
f
is the nonlinear
dynamic mapping.
In many circumstance, numerical methods can be applied to directly solve the set of
equations. The analysis of a nonlinear system imposes challenges. It can also be
approximated by using linearization and put in the state space form as:

( ) ( ) t Bu t Ax x + =
,
( )
0
0 x x =

(5-2)
where A and B are the state matrix and input matrix.
This linear system has analytical solution as:

( )
( ) ( )
( )
}

+ =
t
t
t A t t A
dt t Bu e x e t x
0
0
0
t
,
( )
0
0 x x =

(5-3)
Taylors series expansion utilizes the following approximation about an equilibrium point:

( )
( )
( )
( )
=
=
0
!
n
n
n
a x
n
a f
x f

(5-4)
By taking only the terms before the second order, the linearization in incremental form
can be obtained as:

u
u
f
u
f
u
f
u
f
x
x
f
x
f
x
f
x
f
x
n
m
i
m
n
n
m m
n
A
(
(
(
(
(
c
c
c
c
c
c
c
c
+ A
(
(
(
(
(
c
c
c
c
c
c
c
c
= A
...
: : :
...
...
: : :
...
1
1
1
1
1
1
1

(5-5)
An optimal linearization utilizes the least square method to find the linear expression for
the original nonlinear system. The goal is to minimize the inner product of the error as:

120

( ) c c, = I

(5-6)

( ) ( ) ( ) t t u t x f Bu Ax , , + = c

(5-7)
A global linearization uses global diffeomorphism through the state variables
transformation to solve the nonlinear system equations (Jordan and Nowacki 2003). A
set of new variables, k , are defined so that the new system in terms of the new variables
has a linear format:

Bv k A k + =

(5-8)

( ) x h k =

(5-9)
The new input
v
is a nonlinear map of the original input
u
and the function defined by
the state
x
. With the solution to the new linear system about k , the original system can
be obtained via the inverse transformation:

( ) k h x
1
=

(5-10)
More discussion on the pros and cons of the three common linearization methods can be
found in literatures and goes beyond the scope of this thesis. A local linearization through
Taylor's series is generally adopted for engineering analysis. Choosing a right operating
point is critical to obtain an accurate linear model for a nonlinear system. When a system
works in a wide range with multiple equilibrium points, a constant linearization is needed
to ensure the controller performance.
5.4 MPC architecture and performance evaluation
The hybrid ventilation unit has eight variables that can be continuously manipulated, as
collected in Table 5-5. The condenser fan, compressor rotation speed, and the main unit
fan jointly form the control of the vapor compression unit. The operation of this unit

121
dominates the performance of the hybrid unit since it provides both cooling and
dehumidification to the air stream. The control of the vapor compression unit aims to
provide the required supply air temperature with optimal energy consumption and less
control efforts. The enthalpy recovery wheel is a non-source system. It does not convert
electrical energy to thermal energy. The control is to maximize the heat and moisture
exchange between the outdoor air and exhaust air when such an exchange is beneficial to
the space, and minimize the exchange when it is not. The control of the active desiccant
unit has narrow margin since the performance of the dehumidification is largely
determined by the predefined configuration and operation.
Table 5-5: Manipulated variables of the hybrid ventilation unit
Variables Notes Purpose Max Min Unit
Vapor compression unit
VFD-cd Condenser air flow
rate
Maintain <278 PSI (19 bar)
condenser pressure
5.3 0 kg/s
VFD-co Compressor rotation
speed
Control the outlet air temp
of evaporator
30 100 %
VFD
3
Main unit fan speed Maintain total supply air
flow rate
1.7 0 kg/s
Enthalpy recovery unit
VFD
erv
Recovery wheel
rotation speed
Constant speed 30 0 rpm
VFD
ex
Exhaust air flow rate Match outdoor air flow rate 1.04 0 cfm
Active desiccant unit
Q
g
Gas burner Maintain the regeneration
air temperature
21 0 kW
VFD
reg
Regeneration air flow
rate
Maintain the 0.31 0 kg/s
VFD
dsc
Desiccant wheel
rotation speed
Constant speed 0.42 0 rpm

122
The hybrid ventilation unit controls mainly the space humidity of IWn. The supply air
temperature is also controlled to avoid a large change and disturbance to the space air
which is thermally conditioned by the water-based system. Table 5-6 gives the controlled
variables.
Table 5-6: Controlled variables of the hybrid ventilation unit
Variables Notes Purpose Maximum Minimum Unit
P
c
, T
c
Compressor outlet line
pressure
Maintain < 278 PSI
condenser pressure
19 9 bar
T
4
Cooling coil outlet air
temperature
Common control
practice
26+273 9+273 K
T
7
Regeneration air
temperature
Avoid over-heating of
the regeneration air
80+273 Outdoor air
temp
K
T
8
Supply air temperature Maintain the space
ventilation and
humidity
26+273 13+273 K
omg
8
Supply air humidity Maintain the space
humidity
From
supervisor
-- g/kg
Three MPCs can be structured for the three components individually, as illustrated in Fig.
5-2. With this structure, the supervisory controller supplies the reference set points to the
MPCs. The function of the three components in the hybrid unit is clear:
- The enthalpy recovery unit controls the outdoor air intake. For example, it
maximizes the outdoor air intake for free cooling and minimizes the outdoor air
intake for mechanical cooling and heating;
- The vapor expansion unit tracks the supply air temperature. It is desirable that the
integrated unit uses less electricity or with less control efforts for the same load
condition;
- The active desiccant unit tracks the supply air humidity. It runs only when the
humidity after the vapor expansion unit is high.

123
The decentralized structure has better reliability since the computational load is divided
into three. Another approach can be one MPC for the whole unit as shown in Fig. 5-3.
The computation load is higher with more manipulated variables and constrains. The
tuning effort for the centralized MPC control is also higher. The benefit of using a
distributed MPC structure is more obvious when the equipment complexity grows.
MPC 1
MPC 2
MPC 3
Enthalpy
Recovery Unit
Vapor Expansion
Unit
Active Desiccant
Unit
+
+
+
References Outputs

Figure 5-2: MPC architecture- example 1
MPC
Enthalpy
Recovery Unit
Active Desiccant
Unit
+ References
Outputs
Vapor
Compression Unit
Space

Figure 5-3: MPC architecture- example 2
Evaluation of MPC control over the hybrid ventilation unit is based on the weather data
trended in 2006. Fig. 5-4 plots the weather data by minutes. The unit was newly installed
in 2006. The space air temperature and supply air temperature are plot in Fig. 5-5. Part of

124
the trended data is used as the default set points in the study of MPC control. To increase
the variation of disturbances and investigate the performance of the MPCs, the trended
data are mapped from seconds to minutes. More fluctuation of the external and internal
air conditions present in the study.

Figure 5-4: Trended outdoor air temperature in 2006

125

Figure 5-5: Room air and supply air temperature
5.4.1 Enthalpy recovery wheel
The enthalpy recovery wheel (ERV) is a non-source unit for outdoor air intake
preprocessing. The function is to recover the heat and moisture from the space exhaust air
into the outdoor air intake in mechanical heating seasons. In mechanical cooling seasons,
the process is reversed with heat and moisture transferred from the outdoor air intake into
the exhaust air. Fig. 5-6 shows the variables involved in the MPC of the enthalpy
recovery unit.

Figure 5-6: Abstracted enthalpy recovery unit MPC structure

126
As illustrated in Fig. 5-6, the outdoor air temperature, humidity, and the return air
temperature, humidity are measured disturbances on the first-order system. The enthalpy
recovery wheel unit is a first-order system with varying gain, varying delay and
disturbances. The inputs on the exhaust fan air flow rate and the wheel rotation speed are
constrained. All these conditions remark that MPC is a good choice for the control.
The model of the enthalpy recovery wheel is programmed in Matlab as an S-function.
Figs. 5-7 and 5-8 show the screen shots of the code. The S-function block of the model is
then wrapped in a subsystem with additional inputs and outputs as plotted in Fig. 5-9. To
compare the performance of the MPC with conventional PIDs in terms of the reference
tracking, two models with PID controllers are also built for the unit as illustrated in Fig.
5-10. As a SISO oriented controller, the two manipulated variables, the wheel rotating
speed and the exhaust fan speed, need to be assigned to two references separately to
minimize the coupling issue. The sign of the PI gains in the two PID controllers shall be
adjusted when the process switches from direct action to inverse action, e.g. from winter
to summer.

Figure 5-7: Main S-function for the enthalpy recovery wheel

127

Figure 5-8: Subfunction for the enthalpy recovery

Figure 5-9: ERV model wrapped as a block

128

Figure 5-10: ERV model with PID controllers
The control cost function of MPC is defined as the summation of the weighted quadratic
function of the tracking errors, manipulated variables, and changing rates.

=
+ A +
k
i
R
u
S
Q
y
e u e
1
2
2
2
min

(5-11)
As a non-source equipment, an accurate control of the enthalpy recovery unit to achieve
both the required supply air temperature and humidity is extremely difficult. The
disturbances come from the outside air and return air, including the return air quantity. As
SISO based controller, setting and tuning PIDs for the coupled thermal and mass transfer
process can be challenging. We pick one hour data in winter and control the unit with
both PIDs and MPC to compare the performance. The controllers only track the air
temperature since air humidity is not a concern in the space in winter and there is no
humidifier equipped. Handling the control with PID requires the tuning of the controller
for both the enthalpy recovery wheel rotation speed and the exhaust fan. With MPC
control, it is naturally considered and different control performances are achieved by
2
T_3_out
1
omg_3_out
omg_3_set
omg3
m2
erv_rpm
T3out
To Workspace
T_3_set
T3
m_3
m_2
erv _rpm
omg_1_in
T_1_in
omg_2_in
T_2_in
omg_3
T_3
ERV
T_3_set
T_3
omg_3_set
omg_3
m_2_set
erv _rpm
Control l er
5
T_2_i n
4
omg_2_i n
3
T_1_i n
2
omg_1_i n
1
m_3

129
adjusting the weighting matrix for the reference tracking errors, manipulated variables,
the rates, etc.
For the enthalpy recovery wheel, when the system is operated at the point of 3
m
= 1.13
kg/s,
2
m
= 0.65 kg/s, erv
w
= 20 rpm, e
e
= 11.05 g/kg, e
T
= 302 K,
r
e
= 8.4 g/kg,
r
T
= 298
K , the system matrices A, and B are calculated as follows:

(
(
(
(
=
0124 . 0 0 0 0
0 0105 . 0 0 0
0 0 0565 . 0 0
0 0325 . 0 0 0565 . 0
m
A

(5-12)

(
(
(
(
=
0064 . 0 0 0061 . 0 0 0016 . 0 0073 . 0 0
0 0054 . 0 0 0051 . 0 0007 . 0 0032 . 0 0
0 0 0 0 0 1247 . 0 0717 . 0
0 024 . 0 0 0 0 0645 . 0 0371 . 0
m
B
(5-13)
The eigenvalue of the characteristic matrix is:

(
(
(
(
=
002 240 . 1
002 050 . 1
002 650 . 5
002 650 . 5
e
e
e
e
m

(5-14)
The matrices both have full rank, which indicates the system is observable and
controllable. Meanwhile, the eigenvalues of the linearized dynamic model are shown in
Eq. 5-14 for the operating point. All negative real parts are found for the values, which
means the system is asymptotically stable (Friedland 2005, Ogata 2009). The MPC
controller is programmed in Matlab. Part of the code is illustrated in Fig. 5-11.

130

Figure 5-11: MPC controller for the ERV in Matlab
Fig. 5-12 plots the comparison of the results between two PID controls and MPC control.
PID one (PID-1) has the two controllers both track the supply air temperature. PID two
(PID-2) has the two PID controllers track the set points of the supply air temperature and
the humidity separately. The controllers are tuned for winter season and winter weather
data is used. All the controllers work well to track the supply air temperature set point.
With weight 1 on the supply air temperature tracking, MPC responses a little bit slower
than the two PID controllers. The tracking error of MPC is less than the two PIDs as
zoomed-in in Fig. 5-13, with the response from MPC almost overlapping the set points.
The control actions of MPC are plotted in Fig. 5-14. The MPC reduces the outdoor air
intake quantity when the supply air temperature set point has a step-up disturbance. The
rotation speed of the ERV is correspondingly improved to recycle the heat from the
exhaust air stream.

131

Figure 5-12: Comparison of PIDs and MPC for ERV control- winter 1 hour

Figure 5-13: Comparison zoom in- winter 1 hour

132

Figure 5-14: Outdoor air intake and ERV rotation speed of MPC
The same controllers are applied with the weather data changed from winter to summer.
It is found that the PID controllers do not function well since the action of the unit is
complex and determined by four disturbances. PID-1 drives the outdoor air intake to the
maximum and turns off the enthalpy recovery wheel, while PID-2 drives both the outdoor
air intake and the enthalpy recovery wheel rotation speed to the maximum. This
demonstrates that a PID controller, as a reference tracking only controller, is not an ideal
candidate for the control of the ERV unit. With the same settings, although the set points
cannot be maintained due to the feasibility, MPC drives the outdoor air intake to the
minimum intake almost immediately and turns the ERV rotation speed to the maximum.
The control efforts of MPC for a given target can also be adjusted by changing the
weights on the manipulated variables.
The control weights on the inputs and outputs are the main variables to drive the MPC
toward the desired direction. Four additional scenarios as listed in Table 5-7 are further

133
studied for the MPC control of the ERV unit. Because of the small power capacity,
weight on the ERV rotation speed is set to 0 for all the four scenarios.
Table 5-7: MPC control scenarios for summer and swing weather conditions
Scenario Humidity
set point
(g/kg)
Temp set
point (K)
Weight
on
humidity
Weight
on temp
Weight
on OA
intake
Notes
ERV-1 10 299 to 299.5 1 0 0.5 Humidity control
dominates
ERV-2 10 299 to 299.5 0 1 0 Temp control
dominates
ERV-3 8 290 1 1 0 Typical mechanical
heating and cooling
ERV-4 8 290 0 variable variable Large fluctuation in
swing season

Figure 5-15: Outdoor air intake and ERV rotation speed in summer, ERV- 1

134

Figure 5-16: ERV supply air humidity control in summer, ERV- 1

Figure 5-17: ERV air temperature control in summer, ERV- 1

135
In Scenario ERV-1, the mixed air humidity set point is picked as the most important
control variable. A weight of 0.5 is also set for the outdoor air intake. With the settings,
MPC drives the outdoor air to a low level as long as the humidity set point can be
satisfied. Fig. 5-15 plots the outdoor air intake and the ERV rotation speed. The ERV
rotates at high speed for most of the time since the weight of the rotation speed is zero.
The supply air humidity is maintained well as shown in Fig. 5-16. Except the initial shoot,
the supply air humidity almost overlaps with the set point. The moisture is transferred
from the outdoor air to the exhaust air. Since the supply air temperature gets zero weight,
the response of the supply air temperature as shown in Fig. 5-17 is not controlled.
For the same time period and working conditions, Scenario ERV-2 gives a weight as 1 to
the supply air temperature and 0 to the supply air humidity. The weights on both the
outdoor air intake and the ERV rotation speed are set to zero. The MPC mainly concerns
the supply air temperature. In Fig. 5-18, the humidity set point is ignored by the MPC
controller. The humidity responses naturally to the mass transfer process. Although the
time period is in summer, there are several short periods when the supply air temperature
can possibly be met and MPC makes it as shown in Fig 5-19. Fig. 5-20 plots the
corresponding modulation of the outdoor air intake and the ERV rotation speed under the
weighting settings. The outdoor air intake is maximized when the outdoor air temperature
is lower than the supply air temperature set point, which is desirable for energy savings.

136

Figure 5-18: ERV air humidity control in summer, ERV- 2

Figure 5-19: ERV air temperature control in summer, ERV- 2

137

Figure 5-20: ERV manipulated variables in summer, ERV- 2
In Scenario ERV-3, the supply air humidity set point is set to 8 g/kg and the temperature
set point is set to 290 K. As in Scenario ERV-2, the effort of outdoor air intake and ERV
rotation speed are ignored by setting the weights to zero. However, the set points of the
supply air temperature and humidity are reduced to 290 K and 8 g/kg respectively.
During the time period, the outdoor air humidity is constantly above 10 g/kg, and high to
about 15 g/kg. As plotted in Fig. 5-21, the MPC automatically reduces the outdoor air
intake, and modulates the enthalpy recovery wheel up to the maximum speed to recycle
the heat from the exhaust air. The minimum outdoor air intake to maintain the ventilation
air during the mechanical cooling season is achieved. This indicates that, in summer, a
more realistic set points can drive the MPC toward minimum outdoor intake. The
supervisory level controller can put a low limit on the outdoor air intake to ensure the
ventilation air intake.

138

Figure 5-21: ERV manipulated variables in summer, ERV- 3
It is desired that control of the outdoor air processing unit can bring in more outdoor air
when the space is in cooling mode, and modulate properly the intake when the outdoor
air condition is favorable to maximize the free cooling benefit. When the outdoor air is
warm, the outdoor air intake should be reduced to the minimum. To investigate the
control of MPC, a time slot is picked as Scenario ERV-4. As shown in Fig. 5-22, the
outdoor air temperature fluctuates between 284 K to 302 K and the outdoor air humidity
is around 5 g/kg for most of the time. A dynamic linearization or weighting adjustment is
needed to ensure the linear model represents the original nonlinear model well and the
MPC functions properly, due to the following three different modes in the time period as
marked in Fig. 5-22:
- For the initial time period, condition "a" in Fig. 5-22, the outdoor air temperature
and the return air temperature are close to each other and also the set point;
- For condition "b", the outdoor air temperature is lower than the set point while the
return air temperature is higher than the supply air temperature set point;

139
- For condition "c", both the return air and the outdoor air temperature are higher
than the supply air temperature set point.
For condition "a" and with space in cooling mode, maximum outdoor air intake shall be
encouraged since the outdoor air is favorable. For condition "b", a modulation is probably
needed to maintain the supply air temperature set point. For condition "c", a minimum
outdoor air intake is considered to avoid penalty since the outdoor air temperature is
higher than the return air temperature.

Figure 5-22: Disturbance conditions, ERV- 4

140

Figure 5-23: ERV manipulated variables in swing season, ERV- 4

Figure 5-24: ERV air temperature control in swing season, ERV- 4

141

Figure 5-25: ERV air temperature tracking zoom in, ERV- 4

Figure 5-26: ERV air humidity control in swing season, ERV- 4

142
With the MPC, a conditionally consecutive linearization and dynamic weighting are used
to accommodate the large switches. When the outdoor air temperature is lower than the
return air temperature, the system is linearized around the operation point. The supply air
temperature gets a weight as 1. To prevent the unit from over-tuning, a weight of 0.5 is
imposed on the outdoor air intake modulation rate. When the outdoor air temperature
changes to be higher than the room air temperature and the supply air temperature set
point, a weight as 1 is added to the outdoor air intake and the supply air temperature
output. The simulation results are plotted in Fig. 5-23 to 5-26.
It is shown in Fig. 5-23 that the outdoor air intake is kept to maximum when the outdoor
air temperature is lower than the return air temperature and close to the supply air
temperature. An economizer with maximum outdoor air intake is achieved. When the
outdoor air temperature drops to be lower than the supply air temperature, a modulation
action is taken by the MPC. From Fig. 5-24 and Fig. 5-25, it can be seen that the supply
air temperature is controlled to 290 K as the set point. This strategy is similar to free
cooling operation as in rule-based control. 290 K is selected to prevent over-cooling of
the space and it is adjustable if more free-cooing is desired. When the outdoor air rises to
be higher than the return air temperature, it is a typical cooling season and a minimum
outdoor air intake is desired. As illustrated in Fig. 5-23, MPC reduces the outdoor air
intake to the minimum and turns the ERV rotation speed to full to recycle the heat. The
warm outdoor air intake is cooled down to close to the return air temperature. The
outdoor humidity set point is not weighed since the outdoor air in the season is mild or
dry. Dehumidification is not a concern with temperature control dominating.
5.4.2 Vapor compression unit
The vapor compression unit itself is a highly complex nonlinear system. The components
include the compressor, condenser, thermal expansion valve, evaporator, and main fan.
The thermal expansion valve is a self-regulating device with proportional control. On the
evaporator side, a heat and moisture coupled transfer process happens. To provide
dehumidified air to the space, the compressor varies the speed and coordinates with the

143
condenser fan for cooling capacity modulation. As pointed out in the introduction, a same
marginal cooling capacity can be achieved via the operation on the compressor, supply
air fan, and/or condenser fan. The goal of utilizing an MPC for this unit is to bring in the
benefits of MIMO control along with energy consumption, equipment life cycle cost, and
handling of dynamic coupling and constraints. The system monitors the supply air
temperature and humidity as the controlled variables. The disturbances include the
varying outside air temperature, return air temperature and humidity. To extend the
lifespan of the unit, we also need to penalize frequent changing on the compressor speed
modulation. The structure is illustrated in Fig. 5-27 for the vapor compression unit.

Figure 5-27: Abstracted DX unit MPC structure
Fig. 5-28 shows the integrated model built in Simulink. The model of the direct
expansion (DX) unit includes four blocks: the condenser, evaporator, compressor, and
expansion valve. The inputs and outputs of the components are marked with notes. Fig.
5-29 to Fig. 5-32 plot the individual component model in Simulink. The combined
nonlinear model serves as a system simulator for the MPC controller. Another linearized
model around the operation point is utilized in the MPC for control.

