You are on page 1of 14

Chemical Engineering Science 69 (2012) 4558

Contents lists available at ScienceDirect

Chemical Engineering Science


journal homepage: www.elsevier.com/locate/ces

Energy efcient model predictive building temperature control$


Matt Wallace a, Ryan McBride a, Siam Aumi a, Prashant Mhaskar a,, John House b, Tim Salsbury b
a
b

Department of Chemical Engineering, McMaster University, 1280 Main Street West, Hamilton, Ontario, Canada L8S 4L8
Johnson Controls Inc., 507 E. Michigan Street, Milwaukee, WI 53202, United States

a r t i c l e i n f o

abstract

Article history:
Received 4 March 2011
Received in revised form
5 July 2011
Accepted 8 July 2011
Available online 29 July 2011

Many systems used in buildings for heating, ventilating, and air-conditioning waste energy because of
the way they are operated or controlled. This paper explores the application of model predictive control
(MPC) to air-conditioning units and demonstrates that the closed-loop performance and energy
efciency can be improved over conventional approaches. This work focuses on the problem of
controlling the vapor compression cycle (VCC) in an air-conditioning system, containing refrigerant
which is used to provide cooling. The VCC considered in this work has two manipulated variables that
affect operation: compressor speed and the position of an electronic expansion valve. The system is
subject to constraints, such as the range of permissible superheat, and also needs to regulate
temperature variables to set points. An MPC strategy is developed for this type of system based on
linear models identied from data obtained from a rst-principles model of the VCC. The MPC strategy
incorporates economic measures in the objective function as well as control objectives. Tests are carried
out on a simulated VCC system that is linked to a simulation of a realistic building that is developed in
the U.S. Department of Energy Computer Simulation Program, EnergyPlus. The MPC demonstrated
signicantly better tracking control relative to conventional approaches (a reduction of 70% in terms of
the integral of squared error for step changes in the temperature set-point), while reducing the VCC
energy requirements by 16%. The paper describes the control approach in detail and presents results
from the tests.
& 2011 Elsevier Ltd. All rights reserved.

Keywords:
Vapor compression cycle
Temperature control
Building control
Energy efcient control
Model predictive control
EnergyPlus

1. Introduction
Environmental concerns as well as increased fuel prices have
brought energy efciency to the forefront of research priorities.
Canada currently ranks as the worlds sixth largest user of
primary energy (such as fossil fuels, nuclear fuels, hydro power,
etc.). In Canada, approximately 30% of the energy obtained from
primary sources of energy is consumed in the commercial and
residential sectors of the economy (Behidj et al., 2009). In these
sectors, a signicant portion of the energy is used towards
meeting the thermal and electrical energy demands in buildings.
Recent government reports estimate that through more efcient
building operation, the total energy consumption by the commercial and residential sectors can be reduced by 1520% (Behidj
et al., 2009).
The operating efciency of a building is inuenced by many
factors and can be improved at various points over its lifespan.
Prior to construction, using design standards that incorporate
energy and environmental concerns is often the rst step to
$
This work is a collaborative effort between Johnson Controls Inc. and the
McMaster Advanced Control Consortium.
 Corresponding author.
E-mail address: mhaskar@mcmaster.ca (P. Mhaskar).

0009-2509/$ - see front matter & 2011 Elsevier Ltd. All rights reserved.
doi:10.1016/j.ces.2011.07.023

achieving an energy efcient building design. However, these


design standards alone are not sufcient to ensure that a building
remains energy efcient in response to changing energy and
environmental standards. Once constructed, with the advent of
more energy efcient technology (i.e., EnergyStar certied technology), the building can be appropriately retrotted to meet
more stringent energy and environmental standards. Finally, the
energy efciency of existing buildings can be improved through
better control of their heating, ventilation, and air-conditioning
(HVAC) systems (see American Society of Heating, for a detailed
description of the common components of an HVAC system),
which regulate building comfort (temperature and humidity) and
account for 3050% of the total energy consumption in buildings
(Albieri et al., 2009). The focus of the present work is to
demonstrate (via simulations) this improved efciency achievable
through use of advanced (model based) control techniques. To
this end, we utilize an existing model of a VCC and couple it with
a building model to approximate a (reasonably) realistic scenario
of a roof top unit providing cooling to a room in a building using
air as the only cooling medium.
A vapor compression cycle (VCC) refers to a type of thermodynamic machinery that transfers heat using a compressible uid
referred to as the refrigerant. The most common realization of a
VCC consists of four components: a compressor, condenser,

46

M. Wallace et al. / Chemical Engineering Science 69 (2012) 4558

expansion valve, and an evaporator. In the VCC, the refrigerant


circulates through the four components, undergoing various
thermodynamic changes, which, in turn, inuence external conditions. Mathematically, the dynamics of the refrigerant states
and external conditions are modeled using a set of coupled
nonlinear ordinary differential and algebraic equations, resulting
in a complex differential-algebraic equation system. The control
objectives are typically dened in terms of degrees of superheat
in the refrigerant at the evaporator exit and the air temperature at
the evaporator exit (the supply air temperature). Ensuring that
superheated refrigerant exits the evaporator is of utmost importance in preventing physical damage in the VCC, as liquid
refrigerant can damage the mechanical components used in the
compressor. The manipulated variables include the compressor
speed, the air ow rates and the expansion valve opening. While
research activity has been strong for a long time on the design/
material and, to a smaller extent, modeling, of the various
components of the VCC (see Rasmussen, 2005 for details), there
has been a recent trend towards improved control of the VCC for
improved energy efciency.
Traditional VCC control strategies have included PID/PI decentralized control (i.e., multiple independent single-inputsingleoutput (SISO) controllers) and simple on/off control. The latter
limits overall efciency due to large power requirements and
signicant thermal inertia during start-up transients while the
formers efciency is limited by extensive interactions and nonlinearity in the VCC system dynamics and the presence of input
constraints. Approaches to improve the performance of conventional PID/PI controllers can be categorized into those which
attempt to decouple the VCC dynamics to improve SISO controllability (Keir and Alleyne, 2007; Jain et al., 2010) and adaptive
control approaches which attempt to account for the process
nonlinearity (i.e. time-varying process gains) (Lin and Yeh, 2007;
Zhu et al., 2001).
Among the decoupling approaches, the most straightforward
extension has been to employ linear decouplers to remove
interactions among the individual control loops. However, the
effectiveness of this approach is entirely contingent on the model
accuracy, and a poor model can lead to closed-loop performance
degradation. Another decoupling approach has been to use linear
combinations of available VCC measurements as the controlled
variables instead of traditional controlled variables (Jain et al.,
2010). The adaptive control approaches, on the other hand,
attempt to account for the nonlinear nature of the VCC dynamics
by updating the PI/PID tuning parameters online using a model of
the process. However, these approaches are typically restricted to
linear models (for computational reasons), implying the tuning
parameter updates may be erroneously updated since the true
process is highly nonlinear, leading to poor control performance.
Despite these improvements, PI/PID control designs remain
inherently based on a single-inputsingle-output framework
and do not account for the presence of constraints and optimality.
The control action prescribed by a controller that does not
account for input constraints can result in performance degradation or even closed-loop instability.
One control method well suited to handling constraints and
optimality is model predictive control (MPC). MPC is an optimization-based control approach in which the coupled, multiple-input
multiple-output nature of complex systems can be accounted for in
determining the control action by using a model of the process. In
model predictive control, a nonlinear or linear process model is
used to evaluate the effect of candidate manipulated input trajectories via an objective function, and an optimization problem is
solved to yield the manipulated input trajectory that minimizes the
objective function while satisfying any constraints. Only the rst
piece of the input trajectory is implemented and the problem is