144

Figure 5-28: DX unit model in Simulink

Figure 5-29: Condenser model in Simulink
h_cd
T_cd
T_cdw
T_cda
h_d
p_d
rou_d
m_e
m_c
m_v
h_c
omg_e
T_e
m_c
wco
h_c
compSpeed
rou_a
p_f
p_d
h_f
m_3
m_v
m_f
h_e
omg_3
T_3
Tfprm
omg_4
T_4
T_34a
T_4w
T_ef
h_f
T_fprm
T_ef
rou_d
p_d
h_d
p_e
p_f
rou_f
h_ef
5
p_ef
4
p_cd
3
T_cda
2
omg_4
1
T_4
p_f
p_d
m_v
m_c
h_ef
h_e
h_d
h_cd
h_c
In1 out
expansi on val ve
In1 Out1
evaporator
In1 Out1
condenser
In1 Out1
compressor
T_fprm
T_ef
T_cdw
T_cd
T_4w
T_34a
CompressorPower
7
omg_e
6
T_e
5
omg_3
4
T_3
3
m_3
2
me 1
CompressSpeed
m_e
m_c
h_c
m_v
omg-e
T_e
T_cd
h_d
p_d
rou_d
T_cd_a
hcd
T_2w
m_e
T_e
T_cd_a
T_cd_w
T_cd
T_cd_w
T_cd_a
T_cd_w
T_cd
h_d
h_c
m_c
temp_2
temp_1
T_e
m_e
1
Out1
f(u)
temp2
f(u)
temp1
f(u)
rou_d
f(u)
pd
f(u)
hcd_v
f(u)
hcd
f(u)
h_d
f(u)
dt_cd_w
f(u)
dt_cd_a
f(u)
dt_cd
1
s
dt-cd-w
1
s
dt-cd-a
1
s
dt-cd
Termi nator
Scope7
1
In1

145

Figure 5-30: Evaporator model in Simulink

Figure 5-31: Compressor model in Simulink
m_v
h_e
m_c
T_3
T_ef
omg_4
h_ef
T_4
rou_f
T_4w
h_f
p_f
T_fprm
T_34a
omg-3
m_3
m_3
T_34a
T_4
omg_3
omg_4
T_4w
dT_4
T_4
T_ef
T_4w
T_3
T_34a
T_4
temp1
temp2
m_c
h_e
h_f
T_ef
T_4w
m_3
T_3
T_34a
T_4w
T_ef
T_fprm
T_ef
T_ef
dT_ef
temp-3
m_v
m_3
1
Out1
f(u)
temp-3
f(u)
temp-2
f(u)
temp-1
f(u)
rou-f
f(u)
p_f
f(u)
omg_4
1
s
omg-4
f(u)
h_efv
f(u)
h_efl
f(u)
h_ef
f(u)
dT_fprm
f(u)
dT_ef
f(u)
dT_4w
f(u)
dT_4
f(u)
dT_34a
1
s
dT-fprm
1
s
dT-ef
1
s
dT-4w
1
s
dT-4
1
s
dT-34a
1
In1
compreSpeed
rou_a
p_e
p_d
h_f
h_b
m_c
wco
h_b
1
Out1
f(u)
wco
f(u)
refri gFl ow
f(u)
heat
f(u)
effi ci ency
Scope1
Scope
Saturati on
1
In1

146

Figure 5-32: Expansion valve model in Simulink
As a cooling unit, the typical operation is to maintain a predefined supply air temperature
and allow the supply air humidity drift around. This control can cause the adverse
sensible and latent load coupling issue when the space sensible and latent load does not
match the sensible and latent capacity of the unit. Under most circumstances, the space
humidity is allowed to fluctuate within a range (e.g. 45% to 65%), instead of a fixed set
point. Another possible solution is to allow the supply air flow rate to be modulated
accordingly when the sensible and latent capacity does not match the load. For the
ventilation unit deployed in the IWn, the control is decoupled by adding an additional
dehumidification unit. The DX unit only controls the supply air temperature within a
predefined range, allowing the humidity of the supply air fluctuate. The additional
dehumidification unit monitors the room air humidity.
With conventional PID control, the modulation of the fan, condenser and compressor
needs to be put in a set of rules to ensure a good control on the supply air temperature and
avoid interest conflicts. The component wearing, maintenance and overall unit life span
can hardly be described in the control rules. With optimization based MPC, the
modulation of the compressor and fans can be integrated in the control law by penalizing
the frequent modulation. The coupling among the components and variables are
evaluated by the internal model and the optimization.
m_v
h_e
T_fprm
T_ef
rou_d
p_d
h_d
p_e
1
out
f(u)
vl vopenness
f(u)
refri g fl owrate
mi n
Mi nMax1
max
Mi nMax
0
Constant
-C-
A_max
1
In1

147
In reality, a DX system is designed to operate in the vicinity of a set point for a given
load that does not change significantly. The model of the DX unit built in Matlab is
linearized for the space conditioning in summer around the typical operation point. Here
we pick: 3
T
= 24
o
C, 3
e
= 10 g/kg, e
T
= 30
o
C, e
e
= 14 g/kg, 3
m
= 1.4 kg/s. The reduced
order system has eight states and seven inputs. The compressor rotation speed, the
condenser fan air flow rate, and the main supply fan air flow rate are the three
manipulated variables. The variable e
e
does not have any effect on the states since the air
humidity at the condenser side is not of concern. The supply air temperature, supply air
humidity, condenser air temperature, condenser inline pressure, and evaporator inline
pressure are the five measured outputs.
The system matrices of this DX unit are obtained as follows for the operation condition:

(
(
(
(
(
(
(
(
(
(
(

=
3157 . 1 0 0 0 8956 . 0 2101 . 0 0 0
0 3068 . 1 2495 . 0 8077 . 0 0 0 0 2134 . 0
0 4696 . 0 5985 . 0 0 0 0 0 0
0 2293 . 8 0 0039 . 9 2524 . 0 0 0 0
2909 . 7 0 0 6589 . 0 3625 . 7 0 0 0
1007 . 2 0 0 0 0 8466 . 1 0 0
0 5864 . 0 2163 . 0 0 0 0 1922 . 0 3701 . 0
0 3080 . 1 4825 . 0 0 0 0 4288 . 0 8255 . 0
m
A

(5-15)

(
(
(
(
(
(
(
(
(
(
(

=
0 2101 . 0 0 0 0 7871 . 0 0
0 0 0 0361 . 0 3751 . 2 0 0
0 0 0 1289 . 0 5481 . 0 0 0
0 0 0 0 0 0 0116 . 0
0 0 0 0 0 0 01 . 0
0 2541 . 0 0 0 0 3638 . 0 0
0 0 1922 . 0 0 2053 . 0 0 0
0 0 4288 . 0 0 4580 . 0 0 0
m
B
(5-16)
The eigenvalue of the characteristic matrix is:

148

(
(
(
(
(
(
(
(
(
(
(

+
+
+
+
=
001 6580 . 1
002 8400 . 9
014 8185 . 1
001 0030 . 6
000 4176 . 1
000 0686 . 2
000 2224 . 8
000 8785 . 9
e
e
e
e
e
e
e
e
m

(5-17)
The following two scenarios are selected to investigate the performance of the proposed
MPC. As in the study of MPC for ERV unit, the 2006 trended data is condensed from
seconds to minutes to strengthen the disturbances.
Table 5-8: MPC control scenarios for the DX unit
Scenario Humidity
set point
(g/kg)
Temp set
point (K)
Weight
on
humidity
Weight
on temp
Others
DX-1 7 285 to 282 0 300 Constraints on the manipulated
variables and changing rates. Step
disturbance on the supply air
temperature set point.
DX-2 7 283 0 300 As Scenario 1, with step
disturbance on the inlet air
temperature. Supply air flow rate
is further constrained between 1.1
to 1.5 kg/s.

149

Figure 5-33: Air temperature, DX- 1
In Scenario DX-1, a step input disturbance is imposed on the supply air temperature set
point to investigate the controller's performance. A weight of 300 is assigned to the
temperature set point. From Fig. 5-33, it can be seen that the MPC controls the system to
the predefined set point. The system responses closely to the set point step change. For
the first twenty minutes, when the evaporator inlet temperature is low to 295 K, the MPC
drives the compressor to about 1400 rpm and the condenser fan to around 1.5 kg/s. When
the inlet air temperature increases to 298 K, and the set point steps down to 282 K, the
compressor speed is driven up to 2000 rpm. The condenser fan speed also increases
correspondingly to the load change. The control efforts for the compressor speed and
condenser fan air flow rate are plotted in Fig. 5-34. To accommodate the input step
disturbance, the supply fan air flow rate is reduced from about 1.6 kg/s to 1.2 kg/s as
shown in Fig. 5-35.

150

Figure 5-34: Control variables 1, DX- 1


151

Figure 5-36: Air temperature conditions, DX- 1
Fig. 5-36 shows that the inlet air temperature and the outdoor air temperature swings with
similar trend. The condenser fan air flow rate drops when the outdoor air temperature
decreases.
In Scenario DX-2, a step disturbance is introduced to the inlet air temperature. The
supply air temperature response and inlet air temperature are plotted in Fig. 5-37. When
the inlet air temperature drops suddenly from 295 K to 293 K, the supply air temperature
decreases by up to 0.5 K. The controller then drives the unit back to the set point quickly.
Fig. 5-38 shows the speed drop of the compressor. With the decreased load, the
compressor speed keeps at a medium speed. As in Scenario DX-1, the condenser fan
responses to the outdoor air temperature changing trend. Fig. 5-38 illustrates that the
condenser fan air flow rate decreases when the outdoor air temperature is at the valley. In
this scenario, the supply fan air flow rate decreases when the compressor speed decreases
as shown in Fig. 5-39. The cooling capacity is picked up by the increased condenser fan
speed.

152

Figure 5-37: Air temperature 1, DX- 2


153

5.4.3 Active desiccant unit
An active desiccant unit (ADU) has a similar structure as an enthalpy recovery wheel (i.e.
passive desiccant wheel) in terms of the desiccant matrix. The moisture and heat transfer
differs because of the rotation speed and the air stream ratio. The model is provided in
Chp.4. This unit alone has multiple constraints on the manipulated variables, e.g. the
regeneration air temperature, the bypass air flow rate, etc. The desiccant wheel and the
heat exchanger are also coupled together to provide proper supply air humidity. Because
of the coexistence of heat and mass transfer, the combined desiccant unit does not just
dry the air, but also add unneeded heat into the supply air. The strong coupled MIMO
features make it difficult to apply PID controllers without violating some of the physics.
For example, to dehumidify a humid air stream, the regeneration air might have to be
heated up to more than 100
o
C degree because of the dehumidification efficiency
limitation. To avoid this from happening, an additional algorithm (i.e. trip off) is needed.
As a heating driven device, the temperature and humidity control of the unit can only be

154
possibly achieved in a very narrow range. In this section, the performance of an MPC is
evaluated.
The existing control set by the manufacturer uses the active desiccant unit to follow the
supply air humidity and indirectly maintain the return air humidity. The temperature is
monitored by the cooling DX unit. Fig. 5-40 illustrates the involved inputs and outputs of
the device.

Figure 5-40: Active desiccant dehumidification model in Simulink

Figure 5-41: Active desiccant dehumidification model in Simulink
The model is built in Simulink as illustrated in Fig. 5-41. The MPC for the active
desiccant unit is designed with the linearized nonlinear model. The sensible and latent
conditioning effectiveness is transferred into the effectiveness for enthalpy and relative
5
omg_8
4
T_8
3
T7i n
2
omg5out
1
T5out
f(u)
domg8
f(u)
dT8
m7
q_g_in
T_6
omg6
T_7_in
omg_7_in1
HX
m7
T7_in
omg7_in
T5_in
omg5_in
T5_out
omg_5out
Desi cWheel
7
omg_e
6
T_e
5
q_g_i n
4
omg_4
3
T_4
2
m_7
1
m_3

155
humidity. The effectiveness is near constant and changes only in a very narrow range for
a given active desiccant wheel. The main drive for dehumidification of an active
desiccant wheel is the temperature of the regeneration air (Jeff 1982). The linear model is
obtained around the operation point:
4
T
= 18
o
C,
4
e
= 8 g/kg, e
T
= 25
o
C, e
e
= 14 g/kg, 7
m

= 0.31 kg/s, and g
q
=10kW. The system has 5 states and 2 manipulated variables. The
system matrices of this active desiccant unit are obtained as follows:

(
(
(
(
(
(
=
5649 . 0 0 0 155 . 0 0
0 5649 . 0 0 0 155 . 0
0 0 31 . 0 0 0
0 0 0015 . 0 0167 . 0 0
0 0 0064 . 0 0 0167 . 0
m
A

(5-18)

(
(
(
(
(
(

=
0 0 0 0 4099 . 0 7562 . 1 4819 . 0
0 0 0 4099 . 0 0 7448 . 6 8509 . 1
31 . 0 0 1 0 0 2581 . 32 0
0002 . 0 0042 . 0 0 0017 . 0 0073 . 0 0077 . 0 0
0003 . 0 0084 . 0 0 0122 . 0 0186 . 0 2928 . 0 0
m
B

(5-19)
The system is stable with all the eigenvalues negative:

(
(
(
(
(
(

=
001 100 . 3
002 670 . 1
001 649 . 5
002 670 . 1
001 649 . 5
e
e
e
e
e
m

(5-20)
Three scenarios as collected in Table 5-9 are conducted to test the performance of using
MPC for the unit control.
Table 5-9: MPC control scenarios for the active desiccant unit
Scenario Humidity set
point (g/kg)
Temp set
point (K)
Weight on
humidity
Weight on
temp
Others
ADU-1 5.5 292 to
291
1000 0 Normal constraints.

156
ADU-2 6 291 to
289
0 50 Supply air humidity
between [0, 7].
ADU-3 6 292 0.01 0.001 Rate Constraints on
the gas burner, supply
air humidity [0 6],
supply air temp [290
295].
To force a humidity tracking, a high penalty is imposed on the tracking error in Scenario
ADU-1. As shown in Fig. 5-42, both the regeneration air flow and the heating input are
set to the maximum when the outdoor air humidity is high. A minimum 16 kW is needed
from the heat exchanger to maintain the supply air humidity at 5.5 g/kg level. The supply
air humidity as shown in Fig. 5-43 follows the set point well. Since the air temperature
gets zero weight, there is no trade-off between the additional heat and dehumidification
with the MPC control. The supply air temperature fluctuates around 290 K when the inlet
air is around 280 K as shown in Fig. 5-44. The unneeded heat shall be handled by the
control of the DX unit.

Figure 5-42: Manipulated variables, ADU- 1

157

Figure 5-43: Air humidity, ADU- 1

Figure 5-44: Air temperature, ADU- 1

158
In Scenario ADU-2, the condition is switched with penalty weight added on the air
temperature. A constraint is specified on the supply air humidity to prevent the supply air
humidity from going beyond 7 g/kg. As shown in Fig. 5-45, large modulations on the
heat input exist in the middle and at the end of the period. The supply air humidity close
to 7 g/kg triggers the modulation. The supply air temperature follows the set point for
most of the time as plotted in Fig. 5-47. There are circumstances where the control of the
supply air temperature cannot be satisfied with only the unit. With zero weight on
humidity, Fig. 5-46 shows the supply air humidity fluctuates without following the set
point.

The modulating rate constraints are imposed on the air flow rate in Scenario ADU-3. The
low and high limits on the supply air humidity and temperature are specified as well in
the control law. The simulation results are plotted in Figs. 5-48 to 5-50.

159


Figure 5-47: Air temperature, ADU- 2

160


Figure 5-49: Air Temperature, ADU- 3

161

As shown in Fig. 5-48, the MPC responses with high heat input when the inlet air
humidity is high. Since the weight on the supply air humidity is higher than that on the
supply air temperature, the supply air humidity follows the set point better than the
supply air temperature. However, there are couple of circumstances where the control
causes overshoot on the air humidity and temperature with sudden increase on the
regeneration air flow rate as plotted in Fig. 5-49 and Fig. 5-50.
5.5 Conclusion and discussion
The hybrid ventilation unit in the test bed IWn is a complex air-condition unit. It covers
three common HVAC devices in a diversified building system. In the unit, the coupling
effects include the following two aspects:
- The heat and moisture transfer processes couple and coexist in all the three
components: enthalpy recovery wheel, vapor compression unit, and active
desiccant unit;

162
- In the vapor compression unit, additional coupling effect presents due to the
refrigerant close loop among the compressor, condenser, expansion valve, and
evaporator.
The MIMO nature of the complex device implies that an MPC can be a good candidate
for the control to handle the strong coupling effects and incorporate the constraints on the
inputs and states/controlled variables. The dynamic models for the three units are
established in Matlab/simulink to investigate the transient behavior and the control
performance of MPC. A distributed structure is adopted since the function of the hybrid
unit can be decomposed into three elements:
- The enthalpy recovery unit controls the outdoor air intake. For example, it
maximizes the outdoor air intake for free cooling and minimizes the outdoor air
intake for mechanical cooling and heating;
- The vapor expansion unit tracks the supply air temperature; It is desirable that it
uses less electricity for the same load condition;
- The active desiccant unit tracks the supply air humidity; It runs only when the
humidity after the vapor expansion unit is high.
Compared to a centralized MPC, the distributed structure has higher reliability and lower
computational cost. In an MPC controller, the main tuning parameters are the weights on
the manipulated variables, reference tracking errors, and manipulated variable changing
rates. The relative value of the weights greatly determines the performance of the
controller. For example, to prevent the compressor from frequent ramping up and down,
e.g. to extend the life span of the compressor, a weight can be added on the changing rate.
For a MIMO system, since the manipulated variables are in different units, scaling is very
important in practice. The weights should be scaled based on the expected magnitudes of
disturbances and reference changes, on the allowed magnitude of each input signal, and
on the allowed deviation of outputs. (Sigurd and Ian 2001)
Several scenarios are designed to investigate the performance of classical MPC for the
three components in the hybrid ventilation unit. The nonlinear models are linearized

163
around the operating points to form the internal linear model for the future dynamic
prediction. The MPCs can modulate the MIMO devices with improved performance for
given operation criteria. In addition, the following points are identified with the study:
- The MPC controllers on the enthalpy recovery unit and vapor compression unit
have good disturbance rejection capability. The system response to the step
changes on the inputs and the set points with acceptable stability.
- The MPC controller on the enthalpy recovery wheel can integrate the algorithm to
maximize the use of free cooling air when the outdoor air is favorable and
minimize the outdoor air intake for ventilation only when the its adverse to the
air-conditioning mode. The enthalpy recovery wheel is also correspondingly
modulated to maximize the benefit.
- The MPC controller on the vapor compression unit can effectively coordinate the
multiple inputs and multiple outputs for a given set point. A tuning on the weights
may guide the unit toward higher energy efficiency for a given load.
- The MPC controller on the active desiccant wheel can manipulate the heat input
and regeneration air flow to follow the humidity set point for most of the time and
avoid large penalty on the unwanted heat.
Future study can be conducted in the following aspects:
- An improved dynamic linearization mechanism for a device where the operation
has a wide range should be investigated; the internal model linearized around a
given operation point cannot guarantee the right prediction of the system when the
operation point digresses.
- Simplified data-driven modeling or multi-model approach shall be considered for
some complex system to reduce model development efforts. Meanwhile,
modeling large transients is very difficult with first-principle alone.
- The developed models can also be utilized to couple with the space for the
investigation of simultaneous temperature and humidity control.

164
CHAPTER 6 CONTROL OF SPACE THERMAL
CONDITIONING
6.2 Base control of the space conditioning 6.3 MPC architecture and performance evaluation
- Base control
- Space layout
- Thermal comfort and system model
- Hammerstein-Wiener structure based nonlinear
MPC
6.1 System description
The IWn integrated building energy system is a living laboratory for the research on high
performance buildings and conditioning techniques. Fig. 6-1 shows the outside view of
the space and the surroundings. The building features a reconfigurable open space and
accommodates offices for some thirty graduate students and faculty. According to a DOE
report, the dedicated outdoor air system, radiant cooling/chilled beam, enthalpy recovery
unit, and displacement ventilation are listed as current technology options for commercial
buildings. The utilization of these techniques in commercial buildings could bring in
around 1.6 quadrillion BTUs potential savings to the US annually (DOE 2008). The
integrated system has all the four energy efficient technologies installed. Meanwhile, the
automated outside louver and blinds endow the system potential exploitation of passive
heating and daylighting for energy conservation.

Figure 6-1: External view of the building system

165
Approximate 600 m
2
area is conditioned. Twenty-six sets of water mullions are deployed
along the facade to provide heating and a moderate level cooling to the space. Each set of
water mullion consists of four vertical embedded pipes and is equipped with a
modulating valve on the inlet pipe. A surface temperature sensor is also installed on the
outlet point of each mullion. The water is supplied from the campus grid via a tertiary
pump. A heat exchanger is installed in the basement to covert the steam from the campus
grid into hot water and feed the space in winter. In the heating season, the primary supply
water temperature is adjusted by modulating the steam valve. In the cooling season, the
primary supply water temperature is not adjustable. To avoid condensation on the surface
of the water-based conditioning terminals and also provide addition control flexibility, a
mixing valve is mounted after the tertiary pump. The mixing valve controls the supply
water temperature circulated to the space. The other water-based terminals for local
environment control include suspended radiant ceilings and ductless radiant chilled
beams. When additional or individual conditioning is needed, these terminals can be
controlled individually to provide extra heating and cooling.
As a water-based air-conditioning system, an additional air processing unit is needed for
ventilation air and space humidity control. A conventional chilled water coil can only
remove the moisture to the extent of a dew point close to the chilled water temperature,
which is around 12
o
C degree. In a humid and mild weather, a thermal coupling between
the sensible load and latent load is inevitable with this approach. A hybrid
dehumidification unit with an enthalpy recovery wheel, a vapor compression heat pump,
and an active desiccant wheel is installed to condition the outside air and provide
humidity control. The active desiccant wheel is regenerated by the air through a gas
burner heat exchanger. The conditioned air is evenly distributed into the space via the
floor diffusers. The amount of outside air intake can be adjusted from zero to one
hundred percent of the total air based on the indoor and outdoor conditions.

166
Besides these active conditioning techniques, there are passive techniques for illuminance
and thermal conditioning control. To maximize the utilization of the natural light for day
lighting and passive solar heating purpose, the space has about 70% of the external wall
as two-layer low-E Argon windows. About 30% of the windows are operable.
Daylighting and passive solar heating can be controlled by the openness of the external
louver and internal window blinds. The shading devices are motorized and linked to the
BAS. Ambient lighting can also be provided by the continuously dimmable up ceiling
lights when the daylighting is not sufficient or available. Windows also serve as the
natural ventilation and night cooling control devices. In addition to the manually operable
windows, eight of the transom windows are equipped with motors and linked to the BAS
to provide automated natural ventilation and night cooling to the space. Fig. 6-2
illustrates the mechanical components deployed in the building.

Figure 6-2: Illustration of the integrated building energy system
As a research center for high performance integrated systems, the supply side has a solar
heating and cooling sources installed beside the campus grid. The solar thermal system
consists primarily of 52 m
2
parabolic trough solar collectors and a 16 kW two stage
absorption chiller driven either by hot fluid from the solar receivers or by natural gas fuel.
They serve as supplementary heating and cooling sources and can also feed the surplus
1 campus
thermal grid
integrated
RTU
water system
1. radiant panels
and cool beams
2. water mullions
air system
3. air diffusers
lighting system
4. dimmable
artificial lights
2
3
4
5. operable windows
6. motorized blinds
7. motorized louvers
5
6
7
Active system
Passive system
natural
resources
nature gas
electricity

167
heating and cooling back into the campus grid. The components are to be gradually
integrated into a building wide BAS.
6.2 Base control of the space condition
6.2.1 Base control
As a living and lived in laboratory, the space has been constantly used for experiments
research projects. The control algorithm and logic of the subsystems and equipment are
not always the same. The basic structure of the control architecture includes a layer of
rule-based supervisory control and the local SISO based ON/OFF and PID control.
In general, the space is conditioned by the water mullions deployed along the facade. In
summer, the chilled water is delivered from the campus grid at 44
o
F degree. The
secondary supply water temperature is modulated by the mixing valve to cool down the
space based on the average space temperature. The ventilation unit sitting outside of the
space provides fixed amount of outdoor air intake at a constant supply air rate during the
mechanical cooling season. The control of the ventilation unit is mainly for latent load
accommodation and the supply air temperature is fixed at around 72
o
F. The impact on the
space thermal load due to the ventilation unit can be regarded as a constant disturbance.
In winter, the primary supply water temperature is reset based on the outdoor air
temperature. The mixing valve modulates the secondary supply water temperature to
meet the average space air temperature set point. The ventilation unit is shut down and
the ventilation is provided by the infiltration air.
6.2.2 Space layout
The space is designed for easy layout reconfiguration and accommodates different
experiments. For this study, the north bay is divided into two zones: one along the east
facade, and the other along the west facade. Fig. 6-3 illustrates the space layout.

168

Figure 6-3: Illustration of the space layout
6.3 MPC architecture and performance evaluation
Building mechanical system serves for a comfortable indoor environment with minimum
energy consumption. A good thermal comfort not just increases the occupants'
satisfaction but also improves the productivity and promotes the health. Since thermal
comfort is not a directly measurable variable, it cannot be used with the conventional
control law approach, neither can the energy consumption. The map from the main
measurable variables, e.g. indoor air temperature and humidity to the thermal comfort
condition is not a simple linear function. In addition, to utilize the power of classical
linear MPC, the nonlinearity involved in the modulation of terminals should also be
considered. For instance, a VAV box delivers heat or coolness to the space by changing
the air flow rate and/or supply air temperature. The relationship of the energy to the air
flow rate and supply air temperature is nonlinear either.

169
For MPC control, the main purpose is to utilize the model and available future
information to proactively prepare and arrange the manipulated variables. The building
space usually has much slower dynamics and longer time scale compared to the local
devices. Looking-ahead capability is also of great importance especially when there are
energy storage capacitance, space use difference, and multiple thermal sources. By
properly manipulating the available variables, defining the cost function, and tuning the
weights, MPC can potentially improve the energy performance of a given building
system.
6.3.1 Thermal comfort and system model
An MPC is driven by the internal model, the defined cost function, and the given
constraints on the system states, inputs and outputs. While most constraints are imposed
due to the physics, the cost function is defined more loosely in various formats. In an
optimization based general MPC, the cost function can be almost anything and in any
format. Energy and thermal comfort may both be included to represent the overall cost.
For energy consideration, the cost function can be defined as the summation of the
overall energy consumption over the prediction horizon. For overall cost consideration,
the monetary summation of the different power consumption by the equipment and/or the
power surcharges might be used. When the occupants' thermal comfort is concerned, the
cost function can be rearranged with the price tag added to the predefined comfort index.
An alternative approach is to set the thermal comfort as constraints. It deserves to
mention, that a cost function of an optimization based control is more or less subjective
and only relevant to the specific settings. Utilizing thermal comfort other than fixed air
temperature as a control variable can also bring in energy savings.
6.3.1.1 Thermal comfort evaluation
Building mechanical system maintains an acceptable indoor environment for the
occupants in the space. The performance of a system can be evaluated based on the
calculated comfort at the sampled points. PMV and PPD are two widely disseminated

170
indexes. In an air-conditioned space, the main independent variables in the function for
PMV and PPD are the indoor air temperature and humidity, in addition to other six air
states and personal conditions. The models for PMV and PPD are provided in Chp. 4.
The solution of the models requires iteration. The models are programmed in Matlab as a
function.
To guide a control toward the desired goal, the "cost" for discomfort and the energy
consumption shall be evaluated. It can be combined as a single cost function or organized
as multi-objective function. Liao and Dexter (2004) used the air temperature range to
define the thermal comfort for a heating system. A penalty is imposed when the air
temperature goes beyond the predefined thresholds. The overall comfort cost is then
calculated as percentage based on the comfort violation in the given time period. A
weighting factor is assigned to the cost in different zones. The final comfort cost in
percentage is then combined with an energy efficiency of a boiler to form the system
performance index. Yu et al. (2012) treated thermal comfort violation as a penalty to the
overall energy consumption. PMV values are calculated for the individual zones in a
building. A violation counts when the building is occupied. A weighting factor is
assigned to the different zones based on the function and relative importance to obtain the
overall PMV. The penalty is calculated based on the weighted PMV. Kummert and
Andre (2005) adopted an expression with PPD and energy consumption considered. A
reference zone is picked to evaluate the overall system performance. PPD values are
integrated and regarded as the overall cost due to occupants' discomfort. One arbitrary
coefficient is assigned as the weight of discomfort cost in order to allow direct
summation of comfort cost and energy consumption.
Thermal comfort can be directly treated as a reference, just as the room air temperature,
and get involved in the cost function as a quadratic term. Alternatively, a band of thermal
comfort value may be imposed as constraints to differential the operation variation due to
occupancy changes. For a general optimization based nonlinear MPC, a summation of
weighted thermal comfort and energy consumption can be considered.

171
6.3.1.2 Building dynamic model
The complexity of temperature evolution in a building comes from the thermal
interaction among the zones and between the indoor and outdoor air through the
enclosures. To evaluate the effect, a model that can describe the dynamics is needed. A
RC model has a linear structure and very reliable feature to represent the dynamics of a
building system once established. However, for a large building as studied here, it is a
challenge to deduce and denote the model equations if handled manually. The number of
governing equations grows exponentially as the number of zones increases. The building
model is constructed as a linear state space representation based on the methodology
described in Chp.4.2. A tool is developed in Matlab/Simulink to generated a dynamic RC
model automatically for the studied building. The following deduction (Kun Deng et al.
2010) is utilized to obtain the A and B matrixes in the model:

0 =
ij
A
,
( ) c e = j i j i if , ,

(6-1)

( )
ij i ij
R C A 1 =
,
( ) c e = j i j i if , ,

(6-2)

=
=
i j
ij ij
A A
,
( ) c e = j i j i if , ,

(6-3)
where A is the state matrix,
c
denotes the adjacent relationship, C is the total thermal
capacitance of the given node (kJ/K), R is the total thermal resistance between two
nodes (K/kW). Subscripts, i and j, denote the nodes.
The building dynamic matrix for the studied space is based on the space configuration
and layout. The roof for each space is divided into two orientations as facing south and
north. There are in total seven boundaries, including the four orientations, two direction
roofs, and the floor. The natural infiltration model is adopted from Yu (2012) and treated
as unmeasured disturbance. The heat gain from the solar radiation is modeled based on
the different orientations as described in Chp.4.

172
6.3.2 Hammerstein-Wiener structure based fast nonlinear MPC
For most conventional buildings, there are single mechanical heating or cooling sources
deployed in the space and the controlled variable is typically the space air temperature.
For instance, in a VAV conditioned building, the supply air temperature and the air flow
delivered into the individual room are the manipulated variables. In a fan coil conditioned
space, the supply water temperature and the air flow through each fan coil are the
manipulated variables. In a residential building, the power input to the air-conditioner or
furnace can be considered as the manipulated variable. Under these circumstances, it is
beneficial to design a model-based predictive controller with the building treated as a
linear time-invariant system. Fig. 6-4 illustrates the layout the internal system for a linear
MPC structure with heat flux as the system inputs and direct measurable variables as
outputs.

Figure 6-4: Pure linear classical MPC for building systems
This approach has the following benefits:
- By using a linear model and quadratic cost function, the problem becomes a
highly structured convex problem and can be solved fast. There are mature
solution and analysis methodologies;
- The space system can be decoupled from the mechanical system. The space
model can be re-utilized for other study purposes;

173
- The manipulated variables can be set as the heat flux and the cost function can
directly reflect the overall energy consumption.
Realistic hypotheses are made for the application of this nonlinear approach:
- There is no large frequent disturbance in the space. This is generally true for the
building air-conditioning;
- The operating points do not change frequently or span very widely. A space is
normally conditioned in a narrow room air temperature range to ensure the
occupants thermal comfort.
The control variation due to model mismatch is assumed to be off-set by the feedback
mechanism of the control under the hypotheses.
While the building system dynamics can be reasonably simplified into a linear time-
invariant model as a RC thermal network, the nonlinearity still exists. For example, Freire
et al. (2006) considered the nonlinearity involved in the outputs from measureable
variables to thermal comfort index. Their study described a regression model based
Wiener structure nonlinear MPC for a building and studied the basic performance with
thermal comfort in the cost function and constraint. However, the building is over-
simplified as one zone with only a heat flux as the direct input. This does not usually
exist in real buildings and cannot be implemented directly.
LTI (V) h(W)
f(U)
( ) k U
( ) k V
( ) k W ( ) k Y
nonlinear nonlinear

Figure 6-5: Hammerstein-Wiener nonlinear model
Fig. 6-5 illustrates the typical model structure in reality. The two ends of the linear time
invariant (LTI) system are static nonlinear mapping. A system with only the static
nonlinearity before the LTI is termed a Hammerstein nonlinear system. A system with

174
only the static nonlinearity after the LTI is called a Wiener nonlinear system. A system
with the nonlinearity at both ends is termed Hammerstein-Wiener (HW) nonlinear system
(Rohit et al. 1998). The nonlinearity in a Hammerstein system comes from the fact that
the heat flux into the air space is usually not the direct manipulated variable. In a Wiener
system, the direct measurable variables are not the target control variables.
In IWn space, during mechanical heating and cooling seasons, the water-based system is
the main thermal energy source to the space. The ventilation unit can be regarded as a
constrained white noise with the mean supply air temperature close to the space air
temperature and the mean humidity close to 55%. The manipulated variables are either
the supply water flow rates to the individual terminals or the supply water temperature to
the overall hydronic system. To enable a thermal comfort oriented control, the comfort
index PMV or PPD is "sensed" as the system output.
LTI (V)
h(W)
Linear
MPC
( ) k V
( ) k Y
NL
-1
NL
-1
-
+
( ) k Y
r
Wiener model
( ) k E
W

Figure 6-6: Control structure for Wiener model plant
To utilize the advantage of linear MPC for online implementation and solve the nonlinear
system with static nonlinearity, the system inputs and outputs need to be processed
accordingly. For a Wiener model, as shown in Fig. 6-6, the outputs from the LTI are
mapped into the desired outputs by using the inverse function of the mapping. The signal
is then compared to the inversed reference set point. The errors are fed into the linear
MPC to form a close loop feedback control. In a thermal comfort oriented control, the

175
direct map is from the room air temperature, room air humidity, air velocity, metabolic
rate, external work, clothing insulation to the thermal comfort index.
In a Hammerstein model, as illustrated in Fig. 6-7, the control action generated by the
linear MPC first goes through a static nonlinear inverse function, e.g. from the heat flux
to the water flow rate. The actual signal is then obtained and fed into the system after the
inverse map. In IWn space, the ultimate input nonlinearity comes from the actuators and
valves. In addition, for a complex hydronic system, the coupling among the individual
terminals shall also be considered. A new handling methodology of hydraulic-and-
thermal coupled system control, termed "node-branch-state" approach, is proposed and
studied in Chp 7.
LTI (V) f(U)
Linear
MPC
( ) k V ( ) k Y
NL
-1
-
+ ( ) k Y
r
Hammerstein model
( ) k E
Y
( ) k U

Figure 6-7: Control structure for Hammerstein model plant
Nonlinear Hammerstein-Wiener (HW) model consists of a linear dynamic block and two
static output non-linearity functions on both sides. This structure is more realistic in
representing space conditioning for thermal comfort in most building systems. The
nonlinear HW system can be described by the following equations:

( ) ( ) ( ) k Bu k Ax k x + = +1
(6-4)

( ) ( ) k Cx k y =
(6-5)

( ) ( ) ( ) k y h k z =
(6-6)

176

( ) ( ) ( ) k u f k v =
(6-7)
where A , B , and C are the state, input, and output matrix respectively,
x
is the state
variable vector in the LTI,
y
is the output vector in the LTI,
u
is the input vector in the
LTI, z is the overall system output vector,
v
is the overall system input vector, h and
f are the two static mapping functions, and k is the time step.
For the water-based heating and cooling system as in the IWn, the Hammerstein-Wiener
model and the linear MPC-based control loop is illustrated in Fig. 6-8.
Dynamic LTI
space model
Static nonlinear
map PMV = h (x)
MPC
Static nonlinear
map Q = f(m)
Thermal hydronic system
Static nonlinear
map m = invf (Q)
Static nonlinear
map x =invh (PMV)
Reference
+
x =invh (PMV)

Figure 6-8: IW water based system in Hammerstein-Wiener structure
The mapping from thermal comfort to the space room air temperature and humidity does
not have a one-to-one relationship, with the acceptable room air temperature and
humidity in a narrow range. To avoid the fluctuation due to room air temperature and
humidity set point fluctuation, the room air humidity is used as a measured disturbance.
This change is reasonable since most building mechanical system tracks the room air
temperature set point and allows the humidity fluctuating in a small range.
6.3.2.1 Formulation
Fig. 6-9 illustrates the variables involved in the proposed nonlinear MPC.

177
Linear building system
(LTI)
1
Q
n
Q
1
m
nonlinear
mapping
n
m
oa
T
bd
T
1 , sol
Q
ln so
Q
1 sen
Q
n sen
Q
,
lat
Q
oa
m
oa
g
s
u
r
f
a
c
e
a
i
r

n
o
d
e
a
i
r

n
o
d
e
1 , ra
T
n ra
T
,
ra
g
nonlinear
mapping
1
PMV
n
PMV

Figure 6-9: Illustration of the variables in HW nonlinear MPC
As shown in Fig. 6-9, two heat fluxes from the terminals are the direct inputs to the air
nodes. The outdoor air temperature, boundary surface temperature, and solar radiation
heat flux are the external inputs that affect the space temperature though the construction
surfaces. In addition, the sensible heat gain, latent heat gain, infiltration air flow rates,
etc., are the disturbances to the system. The indoor air temperature and average air
humidity are the direct outputs from the LTI space system. The mathematic expression of
the LTI system for the two room layout 2R1C space can be put as:

( ) ( ) ( ) k Bu k Ax k x + = +1

(6-8)

| |
T
ra i m i ra
g T T x
1 14
11 ... 1 , 2 , 1 ,

= =
=

(6-9)

| |
T
i
Q u
1 2 2 .. 1 =
=

(6-10)

| |
T
i sol lat i sen
Q Q Q d
1 5
2 , 1 , 2 , 1 ,

= =
=

(6-11)

178

| |
T
oa oa og i oa
g m T T w
1 9
6 ... 1 ,

=
=

(6-12)
where u, d and w are the manipulated variable vector, disturbance vector, and external
disturbance vector respectively,
oa
m
is the infiltration air flow rate (kg/s), and Q
is the
heat flux (kW). Subscripts ra, oa, m, sen, lat, sol, og denote the room air, outdoor air,
material, sensible heat, latent heat, solar heat, and ground condition respectively.
In the nonlinear HW-MPC, water flow rates to the space terminals are considered as the
manipulated variables for the space thermal conditioning. Instead, the supply water
temperature may be regarded as one manipulated variable if the whole space is treated as
one zone. With the two-zone space layout, the total number of manipulated variables is
two for each control time internal. In the prediction horizon, e.g. 100 minutes, twenty
variables in total needs to be determined by the MPC controller. The input constraints are
the minimum and maximum water flow rates. With 10 to 20 minutes as one control
horizon, which is long enough to achieve a stable water distribution among the terminals,
the water flow changing rates are not constrained.
The following scenarios in Table 6-1 are considered in this study to evaluate the
performance of classical MPC and proposed HW nonlinear MPC for the hydronic system.
The MPC objectives are selected as the room air temperature, thermal comfort, and
energy consumption respectively. The quadratic cost function follows the format as stated
in Chp. 2 since the internal MPC is linear LTI-based. The simulations are conducted for
five days in the heating season. The control horizon for the study of the space is 1 and the
prediction horizon is 10 unless otherwise declared. Results are plotted and analyzed in
next section. Both Scenarios SC-1 and SC-1' are temperature set point oriented control.
Table 6-1: Control scenarios for classical MPC and HW nonlinear MPC
Scenario Problem description
SC-1 Classical MPC, direct heat flux inputs, temperature oriented control. 24/7 schedule.
SC-1' Classical MPC, direct heat flux inputs, temperature oriented control. Office schedule.

179
SC-2 Wiener nonlinear MPC, thermal comfort oriented control.
SC-3 Wiener nonlinear MPC, energy consumption and thermal comfort oriented control.
Thermal comfort serves as constraints, PMV = [-0.5, 0.5].
SC-4 HW nonlinear MPC, thermal comfort oriented control.
SC-5 HW nonlinear MPC, energy consumption and thermal comfort oriented control.
Thermal comfort serves as constraints, PMV = [-0.5, 0.5].
SC-6-1 As case 5, with variable occupancy schedule. When unoccupied, PMV = [-3, 3];
when occupied, PMV = [-0.5, 0.5]. Band control.
SC-6-2 As case 6-1, dynamic constraints based on outdoor air temperature.
SC-7 As case 4, with variable occupancy schedule and full look-ahead ability. When
unoccupied, PMV = [-3, 3]; when occupied, PMV = 0.
6.3.2.2 Results

Figure 6-10: IWn model without nonlinearity
For pure LTI-based control, there is no nonlinearity included. The system can be depicted
by the linear combination of the inputs and outputs as in a state-space model. The

180
manipulated variables are the heat fluxes added into the air node in the space. The effects
on the outputs from the disturbances are described linearly as well. The outputs are the
direct measurable variables from the system. The model of the two-zone test bed is built
in Matlab/Simulink as depicted in Fig. 6-10.
For a Hammerstein structure nonlinear MPC, a block of radiant hydronic system is placed
before the heat flux. The manipulated variables from the controller are the water flow
rates into the terminals. Fig. 6-11 illustrates the model built in Matlab/Simulink. The
nonlinear map from the air temperature and humidity to thermal comfort is calculated in
Matlab. When control is concerned for HMPC, WMPC and HW-MPC, the inverse
functions are included as shown in Fig. 6-8. The codes are omitted here.

Figure 6-11: IWn model with nonlinear hydronic system
As aforementioned, the disturbances, including the outdoor air infiltration, solar radiation,
etc, are pre-processed based on the models detailed in Chp.4. Fig. 6-12 plots the outdoor
air temperature, relative humidity, solar radiation to the two rooms. The outdoor air
fluctuates between 2C to 18C during the simulation period, which represents a typical
winter.

181

Figure 6-12: Outdoor air condition and solar radiation
A conventional space thermal conditioning is to track a room air temperature reference. It
is typically a fixed set point, e.g. 24C. A base case is set as a 24/7 occupied schedule
with the fixed room air temperature set point at 24C. Fig. 6-13 plots the room air
temperature, relative humidity, PMV value, and heat inputs. The MPC modulates the heat
inputs properly and maintained well the room air temperature close to the reference as
shown in the figure. The overall energy consumption for the 24/7 base case in the five
days is 3091 kWh. The PMV calculation shows that, with a fixed room air temperature,

182
the PMV value fluctuates around 0.2. An energy saving can be achieved if the thermal
comfort can be used as the reference instead of the air temperature.

Figure 6-13: Simulation results, SC- 1
An office occupancy based air temperature control is also simulated for the five day
period serving as an additional base case. The results are collected in Fig. 6-14. The
system has a morning warm-up operation to ensure a smooth transition in the space from
unoccupied status in the night to occupied status in the day. The overall energy
consumption for this base case is 1941 kWh.