re-solved at the next sampling time, using the new measured


values of the process variables. One of the strengths of the MPC
framework is the exibility (and scope) in specifying optimality
objectives through various terms in the objective function, or
through constraints on the variables of interest. This, and the
results available on the stability and feasibility properties of MPC
formulations (see, e.g., Mhaskar et al., 2005, 2006; Mhaskar, 2006)
make MPC a preferred candidate to be evaluated for possible use
within building control structures. Recently, there have been many
examples in the literature of the application of MPC for regulating a
wide range of VCC or HVAC systems. In addition to the nature of
the system being regulated, the major differentiating feature
among these MPC approaches is the complexity of the model used
for predictions. Specically, the predictive model may be a linearized version of a non-linear state-space model (Schurt et al., 2009;
Sandipan et al., 2010; Morosan et al., 2010), an empirically
identied linear (Huang et al., 2009; Ma et al., 2010) or nonlinear
model (Xi et al., 2007), or a rst-principles non-linear model
(Leducq et al., 2006; Ma et al., 2010; Sarabia et al., 2007). The
majority of the MPC application examples have utilized a linear
VCC model (either a linearized state-space model or an empirically
identied inputoutput model). For example, in Sandipan et al.
(2010), an experimental chiller network, consisting of two chillers
and multiple external heat exchangers, is regulated to satisfy the
cooling load in addition to minimizing electricity costs. In Huang
et al. (2009), empirical rst-order time-delay (FOTD) models are
used in a robust MPC formulation for improving temperature
regulation of an air-conditioning system. Specically, several FOTD
models are identied at various operating points, and based on the
current operating conditions, the most appropriate FOTD model is
used in the MPC optimization. Linear MPC applications in the
context of building control include the work in Morosan et al.
(2010) where a distributed MPC design is used to regulate the
temperature of multiple zones in a building and minimize power
consumption. That is, each zone is served by a separate HVAC unit
under the control of a zone-specic MPC design to regulate the
internal zone temperature. Another example of building control
using (linear) MPC is available in Ma et al. (2010) where weather
data is incorporated into the design to determine the building zone
temperature set-point. This allows for pre-cooling during non-peak
periods and reduced power consumption compared to traditional
pre-programmed HVAC unit control strategies. The existing results
notwithstanding, there still exists a lack of results on the application of a MPC design to a detailed model of a VCC unit coupled with
a realistic building model to evaluate the control performance in
the presence of disturbances.
Motivated by the above, this work evaluates the performance
of an integrated temperature control framework via simulations.
Specically, we design a predictive controller for a stand-alone
VCC unit and utilize it in a cascade control structure for temperature control. The proposed control structure is implemented on a
realistic building model that accounts for varying weather conditions and internal heat loads throughout the course of a day and
the results are compared with a PI-based control structure. The
rest of this manuscript is organized as follows. In Section 2, we
give an overview of the VCC and building models used in this
work. Then, in Section 3, we develop a control strategy for
temperature control in a building zone. To this end, we rst
estimate an inputoutput model for the VCC in Section 3.1 and
then design an offset free predictive controller for a stand-alone
VCC unit in Section 3.2. This controller is subsequently incorporated in a cascade control strategy to control a specic room
temperature and then implemented on a realistic building model
in Section 3.3. The performance of the proposed control strategy is
compared with a conventional PI-based control strategy. Finally,
we summarize our results in Section 4.

M. Wallace et al. / Chemical Engineering Science 69 (2012) 4558

2. Preliminaries
In this section, we give an overview of an existing VCC model
used in this work and point out the key limitations of the models
and modications. Next, we describe the building model and the
software used for interfacing the VCC model with the building
model. Note that model development is not the focus of the
present work. A detailed model of the VCC and the building is
used only to illustrate the control design and is described in this
section for completeness.
2.1. VCC model overview
An ideal VCC consists of four processes: isentropic compression in a compressor, isobaric energy dissipation in a condenser,
isenthalpic expansion in an expansion valve and isobaric energy
absorption in an evaporator. An overlay of a VCC and the
corresponding pressure-volume diagram of the refrigerant is
shown in Fig. 1.
In a VCC unit, the refrigerant enters the compressor as a
superheated vapor and is compressed to a higher pressure,
resulting in the superheated vapor having a higher temperature
than the ambient temperature. From the compressor, the superheated refrigerant vapor enters a condenser (typically placed
outdoors), condensing to a sub-cooled liquid at the condenser
exit as a fan blows the ambient air over the condenser. The high
pressure sub-cooled liquid then ows into an expansion valve
which decreases the pressure and temperature of the refrigerant,
causing a liquidvapor mixture to form. Then, the two-phase
refrigerant mixture enters an evaporator that is exposed to the
environment to be cooled. The environment temperature is above
the temperature of the refrigerant, resulting in the evaporation
and subsequent heating of the refrigerant to a superheated vapor
at the evaporator exit. The air, in turn, is cooled and available as
primary air to be distributed for cooling. The superheated vapor
from the evaporator exit then ows into the compressor, completing the cycle.
The VCC model used in this work is adapted from the existing
simulation package, Thermosys, developed at the Air Conditioning
& Refrigeration Center (ACRC) at the University of Illinois at
Urbana-Champaign. In this simulation package, the refrigerant is
R-134a and is assumed to be cooling an air medium. The
simulator consists of dynamic models for the condenser, evaporator, and compressor and static models (i.e., algebraic equations) for the expansion valve and piping. In the following
subsections, a brief overview of the model components is provided followed by a general mathematical representation. For a

Pressure

Heat from refrigerant


Liquid

Condenser
Valve
Evaporator

Wet vapor
(saturated conditions)

Compressor

Gas
Heat from process
Volume

Fig. 1. VCC overlay on a pressurevolume diagram of a typical refrigerant.