183

Figure 6-14: Simulation results, SC- 1'
Scenario SC-2 is constructed to investigate the thermal comfort based MPC control. The
nonlinearity related to the energy consumption is ignored in this control. With a Wiener
structure for the thermal comfort mapping, the internal classical MPC works well with
the overall hourly summation on square PMV equal to 0.07. The PMV tracking output is
plotted in Fig. 6-15, close to the set point 0. The manipulated variables are the heat fluxes.
Compared to Scenario SC-1, the room air temperature drops to be below 24C and
fluctuates along with the room air humidity as shown in Fig. 6-15. The overall energy
consumption is reduced by around 6% to 2895 kWh. With different room air temperature,
the indoor air relative humidity increases slightly.

184

While setting PMV to 0 can maintain a very good indoor thermal comfort, a slightly cool
indoor air condition in winter can bring in more energy savings and is usually acceptable
with winter clothing. Less than 10% occupants vote dissatisfaction when PMV drops to -
0.5 according to the existing study in (ASHRAE 2004). In Scenario SC-3, the MPC is
configured to minimize the energy consumption with PMV constrained between -0.5 and
0.5. As shown in Fig. 6-16, the MPC successfully drives the system to the lower limit
since it causes the lowest energy consumption. Compared to Scenario SC-2, with PMV
equals to -0.5 and the corresponding room air humidity, the indoor air temperature is

185
reduced by about 2C. The overall energy consumption decreases to 1977 kWh, which is
about 36% less than the base condition. Fig. 6-16 also plots the heat input into the space.

From the study of Scenario SC-2 and SC-3, we see that the proposed Wiener nonlinear
MPC can make full use of the advantages of classical linear MPC. It turns the room air
temperature based space conditioning into a thermal comfort oriented control while
achieving additional energy savings. The overall run time for 5 days control takes only 76
to 176 seconds to identify the optimums for the 180 variables. The nonlinearity is
successfully handled with this approach. In reality, a heat flux is usually not directly

186
controllable. The corresponding nonlinearity from the manipulated variables, e.g. water
flow rates, to the heat fluxes needs to be considered. A Hammerstein-Wiener structure is
introduced in this study. Scenarios SC-4 to SC-7 are designed to investigate the
performance of HW nonlinear MPC, including the reference tracking, disturbance
rejection, and cost function minimization.


187

In Scenario SC-4, the water flow rates are used as the manipulated variables in the HW
nonlinear MPC. The controller is designed to minimize the discomfort from the
occupants by letting PMV track 0. From Fig. 6-17, it can be seen that compared to
Scenario SC-2, the HW MPC handles the system well and the PMV tracking error is
close to that of a Wiener nonlinear MPC where the heat fluxes are used as the
manipulated variables. With this system layout, the computational cost is increased by
only 2 seconds than Scenario SC-2. Similarly, Scenario SC-5 is included to evaluate the
performance of HW nonlinear MPC. The controller brings the system to the lowest
acceptable PMV to minimize the energy consumption. The system performance is very

188
similar to Scenario SC-3 as plotted in Fig. 6-18 in terms of PMV. The hot water flow
rates are controlled smoothly. In the following scenarios, HW nonlinear MPC is utilized
to respect the reality.

Figure 6-19: Simulation results, SC- 6-1
Commercial office buildings may have variable occupancy. The change of set point is a
disturbance to the controller in addition to the unmeasured solar heat flux and internal
heat gains. In Scenario SC-6-1, a variable occupancy schedule is studied and a band
control is adopted. When the space is unoccupied, the constraint on PMV is loosened to
[-3, 3]. They are reduced to [-0.5, 0.5] when the space is occupied. The controller is set to

189
minimize the energy consumption during the 5 days. As plotted in Fig. 6-19, the
controller drives the system between the two different constrains with the occupancy
switched between occupied status and unoccupied status. The simulation takes 187
seconds to obtain the optimum.

Figure 6-20: Simulation results, SC- 6-2
However, it can also be seen from Fig. 6-19 that the PMV value during the occupied
hours can be lower than -0.5 and may cause discomfort. It is due to the system heating
capacity limitation, outdoor air condition, and inertia of building structure. With the

190
energy consumption reduced by 56.7% compared to SC-1', the summation of squared
PMV during the occupied hours increases to 39. Additional consideration should be taken
to improve the performance.
In Scenario SC-6-2, the constraints are adjusted dynamically based on the outdoor air
temperature prediction. The modified MPC drives the system to be within the acceptable
PMV constraints with minimum energy consumption. Fig. 6-20 illustrates the simulation
results. Compared to Scenario SC-6-1, the overall energy consumption is increased by
293 kWh but the summation of squared PMV is reduced to 13.44. While energy
consumption is of our concern, the thermal comfort shall not be compromised. Compared
to the office schedule base case SC-1', the energy saving in SC-6-2 is about 41.6%. The
strategy adopted in Scenario SC-6-2 is regarded proper to improve the overall
performance.
current i
Occupied hours
PMV=0
P
M
V
-3
0
Time
heavy weight
light weight
Unoccupied hours
Prediction horizon

Figure 6-21: Illustration of desired transition from unoccupied condition to
occupied condition in winter

191
When a space is switched from unoccupied hours to occupied hours, a smooth transition
of the indoor air thermal condition is needed to ensure the space thermal comfort when
people come back to the building. It is termed morning warm-up in winter or cool-down
in summer. The strategy is also to minimize the energy consumption while the space is
unoccupied. Fig. 6-22 plots the desired set points in a receding horizon manner. The set
point guides the system gradually toward the comfort condition with PMV set to 0 during
the day time.


192
Unless a heuristic and/or iterative search is conducted, we don't know what a proper
profile shall be used to guarantee the smooth transition with the changing internal and
external conditions. With MPC, the desire is fulfilled by penalizing the reference tracking
errors and utilizing look-ahead measure on the references and constraints. The look-ahead
technique is similar to a feed-forward strategy used in conventional SISO PID control.
Scenario SC-7 is designed to evaluate the performance with the techniques. The control
horizon is set as 1 and the prediction horizon is set as 10. Fig. 6-22 plots the simulation
results. As can be seen in the plot, the controller successfully brings the system back to
the desired comfort with PMV equal 0 when the space starts to be occupied. The
summation of squared PMV during the occupied hours is reduced to 0.45. The overall
energy consumption compared to the based condition SC-1' decreases by about 7%. The
modification on the HW nonlinear MPC increases the evaluation time to about 460
seconds for a 5-day simulation.
Table 6-2: Comparison of energy consumption, thermal comfort, and simulation
duration
Case
number
Heating Energy
Consumption
(kWh)
Mean
(PMV+0.5) in
Occupied
Hours
Energy Savings
(%)
Simulation
Duration
(s)
SC-1 3091.7 0.72 0 36.1
SC-2, 4 2895 0.56 6 76.9 - 78.1
SC-3, 5 1977 0 36.1 175.8 - 198.1
SC-1' 1941 0.69 0 40.6
SC-6-1 840.8 -0.25 56.7 187.4
SC-6-2 1134 0.01 41.6 205.3
SC-7 1800 0.51 7 460
The simulation results in terms of the energy consumption, thermal comfort, and
simulation duration for the scenarios are collected in Table 6-2. In addition to the
simulation duration and energy savings, an additional index based on PMV value in the

193
occupied hours is also included. We assume -0.5 should be adopted in winter, the results
show that the higher the mean PMV value, the lower the energy savings. Scenario SC-6-1
stands out since cool morning hours happen which is not acceptable. With a 24/7
schedule, comfort oriented MPC control in the test bed can bring in from 6 to 36 percent
energy savings. With an office schedule, comfort oriented MPC control can offer from 7
to 41.6 percent energy savings.
In this chapter, the space thermal conditioning is investigated in a model based predictive
control environment. The purpose is to utilize the capability of MPC on handling of
MIMO coupling, system constraints, and multiple criteria on system performance. This is
useful when a space thermal conditioning is not just to maintain a single temperature
reference, but a virtual thermal comfort defined by multiple variables. The application
can potentially coordinate the system for the consideration on comfort, energy, or both.
To reduce a high computational cost and ease the online implementation, Wiener and
Hammerstein-Wiener nonlinear MPC for building systems are proposed and investigated.
With a Wiener structure-based MPC, the nonlinearity involved on the controlled variable
side in a system is handled by decomposing the system into a linear system plus a static
nonlinear map. With a Hammerstein-Wiener structure-based MPC, the nonlinearity on
the input and output side is processed by using a linear system plus two static nonlinear
maps. For space thermal conditioning in most HVAC systems, the two innovative
approaches are practical and adequate in order to improve the performance with optimal
control.
Several scenarios with various MPC structures and different techniques are designed for
the study. The aspects, including reference tracking, disturbance rejection, computational
cost, energy savings, etc., are compared. With the study, the following points can be
concluded about the proposed MPC:

194
- The space thermal conditioning system can be controlled toward a minimum
energy consumption condition, tracking a desired thermal comfort condition, and
handling the nonlinearity involved in the system. Thermal comfort in a band
offers more energy savings than a fixed thermal comfort reference approach;
- The rejection capability on the unmeasured disturbances from the infiltration,
internal heat gains, etc., is acceptable. The controller can proactively guide the
system transit from unoccupied condition to occupied condition in an energy
conservative manner;
- Dynamic look-ahead on both the set points and the weights improves the overall
performance since the future information is utilized for the preparation of a
building system where the time scale is of hours level;
- The computational cost of the simplified approach with linear MPC as the core is
low which makes it suitable for online implementation in conventional building
systems for space thermal conditioning. All the scenarios for one week control
can be accomplished in less than eight minutes;
- By applying thermal comfort oriented control, in a 24/7 schedule condition, an
MPC helps acquire about six to thirty six percent of energy savings. In an office
schedule condition, about seven to forty two percent of energy may be saved by
using MPC empowered thermal comfort control.
- Expand the variables involved in the thermal comfort determination. In this study,
only the space air temperature and air humidity are considered. For personal
comfort control, the occupants' preference, activity, clothing, etc., can be
embedded into the algorithm;
- Investigate the performance of HW-MPC for various thermal conditioning
terminals. Variable air volume terminal boxes, fan coils, and so forth can be
experimentally studied.
- Investigate the implementation architecture and performance of HW-MPC for a
building with a large number of terminals. Centralized, distributed, and/or a

195
derivative combination structure may be needed for a given circumstance. The
pros and cons can be compared to provide guidance for real application.

196
CHAPTER 7 CONTROL RELATED ISSUE ON
HYDRONIC SYSTEM
7.2 Hypotheses 7.3 Modeling methodology 7.4 Model and simulating
structure
7.5 Evaluation cases
- Hydraulic system
simulation
hypotheses
- Hydraulic network
- Resistor components
- Pumps
- Serve as virtual sensing
in optimal control
- Case 1
- Case 2
7.1 Introduction
The main purpose of heating, ventilation and air-conditioning system is to obtain a
comfortable indoor climate. At the mean time, minimum energy consumption and little or
no operational problems are desired ideally. Compared to air-based systems, radiant
heating and cooling systems are well known for the feature of low energy consumption
and high thermal comfort (Gong 2010, Yu 2010, Laouadi 2004). The overall performance
of an air conditioning system and the energy efficiency rely not just on the system design
and construction, but more importantly on the sensing, controlling, and actuating
systems. A lot of research described that despite the improved efficiency of individual
components, the faults and inferior operation waste about 20-50% of the current energy
use in buildings. With existing buildings dominating the building energy consumption,
the research focuses to achieve tangible high energy efficiency of building systems
should be operation oriented. Mean while, in practice, due to the lack of sensing or zero-
sensing redundancy in building mechanical system, even the most sophisticated
controllers do not always function efficiently. This is especially true to a hydronic heating
and cooling system where the water distribution is influenced by the pipe pressure, the
actuator nonlinearity, and the valve nonlinearity.
Unlike a variable air volume terminal box, where the airflow rate and discharge air
temperature are measured and cascaded proportional integral (PI) controllers are
employed, the terminals in a hydronic system typically do not have measurements on the
water flow rates. The controllers are relatively simple as one-time tuned PI controllers for
the low cost on initiating and maintenance. It becomes difficult to evaluate the

197
performance and therefore conduct proper control actions on the valves in advance if a
prediction control is desired. The situation could be even more complex when there are
multi-pump powered branches and compact piping present in the system. In addition, as
Macleod (2008) pointed out that, while the water-based central heating system is
relatively cheap to run, it is also tended to be wasteful. Compared to the electric
counterpart, the control of water-based system is often rudimental and energy usage
information is not well aware by the end users to the same extent as in electric system.
Models and simulation could help control and operation by providing extra information
than basic sensors in the circumstance of improper or lack of sensing in the system. Yu et
al. (2011a) proposed a virtual calibration of supply air temperature sensor for rooftop
units to obtain the correct information on the measurement. The real value of supply air
temperature is critical in many aspects closely linked to energy efficiency, e.g. stage
control of heating and cooling, economizer, and outside air intake control, fault detection
and diagnostics etc. The model was also found to exhibit excellent robustness in terms of
the knowledge of the airflow rate as a low cost virtual sensor (Yu et al. 2011b).
Many researchers have investigated the approaches to improve the energy efficiency and
thermal comfort of a hydronic system for heating. Yu et al. (2010) utilized a combination
of system identification and numerical optimization on a floor radiant heating system. It
is found that a significant amount of energy could be saved if a dynamic warming-up
process can be utilized before the space is occupied. Liao and Dexter (2004) studied the
energy saving potential of a heating system by improving the controls of boilers.
Kulkarni and Hong (2004) utilized a state-space sensible heat transfer model to optimize
the control of a residential conditioning system. However, none of the aforementioned
study considered the influence of hydraulic coupling and distribution on the thermal
performance of the hydronic system.
There have been few studies that investigated the thermal performance of a hydronic
system by coupling the hydraulic distribution network and thermal system together.
Gamberi et al. (2009) utilized steady-state model to investigate the performance of a

198
heating system by solving the coupled hydraulic distribution and thermal system.
However, their study did not consider the discontinuity caused by pump ON/OFF and
valve OPEN/CLOSE due to the room air temperature control. Henze and Floss (2011)
studied the temperature degradation features in hydronic cooling system by solving the
coupled water distribution and thermal systems. Similarly, the approach of using
Kirchhoff circuit law does not suit the situation when part of the loop is broken with any
included valve at the OFF position. KN Rhee et al. (2010) conducted a simulation
investigation of hydraulic balancing to improve the room temperature control for a
radiant floor heating system. In their study, a hardware-in-the-loop approach is adopted.
While the hydraulic distribution information is more reliable, the cost of construction,
sensing and data acquisition is prohibitive that makes the approach impractical in real
applications.
There are two challenges to address before we utilize an integrated water network and
thermal system simulation:
1. For a given hydraulic network, the model should be relatively easy to acquire.
2. The hydraulic continuity should remain even the pumps or branch valves are off or
close to the off position.
With a conventional node-mass and loop-pressure-drop approach, the user holds the
responsibility to identify the necessary loops to constitute the hydraulic network. With a
relatively complex hydraulic network, it is close to mission impossible. Meanwhile, since
the definition of a loop holds true only when at least one of the element branches in series
is continuous, this approach does not have the flexibility to configure part of the loop
offline or totally loses the continuity when the valves or pumps on any included branches
are off.
Here we propose a "node-branch-state" based approach to overcome the two challenges
and simultaneously solve the problem for both pressures and flow rates. The method can
be utilized to constitute a stand-alone hydraulic network model for commissioning,

199
additional sensing, advanced control, etc. With minor modification and expanding, the
model can also be easily utilized and incorporated with thermal model to conduct a
thermal-hydraulic coupled simulation. While there are many studies conducted in the past
on constructing and solving water distribution networks, none has discussed the handling
of the discontinuity resulted from the valves and flow movers (Ormsbee 2006, Strafaci
2003). Commercial software packages are usually standalone and have many limitations
on or are not possible at all for thermal coupled co-simulation.
7.2 Hypotheses
While in this chapter a thermal-hydraulic coupled simulation is conducted, modeling of
thermal part is not of the major concern. The hypotheses of formulating control-oriented
thermal models could be found in the previous chapters and literatures (Thomas 1999).
The following hypotheses are made for the modeling and simulation of hydraulic network
part:
1. The flows in the network are continuously evolving steady states;
The time constants associated with the setting of flow in a hydraulic network for building
applications are usually very short compared with the thermal part. Therefore, the flows
can be assumed to settle instantaneously by using continuous evolving steady-state
models based on mass and energy conservation laws (Bhave 1988).
2. The pressure drop along a pipe line and across a component can be expressed by
power law to the flow rates;
The pressure drop of a given component depends on the flow status, e.g., laminar flow,
developing flow, turbulent flow (Warsi 1998). Under most circumstances in the
application of water-based heating and cooling systems, the flow can be regarded as fully
developed and the representative resistance coefficients can be applied. The variations of
fluid temperature, density, friction coefficients, and so forth, along the lumped
components are not considered.

200
7.3 Modeling methodology
Hydraulic Network:
The basic categories of the components existing in a large hydraulic network can be
divided into two as: joint nodes and through branches to establish the governing
equations. Fig. 7-1 illustrates the liquid flow network with one joint node and three
through branches. The balance equations serve as the base to the modeling of a hydraulic
network that links multiple components. For a joint node, where there are greater than
two through pipes linked to it, a mass conservation must hold that the summation of inlet
flow rates at any given node (e.g. i th node) from 1 to N equals to the summation of outlet
flow rates. N is the total number of the joint nodes in the hydraulic network.
i j
p
,
A
i j
m
,
node j
branch
joint node i
j
p
i
p
node
node

Figure 7-1: Liquid flow network with one joint nodes

e
=
c j i
i j
m
,
,
0
(7-1)
where
i j
m
,
is the mass flow rate from node j to i ( i j < ) (kg/s), i is the joint node of the
network, from 1 to N, j is the neighbor node to joint node i, c indicates the neighbor
relationship between i and j. The mass flow rates are directional.
The mass flow rates are directional in the formulation of mass equations. It does not
matter which direction is defined as positive flow, as long as the signs do not conflict for
the flow through the same branch for two neighbor nodes. For automated model
generation, for example, all nodes are considered as sinks by default with all flows

201
toward the joint node. A negative sign should be added to the same branch if it appears
second time in the mass balance equations. While with this balance equation any physical
node in a network can be treated as a node in a model, simplification is encouraged to
combine the trivial ones by using series or parallel analysis. The same holds for the
processing of loops, branches and components.
The total number of mass conservation equation equals to the total number of joint nodes.
Besides the joint nodes, there might also exist nodes that define the boundary of a
subsystem in a loop. One example will be the inlet and outlet of a secondary loop that
link to the primary heating and cooling system. The total flow balance equations equal to
the total number of the nodes in the network. The total number of the flow rates that need
to be solved in a hydraulic network equals to the total number of the branches.
In addition to node mass conservation, the mechanical energy conservation holds for all
the through branches with continuous fluid. Using the energy conservation across a
branch is more convenient and intuitive than applying the energy conservation law of
loops. The necessary number of loops for solving the hydraulic unknowns is difficult to
identify for a large scale network (e.g. as the system in Fig. 7-2). Meanwhile, a valve at
fully close position disables all the predefined loops that have the valve included. For
example, in Fig. 7-2, valve DV2 at close position deactivates all the nodes and branches
up and down stream between node 12 to node 20. A state in node-branch-state based
hydraulic model approach refers to the ON/OFF condition of any pumps and valves in the
network, whose change can lead to totally different network continuity and the
effectiveness of the governing equations.

202
w
a
t
e
r

m
u
l
l
i
o
n
s
w
a
t
e
r

m
u
l
l
i
o
n
s
w
a
t
e
r

m
u
l
l
i
o
n
s
R
a
d
i
a
n
t
C
e
i
l
i
n
g
P
2
C
o
o
l
W
a
v
e
P
3
P
1
m
a
i
n

b
y
p
a
s
s
Water Supply
Water Return
2
3
9
10 11 12
13 18 19
20
23 24
25
26
27
28
29
30
31
32
33
34
36
37
40
V1 V7 V8
BV1
BV2 BV3
DV2
DV3
DV0
DV1
21 22
1
35

Figure 7-2: Illustration of a secondary thermal hydronic network
Bernoulli's principle states that an increase in the speed of the fluid occurs simultaneously
with a decrease in pressure or a decrease in the fluid's potential energy (Batchelor, 1967).
Accordingly, the general expression of the energy balance for a continuous fluid without
power component is:
2
,
2
2
1
2
1
j j j i j i i i
v h g p p v h g p + + = A + + +

(7-2)
where p is the static pressure on the fluid (Pa),
is the fluid density (kg/m

3
),
g
is the
gravity (m/s
2
), h is the height of the point to the reference point (m), v is the volumetric
speed of the fluid at the point (m/s), i j
p
,
A
is the total pressure loss from j to i (Pa).
A pressure drop across the components on the branch with continuous fluid needs to be
identified to get the gauge pressure at the individual node of the hydraulic system. In
most building hydronic system, where a closed loop is applicable, the energy due to the
height difference and the kinetic element in a branch can be ignored for simplification. A
characteristic coefficient is employed to represent the major and minor pressure loss
across the branch for the pipe, fittings, coils, etc:

o
i j i j i j
m S p
, , ,
= A

(7-3)

203
where
i j
S
,

is the characteristic coefficient of the pressure loss across branch from node j
to node i, o is a constant depending on the flow condition of the branch.
There are many different components that exist on a hydraulic branch, including resistor
components (e.g. valves, pipes, coils, fittings, etc) and power components (pumps).
Valves can be controlled to modulate the flow by either reducing or adding additional
resistance to the branch. Pumps are the power components that circulate the flow by
adding head to the fluid. When a pump changes from ON to OFF, it shifts from a power
component to a resistor component. The handling of the changes is considered in this
proposed node-branch-state modeling approach.
Resistor components:
The static part of the resistor components of a branch, including pipes, coils, and fittings,
can be expressed as one consolidated component with the resistance calculated and
summed altogether. The value of the coefficient once identified remains constant:

=
k k
k
k
k
k
s i j
d
l
d
S
4 2 5 2 , ,
8 8
t

t

(7-4)
where is the D'Arcy-Weisback friction coefficient, d is the hydronic diameter, is the
length of the pipe, k is the counting number of the static resistor component of the branch.
The valves should be considered separately since the different valve openness incurs
different pressure drop across the valve. The expression is given by:

) (
2 , ,
0
v
v i j
k
g
S =
(7-5)
where 0 is the valve openness, k
v
is the valve conductance coefficient, which is a
function of the valve openness.
The total pressure drop of a branch with continuous fluid is the summation of that of the
static components and the modulating valves. Since the flow rate are considered the same

204
through all the components on the branch, the total pressure loss can be obtained by using
the following coefficient equation:

= =
+ =
m
l
v i j
n
k
k i j i j
S S S
1
, ,
1
, , ,

(7-6)
Similarly, the total flow rate through parallel pipes, which share the same inlet node and
outlet node, is the summation of the flow rate through the individual pipe. The pressure
drop across the shared nodes is the same to all the member branches. The relationship can
be put as:

=
=
n
k
k i j i j
S S
1
, , ,
1 1
(7-7)
The series and parallel flow and pressure drop relationships can be used to reasonably
simplify the original hydraulic network before resorting to a numerical model.
pumps:
When there is a pump operating in a branch, the additional head should be added into the
proper side of Eq. 7-2 to account for the external mechanical energy. The relationship
between the pump head and volumetric flow rate can be regressed from the manufacturer
data sheet or onsite measurements as:

2
2 1 0
m a m a a H + + =

(7-8)
where H is the pump head (Pa), a
0
to a
2
are the curve coefficients.
If a variable frequency drive (VFD) is installed on a pump, applying the pump law on the
flow and pump head (Yu 2010), a final equation that represents the relationship under
partial speed can be obtained:

205

|
.
|
\
|
=
e
e
m
f H

2
(7-9)
where e is the motor rotation ratio.
To this end, a basic hydraulic model can be built with the proposed "node-branch-state"
approach. The number of variables to be solved is determined by the complexity of
system. For the simple one node system as in Fig. 7-1, there are three flow rate variables
and three pressure drop variables, including the boundary conditions. The number of
variables, mass balance equations, and energy balance equations increase greatly for a
complex network. For example, in the illustrated thermal heating and cooling system in
Fig. 7-2, there are eight water mullion terminals, four radiant ceilings, and four cool
waves. The total number of nodes in the network without simplification is 40. Node 1 and
24 are at the boundary. The total number of pressure balance equations is 58 for the
branches in the network. An automated model generation algorithm is needed for large
scale network to avoid the mission impossible and potential mistakes of manual counting.
The method is described in the next section.
7.4 Model and simulating structure
With the prescribed simplified modeling method, it is not difficult to generate a model for
a complex hydraulic network system. A handy processing tool is developed in Matlab to
automatically construct the large scale hydraulic network. The size of the network matrix
is node by node (40 by 40 for the network shown in Fig. 7-2) for model preparation. The
obtained model contains the identification of the variables, the map of branches to the
node connections. The condition of the boundary nodes can be a fixed pressure or mass
flow rate and should be provided in addition in order to solve for the unknown flow rates
across the branches and the static pressures at the nodes. The model is then encapsulated
as a stand-alone block that can be called by other functions in Matlab or model blocks in
Simulink. The inputs to the model are the pump states ( or, in addition, speeds if VFDs
are installed) and valve openness. The outputs are the flow rates through the branches and

206
the pressures at the nodes. Fig. 7-3 illustrates the program execution flow of the proposed
modeling approach.
Generate the basic nonlinear equation set:
1. node mass balance equations
2. branch energy conservation equations
Obtain the pump operation
and valve openess information
Call solvers to solve the equation set
pump is on?
Valve openness is
above the
threshhold?
Yes
Apply pump
head equation
No Apply resistance
equation
No Cut off the
branch
Yes
Apply valve
characteristics
Update the resistance coefficients and
parameters of the hydraulic network
Output the flowrates and pressures
Save as
initial values
Start
End
Update the states and
auto-write the equation set.
Pump states change or valve
states change?
Yes
Use previous equation set
as a function
Save pump states
and valve states
No

Figure 7-3: Node-branch-state hydraulic model process flowchart

207
Valve
openness
Controller Valve
Pump states
or speed
Controller Pump
Thermal response
of the multi-zone
space
Temperature sensors
Temperature
Controller
Hydraulic network
simulator to estimate
the states
+ +
Temperature
setpoints
Hydronic network
Thermal system

Figure 7-4: Example integration of the hydraulic network simulator
To conduct a cascaded control or model-based predictive control, where the thermal
water flow rate though the terminals are needed, the model can be further calibrated and
serve as a virtual sensor that provides all the needed hydraulic information back to the
controller. Fig. 7-4 illustrates such a control-oriented utilization structure. The hydraulic
network simulator is embedded to take the valve and pump states from the inner
controllers' outputs and provide the flow rate information back to the controllers.
7.5 Evaluation cases
To evaluate the proposed modeling and simulation approach, two simulation cases, a
hydraulic distribution simulation and a thermal-hydraulic coupled simulation, based on
the method are conducted for the system as shown in Fig. 7-2. The detailed description of
the building system and the thermal components are included in Chp 4. The hydraulic
system is taken from part of an existing water based thermal network in the test bed. The
pumps for the main circulation and the two local branches are constant speed pumps. The
pump curves are shown in Fig. 7-5.

208

Figure 7-5: Pump curves for the branches (U) and main pump (D)
7.5.1 Case 1- Hydraulic distribution
Case 1_0: all pumps are on

209

Figure 7-6: Water flow rate through each branch (U) and node balance error (D)
(case 1_0)
In this case, we leave all the valves, including the three by-pass valves but DV1, open.
The simulation results of the water flow rates and the error of flow balance at each node
are collected in Fig. 7-6. As shown in the figure, the water flow rates through branch 1
(between node 1 and node 2) and 2 (between node 2 and node 3) are of the same flow
value with no bypass flow through DV1. Meanwhile, with the adopted hydraulic

210
network, the water flow rate at the center through branch 20 (between node 11 and node
12) is even higher than the main inlet. The reason is that branch 20 is at the inlet of two
parallel pumps. Due to the pump suction, and the balance issue with all the by-pass valve
open, there are reverse flow through branch 21 (bypass pipe between node 11 and node
22), which is shown as negative flow in the chart. Also, due to the local recirculation, the
flow rate through branch 33 (between node 21 and node 22) is even higher than the flow
rate in the main pipe.
The information provided by the model is rich for a detailed hydraulic network balancing
analysis and control where no or very few flow rate meters are installed. The similar
reverse flow also happens at a couple of other locations in the middle of the network (e.g.
branch 13, 17, and 19). The solutions to the issue could be shutting off bypass valves,
installing constant flow rate valves, or installing check valves, etc. The node balance
error shows that the mass at each node has good convergence. In Fig. 7-6, all the errors
are within the magnitude of 10
-13
gpm and the majority is close to 0. The other simple
checking is to compare the flow rate value for node 2-3(branch 2) and node 22-23
(branch 35), node 12-25 (branch 22) and node 20-32 (branch 32), and node 12-33 (branch
23) and node 21-40 (branch 34), as three inlet and out let pairs.
Case 1_1: the main pump is on, the other pumps are off

211

Figure 7-7: Water flow rate through each branch (U) and node balance error (D)
(case 1_1)
In this sub-case, only the main pump is on and the other two branch pumps are off. All
the valve openness are the same as in Case 1_0. The simulation results are shown in Fig.
7-7. From the water flow rate through each branch, it can be seen, that without the two
branch pumps, there is no reverse flow existing in the pipe line. The water flow rates drop
along the way of the flow. The branches 5, 7, 9, 11, 13, 15, 17, and 19 (the water mullion

212
terminals) have almost identical water flow rates at about 3 gpm each. The mass balance
errors for the nodes can be regarded as zero as shown in the lower chart. Meanwhile, it
can be seen that by comparing Fig. 7-6 and 7-7., without the branch pumps, the water
flow rates through the branch terminals (e.g. branch 40, 42, etc) are much less than the
situation with branch pumps. The water flow through the bypass pipe is as normal
without reversing. The mass calculation errors for the branches close to the pump are
close to zero. The overall convergence is within the magnitude of 10
-15
gpm, which
indicates a good simulation. For the purpose of better allocating the water flow in the
network, especially to get more water flow to the radiant ceilings and cool waves close to
the end of the network, the valves of the water mullions and the bypass should be close
more or properly modulated.
7.5.2 Case 2- Thermal- hydraulic co-simulation

Figure 7-8: Simulink model for thermal-hydraulic co-simulation
The hydraulic model is embedded in a thermal model to conduct a co-simulation as
shown in Fig. 7-8. For a real thermal heating cooling system, since the hydraulic network
model provides additional sensing information about the water distribution, the operation
strategy can be correspondingly improved to reduce the potential adverse influences. To
distinguish the case from the complex hydraulic system as in the real facility, only the
representative water based terminals are used for the seven zone space. The heating

213
capacities of the lumped terminals are adjusted correspondingly to match the building
load condition. The simulation is conducted for one day in January with the real trended
weather data.

Figure 7-9: Simulation results for 7 zones in the space
(U: room air temp, D: water flow rates)

214
From the results shown in Fig. 7-9, it can be seen that the water network has some
coupling issue. The pipe in room 5 and 7 are in the disadvantage location, where the
valves get starved and the room air temperature cannot be satisfied during almost the
whole simulation period. The right chart gives the water flow distribution among the
terminals in the seven zones. Room 1 does not need much water to satisfy the load, while
room 4 consumes the most. The water distribution simulation can tell more about the
network if the bypass control and the mixing valves are further tuned.
For water based thermal heating and cooling system analysis, a hydronic network
simulation becomes necessary in order to gain more insight of the water distribution. The
more detailed water distribution information through the terminals serves as "virtual
sensor" for advanced control and diagnostics. In this chapter, a simplified "node-branch-
state" approach for hydraulic system simulation is proposed. A handy processing tool is
developed in Matlab to automatically construct the model for a large scale hydraulic
network. The hydraulic system modeling method can be modified, expanded and
integrated into a detailed thermal model to conduct co-simulation.
With the study, the following points can be concluded:
- This approach does not require the users to identify the loops. With the method,
the users only need to provide the information about the nodes and the branches.
Therefore, a large scale of hydronic system can be handled easily;
- The model is capable of dealing with the discontinuity issues due to pump or
valve on/off operation and is suitable for the integration into control. Pump
On/Off, valve Close/Open, and branches On/Off often present a challenge to a
control and simulation since part of the dynamics can be totally different when the
discontinuity happens.

215
- Solving a complex hydraulic network involves the iterations or optimization to
obtain the solution of large-scale nonlinear equations simultaneously, which can
be time consuming. More study is needed to investigate the solver computation
reliability and speed for a thermal-hydraulic coupled simulation or assist the
online operation strategies;
- Utilization and the architecture of the virtual sensing capability of hydronic
network in advanced control and other potential applications should be further
studied to explore the value and experimentally assessed in the future.

216
CHAPTER 8 CONCLUSIONS AND DISCUSSION
Buildings account for about 40% of the total energy consumption in the U.S. (DOE
2008), and most of the energy is used to heat, ventilate, and air-conditioning the space.
To improve the management effectiveness and efficiency, a building automation system
(BAS) is typically deployed in modern buildings to interconnect and automate the various
systems and devices. One of the important goals of applying BAS is to use the control
functions to satisfy occupants comfort requirements with minimal energy use (Mathew
2001). Considering the weight of energy use by buildings, improving the performance of
controls is thus critical for both energy and indoor environmental sustainability.
Utilizing models in control system and applying optimization theory provide us chances
for improving the building energy system performance. A building system and/or
equipment model can be integrated into real time control or offline evaluation, in a
defined time interval, with given inside and outside conditions and generate future
control actions. In the thesis, the research on general model/simulator assisted operation,
model based predictive control, and thermal oriented control is systematically surveyed.
A couple of issues identified in the area that deserve further study are investigated in this
thesis. The theory, including the problem formulation and solving methods, of classical
linear model predictive control and general optimization based predictive control is
summarized. For the purpose of online implementation, the mature linear model
predicative control and related issues are of the main concern.
This thesis systematically investigates the utilization of models in building system control
and/or operation at both local equipment level and space thermal conditioning level. The
following conclusions can be made based on the thesis work:
1) A set of multivariate control-oriented dynamic models for the diversified building
energy systems in the test bed are developed. A lumped-parameter approach based on
mass and energy conservation is applied. The buildings are modeled as a thermal
resistance and capacitance network. A tool is developed to generate the model

217
automatically with a user defined parameter datasheet. Both sensible heat and latent heat
can be modeled with the approach. In addition, two type of first-principle models for the
hybrid ventilation system, including an enthalpy recovery wheel unit, a vapor
compression unit, and an active desiccant unit, are deduced. The full dynamic models
require higher computational cost and can produce more detailed information about the
devices. The simplified dynamic models keep the accuracy of the concerned variables
with reduced computational cost. The latter one is suitable for online implementation.
The models are provided in explicit equations with heat and moisture transfer considered.
They can be easily adopted for different research purposes on system dynamics. The
simulators of the diversified building systems are established based on the models.
2) Linear model predictive controllers are designed and evaluated in various scenarios for
the mechanical components in the diversified system. The nonlinear simulators for the
complex hybrid ventilation unit are linearized around the operation points to obtain the
internal models. It is found that, with the flexible formulation of MPC, the controllers
guide the devices toward the goal with minimum efforts. For example, in the swing
season, the MPC for the enthalpy recovery unit successfully coordinates the outdoor air
intake and the enthalpy recovery wheel to maximize the benefit. When the space is in
need of free cooling outdoor air, the MPC deactivates the wheel and speeds up the
exhaust fan. The MPC for the vapor compression unit coordinates the condenser fan and
compressor to track the supply air temperature set point or the humidity set point. Control
of the active desiccant wheel imposes a challenge since the dehumidification
performance is highly dependent on the regeneration air temperature with less
manipulation allowed. The designed MPCs facilitate the handling of constraints, enables
the diversified control criteria and eases the managing of coupling effects in a multiple-
input-multiple-output system.
3) A Wiener- and a Hammerstein-Wiener nonlinear MPC are proposed and studied for
the space thermal conditioning. It keeps the linear dynamics of the building thermal
system and takes the advantages of linear MPC for fast online implementation. The
original nonlinearities included in a dynamic building mechanical system can be handled

218
through the combination of a linear dynamic model and one or two nonlinear static
mappings. The performance with pure linear MPC, Wiener nonlinear MPC with thermal
comfort nonlinear mapping, and Hammerstein-Wiener nonlinear MPC with both input
and output mapping is evaluated. Cases with different formulation of the optimization
problem are studied. The results based on the simulations show the benefits of using the
proposed control for both thermal comfort and energy conservation with minimum
computational cost. Thermal comfort in a band offers more energy savings than a fixed
thermal comfort approach. By applying thermal comfort oriented control, in a 24/7
schedule condition, the MPC helps acquire about six to thirty six percent of energy
savings. In an office schedule condition, about seven to forty two percent of energy may
be saved by using MPC empowered thermal comfort control.
4) A simplified "node-branch-state" modeling approach is proposed and studied for
hydraulic systems to improve the control performance. A handy processing tool is
developed in Matlab to automatically construct the model for a large scale hydraulic
network. The modeling method can be utilized as virtual sensing in the various
circumstances for the knowledge of flow distribution, coupled simulation, and evaluation
of control, fault detection and diagnostics, etc. The method focuses on water-based
heating and cooling systems, which generally have high energy efficiency in design but
can perform poorly in reality due to the under-sensing condition and strong thermal-
hydraulic coupling. The approach can handle the discontinuity involving valve
open/close and pump on/off, which often presents a challenge to a control and simulation
since part of the dynamics can be totally different. The node-branch-state modeling
approach can be easily modified, expanded and integrated into a detailed thermal model.
With the IWn space as the test bed, the enabling techniques presented in this thesis make
important contributions to the research of model predictive control of air-conditioning
systems in office buildings. Unlike many other studies, this thesis is the first one that
integrally presents the process and approach of constructing a model predictive control
for complex equipment, space thermal conditioning, and virtual sensing in optimal
control. The dynamic control-oriented models for the buildings, vapor compression units,

219
passive and active desiccant units, etc., can be utilized for the study of many other
advanced control as well. Diversified control criteria, such as thermal comfort, energy
conservation, and equipment life span, become possible with the optimal control
strategies. With the growing concern on energy and comfort, increasing computation
capability of building control systems, better prediction of future information, and lower
cost modulation devices, the application of optimal multivariate control in office
buildings can be expected in the near future.
Many research points are also identified in the thesis that may be considered for future
work. For example:
- An improved dynamic linearization mechanism for a device where the operation
has a wide range should be investigated; The internal model linearized around a
given operation point cannot guarantee the right prediction of the system when the
operation point digresses greatly.
- More variables involved in the thermal comfort determination can be explored for
personal thermal comfort control. In this study, only the space air temperature and
air humidity are considered. For personal comfort control, the occupants'
preference, activity, clothing, etc., can be embedded into the algorithm.
- Investigate the implementation architecture and performance of HW-MPC for a
building with a large number of terminals. Centralized, distributed, and/or a
derivative combination structure may be needed for a given circumstance. The
pros and cons can be compared to provide guidance for real application.
- Utilization and the architecture of the virtual sensing capability of hydronic
network in advanced control and other potential applications should be further
studied to explore the value and experimentally assessed in the future.
The model based predictive control presented in this thesis work can also be extended to
investigate the energy saving potentials of the primary system in the Intelligent
Workplace as Energy Supply System. The coupling of combined heat and power, solar
heating and cooling, and active desiccant ventilation unit, etc., deserve more study with

220
the advanced control technologies. Finally, to bring the technology to wide application,
an intuitive operation interface, quick obtaining of model structure and parameters, and
the acceptance of the integration from the onsite engineers should be investigated in the
future.

221
REFERENCE
[1] Abushakra, B. and Knebel, D.E. 2008. Modeling Office Building Occupancy in
Hourly Data-Driven and Detailed Energy Simulation Programs, ASHRAE
Transactions, pp: 472-481.
[2] Bejan Adrian. 1984. Convection Heat Transfer, John Wiley and Sons, New York,
NY, 1984
[3] Rafael Alcala, Jorge Casillas, et al. 2005. A genetic rule weighting and selection
process for fuzzy control of heating, ventilating and air conditioning systems.
Engineering Applications of Artificial Intelligence, Vol. 18, pp: 279-296.
[4] Frank Allgower, Rolf Findeisen, and Zoltan K. Nagy. 2004. Nonlinear model
predictive control: From theory to application. J. Chin. Inst. Chem. Engrs, Vol.
35(3), pp: 299-315.
[5] ARI. 2003. Guidline: Calculating the efficiency of energy recovery ventilation and
its effect on efficiency and sizing of building HVAC systems.
[6] ASHRAE. 1997. ASHRAE Handbook- Fundamentals. 1997.
[7] ASHRAE. 2008. ASHRAE Handbook- HVAC Systems and Equipment, Chapter 8.
American Society of Heating, Refrigeration and Air-conditioning Engineers, Inc.,
Atlanta, GA.
[8] ASHRAE. 2004. ANSI/ASHRAE Standard 55-2004 Thermal environmental
conditions for human occupancy. 2004.
[9] P.J. Banks. 1972. Coupled equilibrium heat and single adsorbate transfer in fluid
flow through porous medium. I. Characteristic potentials and specific capacity ratios.
Chemical Engineering Science, 27, pp: 1142-1155.
[10] P.J. Banks. 1985. Prediction of heat and mass regulator performance using nonlinear
analogy method. Part 2: Comparison of methods, Transaction of the ASME- Journal
of Heat Transfer, 107, pp: 230-238.
[11] G.K. Batchelor. 2000. An Introduction to Fluid Dynamics. Cambridge University
Press.
[12] Satyam Bendapudi, James E. Braun. 2002. A review of literature on dynamic models
of vapor compression equipment, ASHRAE 1043-RP final report.
[13] B. Wayne Bequette. 2007. Nonlinear model predictive control: a personal
retrospective. The Canadian Journal of Chemical Engineering, 85, pp:408-415.
[14] Fred Betz. 2009. PhD thesis: Combined Cooling, Heating, Power, and Ventilation
System Integration. Carnegie Mellon University.
[15] P.R.Bhave. 1988. Extended Period Simulation of Water Systems- Direct Solution.
Journal of Environmental Engineering, Vol. 114 (5), pp: 1146-1159.