47

full description of the model components and a complete list


of the equations and parameters, the reader is referred to
Rasmussen (2005).
2.1.1. Compressor
The compressor in the VCC model is a reciprocating compressor dened by its isentropic efciency, Zk , which is the ratio of the
work required for ideal adiabatic compression to the work
required for actual compression, and its volumetric efciency,
ZV , which is the ratio of induced gas volume to the discharged gas
volume (swept volume). Note that in the present work we use a
model of a variable speed compressor to illustrate the improved
efciency achievable by model-based control designs and the key
interpretations remain applicable for other compressor types (on/
off compressors, etc.). Demonstrating improved energy efciency
on other cooling units is the subject of future work and outside
the scope of the present manuscript.
The efciency of the compressor depends on the pressure ratio
of the outlet to inlet stream and the compressor RPM, ok . They
are obtained via (experimentally obtained) lookup tables of
efciencies at various operating conditions. The mass ow rate
in the compressor is modeled using the following static equation:
_ k ok Vk rk ZV
m

where Vk denotes the swept volume (a compressor parameter)


and rk is the refrigerant density at the inlet. The term, Vk rk ZV ,
characterizes the compressor capacity in terms of inlet refrigerant
conditions. For the compressor energy balance, the dynamics of
the heat transfer during the transport of the refrigerant from the
compression cylinder (where the compression takes place) to the
shell (where the refrigerant exits) are taken into consideration.
Specically, the energy dynamics are modeled as a linear (in K),
rst-order differential equation:
o

tk h_ k hok KZk ,hik ,h k

hok

is the enthalpy of the outlet refrigerant, tk is a time


where
constant (a compressor parameter), and K is the gain. The gain
is constant during integration and a nonlinear function of the
isentropic efciency, inlet refrigerant enthalpy, hik , and the ideal
o
isentropic outlet enthalpy, h k , which is determined by the
refrigerant thermodynamic properties and inlet enthalpy.
Remark 1. In general, the compressor type (variable speed or on/
off) is dependent on the specic application of the VCC (most
existing compressors utilize an on/off strategy). Due to the
limited range of operating speeds for an on/off compressor, the
startup and shutdown of a setup equipped with this compressor
type can draw considerably more energy during these operating
conditions than a setup equipped with a variable speed compressor. Demonstrating the improvement over traditional onoff
setups (using a good model of such a unit) does remain an
objective of future work, but is outside the scope of this manuscript. If an on/off type compressor is used in the VCC, modications can be made to the model (as presented in this section), as
well as to the control design to ensure the implementation of the
MPC control structure is feasible. In particular, any proposed
control design for the VCC must account for the discrete nature of
the compressor operation. For instance, in the present work, the
compressor RPM is a manipulated variable for VCC control and
treated as a continuous variable. With an on/off type compressor,
the compressor RPM is xed when it is on and zero otherwise. In a
model predictive control framework, this can be modeled either
indirectly (by choosing the on/off durations as input variables) or
explicitly using binary variables in the optimization problem. On
the other hand, classical control approaches, such as PI control,
have limited options in handling units with discrete operation.

48

M. Wallace et al. / Chemical Engineering Science 69 (2012) 4558

Another limitation of the current VCC model is the range of


compressor speeds over which the model is valid. In particular,
the model does not remain valid for low RPM values, and as a
result, the simulations have been carried out with the RPM
restricted to this range to preserve the validity of the results.
Enhancing the range of validity of the model also remains another
direction of future work.
2.1.2. Expansion valve
The expansion valve is modeled as an isenthalpic process,
meaning the inlet and outlet enthalpies are identical, with a mass
ow rate given by the following expression:
q
_ v Cd rv DPv
m
3
where Cd is the valve discharge coefcient, DPv is dened as the
pressure difference between the inlet and outlet refrigerant, and
rv is the maximum of either the sub-cooled liquid density or the
saturated liquid density at the inlet operating conditions. The
discharge coefcient is determined by the valve opening and
pressure differential and obtainable from experimental lookup
tables. The sub-cooled and saturated liquid refrigerant densities
are also obtained from lookup tables of the refrigerants thermodynamic properties.
Remark 2. An electronic expansion valve (EEV) is used in this
VCC model, where the valve position is adjustable and in agreement with electronic expansion valves used in practice. The valve
position of an actual EEV is proportionally adjusted through
varying the frequency of a built-in step-motor, where the energy
draw of this motor is minimal relative to the other energyconsuming components of a practical VCC (i.e., compressor and
fan motors).
2.1.3. Heat exchangers
The dynamics of the VCC are dominated by the condenser and
evaporator. Both heat exchangers are modeled as a long thin
horizontal tube with one-dimensional uid ow, negligible pressure drop (due to momentum change and viscous friction), and
negligible axial conduction.
In both heat exchangers, the refrigerant may undergo multiple
phase changes; accordingly, the refrigerant is modeled using a
lumped parameter, moving boundary approach which accounts
for different uid regions (superheated vapor, saturated vapor
liquid, or sub-cooled liquid) and their time varying boundaries. In
this approach, each uid region is represented as a separate
control volume (see Fig. 2) with corresponding states and parameters. For the evaporator, there are two uid regions: a twophase region followed by a superheat region while the condenser
has three uid regions: a vapor region followed by a two-phase
and a sub-cooled region. In each two-phase region, the refrigerant
uid properties are taken as the weighted combination of the
saturated liquid and vapor properties. The mean void fraction, g,
which is dened as the ratio of the vapor volume in a region to
the total region volume, is used to weight the properties. For
instance, the refrigerant density in a two-phase region is given by
grf 1gr where rf and r are the saturated vapor and liquid

Two-phase region

Superheat region

densities, respectively. In the superheated and sub-cooled


regions, the refrigerant properties, such as density and temperature, are determined using the heat exchanger pressure (assumed
constant) and the average regional enthalpy (the average of the
inlet and outlet enthalpies).
The mass and energy balance ordinary differential equations
(ODEs) for each uid region are derived from the governing
partial differential equations (PDEs) for uid ow in a tube. To
yield a set of ODEs from the PDEs, the spatial dependence from
the PDEs is removed after applying simplifying assumptions and
Leibnitzs rule on any differential with respect to the spatial coordinate, z. The full details of the modeling approach are available
in Rasmussen (2005). Eqs. (4) and (5) represent the governing
refrigerant mass and energy balance PDEs (respectively) of a
specic uid region.
_
@rAc @m