222
[16] Brager, G.S. and de Dear, R.J. 1998. Thermal adaptation in the built environment: A
literature review. Energy and Buildings, V.27(1), pp.83-96.
[17] E.F. Camacho, Carlos Bordons. 1995. Model Predictive Control in The Process
Indusrty. Springer-Verlag.
[18] Carrier, 2004, Scroll compressors: High efficiency compression for commercial and
industrial applications. SEMCO datasheet.
[19] Mark A. Cascia. 2000. Implementation of a near-optimal global set point control
method in a DDC controller. ASHRAE Trans, pp: 249-263.
[20] CBECS. 2003. U.S. Energy Information Administration.
[21] Chen W. 2005. PhD Thesis: Modeling and control of a direct expansion variable-air-
volume air conditioning system. Hong Kong Polytechnic University.
[22] T.Y. Chen. 2001. Real-time predictive supervisory operation of building thermal
systems with thermal mass. Energy and Buildings. Vol. 33, pp: 141-150.
[23] Chen W.Y., Chen Z.J., Zhu R.Q. and Wu Y.Z., 2001, Control algorithm of electronic
expansion valve in evaporator superheat regulation. Journal of Shanghai Jiaotong
University, 35(8), pp: 1228-1232.
[24] J. Chi, D. Didion. 1982. A simulation model of the transient performance of a heat
pump. Int. J. Refrigeration, pp: 176-184.
[25] Jiri Cigler and Samuel Privara. 2010. Subspace identification and model predictive
control for buildings. 11th conf. Control, Automation, Robotics and Vision.
Singapore, pp: 751-755.
[26] Clark J.A., Cockroft, J, et al. 2001. Control in building energy management systems:
the role of simulation. Building Simulation proceedings of 7th IBPSA conference,
2001, pp: 99-106.
[27] Shiming Deng. 2000. A dynamic mathematical model of a direct expansion water-
cooled air-conditioning plant, Building and Environment, 35, pp: 603-613.
[28] Kun Deng, Prabir Barooah, Prashant G. Metha, and Sean P. Meyn. 2010. Building
Thermal Model Reduction via Aggregation of States. Proceedings of ACC 2010.
[29] Guoliang Ding, 2007, Recent development in simulation techniques for vapor-
compression refrigeration systems. International Journal of Refrigeration, 30, pp:
1119-1133.
[30] David Daum, Frederic Haldi, and Nicolas Morel. 2011. A personalized measure of
thermal comfort for building controls. Building and Environment, Vol. (46), pp: 3-
22.
[31] de Dear R.J. and G.Schiller Brager. 1998. Developing an adaptive model of thermal
comfort and preference. ASHRAE Trans., V.104(1a), pp.145-167.
[32] DOE. 2008. Energy efficiency trends in residential and commercial buildings. The
Office of Energy Efficiency and Renewable Energy.

223
[33] Donald E. Kirk. 1998. Optimal control theory: an introduction. Dover Pulications,
Inc. Mineola, New York.
[34] EES, http://www.fchart.com/ees/add-ons.php.
[35] EnergyPlus documentation. 2010.
[36] P.O. Fanger. 1972. Thermal comfort. Analysis and applications in environmental
engineering. McGraw-Hill, New York.
[37] Fatemeh Esfandiari Nia,Dolf Van Paassen. 2006. Modeling and simulation of
desiccant wheel for air conditioning. Energy and Buildings, 38: 1230-1239.
[38] Rolf Findeisen and Frank Allgower. 2002. An introduction to nonlinear model
predictive control. 21st Benelux Meeting on Systems and Control, Veidhoven.
[39] John Fischer. 2004. Desiccant-Based Combined Systems: Integrated Active
Desiccant Rooftop Hybrid System Development and Testing.
[40] John Fischer. 2007. Carnegie Mellon Intelligent Workplace Institute and big box
retial facility field verification sites: revolution active desiccant-vapor compression
hybrid rooftop system with heat pump integration.
[41] K.F. Fong, V.I. Hanby, and T.T. Chow. 2006. HVAC system optimization for
energy management by evolutionary programming. Energy and Buildings. Vol.38,
pp: 220-231.
[42] Roberto Z. Freire, Gustavo H. Oliveira, and Nathan Mendes. 2008. Predictive
controllers for thermal comfort optimization and energy savings. ABCM Symposium
Series in Mechatronics, Vol. (3), pp: 839-848.
[43] Bernard Friedland. 2005. Control System Design: An Introduction to State-Space
Methods. ISBN-10: 0486442780
[44] Monika Frontczak, Pawel Wargocki. 2011. Literautre survey on how different
factors influence human comfort in indoor environments. Building and Environment,
Vol.(46), pp: 922-937.
[45] M. Gamberi, R. Manzini and A. Regattieri. 2009. Simulink Simulator for Building
Hydraulic Heating Systems Using the Newton-Raphson Algorithm. Energy and
Buildings, Vol. 41, pp: 848-855.
[46] M. Gamberi, R. Manzini and C.P. Underwood. 1998. HVAC Control Systems:
Modeling, Analysis and Design. E & FN Spon.
[47] T.S. Ge, F. Ziegler, R.Z. Wang. 2010. A mathematical model for predicting the
performance of a compound desiccant wheel. Applied Thermal Engineering, 30, pp:
1005-1015.
[48] Geng, G., G. M. Geary. 1993. On performance and tuning of PID controllers in
HVAC systems. Proceedings of the IEEE Conference on Control Applications. Vol.
2, pp: 819-824.
[49] Carlan Graham. 2009. High-performance HVAC. Whole Building Design Guide.

224
[50] Ravi Gondhalekar, Frauke Oldewurtel, Colin N. Jones. 2010. Least-restrictive robust
MPC of periodic affine systems with application to building climate control. 49th
IEEE Conference on Decision and Control, pp: 5257 - 5263.
[51] Xiangyang Gong. 2007. PhD Thesis: Investigation of a radiantly heated and cooled
office with an integrated desiccant ventilation unit. Texas A&M.
[52] Grald, E.W. and MacArthur, J.W. 1992, A moving boundary formulation for
modeling time-dependent two-phase flows. Int. J. Heat and Fluid Flow, Vol. 13(3),
pp: 266-272.
[53] Markus Gwerder, Jurg Todtli. 2005. Predictive control for integrated room
automation. 8th REHVA World Congress for Building Technologies, CLIMA 2005.
[54] Yogesh M. Harsho, Rangjeet P. Utikar, et al. 2005. Modeling of Rotary Desiccant
Wheels. Chem. Eng. Technol, Vol. 28(12), pp: 1473-1479.
[55] G. Henze, A. Floss. 2011. Evaluation of Temperature Degradation in Hydraulic
Flow Networks. Energy and Buildings, Vol. 43, pp: 1820-1828.
[56] Xiangdong He. 1985. PhD Thesis: Dynamic modeling and multivariable control of
vapor compression cycles in air conditioning systems.
[57] He XD, S. Liu, and H.H. Asada, 1997. Modeling of vapor compression cycles for
multivariable feedback control of HVAC systems. Journal of dynamic systems,
measurement, and control, Vol. 119(2), pp: 183-191.
[58] He XD, Liu S, Asada HH, Itoh H. 1998. Multivariable control of vapor compression
systems. HVAC&R Research, Vol. 4(3), pp: 205-230.
[59] Michael A. Henson. 1998. Nonlinear model predictive control: current status and
future directions. Computers and Chemical Engineering, 23, pp: 187-202.
[60] Henze, G.P., R.H. Dodier, and M. Krarti. 1997. Development of a predictive optimal
controller for thermal energy storage systems. HVAC&R Research, Vol.3(3), pp:
233-264.
[61] Hu Mingjiang, Meng Jucai, 2011, Simulating the performance on electronic
expansion valve, Transaction of Electrical and Control Engineering (ICECE), 201:
379-382.
[62] Incropera, F.P. and D.P. Dewitt 1996, Fundamentals of Heat and Mass Transfer, 4th
edition, Wiley, Toronto.
[63] A. Jakobsen, B. Rasmussen, M.J. Skovrup, and J. Fredsted. 2000. Development of
Energy Optimal Capacity Control in Refrigeration Systems, International
Refrigeration and Air Conditioning Conference, pp: 329-336.
[64] J.Y. Jang and J. T. Lai. 1997. Numerical and experimental analysis of heat transfer
and fluid flow in 3-D circular finned-tube heat exchangers, Inter Conf on CFD in
mineral & metal processing and power generation.

225
[65] Kun Ji, Yan Lu, et al. 2012. Energy management control framework of optimization
control and operation for energy efficient buildings. ACEEE 2012. Accepted.
[66] A.J. Jordan. 2006 Linearization of nonlinear state equations, Bulletin of the polish
academy of sciences- Technical sciences, 54(1): 63-73.
[67] A. Jordan and J.P. Nowacki. 2003. Global linearization of non-linear state equations,
Int. J. Applied electromagnetic and mechanics, 19, pp: 637-642.
[68] Jeff Jermoe Jurinak. 1982. PhD Thesis, Open Cycle Solid Desiccant Cooling-
Component Models and system Simulations. University of Wisconsin--Madison.
[69] Kays, W.M. and London, A.L., Compact heat exchanger, McGraw-Hill, NewYork,
1964.
[70] A. Keblawi, N. Ghaddar, and K. Ghali. 2011. Model-based optimal supervisory
control of chilled ceiling displacement ventilation system. Energy and Buildings,
Vol. 43, pp: 1359-1370.
[71] George E. Kelly. 1988. Control system simulation in north America. Energy and
Buildings. Vol. 10, pp: 193-202.
[72] Kohl, R. 2001. Commissioning HVAC controls systems. ASHRAE Journal,
Vol.43(12), pp: 27-30.
[73] D. Kolokotsa, A. Pouliezos, G. Stavrakakis, C. Lazos. 2009. Predictive control
techniques for energy and indoor environmental quality management in buildings.
Building and Environment. Vol. 44, pp:1850-1863.
[74] Krakow KI, Lin S, Zeng ZS. 1995. Temperature and humidity control during cooling
and dehumidifying by compressor and evaporator fan speed variation. ASHRAE
Transactions, 1995. Vol. 101(1), pp: 292-304.
[75] Michael Kummert and Philippe Andre, 2005. Simulation of a Model-Based Optimal
Controller for Heating Systems Under Realistic Hypotheses. Building Simulation
2005.
[76] A. Laouadi. 2004. Development of a Radiant Heating and Cooling Model for
Building Energy Simulation Software. Building and Environment, Vol.39, pp: 421-
431.
[77] Denis Leducq, Jacques Guilpart, Gilles Trystram. 2006. Nonlinear predictive control
of a vapor compression cycle, International Journal of Refrigeration, 29, pp: 761-
772.
[78] Haorong Li, James Braun. 2009. Decoupling features for diagnosis of reversing and
check valve faults in heat pumps. International Journal of Refrigeration 32, pp: 316-
326.
[79] Haorong Li, Daihong Yu, James, E. Braun. 2011. A Review of Virtual Sensing
Technology and Application in Building Systems. HVAC&R Research, 17 (5): 619-
646.

226
[80] Z. Liao, A.L. Dexter. 2004. The potential for energy saving in heating systems
through improving boiler controls. Energy and Buildings, Vol. 36, pp: 261-271.
[81] Benjamin.Y.H. Liu and Richard C. Jordan. 1963. A rational procedure for predicting
the long-term average performance of flat-plate solar-energy collectors. Solar
Energy. Vol. 7(2): 53-74.
[82] Yudong Ma, Garrett Anderson, and Francesco Borrelli. 2010. A distributed
predictive control approach to building temperature regulation.
[83] MacArthur, J.W, and Grald, E. W. 1989. Unsteady compressible two-phase flow
model for predicting cyclic heat pump performance and a comparison with
experimental data, Int. J. Refrigeration, 12, pp: 29-41.
[84] I.L. Maclaine-Cross, P.J. Banks. 1972. Coupled heat and mass transfer in regerators-
prediction using an analogy with heat transfer. International Journal of Heat and
Mass Transfer, 15: 1225-1242.
[85] A. Macleod. 2008. The Heating System as a Distributed Sensor Network. Energy
and Buildings, Vol. 40, pp: 1547-1552.
[86] Ardeshir Mahdavi. 1997. Toward a simulation-assisted dynamic building control
strategy.
[87] Ardeshir Mahdavi and Claus Proglhof. 2008. A model-based approach to natural
ventilation. Building and Environment, Vol. 43, pp: 620-627.
[88] Sophie Masson. 2007. IW internal report.
[89] Mathews, E.H., C.P. Botha, et al. 2001. HVAC control strategies to enhance comfort
and minimize energy usage. Energy and Buildings, Vol.(33), pp: 853-863.
[90] Pietro Mazzei, Francesco Minichiello, and Daniele Palma. 2005. HVAC
dehumidification systems for thermal comfort: a critical review. Applied Thermal
Engineering, Vol.25, pp: 677-707.
[91] Yuji Miyajima, Masayoshi Sakuma, et al. 2007. Field testing of a supervisory
optimal controller for an air-conditioning system. ASHRAE Transactions, 2007, pp:
323-333.
[92] Petru-Daniel Morosan, Romain Bourdais, et al. 2010a. Distributed model predictive
control based on Benders decomposition applied to multisource multizone building
temperature regulation. Transactions of 49th IEEE Conference on Decision and
Control:, pp: 3914- 3919.
[93] E. Morishita, M. Sugihara, et al. 1984. Scroll compressor analytical model.
International Compressor Engineering Conference. 1984: 487-495.
[94] Petru-Daniel Morosan, Romain Bourdais, et al. 2010b. Distributed model predictive
control for building temperature regulation. Transactions of 2010 American Control
Conference, pp: 3174- 3179.

227
[95] Kwok Wai Horace Mui, Wai Tin Daniel Chan. 2003. Adaptive comfort temperature
model of air-conditioned building in Hong Kong. Building and Environment, Vol.38,
pp: 837-852.
[96] S. Neti, E.I. Wolfe. 2000. Measurements of effectiveness in a silica gel rotary
exchanger. Applied Thermal Engineering, 20: 309-322.
[97] Fatemeh Esfandiari Nia. 2011. PhD Thesis: Sustainable air handling by evaporation
and adsorption. Sharif University of Technology.
[98] Michael Nikolaou, 2010. Report: Model predictive controllers: a critical synthesis of
theory and industry needs. University of Houston.
[99] Katsuhiko Ogata. 2009. Modern control engineering. ISBN-10: 0136156738.
[100] F. Oldewurtel, A. Parisio, et al. 2010. Energy efficient building climate control
using stochastic model predictive control and weather predictions. Transactions of
American Control Conference 2010, pp: 5100 - 5105.
[101] L. Ormsbee. 2006. The History of Water Distribution Network Analysis: The
Computer Age. 8th Annual Water Distribution Systems Analysis Symposium.
Cincinnati.
[102] G. Panaras, E. Mathioulakis, V. Belessiotis, and N. Kyriaks. 2010. Experimental
validation of a simplified approach for a desiccant wheel model. Energy and
Buildings, 42, pp: 1719- 1725.
[103] Curtis O. Pedersen, Daniel E. Fisher, et al. 2003. ASHRAE Toolkit for Building
Load Calculations. ASHRAE. Transactions. 109(1), pp: 583- 589 .
[104] Samuel Privara, Jan Siroky, et al. 2011. Model predictive control of a building
heating system: the first experience. Energy and Buildings, Vol. 43, pp: 564-572.
[105] Qi Qi, Shiming Deng, 2009, Multivariable control of indoor air temperature and
humidity in a direct expansion air conditioning system, Building and Environment,
44, pp: 1659 - 1667.
[106] S. Joe Qin, Thomas A. Badgwell. 2003. A survey of industrial model predictive
control technology. Control Engineering Practice, Vol. 11, pp: 733-764.
[107] B.P. Rasmussen, R. Shah, et al. 2002. Report: Control-oriented modeling of
transcritical vapor compression systems. University of Illinois-Urban Champion.
[108] James B. Rawlings. 2000. Tutorial overview of model predictive control. Control
Systems, IEEE, Vol. 20(3), pp: 38-52.
[109] Rawlings, J.B. and Mayne, D.Q. 2009. Model predictive control: Theory and
design. Nob Hill Publishing.
[110] K. Rhee, S. Ryu, et al. 2010. Simulation Study on Hydraulic Balancing to
Improve Individual Room Control for Radiant Floor Heating System. Building
Services Engineering Research and Technology, Vol. 31(1), pp:57-73.

228
[111] K. Rhee, M. Yeo and K. Kim. 2011. Evaluation of the Control Performance of
Hydraulic Radiant Heating Systems Based on the Emulation Using Hardware-in-the-
Loop Simulation. Building and Environment, Vol. 46, pp: 2012-2022.
[112] Rohit S. Pathwardhan, S. Lakshminarayanan, and Sirish L. Shah. 1998.
Constrained nonlinear MPC using Hammerstein and Wiener Models: PLS
Framework. AIChE Journal, Vol.44(7), pp: 1611-1622.
[113] Carina Sagerschnig, Dimitrios Gyalistras, et al. 2011. Co-simulation for building
controller development: the case study of a modern office building.
[114] Timothy I. Salsbury. 2005. A survey of control technologies in the building
automation industry. IFAC 16th Triennial World Congress, pp: 90-100.
[115] Arvid Skartveit, Jan Asle Oleseth, and Marit Elisabet Tuft. 1998. An hourly
diffuse fraction model with correction for variability and surface albedo. Solar
Energy. Vol. 63(3), pp: 173-183.
[116] Arvid Skartveit and Jan Asle Oleseth. 1987. A model for the diffuse fraction of
hourly global radiation. Solar Energy, Vol. 38(4), pp: 271-274.
[117] Sigurd Skogestad, and Ian Postlethwaite. 2001. Multivariable Feedback Control-
Analysis and Design. John Wiley & Sons.
[118] Gunnar Steich. 1994. Msc Thesis: Performance of rotary enthalpy exchangers.
University of Wisconsin- Madison.
[119] A. Strafaci, 2003. Advanced Water Distribution Modeling and Management,
Haestad Press, Waterbury.
[120] P. Thomas, 1999. Simulation of Industrial Processes for Control Engineers.
Elsevier Science & Technology Books.
[121] TomLab manuals, 2012. TomLab_PROPT.
[122] J.V.C.Vargas, J.A.R. Parise. 1995. Simulation in transient regime of a heat pump
with closed loop and on-off control, Int. J. Refrig, 18(4), pp: 235-243.
[123] Liuping Wang. 2009. Model predictive control system design and implementation
using Matlab. Springer, Australia.
[124] Shengwei Wang and Xinqiao Jin. 2000. Model-based optimal control of VAV air-
conditioning system using genetic algorithm. Building and Environment, Vol. 35,
pp: 471-487.
[125] Shengwei Wang and Zhenjun Ma. 2008. Supervisory and optimal control of
building HVAC systems: a review. HVAC&R Research. Vol. 14(1), pp: 3-32.
[126] Z. Warsi, 1998. Fluid Dynamics: Theoretical and Computational Approaches.
CRC Press.
[127] Westphalen and Koszalinksi 2001, Energy consumption characteristics of
commercial building HVAC systems, ADL reference No. 36922-00, Arthur D.
Little, Inc., 20Acorn Park, Cambridge, MA 2140-2390, pp 1-2 - 1-3.

229
[128] Shui Yuan, Ronald Perez. 2006a. Multiple-zone ventilation and temperature
control of a single-duct VAV system using model predictive strategy. Energy and
Buildings, Vol. 38, pp: 1248-1261.
[129] Ydstie B. Erik, 2012. Class notes: Advanced process dynamics and control.
Carnegie Mellon University.
[130] Kenzo Yonezawa, Fumio Yamada, et al. 2000. Comfort air-conditioning control
for building energy-saving. Trans. IEEE, 2000, pp: 1737-1742.
[131] D. Yu, H. Li, and M. Yang. 2011. A virtual supply airflow rate meter for rooftop
air-conditioning units, Building and Environment, 46, pp: 1292-1302.
[132] D. Yu, H. Li, et al. 2011a. Virtual Calibration of a Supply Air Temperature
Sensor in Rooftop Air Conditioning Units. International Journal of HVAC & R
Research, Vol. 17(1), pp: 31-50.
[133] D. Yu, H. Li, and Y. Yu. 2011b. A Gray-box Based Virtual Supply Airflow Meter
in Rooftop Air-Conditioning Units. Journal of Thermal Science and Engineering
Applications, Vol. 3(1), pp: 1-7.
[134] Yuebin Yu. 2010. Msc Thesis: Integrating Air Handling Units in Office Buildings
for High Performance. University of Nebraska-Lincoln.
[135] Yuebin Yu, Bing Dong, et al. 2009. Application of system identification and
numerical optimization to a floor radiant heating control in a solar house.
Proceedings of CLIMA 2010. Turkey.
[136] Yuebin Yu, Mingsheng Liu, et al. 2012b. Synergization of air handling units for
high energy efficiency in office buildings: the theory and analysis. Energy and
Buildings, under review.
[137] Yuebin Yu, Ke Xu, YoungHum Cho and Mingsheng Liu. A smart logic for
conference room terminal box of single duct VAV system. Proceedings of ICEBO
2007. California.
[138] Yuebin Yu, Vivian Loftness, Daihong Yu, Yan Lu, Zhen Song. 2012a.
Development of Innovative Energy System and Its Advanced Control in EnergyPlus.
Proceedings of eSim 2012.
[139] Shui Yuan, Ronald Perez. 2006b. Model predictive control of supply air
temperature and outside air intake rate of a VAV air-handling unit. ASHRAE Trans.
Vol. 112, pp:145-161.
[140] Zhang, Y., and V.I. Hanby. 2006. Model-based control of renewable energy
systems in buildings. HVAC&R Research, Vol.12(3a), pp: 739-760.
[141] Chaoqin Zhai. 2008. PhD Thesis: Performance modeling of desiccant wheel
design and operation. Carnegie Mellon University.
[142] Tieshan Zhao, Zengzhi Li, and Bo Gao. 2005. Research on Simulation Models for
Air Conditioning System. Computer Simulation, Vol.22(11), pp: 65-68.

230

PUBLICATION
This section collects the major publications during the thesis author's study in the Center
for Building Performance and Diagnostics at Carnegie Mellon University.
Journal Articles

2012 Yu, Y., Liu, M., Li, H., Yu, D., Loftness V. Synergization of Air Handling Units
for High Energy Efficiency in Office Buildings: Implement Methodology and
Performance Evaluaion. Energy and Buildings. 2012. Vol. 54, pp: 426-435.
2012 Yu, D., Li, H., Kibrya, G., Yu, Y. A Robust Cooling-Based Virtual Supply
Airflow Meter of Rooftop Air Conditioning Units. International Journal of HVAC
& R Research. Submitted and under review.
2012 Yu, Y., Gross, D.M., and Yu, D. Pegi: Motivating Exercise with a Pedometer-Gift
Box. Proceedings of APCHI2012 (10th Asia Pacific Conference on Computer
Human Interaction). HCD-Net Journal (ISSN: 1882-9635).
2011 Yu, Y., Liu, M., Li, H., Yu, D., Loftness V. Synergization of Air Handling Units
for High Energy Efficiency in Office Buildings: The Theory and Analysis. Energy
and Buildings. 2012. Vol. 45: 264-273.
2011 Yu, D., Li, H., and Yu, Y. A Gray-Box Based Virtual Supply Airflow Meter in
Rooftop Air-Conditioning Units. Journal of Thermal Science and Engineering
Applications, 2011, 3(1): 011005.1-7.
2011 Yu, D., Li, H., Yu, Y., Xiong, J. Virtual Calibration of a Supply Air Temperature
Sensor in Rooftop Air Conditioning Units. International Journal of HVAC & R
Research. 2011, 17 (1): 31-50.
Conference Papers

2013 Yu, Y., Lasternas, B., Lu., Y., and Hartkopf, V. A Study of Rule and Model
Cascaded Predictive Control for an Office Building. CLIMA 2013, Czech.
(Abstract Accepted)
2012 Yu, Y., Loftness, V., Yu, D., Lu, Y., and Hartkopf, V. Development and
Evaluation of a Simplified Modeling Approach for Hydraulic System.
Proceedings of 2nd International High Performance Building Conference 2012.
Purdue.
2012 Ji, K., Lu, Y., Song, Z., Yu, Y. Energy Management Control Framework of
Optimization Control and Operation for Energy Efficient Buildings. Proceedings
of ACEEE 2012.

231
2012 Yu, Y., Loftness, V., Lu, Y., Yu, D. Development of an Innovative Energy
System and its Advanced Control in EnergyPlus. Proceedings of eSim 2012.
Canada.
2011 Song, Z., Lu, Y., Ji, K., Marion, F., Yu, Y., Kerl, M., Liao, L., Wei, D.
Experimental Study on Control-Oriented Simulation Models for Building Control
and Energy Management. Building Simulation 2011.
2011 Yu, Y., Liu, M., Yu, D., Loftness, V. An Integrated Outside Air Control of Air
Handling Units in Office Buildings for Energy Efficiency. Proceedings of 7th
International Symposium of Heating Ventilation and Air-Conditioning
(ISHVAC), 2011. Shanghai.
2010 Yu, Y., Dong, B., Hu, Y. Simulation-Based Hybrid Ventilation System Design
and Evaluation. Proceedings of 1st International High Performance Buildings,
2010. Purdue.
2010 Yu, Y., Dong, B., Yu, D., Karaguzel, O. Application of System Identification and
Numerical Optimization to a Floor Radiant Heating Control in a Solar House.
Proceedings of CLIMA 2010. Turkey.

232
APPENDIX A:
A-1. Building structure and materials
Table A-1: Construction and thickness
Envelope Layers Thickness (m)
Vertical wall
Aluminum siding 0.003
R-19 bat insulation 0.064
Aluminum siding 0.003
Roof
Steel deck 0.01
R-38 bat insulation 0.14
Build up roofing 0.035
Floor Concrete stab 0.08

Table A-2: Fenestration parameters
CEN test Type
U value
(W/m
2
K)
Shading
coefficient
SHGC Tvis
Thickness
(m)
Fixed glazing
Viracon VE-2E.
VE132m.vir
5.743 0.517 0.449 0.799 0.0057
Argon 0.0127m thickness
Viracon VE-2E.
VE132m.vir
5.743 0.517 0.449 0.799 0.0057
Combined 1.224 0.407 0.354 0.64 0.0241
Operable
glazing
Interprane, ip-
ipsun.ipe
5.798 0.547 0.476 0.788 0.004
Air 0.0127m thickness

233

A-2. Building thermal conditioning experiments
A set of experiments were conducted together with Dr. Zhen Song from Siemens
Cooperate Research in August 2011 to evaluate the terminal thermal conditioning
capacity. Part of the results are illustrated as below. More details can be found in the
project report.

Figure A-1: Mote sensor installed in the north-east bay
Interprane, ip-
ipsun.ipe
5.798 0.547 0.476 0.788 0.004
Combined 1.58 0.427 0.371 0.622 0.0207
Roof glazing
Viracon VE-2E.
VE132m.vir
5.743 0.517 0.449 0.799 0.0057
Argon 0.0127m thickness
Viracon VE-2E.
VE132m.vir
5.743 0.517 0.449 0.799 0.0057
Combined 1.224 0.407 0.354 0.64 0.0241

234

Figure A-2: Underfloor surface temperature measurement

Figure A-3: Mote sensor array in north-east bay

235

Figure A-4: Mote sensor array in the meeting area

Figure A-5: Illustration of space temperature distribution- initial

236

Figure A-6: Illustration of space temperature distribution- final
A-3. Building model calibration

Figure A-7: Simulated air temperature vs real temperautre in the plenum

237

Figure A-8: Simulated air temperature vs real temperautre in the space
A-4. Hybrid ventilation unit

238

Figure A-9: Hybrid ventilation unit control algorithm air temperature set point to
compressor to percentage

239

Figure A-10: Hybrid ventilation unit control algorithm compressor output
percentage to VFD

Figure A-11: Spreadsheet based manufacturer model of the hybrid ventilation unit

240
Table A-3: Enthalpy recovery unit efficiency
Vair
fpm (m/s)
Base Eff
(%)
CFMsup/
CFMret
Base effectiveness
68 70 72 74 76 78 80
300 (1.524) 80.3 0.7 82.6 84.0 85.5 87.0 88.5 90.0 91.4
400 (2.032) 77.8 0.8 78.0 79.8 81.5 83.2 84.9 86.7 88.4
500 (2.54) 75.3 0.9 73.2 75.1 77.0 78.9 80.8 82.7 84.7
600 (3.048) 72.8 1.0 68.0 70.0 72.0 74.0 76.0 78.0 80.0
700 (3.556) 70.3 1.1 66.1 67.9 69.6 71.4 73.1 74.8 76.6
800 (4.064) 67.8 1.25 62.4 63.8 65.2 66.6 68.0 69.4 70.8
900 (4.572) 65.3 1.4 55.9 58.5 60.7 61.8 62.9 63.9 65.1

Figure A-12: Transfer efficiency vs wheel face velocity
y = -0.025x + 87.8
40
45
50
55
60
65
70
75
80
85
0 200 400 600 800 1000 1200
E
f
f
i
c
i
e
n
c
y
(
%
)
WheelFaceVelocity(fpm)

241

Figure A-13: Efficiency Comparison
R = 1
60
65
70
75
80
85
90
60 65 70 75 80 85 90
E
f
f
i
c
i
e
n
c
y
B
a
s
e
d
o
n
F
o
r
m
u
l
a
(
%
)
EfficiencyBasedonCatalogue

242
APPENDIX B:

Figure B-1: Thermal properties of R22

243

Figure B-2: Temperature vs enthalpy of saturated liquid

Figure B-3: Temperature vs enthalpy of saturated vapor
y=0.0032x
2
0.6142x+131.43
100
150
200
250
300
350
240 260 280 300 320 340 360 380
E
n
t
h
a
l
p
y
o
f
s
a
t
u
r
a
t
e
d
l
i
q
u
i
d
(
K
J
/
k
g
)
Temp(K)
y=0.0045x
2
+2.9045x 53.407
300
320
340
360
380
400
420
440
240 260 280 300 320 340 360 380
E
n
t
h
a
l
p
y
o
f
s
a
t
u
r
a
t
e
d
v
a
p
o
r
(
K
J
/
k
g
)
Temp(K)

244

Figure B-4: Temperature vs pressure of two-phase mixture

Figure B-5: Temperature vs density of saturated liquid
y=0.0033x
2
1.6561x+213.17
0
5
10
15
20
25
30
35
40
45
240 260 280 300 320 340 360 380
P
r
e
s
s
u
r
e
o
f
t
w
o
p
h
a
s
e
m
i
x
t
u
r
e
(
b
a
r
)
Temp(K)
y=0.0236x
2
+10.01x+310.46
600
700
800
900
1000
1100
1200
1300
1400
240 260 280 300 320 340 360 380
D
e
n
s
i
t
y
o
f
s
a
t
u
r
a
t
e
d
l
i
q
u
i
d
(
k
g
/
m
3
)
Temp(K)

245

Figure B-6: Temperature vs density of saturated vapor

Figure B-7: Temperature vs specific heat of superheated vapor
y=0.0152x
2
7.7498x+1003.4
0
20
40
60
80
100
120
240 260 280 300 320 340
D
e
n
s
i
t
y
o
f
s
a
t
u
r
a
t
e
d
v
a
p
o
r
(
k
g
/
m
3
)
Temp(K)
y=0.0001x
2
+0.0238x+0.6221
0
0.2
0.4
0.6
0.8
1
1.2
0 5 10 15 20 25
S
p
e
c
i
f
i
c
h
e
a
t
o
f
s
u
p
e
r
h
e
a
t
e
d
v
a
p
o
r
(
k
J
/
k
g
K
)
Pressure(bar)

246

Figure B-8: Temperature vs humidity of air

Figure B-9: Air density vs air density
y=2E05x
2
0.0087x+1.1733
0
0.002
0.004
0.006
0.008
0.01
0.012
0.014
275 280 285 290 295
H
u
m
i
d
i
t
y
r
a
t
i
o
(
k
g
/
k
g
)
Airtemperature(K)
y=0.0032x+1.2625
0.6
0.7
0.8
0.9
1
1.1
1.2
1.3
0 20 40 60 80 100 120
A
i
r
d
e
n
s
i
t
y
(
g
/
k
g
)
Airtemperature(C)

247
APPENDIX C:
C-1. Building model matrix
IW space linear time invariant space model is deduced based on the RC electric analog
methodology. The outdoor air temperature, internal heat gain, and the solar heat flux into
the space are considered as the basic inputs to the model.
%% start of the expression for the model
C1=69.92*gam_air1;C1g1=1528.704*alfg;C1e1=2886.8614*0.6*alfroof;C1f1=28
86.8614*0.4*alfroof;...
C1b1=815.6615*alfw;C1a1=792.2405*alfw;C2=82.8*gam_air2;C2g1=1528.704*al
fg;C2e1=2760.493*0.6*alfroof;...
C2f1=2760.493*0.4*alfroof;C2b1=835.4457*alfw;C3=92.52*gam_air3;C3g1=152
8.704*alfg;C3e1=2919.1613*0.6*alfroof;...
C3f1=2919.1613*0.4*alfroof;C3b1=815.6615*alfw;C4=481.77*gam_air4;C4g1=9
341.424*alfg;C4e1=16138.8638*0.6*alfroof;...
C4f1=16138.8638*0.4*alfroof;C4a1=2154.2948*alfw;C4d1=1777.6678*alfw;C5=
451.23*gam_air5;C5g1=7814.664*alfg;...
C5e1=11311.3831*0.6*alfroof;C5f1=11311.3831*0.4*alfroof;C5b1=815.6615*a
lfw;C5c1=1513.9271*alfw;...
6.337*0.4*alfroof;C6d1=807.6605*alfw;...
0.2646*0.4*alfroof;...
C7d1=807.6605*alfw;R1_1g1=8.6976*betg1;R1g1_g=8.6976*betg2;R1_1e1=135.5
347*0.6*betroof1;R1e1_e=135.5347*0.6*betroof2;...
R1_1f1=135.5347*0.4*betroof1;R1f1_f=135.5347*0.4*betroof2;R1_1b1=285.80
69*betw1;R1b1_b=285.8069*betw2;...
R1_b=89.6219*alfwnde;R1_1a1=294.2562*betw1;R1a1_a=294.2562*betw2;R1_a=1
04.2749*alfwndn;R1_4=12.2656*gam_air1_4 ;...
R1_2=18.2229*gam_air1_2 ;R2_2g1=8.6976*betg1;R2g1_g=8.6976*betg2;R2_2e1
=141.7392*0.6*betroof1;...
R2e1_e=141.7392*0.6*betroof2;R2_2f1=141.7392*0.4*betroof1;R2f1_f=141.73
92*0.4*betroof2;R2_2b1=279.0387*betw1;...

248
R2b1_b=279.0387*betw2;R2_b=90.9691*alfwnde;R2_4=12.2656*gam_air2_4 ;R2_
3=18.2201*gam_air2_3 ;...
R3_3g1=8.6976*betg1;R3g1_g=8.6976*betg2;R3_3e1=134.0351*0.6*betroof1;R3
e1_e=134.0351*0.6*betroof2;...
R3_3f1=134.0351*0.4*betroof1;R3f1_f=134.0351*0.4*betroof2;R3_3b1=285.80
69*betw1;R3b1_b=285.8069*betw2;...
R3_b=89.6416*alfwnde;R3_5=7.3311*gam_air3_5 ;R4_4g1=1.4233*betg1;R4g1_g
=1.4233*betg2;R4_4e1=24.244*0.6*betroof1;...
R4e1_e=24.244*0.6*betroof2;R4_4f1=24.244*0.4*betroof1;R4f1_f=24.244*0.4
*betroof2;R4_e=58.7554*0.6*alfwndr1;...
R4_f=58.7554*0.4*alfwndr2;R4_4a1=108.2125*betw1;R4a1_a=108.2125*betw2;R
4_a=31.5771*alfwndn;...
R4_4d1=131.1391*betw1;R4d1_d=131.1391*betw2;R4_d=47.4334*alfwndw;R4_5=8
.8619*gam_air4_5 ;R4_6=18.2352*gam_air4_6 ;...
R5_5g1=1.7014*betg1;R5g1_g=1.7014*betg2;R5_5e1=34.5908*0.6*betroof1;R5e
1_e=34.5908*0.6*betroof2;...
R5_5f1=34.5908*0.4*betroof1;R5f1_f=34.5908*0.4*betroof2;R5_e=25.1809*0.
6*alfwndr1;R5_f=25.1809*0.4*alfwndr2;...
R5_5b1=285.8069*betw1;R5b1_b=285.8069*betw2;R5_b=89.6416*alfwnde;R5_5c1
=153.9847*betw1;R5c1_c=153.9847*betw2;...
R5_c=66.3737*alfwnds;R5_6=12.2656*gam_air5_6 ;R5_7=12.2656*gam_air5_7 ;
R6_6g1=8.7033*betg1;...
R6g1_g=8.7033*betg2;R6_6e1=135.5594*0.6*betroof1;R6e1_e=135.5594*0.6*be
troof2;R6_6f1=135.5594*0.4*betroof1;...
R6f1_f=135.5594*0.4*betroof2;R6_6d1=288.6382*betw1;R6d1_d=288.6382*betw
2;R6_d=89.0942*alfwndw;...
R6_7=18.2352*gam_air6_7 ;R7_7g1=8.7033*betg1;R7g1_g=8.7033*betg2;R7_7e1
=163.0112*0.6*betroof1;...
R7e1_e=163.0112*0.6*betroof2;R7_7f1=163.0112*0.4*betroof1;R7f1_f=163.01
12*0.4*betroof2;R7_e=176.2661*0.6*alfwndr1;...
R7_f=176.2661*0.4*alfwndr2;R7_7d1=288.6382*betw1;R7d1_d=288.6382*betw2;
R7_d=89.0942*alfwndw;...
%% start of the A_Matrix for the model
%
{'T1','T2','T3','T4','T5','T6','T7','T1a1','T1b1','T1e1','T1f1','T1g1',
'T2b1','T2e1','T2f1','T2g1','T3b1','T3e1','T3f1','T3g1','T4a1','T4d1','
T4e1','T4f1','T4g1','T5b1','T5c1','T5e1','T5f1','T5g1','T6d1','T6e1','T

249
6f1','T6g1','T7d1','T7e1','T7f1','T7g1','Ta','Tb','Tc','Td','Te','Tf','
Tg';}
AA=[-
1*(1/(R1_2*C1)+1/(R1_4*C1)+1/(R1_1a1*C1)+1/(R1_1b1*C1)+1/(R1_1e1*C1)+1/
(R1_1f1*C1)+1/(R1_1g1*C1)+1/(R1_a*C1)+1/(R1_b*C1)),...
1/(R1_2*C1), 0,1/(R1_4*C1), 0, 0,
0,1/(R1_1a1*C1),1/(R1_1b1*C1),1/(R1_1e1*C1),1/(R1_1f1*C1),1/(R1_1g1*C1)
,...
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0,;1/(R1_2*C2),-
1*(1/(R1_2*C2)+1/(R2_3*C2)+1/(R2_4*C2)+1/(R2_2b1*C2)+1/(R2_2e1*C2)+1/(R
2_2f1*C2)+1/(R2_2g1*C2)+1/(R2_b*C2)),...
1/(R2_3*C2),1/(R2_4*C2), 0, 0, 0, 0, 0, 0, 0,
0,1/(R2_2b1*C2),1/(R2_2e1*C2),1/(R2_2f1*C2),1/(R2_2g1*C2), 0, 0, 0,
0,...
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,; 0,1/(R2_3*C3),-
1*(1/(R2_3*C3)+1/(R3_5*C3)+1/(R3_3b1*C3)+1/(R3_3e1*C3)+1/(R3_3f1*C3)+1/
(R3_3g1*C3)+1/(R3_b*C3)),...
0,1/(R3_5*C3), 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0,1/(R3_3b1*C3),1/(R3_3e1*C3),1/(R3_3f1*C3),1/(R3_3g1*C3), 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0,...
0, 0, 0, 0, 0, 0, 0,;1/(R1_4*C4),1/(R2_4*C4), 0,-
1*(1/(R1_4*C4)+1/(R2_4*C4)+1/(R4_5*C4)+1/(R4_6*C4)+1/(R4_4a1*C4)+1/(R4_
4d1*C4)+1/(R4_4e1*C4)+1/(R4_4f1*C4)+1/(R4_4g1*C4)+1/(R4_a*C4)+1/(R4_d*C
4)+1/(R4_e*C4)+1/(R4_f*C4)),...
1/(R4_5*C4),1/(R4_6*C4), 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0,1/(R4_4a1*C4),1/(R4_4d1*C4),1/(R4_4e1*C4),1/(R4_4f1*C4),...
1/(R4_4g1*C4), 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,; 0,
0,1/(R3_5*C5),1/(R4_5*C5),-
1*(1/(R3_5*C5)+1/(R4_5*C5)+1/(R5_6*C5)+1/(R5_7*C5)+1/(R5_5b1*C5)+1/(R5_
5c1*C5)+1/(R5_5e1*C5)+1/(R5_5f1*C5)+1/(R5_5g1*C5)+1/(R5_b*C5)+1/(R5_c*C
5)+1/(R5_e*C5)+1/(R5_f*C5)),...
1/(R5_6*C5),1/(R5_7*C5), 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0,1/(R5_5b1*C5),1/(R5_5c1*C5),1/(R5_5e1*C5),1/(R5_5f1*C5),...
1/(R5_5g1*C5), 0, 0, 0, 0, 0, 0, 0, 0,; 0, 0,
0,1/(R4_6*C6),1/(R5_6*C6),-
1*(1/(R4_6*C6)+1/(R5_6*C6)+1/(R6_7*C6)+1/(R6_6d1*C6)+1/(R6_6e1*C6)+1/(R
6_6f1*C6)+1/(R6_6g1*C6)+1/(R6_d*C6)),...
1/(R6_7*C6), 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0,1/(R6_6d1*C6),1/(R6_6e1*C6),1/(R6_6f1*C6),1/(R6_6g1*C6),...

250
0, 0, 0, 0,; 0, 0, 0, 0,1/(R5_7*C7),1/(R6_7*C7),-
1*(1/(R5_7*C7)+1/(R6_7*C7)+1/(R7_7d1*C7)+1/(R7_7e1*C7)+1/(R7_7f1*C7)+1/
(R7_7g1*C7)+1/(R7_d*C7)+1/(R7_e*C7)+1/(R7_f*C7)),...
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0,
0,1/(R7_7d1*C7),1/(R7_7e1*C7),1/(R7_7f1*C7),1/(R7_7g1*C7),;1/(R1_1a1*C1
a1),...
0, 0, 0, 0, 0, 0,-1*(1/(R1_1a1*C1a1)+1/(R1a1_a*C1a1)), 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0,;1/(R1_1b1*C1b1),...
0, 0, 0, 0, 0, 0, 0,-1*(1/(R1_1b1*C1b1)+1/(R1b1_b*C1b1)), 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0,;1/(R1_1e1*C1e1),...
0, 0, 0, 0, 0, 0, 0, 0,-1*(1/(R1_1e1*C1e1)+1/(R1e1_e*C1e1)), 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0,;1/(R1_1f1*C1f1),...
0, 0, 0, 0, 0, 0, 0, 0, 0,-1*(1/(R1_1f1*C1f1)+1/(R1f1_f*C1f1)), 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0,;1/(R1_1g1*C1g1),...
0, 0, 0, 0, 0, 0, 0, 0, 0, 0,-1*(1/(R1_1g1*C1g1)+1/(R1g1_g*C1g1)), 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0,; 0,1/(R2_2b1*C2b1),...
0, 0, 0, 0, 0, 0, 0, 0, 0, 0,-1*(1/(R2_2b1*C2b1)+1/(R2b1_b*C2b1)), 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,;
0,1/(R2_2e1*C2e1),...
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,-1*(1/(R2_2e1*C2e1)+1/(R2e1_e*C2e1)),
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,;
0,1/(R2_2f1*C2f1),...
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,-
1*(1/(R2_2f1*C2f1)+1/(R2f1_f*C2f1)), 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,; 0,1/(R2_2g1*C2g1),...
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,-
1*(1/(R2_2g1*C2g1)+1/(R2g1_g*C2g1)), 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0,; 0, 0,1/(R3_3b1*C3b1),...
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,-
1*(1/(R3_3b1*C3b1)+1/(R3b1_b*C3b1)), 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0,; 0, 0,1/(R3_3e1*C3e1),...
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,-
1*(1/(R3_3e1*C3e1)+1/(R3e1_e*C3e1)), 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0,; 0, 0,1/(R3_3f1*C3f1),...