0
@t
@z

_
@rAc hAc P @mh

pi ai Tw Tr
@t
@z

_ and h denote the refrigerant density, mass ow rate,


where r, m,
and specic enthalpy (respectively), Ac is the heat exchanger
cross-sectional area, P is the uid region pressure, pi is the inner
perimeter of the heat exchanger, ai is the heat transfer coefcient
between the refrigerant and the heat exchanger inner wall, Tw is
the wall temperature, and Tr is the refrigerant temperature. These
PDEs are coupled with the following wall energy balance for each
region:
cp rAw T_ w pi ai Tr Tw po ao Ta Tw
where cp rAw is the thermal capacitance of the tube wall per unit
length, po is the outer perimeter of the heat exchanger, and ao is
the heat transfer coefcient between the tube wall and the
_
surrounding air with temperature Ta. After integrating @m=@z
_
and @mh=@z
along the length of the tube using Leibnitzs rule,
the nal set of ODEs for the heat exchangers can be arranged in
the following matrix form:
Zh xh ,ux_ h f h xh ,u,d
where Zh  and f h  are a matrix and vector, respectively, and the
heat exchanger state variables, xh , include: the length of the
superheat, Lc,1 , and two-phase, Lc,2 , regions in the condenser,
the condenser wall temperatures in all three regions, Tw,c,1 , Tw,c,2 ,
and Tw,c,3 , the constant condenser pressure, Pc, the condenser
outlet refrigerant enthalpy, hoc, the length of the two-phase region
in the evaporator, Le,1 , the evaporator wall temperatures for both
regions, Tw,e,1 , Tw,e,2 , the constant evaporator pressure, Pe, the
evaporator outlet refrigerant enthalpy, hoe, and the compressor
outlet refrigerant enthalpy, hok. The input vector, u, elements are
the compressor RPM, ok , and valve opening, vo. Note that in some
VCC systems, the fan speeds for the air being blown over the
evaporator and condenser (and therefore the mass ow rate of
air) may also be available for adjustment; however, for this VCC
model, these are assumed constant. The disturbance vector, d, is
constituted of two measurable temperatures: the temperature of

Superheat region

Two-phase region

Sub-cooled region

Fig. 2. Heat exchanger schematics showing the different uid regions (adapted from Rasmussen, 2005) which are modeled using the moving boundary approach.
(a) Evaporator with two uid regions. (b) Condenser with three uid regions.

M. Wallace et al. / Chemical Engineering Science 69 (2012) 4558


i
the air to be blown over the evaporator, Ta,e
, which is commonly
referred to as the mixed air temperature, and the air temperature
at the inlet of the condenser. The latter temperature is simply the
ambient air temperature since the condenser is assumed to be
outdoors and henceforth will be denoted by Tamb . The air supplied
to the VCC evaporator is typically a mix of the zone (i.e., room)
and ambient air. For example, the mixed air may be a mixture of
80% zone air and 20% ambient air.

2.1.4. Mathematical representation


The Thermosys VCC model comes in the form of a Simulinkbased toolbox in Matlab. For this work, we extracted the VCC
model ODEs and algebraic expressions from the source les and
expressed the model as a differential-algebraic equation (DAE)
system. In order to integrate the DAE system, algebraic equations
are required to be satised at all integration steps. Integrating the
DAE system in Matlab as opposed to running the Simulink model
les yielded signicant computational benets with execution
times for the same test period being reduced by over 70%. The
VCC DAE can be expressed in the following general form:
Zx,ux_ f x,u,d
gx,u 0
y hx,d
where Z and f  again denote a matrix and vector, respectively,
x is the VCC state vector, u and d were previously dened in
Section 2.1.3, gx,u represents the algebraic expressions (used for
modeling the piping and expansion valve), and y denotes the VCC
outputs. The VCC outputs are dened to be the superheat of the
refrigerant exiting the evaporator, Ts,e , and the air temperature at
o
the evaporator exit, the so-called supply air temperature, Ta,e
. The
superheat is dened as the number of degrees by which the
refrigerant temperature at the evaporator exceeds its saturation
temperature. As mentioned in Section 1, Ts,e is required to be
maintained above 0 1C to protect against any liquid refrigerant
entering the compressor and therefore required for safe and
reliable compressor operation. In practice, the superheat is maintained above 0 1C with a safety margin. The vector, h, denotes
the (nonlinear) output mapping function. The mapping function
for the superheat is relatively straightforward whereas the function to compute the supply air temperature consists of nding the
root of a nonlinear equation as discussed next.
The following discussion contains modications of the supply
air temperature calculation procedure found in the original
Thermosys model. Specically, we make corrections to the procedure in the event of any condensation of the water vapor
o
content in the air. To obtain the supply air temperature, Ta,e
, an
energy balance for the wall side of the evaporator is solved.
Assuming no energy accumulation in the evaporator walls, the
heat absorption by the evaporator wall must equal the heat loss of
the air:

ao Ac

2
X
Le,i
i1

Le

T a,e Tw,e,i Hloss

P
The term, ao Ac 2i 1 Le,i =Le T a,e Tw,e,i , represents the energy
absorption by the evaporator where T a,e is the average air
temperature around the evaporator:
i
o
T a,e 12Ta,e
Ta,e

In the most general case (i.e., assuming there is condensation of


the water vapor content in the air), the heat loss of the air, Hloss , is
given by
i
o
_ a,e cp,a,e Ta,e
_ a,e wia,e hiW,e woa,e hoW,e h,e
Ta,e
m
Hloss m

49

_ a,e and cp,a,e denote the mass ow rate and specic heat
where m
capacity of the dry air being blown over the evaporator (assumed
to remain constant), wia,e and woa,e denote the humidity ratio,
which is dened as the ratio of the mass of water vapor in the air
to the total dry air mass, of the inlet and outlet air (respectively),
and hiW,e and hoW,e denote the specic water vapor enthalpy at the
inlet and outlet air conditions (respectively). The rst term in Eq.
(7) is the heat loss of the dry air and the only unknown variable in
o
this term is Ta,e
(the variable of interest). The second term is the
energy loss of the water vapor content in the air. In this term, the
inlet humidity and water vapor enthalpy are readily computable
from the known temperature (and pressure). If no condensation
occurs, there is no change in the humidity ratio and wia,e woa,e . In
the case of condensation, the outlet air is saturated, implying the
o
relative humidity at the outlet, fa,e , is 1.1 To compute the
humidity ratio, its relationship with the relative humidity can
be used to derive (Dincer and Rosen, 2007, Chapter 6):
woa,e 0:622

o
foa,e PW,e,sat
o
o
Pa fa,e PW,e,sat

o
where Pa is the known air pressure and PW,e,sat
is the saturation
o
pressure of water at Ta,e , which can be computed using Antoines
equation. Meanwhile, the outlet water vapor enthalpy is computed using the standard formula:
Z Ta,e
o
hoW,e hf ,sat
cp,W T dT
Tref

where hf ,sat is the heat of saturated water vapor at the air


pressure, Tref is a reference temperature, and cp,W T is the
(possibly) temperature-dependent specic heat capacity of water
vapor.2 The third term in Eq. (7), h,e , represents the heat content
in the condensed water if condensation occurs. Note that the
negative sign is required in front of h,e since heat losses are
written as positive energies in Eq. (7). The heat content in the
water is given by
o
o
_ a,e wia,e woa,e cp,W Ta,e
_ ,e cp,W Ta,e
m
h,e m