251
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,-
1*(1/(R3_3f1*C3f1)+1/(R3f1_f*C3f1)), 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0,; 0, 0,1/(R3_3g1*C3g1),...
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,-
1*(1/(R3_3g1*C3g1)+1/(R3g1_g*C3g1)), 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0,; 0, 0, 0,1/(R4_4a1*C4a1),...
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,-
1*(1/(R4_4a1*C4a1)+1/(R4a1_a*C4a1)), 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0,; 0, 0, 0,1/(R4_4d1*C4d1),...
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,-
1*(1/(R4_4d1*C4d1)+1/(R4d1_d*C4d1)), 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0,; 0, 0, 0,1/(R4_4e1*C4e1),...
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,-
1*(1/(R4_4e1*C4e1)+1/(R4e1_e*C4e1)), 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0,; 0, 0, 0,1/(R4_4f1*C4f1),...
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,-
1*(1/(R4_4f1*C4f1)+1/(R4f1_f*C4f1)), 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0,; 0, 0, 0,1/(R4_4g1*C4g1),...
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,-
1*(1/(R4_4g1*C4g1)+1/(R4g1_g*C4g1)), 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0,; 0, 0, 0, 0,1/(R5_5b1*C5b1),...
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,-
1*(1/(R5_5b1*C5b1)+1/(R5b1_b*C5b1)), 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0,; 0, 0, 0, 0,1/(R5_5c1*C5c1),...
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,-
1*(1/(R5_5c1*C5c1)+1/(R5c1_c*C5c1)), 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,;
0, 0, 0, 0,1/(R5_5e1*C5e1),...
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,-
1*(1/(R5_5e1*C5e1)+1/(R5e1_e*C5e1)), 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,; 0,
0, 0, 0,1/(R5_5f1*C5f1),...
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,-
1*(1/(R5_5f1*C5f1)+1/(R5f1_f*C5f1)), 0, 0, 0, 0, 0, 0, 0, 0, 0,; 0, 0,
0, 0,1/(R5_5g1*C5g1),...
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0,-1*(1/(R5_5g1*C5g1)+1/(R5g1_g*C5g1)), 0, 0, 0, 0, 0, 0, 0, 0,; 0, 0,
0, 0, 0,1/(R6_6d1*C6d1),...
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0,-1*(1/(R6_6d1*C6d1)+1/(R6d1_d*C6d1)), 0, 0, 0, 0, 0, 0, 0,; 0, 0, 0,
0, 0,1/(R6_6e1*C6e1),...

252
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0,-1*(1/(R6_6e1*C6e1)+1/(R6e1_e*C6e1)), 0, 0, 0, 0, 0, 0,; 0, 0, 0, 0,
0,1/(R6_6f1*C6f1),...
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0,-1*(1/(R6_6f1*C6f1)+1/(R6f1_f*C6f1)), 0, 0, 0, 0, 0,; 0, 0, 0, 0,
0,1/(R6_6g1*C6g1),...
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0,-1*(1/(R6_6g1*C6g1)+1/(R6g1_g*C6g1)), 0, 0, 0, 0,; 0, 0, 0, 0,
0, 0,1/(R7_7d1*C7d1),...
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0,-1*(1/(R7_7d1*C7d1)+1/(R7d1_d*C7d1)), 0, 0, 0,; 0, 0, 0, 0, 0,
0,1/(R7_7e1*C7e1),...
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0,-1*(1/(R7_7e1*C7e1)+1/(R7e1_e*C7e1)), 0, 0,; 0, 0, 0, 0, 0,
0,1/(R7_7f1*C7f1),...
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0,-1*(1/(R7_7f1*C7f1)+1/(R7f1_f*C7f1)), 0,; 0, 0, 0, 0, 0,
0,1/(R7_7g1*C7g1),...
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0,-1*(1/(R7_7g1*C7g1)+1/(R7g1_g*C7g1)),;];
% end of the A_Matrix for the model
% start of the B_Matrix for the model, the input is the outdoor air
temperature, C
BB=[1/(R1_a*C1),1/(R1_b*C1), 0, 0, 0, 0, 0,; 0,1/(R2_b*C2), 0, 0, 0, 0,
0,; 0,1/(R3_b*C3), 0, 0, 0, 0, 0,;1/(R4_a*C4), 0,
0,1/(R4_d*C4),1/(R4_e*C4),...
1/(R4_f*C4), 0,; 0,1/(R5_b*C5),1/(R5_c*C5), 0,1/(R5_e*C5),1/(R5_f*C5),
0,; 0, 0, 0,1/(R6_d*C6), 0, 0, 0,; 0, 0, 0,1/(R7_d*C7),...
1/(R7_e*C7),1/(R7_f*C7), 0,;1/(R1a1_a*C1a1), 0, 0, 0, 0, 0, 0,;
0,1/(R1b1_b*C1b1), 0, 0, 0, 0, 0,; 0, 0, 0, 0,1/(R1e1_e*C1e1), 0, 0,...
; 0, 0, 0, 0, 0,1/(R1f1_f*C1f1), 0,; 0, 0, 0, 0, 0, 0,1/(R1g1_g*C1g1),;
0,1/(R2b1_b*C2b1), 0, 0, 0, 0, 0,; 0, 0, 0, 0,1/(R2e1_e*C2e1), 0, 0,; 0,
0,...
0, 0, 0,1/(R2f1_f*C2f1), 0,; 0, 0, 0, 0, 0, 0,1/(R2g1_g*C2g1),;
0,1/(R3b1_b*C3b1), 0, 0, 0, 0, 0,; 0, 0, 0, 0,1/(R3e1_e*C3e1), 0, 0,; 0,
0, 0, 0,...
0,1/(R3f1_f*C3f1), 0,; 0, 0, 0, 0, 0,
0,1/(R3g1_g*C3g1),;1/(R4a1_a*C4a1), 0, 0, 0, 0, 0, 0,; 0, 0,
0,1/(R4d1_d*C4d1), 0, 0, 0,; 0, 0, 0, 0,1/(R4e1_e*C4e1),...

253
0, 0,; 0, 0, 0, 0, 0,1/(R4f1_f*C4f1), 0,; 0, 0, 0, 0, 0,
0,1/(R4g1_g*C4g1),; 0,1/(R5b1_b*C5b1), 0, 0, 0, 0, 0,; 0,
0,1/(R5c1_c*C5c1), 0, 0, 0, 0,...
; 0, 0, 0, 0,1/(R5e1_e*C5e1), 0, 0,; 0, 0, 0, 0, 0,1/(R5f1_f*C5f1), 0,;
0, 0, 0, 0, 0, 0,1/(R5g1_g*C5g1),; 0, 0, 0,1/(R6d1_d*C6d1), 0, 0, 0,; 0,
0,...
0, 0,1/(R6e1_e*C6e1), 0, 0,; 0, 0, 0, 0, 0,1/(R6f1_f*C6f1), 0,; 0, 0,
0, 0, 0, 0,1/(R6g1_g*C6g1),; 0, 0, 0,1/(R7d1_d*C7d1), 0, 0, 0,; 0, 0, 0,
0,...
1/(R7e1_e*C7e1), 0, 0,; 0, 0, 0, 0, 0,1/(R7f1_f*C7f1), 0,; 0, 0, 0, 0,
0, 0,1/(R7g1_g*C7g1),;];
% end of the B_Matrix for the model
% start of the C_Matrix for the model
CC=[1,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
,0,0,0,0,;0,1,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
,0,0,0,0,0,0,0,;0,0,1,0,0,0,0,0,0,0,...
0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,;0,0,0,1,0,0,0,
0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,;0,0,0,0,
1,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,...
0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,;0,0,0,0,0,1,0,0,0,0,0,0,0,0,0,0,0,
0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,;0,0,0,0,0,0,1,0,0,0,0,0,0,0,
0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,...
0,0,0,0,0,0,0,0,;];
% end of the C_Matrix for the model
% start of the D_Matrix for the model
DD=[0,0,0,0,0,0,0,;0,0,0,0,0,0,0,;0,0,0,0,0,0,0,;0,0,0,0,0,0,0,;0,0,0,0
,0,0,0,;0,0,0,0,0,0,0,;0,0,0,0,0,0,0,;];
% end of the D_Matrix for the model

% sensible heat gain into the room air node
Croomair=[C1 C2 C3 C4 C5 C6 C7];
Bb=diag(1./Croomair);
Bb=[Bb;zeros(length(AA)-7,7)];
BB=[BB,Bb];

254
Ba=zeros(length(AA),7);
Ba(12,1)=1/C1g1;Ba(16,2)=1/C2g1;Ba(20,3)=1/C3g1;Ba(25,4)=1/C4g1;Ba(30,5
)=1/C5g1;Ba(34,6)=1/C6g1;Ba(38,7)=1/C7g1;
BB=[BB,Ba];
C-2. Hybrid ventilation unit lookup table
function tableSemco= lookupSem()

yOAT=[-2 5 11 14 19 23 27 34]; % 1st dimension is based on temperature
(C)
xOAH=[20 30 40 50 60 70 80 95]; % 2nd dimension is based on the
relative humidity (RH)

satSemco=[20 23 22.5 22.7 22 22.5 22 22.5;...
20 20.8 20 20.8 20.5 20.8 20 20.8;...
21 21 20 20 19 20.4 20.8 22;...
21 21 20 20 19 20.4 20.8 22;...
17.5 14 16.6 20.1 18.4 17.3 18 17;...
21.5 16.8 17.2 16.3 16 16.7 16.5 17;...
21 16 16.5 18.5 18.5 18.5 18.7 18.7;...
19 16 15.7 18.5 18 18 18 18];

sarhSemco=[21 10 13.5 12 13 13 13 13;...
21 20.3 21 20.3 25 25 25 25;...
21 21 21 30 31 48.3 36.4 42.5;...
21 21 21 30 31 48.3 36.4 42.5;...
38.4 50.6 47.9 45 57 48.7 49 55;...
28 51.7 55 46 49 53 54 52;...
25 50 47 50 50 52 54 54;...
69 45 45 50 52 52 52 52];

cfmSemco=[1500 3000 3000 3000 3000 3000 3000 3000;...
1500 1500 1500 1500 1500 1500 1500 1500;...
1500 1500 1500 1500 1500 1500 1500 1500;...
1500 1500 1500 1500 1500 1500 1500 1500;...
2000 2000 2000 2000 2000 2000 2000 2000;...
1500 1500 2000 2000 2000 2000 2000 2000;...
1500 3000 2000 2000 2000 2000 2000 2000;...
1500 2000 2000 2000 2000 2000 2000 2000];

eleSemco=[1.1 4.2 10 11.4 4.1 4 4.1 4;...
1.1 1.1 1.1 1.1 1.1 1.1 1.1 1.1;...
1.2 1.2 1.15 1.2 1.2 1.15 1.2 1.2;...
1.2 1.2 1.15 1.2 1.2 1.15 1.2 1.2;...
9.05 6.83 0.71 1.2 2.2 8.22 6.4 8.22;...

255
1.1 2.52 5.9 8.47 6.82 9.78 8.44 8.65;...
7.75 13 10.2 10.15 10.15 10.25 9.86 9.86;...
9.2 12.2 12.53 11.88 12.57 12.57 12.57 12.57];

gasSemco=[0.7 0.7 0.7 0.6 0.7 0.7 0.7 0.7;...
0.7 2.5 0.7 10 13 20 23 23;...
2.8 2.76 6.99 19 25.02 5.77 13.81 26.2;...
2.76 2.76 6.99 19 25.02 5.77 13.81 26.2;...
0.7 1.01 0.98 2.3 22.3 24.85 24.85 22.8;...
0.6 0.67 0.7 22.5 21.85 21.64 21.43 22.81;...
0.7 10.82 18 20.08 20.07 23.38 22.31 22.31;...
0.7 21.86 21.33 19.69 20.81 20.81 20.81 20.81];

tableSemco.xOAH = xOAH;
tableSemco.yOAT = yOAT;
tableSemco.satSemco = satSemco;
tableSemco.sarhSemco = sarhSemco;
tableSemco.cfmSemco = cfmSemco;
tableSemco.eleSemco = eleSemco;
tableSemco.gasSemco = gasSemco;
% use of the table:
% w=interp2(xOAH,yOAT,satSemco,xOAHreal,yOATreal);

end
C-3. Hydraulic system network
beta=2; % assume P=SQ^beta
Svlv_type1=56; % the mullion valve, S to valve openning, from gpm to
inch water
Svlv_type2=57; % the main bypass valve, S to valve openning, from gpm
to inch water
Svlv_type3=58; % the radiant ceiling and coolwave bypass valve, S to
valve openning, from gpm to inch water
Svlv_type4=59; % the mixing valve, S to valve openning, from gpm to
inch water
Svlv_type5=60; % the radiant ceiling valve, S to valve openning, from
gpm to inch water
Spip_type1=30; % the 2 inch pipe friciton factor, from gpm to inch
water
Spip_type2=40; % the 1.5 inch pipe friciton factor, from gpm to inch
water
Spip_type3=50; % the 1 inch pipe fricition factor, from gpm to inch
water
Spip_type4=60; % the 0.5 inch pipe friction factor, from gpm to inch
water

256
Q=NaN(88,88); p=NaN(88,1);% use NaN to indicated that those NaN does
not have values or joints in the pipe network
% S=NaN(88,88);

vlvopenness=zeros(88,88);
%-------- get the valve openness, 1-26 the mullion valves
vlvopenness(3,31)=openness(26);vlvopenness(4,31)=openness(25);vlvopenne
ss(5,32)=openness(24);
vlvopenness(9,36)=openness(20);vlvopenness(10,37)=openness(19);vlvopenn
ess(11,39)=openness(18);
vlvopenness(12,40)=openness(17);vlvopenness(13,42)=openness(16);vlvopen
ness(14,43)=openness(15);
vlvopenness(24,51)=openness(8);vlvopenness(25,52)=openness(7);

%--- bypass valves, 27- 29 the bypass valves

%--- radiant ceiling valves, 30-33 the radiant ceiling valves
ness(64,67)=openness(32);vlvopenness(64,68)=openness(33);

%--- cut off valves on , 34, 35, the two branch cut off valves
vlvopenness(38,68)=openness(34);vlvopenness(41,88)=openness(35);

%--- mixing valves, 36, 37, the mixing valves
vlvopenness(1,2)=openness(36);vlvopenness(2,59)=1-openness(36);

Spipe=zeros(88,88);
%-------- calculate the pipe friction drop
for i=1:1:88
for j=i+1:1:88
if B(i,j)==1
if Pipe_R(i,j)<=Pipe_type4 && Pipe_R(i,j)>0
Spipe(i,j)=Spip_type4*Pipe_Lg(i,j);
end
if Pipe_R(i,j)<=Pipe_type3 && Pipe_R(i,j)>Pipe_type4
end
end

257
end
end
end
end

%-------- define the minor friction drop due to the fittings
Sminor=zeros(88,88);
for i=1:1:88
end

%-------- define the control valve friction drop due to the valve
opennes
%-------- if the valve is fully closed, the S value is NaN, the branch
%-------- flow is determined as a one line equation as 0;
%-------- the corresponding pipe energy should not be calculated
Svlv=zeros(88,88);
for i=1:1:88
for j=i+1:1:88
if B(i,j)==1
if Vlv_ind(i,j)==vlv_type1
Svlv(i,j)=SvlvCal(vlv_type1,vlvopenness(i,j));
end
end
end
end
end
end
end
end

%-------- additional for the two branches that could be closed totally
if vlvopenness(38,68)==0
Svlv(20,61)=Svlv(38,68);
for i=61:1:68
for j=i+1:1:68
if B(i,j)==1
Svlv(i,j)=Svlv(38,68);
end
end
end
end
if vlvopenness(41,88)==0
Svlv(23,69)=Svlv(41,88);
for i=69:1:88
for j=i+1:1:88

258
if B(i,j)==1
Svlv(i,j)=Svlv(41,88);
end
end
end
end

%-------- calculate the total pressure drop coefficient
% S=zeros(88,88);

S=Svlv+Sminor+Spipe;

%-------- generate the nonlinear equations for optimization

%--- first, make a map about the mass flow, so it becomes a vector in
the
%--- object function and constraint function, instead of being a matrix
K=0; % check how many flows or pipes are included
pipe={};
for i=1:1:88
for j=1:1:88
if B(i,j)==1
K=K+1;
pipe(K,1)={strcat(num2str(i),'_',num2str(j))};
end
end
end
% check the pipe connection first
flowmap=cell(K,1); % this one use from main function to obj and con
flowmapreverse=cell(K,1); % this one makes the result back to the
matrix
flow=cell(88,88); % this one keeps the flow data
count=0;
for i=1:1:88
for j=i:1:88
if B(i,j)==1
count=count+1;
k1=min(i,j);k2=max(i,j);

flowmap(count,1)=strcat(flowmap(count,1),'flow','(',num2str(count),')',
'=','Q','(',num2str(k1),',',num2str(k2),')',';');

flowmapreverse(count,1)=strcat(flowmapreverse(count,1),'Q','(',num2str(
k1),',',num2str(k2),')','=','flow','(',num2str(count),')',';');
flow(i,j)=strcat(flow(i,j),'flow','(',num2str(count),')');
end
end
end

%--- second, the node flow balance equations
mbal=cell(88,1);
msum=cell(88,1);
nodes=sum(A,2);
jointnum=0;

259
boundaryflag=0;
boundarynum=0;
for i=1:1:88
% the node connection information for each node is stored in the A
% matrix, not B matrix. B matrix is the one reduced the duplication
% joint number is not necessary equal to nodes number, some nodes
are
% bourdary condition, for example 1 and 60 in this example
if nodes(i)>1
jointnum=jointnum+1;
tempcount=nodes(i);
for j=1:1:88
if A(i,j)==1 % use A matrix, because it truly indicated the
connection with duplication
% must make sure the flow vector and matrix match to each
other
mbal(i,1)=strcat(mbal(i,1),flow(k1,k2));

msum(i,1)=strcat(msum(i,1),'Q','(',num2str(k1),',',num2str(k2),')');
tempcount=tempcount-1;
if tempcount>=1
mbal(i,1)=strcat(mbal(i,1),'+');
msum(i,1)=strcat(msum(i,1),'+');
end
%
msum(i,1)=strcat(msum(i,1),'Q',num2str(k1),'_',num2str(k2),'+');

end
end
mbal(i,1)=strcat('f(',num2str(jointnum),')','=',mbal(i,1),';');
msum(i,1)=strcat('f(',num2str(jointnum),')','=',msum(i,1),';');
else if nodes(i)==1 % boundary condition for the flow in pipe 1-2
and pipe 59-60
if boundaryflag==0
jointnum=jointnum+1;
boundaryflag=1;
boundarynum=i;

mbal(boundarynum,1)=strcat(mbal(boundarynum,1),'f(',num2str(jointnum),'
)','=',flow(boundarynum,boundarynum+1));
else
mbal(boundarynum,1)=strcat(mbal(boundarynum,1),'-',flow(i-
1,i),';');
end
end
end
end

%--- third, the energy balance equations
pbal=cell(K,1); % pressure balance
count=0;

260
for i=1:1:88
for j=i:1:88
if B(i,j)==1
count=count+1;
% remember to check the Qi,j, if negative, then the
% pressure should be reversed; otherwise positive
%
pbal(count,1)=strcat(pbal(count,1),'abs','(','p','(',num2str(k1),')','-
','p','(',num2str(k2),')',')','-','S',...
%
'(',num2str(k1),',',num2str(k2),')','*','Q','(',num2str(k1),',',num2str
(k2),')','^',num2str(beta));

pbal(count,1)=strcat(pbal(count,1),'abs','(','p','(',num2str(k1),')','-
','p','(',num2str(k2),')',')','-','S',...

'(',num2str(k1),',',num2str(k2),')','*',flow(k1,k2),'^',num2str(beta));
if Pump_ind(i,j)>0
% if there is a pump, the pump energy should be
counted
pumpc=Pumpcoef(Pump_ind(i,j));
%
pbal(count,1)=strcat(pbal(count,1),'+',num2str(pumpc(1)),'*','Q','(',nu
m2str(k1),',',num2str(k2),')','^','2',...
%
'+',num2str(pumpc(2)),'*','Q','(',num2str(k1),',',num2str(k2),')','+',n
um2str(pumpc(3)));

pbal(count,1)=strcat(pbal(count,1),'+',num2str(pumpc(1)),'*',flow(k1,k2
),'^','2',...

'+',num2str(pumpc(2)),'*',flow(k1,k2),'+',num2str(pumpc(3)));
end

pbal(count,1)=strcat('pbal','(',num2str(count),')','=',pbal(count,1),';
');
end
end
end

%--- forth, generate the vector of friction loss from the S vector
sval=cell(K,1); % the vector of friction coefficient
count=0;
for i=1:1:88
for j=i:1:88
if B(i,j)==1
count=count+1;
if isnan(S(i,j))

sval(count,1)=strcat(sval(count,1),'S','(',num2str(k1),',',num2str(k2),
')','=','0',';');
else

261
% remember to check the Qi,j, if negative, then the
% pressure should be reversed; otherwise positive

sval(count,1)=strcat(sval(count,1),'S','(',num2str(k1),',',num2str(k2),
')','=',num2str(S(k1,k2)),';');
end

end
end
end

Model Predictive Control Thesesgood PDF

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Model Predictive Control Thesesgood PDF

Uploaded by

Copyright:

Available Formats

PH.D.

is the air density (kg/m

is the relative humidity of desiccant matrix,

is the air density (kg/m

is the constant air density (kg/m

is the density of the

is the density of the refrigerant (kg/m

is the friction coefficient,

is the density of air, a cd

is the mass flow rate of

are the liquid temperature (K) and density (kg/m

is the air flow rate

is the vapor density entering the compressor (kg/m

is the fluid density (kg/m

You might also like