_ a,e wia,e woa,e equals the mass of condensed


where the product m
_ ,e and
water (follows from the denition of humidity ratio) or m
cp,W is the constant heat capacity of liquid water.
Having dened all the terms/variables in Eq. (6) and their
o
dependence on the unknown supply air temperature, Te,a
, a root
o
nding algorithm can be applied to Eq. (6) to compute Te,a
.
Alternatively, an iterative (i.e., direct-substitution) procedure
can be used where an initial guess for the supply air temperature
is made, Hloss is computed, and then the left hand side of Eq. (6) is
o
solved for Ta,e
. If the difference between the newly computed
supply air temperature and the initial guess exceeds a pre-dened
tolerance, the newly computed value can be used to initialize the
next iteration. Note also that when solving Eq. (6), an assumption
regarding the occurrence of condensation has to be made. In this
work, we rst solve Eq. (6), assuming no condensation (i.e., with
o
no h,w term), which is correct only if Ta,e
4 Tdp where Tdp is the
dew-point temperature (computable from the air pressure). If
o
Ta,e
rTdp , the necessary condensation term is added to Hloss prior
to solving Eq. (6) and the equation system is re-solved.
2.1.5. VCC cooling capacity
For proper regulation of a building zone temperature, the
corresponding VCC unit for the zone must meet the cooling
1
The relative humidity is dened as the ratio of the partial pressure of the
water vapor to the saturation pressure of water at the system temperature.
2
The same reference temperature is used for the computation of how,e and hiw,e
such that Tref disappears in Eq. (7).

50

M. Wallace et al. / Chemical Engineering Science 69 (2012) 4558

capacity dictated by the highest possible ambient conditions and


heat load disturbances. The nominal Thermosys VCC model has a
maximum cooling capacity of 1127 W or 0.32 ton of refrigeration.
This capacity is in agreement with that of a small experimental
VCC used to validate the nominal model. Exploratory simulations
revealed that this cooling capacity is insufcient (even with
perfect control) to achieve the desired control objectives in terms
of temperature control (see Section 3.3) for the ambient conditions and heat load disturbances considered in the simulations. As
a result, the VCC model parameters are re-scaled such that the
cooling capacity increases to 0.85 ton. Specically, the mass ow
_ r , along with the compressor
rate of the circulating refrigerant, m
volume, Vk, are rst increased. Next, the length of the evaporator,
Le, is increased to allow longer contact of the supply air with the
evaporator wall. The inner and outer cross-sectional areas of the
evaporator, Ae,i and Ae,o respectively, and the mass of the evaporator, Me, are then increased by the same factor. Then, the
diameter of the evaporator pipe, De, and (dry) air mass ow rate,
_ a,e , are increased to allow for more heat transfer from the air
m
passing over the evaporator. To ensure that the additional heat
absorbed by the refrigerant in the evaporator could be dissipated
into the surrounding environment at the condenser, the same
parameters for the condenser are increased by the same factors.
Table 1 lists the nominal model parameters and the new rescaled parameters.
Remark 3. While the VCC model includes an EEV and variable
speed compressor, enabling the VCC to achieve a varying cooling
capacity, our current model does not capture the total cooling
range associated with either an experimental or an industrial
cooling unit. Specically, the operating conditions corresponding
to operating the VCC near its upper and lower cooling extremes
are not captured in the scaled VCC model, as inaccuracies arose
due to two factors: (1) the experimentally populated lookup
tables for certain component and thermodynamic parameters
(ZV , Zk , Cd, etc.) corresponded to different operating conditions
for the scaled and original VCC, resulting in a smaller feasible
operating range for the scaled system, and (2) the limiting nature
of the EEV caused the minimal operating conditions associated
with the compressor to be higher. Further reductions in the
compressor RPM cause the refrigerant mass ow rate to decrease,
however, the refrigerant mass ow rate will only converge to a
steady-state value as long as the static valve opening is able to
achieve the same decrease in ow. Eventually, the compressor
RPM will reach a value where the static valve opening will not be
able to reduce the refrigerant mass ow rate to the exit conditions
corresponding to the specic RPM value, causing the refrigerant
to never reach a steady-state value throughout the cycle, which
will eventually result in liquid owing into the compressor (i.e.,
Table 1
Nominal and re-scaled VCC model parameters to allow for a greater cooling
capacity.
Parameter

Nominal

Re-scaled

Units

_r
m
Vk
Le
Ae,o
Ae,i
Me
De
_ e,a
m
Lc
Ac,o
Ac,i
Mc
Dc

7.76  10  3
3.04  10  5
11.46
3.07
0.32
1.55
8.90  10  3
0.243
10.7
2.79
0.28
4.66
8.10  10  3

1.13  10  2
1.52  10  4
57.29
15.34
1.60
7.74
3.57  10  2
2.43
53.5
13.97
1.38
23.30
3.24  10  2

kg/s
m3
m
m2
m2
kg
m
kg/s
m
m2
m2
kg
m

evaporator superheat region going to zero). This factor is solely a


contribution of the choice of valve opening in the VCC and not
affected by the current VCC model used. Future work will explore
the potential benet of using a non-adjustable valve in the VCC.
2.2. Building model
The key disturbances in the VCC model are the ambient air
temperature (the air temperature at the condenser inlet) and
mixed air temperature (the air temperature at the evaporator
inlet). The ambient air temperature is naturally dictated by the
outdoor weather conditions while the mixed air temperature is
inuenced by a variety of interacting factors including the degree
of active heating/cooling in the room, the heating/cooling in
adjacent rooms (if any), and various heat load disturbances,
including the ambient air temperature. In this work, we utilize
the EnergyPlus simulation package to provide realistic mixed and
ambient air conditions based on a detailed building model and
actual weather data.
The building model in EnergyPlus accounts for building construction, surface geometries, and HVAC systems with the details
based on the U.S. Department of Energy reference small ofce
building model (U.S. Department of Energy). An important feature
of the EnergyPlus building model is that it accounts for the typical
daily variation of the internal gains in a building. Internal gains
capture heat variations caused by a variety of realistic heat loads
such as the movement of people and lighting schedules. As the air
in a building is exposed to these internal gains, varying amounts
of heat transfer occur from/to the air, causing uctuations in the
temperature and humidity of the zone temperature. This is
reected in the VCC unit as variations in the mixed air conditions.
Recall that the mixed air is a mixture of the zone temperature and
the ambient air temperature.
The EnergyPlus building model used in this work considers a
small (511 m2) single story ofce building in Chicago, Illinois, on a
typical July day. The building is assumed to be divided into ve
occupied thermal zones, in which there is a conditioned oor area
of 150 m2 in the core zone, 113 m2 in perimeter zones 1 and 3,
and 67 m2 in perimeter zones 2 and 4. The ground-to-ceiling
height in all zones is 3 m. In total, the building houses 28 people
at a standard occupant density of 5:38=100 m2 per zone. During
peak operation, the building is occupied between the hours of
8:00 and 18:00 with the highest levels of occupancy. In this work,
we assume that the thermal environment of perimeter zone 2 is
regulated by the detailed VCC model described in Section 2.1. All
remaining thermal zones are assumed to be controlled by
separate air-conditioning units (pre-)modeled in EnergyPlus,
and their zone temperatures are maintained at a constant setpoint temperature of 24 1C (to minimize inter-zone heat transfer).
Fig. 3 shows the ambient temperature, Tamb , relative humidity,
and the zone 2 internal gains over the course of the July day
considered for the building model. The ambient conditions are
obtained from historical data (in a data le) consisting of hourly
measurements of the temperature and relative humidity.
2.2.1. VCC-building model interface
To link the building model in EnergyPlus with the VCC unit
model in Matlab, data is exchanged between the two environments over sockets using the Building Controls Virtual Test Bed
(BCVTB) middle-ware (Wetter and Haves, 2008). This exchange is
accomplished using a Matlab script le (exchangeDoublewithSocket.m), which is included in the BCVTB library (see Fig. 4).
In the EnergyPlus client, the ambient air and the air of zone two
are mixed (80% zone air with 20% ambient air) to form the mixed
air conditions for the VCC. Ideally, the data exchange sequence

Ambient relative humidity (%)

Ambient temperature, Tamb (C)

M. Wallace et al. / Chemical Engineering Science 69 (2012) 4558

30
28
26
24

10

12
14
Time (h)

16

18

51

90

80

70

60
8

10

12
14
Time (h)

16

18

Internal Gains (W)

500

400

300

200

10

12
14
Time (h)

16

18

Fig. 3. Variations in the ambient temperature, relative humidity, and internal gains, which act as disturbances in the zone 2 EnergyPlus building model.

Read values

ExchangeDoublewithSocket.m

Cooling load

Matlab

BCVTB

Integrate
VCC model

Read values

Integrate
building model

Output
variables
EnergyPlus

Fig. 4. Schematic of the energy Plus-Matlab interface.

should be as follows: (1) subsequent to computing the mixed air


conditions, the EnergyPlus client is paused momentarily, (2) the
mixed and ambient air conditions are sent to Matlab, (3) in Matlab,
the VCC model is integrated (with computed input values from a
given control algorithm) and the corresponding cooling load is
computed, (4) the cooling load is sent to the EnergyPlus model, and
(5) the EnergyPlus model is un-paused and integrated forward
using the newly computed cooling load. However, one limitation of
the interfaced environment is that concurrent to sending the mixed
and ambient air conditions to Matlab, the BCVTB software requires
a cooling load from Matlab. That is, data is sent to and read from
Matlab simultaneously because there is no effective way to pause
the EnergyPlus model until the newly computed cooling load
(corresponding to the sent data) is computed. Instead, during the
simultaneous data exchange, the cooling load from the previous
time step is read and implemented in EnergyPlus, thereby introducing an input delay (of one sampling instant). To minimize the
effects of this delay, the fastest available sampling time of 60 s is
used for the EnergyPlus model.

The cooling load of a VCC quanties the heat absorption by the


refrigerant in the evaporator from the inlet air. The accurate
computation of the cooling load is essential for properly interfacing the VCC model. The total VCC cooling load is the sum of the
sensible and latent cooling loads, which correspond to changes in
the evaporator dry air temperature and humidity, respectively.
The sensible cooling load is equivalent to the rst term in Hloss in
Eq. (7). If the inlet air has a sufciently high water vapor content,
condensation may result, causing a humidity change and therefore a non-zero latent cooling load. If no condensation occurs, the
humidity ratio of the air does not change (as mentioned in Section
2.1.4); thus, the sensible cooling load equals the total cooling
load. The energy change associated with the humidity change (or
equivalently the condensation of the water vapor content in the
air) is the latent cooling load, X , and is given by
_ ,e Dhf
X m
_ ,e
where Dhf is the enthalpy of water condensation and m
represents the mass of condensed water and depends on the
supply air temperature, which can be computed using the
procedure described in Section 2.1.4.

3. Temperature control
In this section, we propose a temperature control framework
for regulating the air temperature of zone 2 in the EnergyPlus
building model (interfaced with Matlab). To this end, we rst
identify an auto-regressive exogenous (ARX) model for the VCC
outputs using simulation data. Next, we utilize the model in an
offset free predictive control design for the stand-alone VCC unit
and compare its performance against PI control. Finally, we
integrate the proposed predictive controller in a cascade control
structure for regulating the zone temperature and implement the
control structure on the interfaced building model.

52

M. Wallace et al. / Chemical Engineering Science 69 (2012) 4558

3.1. ARX VCC model


In the ARX type modeling approach, the process outputs at a
specic sampling instant are assumed to depend linearly on the
previous process conditions (dened by the process outputs and
inputs). Mathematically, ARX models are dened as
yk

ny
X
i1

Ai yki

nu
X

Bi uki

i1

nd
X

Ci dki vk

i1

where yk and uk are the process output and input vectors at


sampling instant k (respectively), dk is a vector of measurable
disturbances, Ai , Bi , and Ci are model coefcient matrices (that are
estimated using least-squares regression), vk is the noise vector,
and ny, nu, and nd denote the (maximum) number of time lags in
the outputs, inputs, and disturbances (respectively) and dene
the order of the ARX model. For specic outputs, inputs, or
disturbances which do not require the maximum number of lags,
the appropriate elements in the coefcient matrices can be set to
zero. For the VCC, the outputs, inputs, and measurable disturbances were previously dened in Section 2.1.4 as follows:
o T
i T
y Ts,e Ta,e
 , u ok vo T , and d Tamb Ta,e
 .
To identify the ARX model coefcient matrices, pseudo random binary sequences (PRBS) are generated for the inputs and
disturbances for the typical operating range (see Fig. 5 for a
portion of the PRBS data) and subsequently implemented on the
nonlinear stand-alone VCC model. Using the System Identication
Toolbox in Matlab (which essentially solves the linear regression
problem to compute the model coefcient matrices), the ARX
model coefcient matrices for numerous lag choices are estimated. Among these models, the lag choice representing a good
trade-off between the prediction accuracy and number of model
parameters is summarized in Table 2. Fig. 6 compares the output
prediction by the ARX model with the training data from the
nonlinear model, demonstrating the prediction capability of the
identied model.

Remark 4. A key objective of this work was to study the


applicability of a predictive control based scheme for temperature
control in the presence of realistic disturbances. In this work, we
opt for an empirically identied inputoutput VCC model as the
predictive model in the control design instead of a linearized
state-space model (coupled with a state estimator). In general, a
linearized state-space model of a nonlinear system at a specic
operating point only captures the local dynamics around the
linearization point and therefore calls for successive linearization
techniques when used in an MPC framework to maintain reliable
predictions. Another limitation of using a rst-principles (deterministic) model as the foundation of the control design is that the
models reliability is subject to the accuracy of numerous physical
parameters (i.e., thermodynamic properties, etc.), which may not
be known accurately. Additionally, many of the simplifying
assumptions made during the model development can be violated
in practice, further decreasing its validity. These reliability issues
together with the inherent error introduced by linearizing a
nonlinear model motivated the use of an empirical model for this
work. From an industrial perspective, if a sufciently large
number of identical packaged units are produced, it may make
economic sense to invest in the effort to develop a dedicated rst
principles model, or alternatively, generation of enough data to
capture the model characteristics in a data-driven model.

Table 2
Final ARX model lag structure.
Output

Lags

Ts,e
o
Ta,e

Ts,e

o
Ta,e

ok

vo

Tamb

i
Ta,e

2
1

2
1

2
1

2
1

2
1

2
1

103

Valve opening, vo (%)

12.5

RPM, k

1.5

12

11.5

11
10

20

30
Time (h)

40

50

Ambient air temperature, Tamb (C)

Mixed air temperature, Tia,e (C)

26
24
22
20

10

20

30
Time (h)

40

50

10

20
30
Time (h)

10

30
20
Time (h)

40

50

25

20

40

Fig. 5. Portion of the input proles used to generate output data for ARX model identication.

50

Nonlinear model
ARX model

Superheat, Ts,e (C)

25

20

15

10
0

10

20
30
Time (h)

40

50

Supply air temperature, T oa,e (C)

M. Wallace et al. / Chemical Engineering Science 69 (2012) 4558

53

Nonlinear model
ARX model

22
20
18
16
14
0

10

20
30
Time (h)

40

50

Fig. 6. Comparison of the output prediction by the ARX model with the nonlinear model for the input and disturbance proles in Fig. 5.

Remark 5. Using an ARX model with a measurable disturbance


vector as one of the predictors in a predictive control design
effectively incorporates an element of feed-forward control into
the design. That is, the control algorithm utilizes the measured
disturbance vector to anticipate its effect and takes corrective
action and further improve upon the achieved energy efciency.
However, for an MPC design with a prediction horizon greater
than one, future disturbance measurements are required to make
predictions over the horizon. In this case, the current measurements of the disturbances can be assumed to hold for the length
of horizon. This is a common assumption used in MPC formulations that utilize disturbance measurements and is meaningful in
the present context due to the different time scales at which the
VCC evolves (small time scale) and the disturbance variables
change (larger time scale).

energy efciency by minimizing the compressor energy consumption (the largest energy consumer in the VCC). The closed-loop
performance is evaluated in terms of the integral of squared error,
o
ISESA , between the supply air temperature, Ta,e
, and its set-point
o
trajectory, Ta,e,SP
:

Remark 6. The estimated ARX model predicts the VCC output


behavior relatively well; however, the nonlinear nature of the
process dynamics (i.e., varying process gains) cannot be fully
captured using a single linear model. One approach to capture
this nonlinearity is to identify multiple local linear models at
various operating points and combine them with an appropriate
weighting function during prediction. Recently, in Aumi and
Mhaskar (in press), a data-driven modeling methodology was
proposed that unies the concepts of ARX modeling, latent
variable regression techniques, fuzzy c-means clustering, and
multiple local linear models in an integrated framework capable
of capturing process nonlinearities. Specically, plant data is rst
clustered using fuzzy c-means clustering to identify the most
suitable points for linearization and come up with a continuous
weighting function for the individual models. Using this weighting function, the local linear model coefcients are simultaneously
estimated using latent variable regression tools, which allow for
dimensionality and noise reduction. The same weighting function
is then utilized to weight the individual models given an initial
condition and inputs. The proposed modeling methodology has
been shown to be effective for identifying accurate models for use
in MPC formulations (Aumi and Mhaskar, in press; Aumi et al.,
submitted), and represents one possibility for developing
improved data-based models for use in the control design.

TEC Dt

ISESA Dt

K
X

o
o
Ta,e,SP
iTa,e
i2

i1

where i indexes the sampling instant, Dt is the sampling period


(60 s), and K is the total number of sampling instants in the
simulation. To quantify the energy demand associated with a
control design, the instantaneous compressor power is summed
over the simulation time, yielding a measure of the total energy
consumption, TEC (see Section 2.1.1 for the variable denitions):
K
X
_ k ihok ihik i
m

i1

where Z is the combined total efciency of the compressor, which


is the product of the power and the mechanical efciencies (known
parameters).
3.2.1. MPC control design and implementation
Consider a VCC system for which the ARX model for its outputs
has been computed. For the proposed predictive control design,
the inputs to the VCC at sampling instant i are computed by
solving the following constrained optimization problem:
P
X

min

umin r uk r umax

subject to :
^
yk

Jy^ 2 ky2,SP kJQ Ju1 Jr JDuJR

k1

ny
X
i1

Dumin r Duk r Dumax


^

Ai yki

nu
X

Bi uki

i1

nd
X

Ci dki

for k 1, . . . ,P

i1

n
^ a bi
y^ k yk
n

y1,min r y^ 1 k ry1,max
^
a ky0y0

b1 i b1 i1 g1 maxf0,y1,min y1 0g g2 maxf0,y1 0y1,max g

b2 i b2 i1 f y2 0y2,SP 0

3.2. Stand-alone VCC control


In this section, we design and implement a predictive controller
on the nonlinear VCC model using the model identied in the
previous section and compare the closed-loop simulation results
against PI control. The control objectives for stand-alone VCC
control are to track a given set-point trajectory of the supply air
temperature, to maintain reliable/safe operation by maintaining
the superheat above 0 1C (see Section 2.1.4), and to maximize the

where the notation, J  JQ , refers to the weighted norm, dened by


JxJQ xT Q x and Du denotes a vector in which each element is the
difference between successive input moves. The weighting
matrices are diagonal and used to trade-off the relative importance of the different control objectives. The plant measurement
at the current sampling instant i corresponds to k0 or y0.
In this MPC formulation, the control objective of supply air
temperature set-point tracking is addressed by penalizing the

M. Wallace et al. / Chemical Engineering Science 69 (2012) 4558

deviation between the predicted supply-air temperature from its


set-point, y2,SP k, weighted by Q . The predicted superheat is also
bounded between y1,min and y1,max . To reduce the energy consumption associated with the control action, the absolute value of RPM is
also penalized using the weight r. The inputs are constrained in a
range for which the nonlinear VCC model is known to be valid. In
addition to using hard constraints for the input rates, excessive
input movements are penalized in the objective function using a
move suppression factor with the weighting matrix, R. When
tuning the different weighting matrices, the highest importance
was initially given to tracking the supply air set-point. Subsequently, the remaining weighting matrices were adjusted appropriately to achieve relatively smooth input behavior.
To achieve offset free performance, a disturbance/bias term is
added to the model predictions that is expressed by combining
two constant terms, a and bi. The rst term, a, is the disturbance
due to plant-model mismatch at the current sampling instant,
multiplied by a tuning parameter, k. Specically, a is dened as
the difference between the predicted outputs at sampling instant
i from the previous control calculation and the measured outputs
at i. The bi term is the summation of tracking errors up to and
including sampling instant i. For the superheat (output 1 or y1), a
non-zero tracking error at i is used only if the current measurement exceeds the minimum or maximum superheat. The b term
essentially persists and inuences the control action until the
offset is eliminated. It can be understood as operating the same
way as the integral mode in a PI controller. The tuning parameters, g1, g2, and f, are used to trade-off the input aggressiveness
and the amount of offset. The list of tuning parameters which
resulted in offset free performance while maintaining relatively
smooth input behavior is tabulated in Table 3 along with the
constraint bounds. Fig. 7 demonstrates the effect of the a and b. In
the nominal case (no corrections), there is considerable offset in
the supply air temperature. After adding the feedback term to
account for plant-model mismatch, this offset is signicantly
reduced but not eliminated. Zero offset is only achieved after
also including the error accumulation term in the formulation.
Next, closed-loop simulation results for MPC and PI control are
compared. For these simulations, constant disturbances are
assumed. That is, the ambient air conditions (temperature and
humidity) and the inlet air temperature to the evaporator (the
mixed air temperature) are maintained at constant values. Using
the results in Keir and Alleyne (2007), for the PI loop pairing, the
supply air temperature is paired with the compressor RPM while
the superheat is paired with the expansion valve opening. The
superheat set-point for the PI controller is specied to be 10 1C
(see Remark 7). The PI controllers are initially tuned using the
internal model control tuning method and ne-tuned to minimize
the integral of absolute error while maintaining relatively smooth
input trajectories.

Table 3
MPC tuning parameters.
Parameter

Value

P
Q
R
r
fy1,min ,y1,max g
fumin ,umax g
fDumin , Dumax g

4
950
diag{0.004,0.5}
350/17002
{3.5, 20}
{[678.8 6]T, [1700 15]T}

k
fg1 ,g2 g
f

0:2 0:50T
{6, 0.3}
0.01

f200 1T ,200 1T g

23.5

Supply air temperature, T oa,e (C)

54

23.4

23.3

23.2

23.1
SP
Nominal

 +

23
0

4
Time (h)

Fig. 7. Supply air temperature responses using various combinations of the bias
terms in the proposed MPC design.

Fig. 8 displays the closed-loop VCC input and output variable


responses for the two control strategies and Table 4 summarizes
their control performances using the metrics previously discussed
sett
and also the settling times for the supply air temperature, tSA
, for
the different set-point step changes. As shown in Fig. 8, the
proposed MPC design is able to provide better tracking performance
of the supply air temperature for the different set-point changes
with similar settling times and lower energy consumption. The
third supply air set-point change (to approximately 23.7 1C) is an
infeasible set-point for the VCC cooling capacity, but note that the
predictive controller is able to drive the supply air temperature
closer to this set-point compared to the PI controller. Note, however, that the infeasibility is merely a result of the model not being
valid at low RPM (or as low as required) to provide less cooling.
For the MPC design, the superheat is permitted to oat
between its minimum and maximum value whereas for PI
control, the superheat is maintained at the constant safety margin
of 10 1C. This additional degree of freedom for the predictive
controller leads to more accurate tracking and better overall
control performance. Note that if the superheat was prescribed
to be maintained at a constant value of 10 1C for the MPC design
as well, the corresponding closed-loop results would be similar to
those obtained when using the PI controller. In regard to the
energy efciency, the MPC design required 8% less energy
compared to the PI controller. This is a consequence of using
higher valve openings and lower RPM values resulting from the
multivariable nature of the MPC controller and the ability to allow
the superheat to oat between acceptable values.
Remark 7. For the PI closed-loop simulation, the safety margin for
the superheat is specied to be 10 1C. This represented a rough
lower bound for the superheat set-point for reliable simulations
under PI control. When the VCC model is interfaced with the
building model and the superheat set-point is prescribed to be less
than 10 1C, the PI controller drives the superheat to a negative
value, resulting in a failed simulation. Note that in practice, a VCC
unit has protections to ensure against negative superheat values;
however, such protections are not considered in the existing VCC
model. Simulation studies also revealed that by increasing the
superheat set-point to 20 1C, the supply air tracking performance
can be substantially improved. However, maintaining the superheat
at a higher safety margin requires lower valve openings, which, in
turn, results in the PI controller prescribing higher RPM values to

MPC
PI

20
15
10
5
0

1.2

10
Time (h)

23.6
23.4
23.2
SP
MPC
PI

23
0

10
Time (h)

15

103
MPC
PI

1.1
RPM, k

55

23.8

15

1
0.9
0.8
0.7

Valve opening, vo (%)

Superheat, Ts,e (C)

25

Supply air temperature, T oa,e (C)

M. Wallace et al. / Chemical Engineering Science 69 (2012) 4558

14
12
10
8
MPC
PI

6
0

10
Time (h)

15

10
Time (h)

15

Fig. 8. Closed-loop output and input proles for the VCC under MPC and PI control.

Table 4
Stand-alone VCC closed-loop performance metrics.
Metric

Control strategy
PI control
2

ISESA (s 1C )
sett
tSA
(s)
TEC (kJ)

MPC

837

222

1800, 1800, 900, 1620

1020, 1800, 1140, 4440

10017

9217

track the supply air temperature. Thus, there is a trade-off between


the improved tracking performance and increased power consumption. It is also worth noting that the MPC design still offered better
tracking performance (in addition to being more energy efcient by
23%) than the PI controller by 10% even with the increased superheat set-point of 20 1C.
Remark 8. Note that if the model is allowed to operate over the
entire range of RPM, it is possible that the PI controller could be
used to keep the superheat value at a xed, low set-point, this
would result in the RPM being able to change over the entire
range to provide minimum cooling where required, and at other
times, providing additional cooling using minimal RPM. In such a
scenario, the energy efciency of a PI control structure would be
comparable with the MPC in the current form. In such a scenario,
however, possible nonlinear (and more importantly, non-monotonic) dependence of energy efciency on the RPM would (and
could) be incorporated within the MPC controller to provide more
efcient operation over conventional control structures.
3.3. Energy efcient temperature control framework
In this section, we integrate the proposed MPC design for
stand-alone VCC control in a cascade control structure for energy
efcient temperature control in zone 2 of the EnergyPlus building

model. Note that the main purpose of the interfacing is to


demonstrate superior control of the cooling device subject to
realistic disturbances (induced by the interfacing and use of
weather data). The primary control objective we consider is to
maintain the zone 2 temperature, Tzone , within acceptable comfort
standards in the presence of disturbances brought on by varying
ambient conditions, changes in the internal gains, and zone
interactions (see Section 2.2). The secondary control objectives
are the stand-alone VCC control objectives listed in Section 3.2.
The comfort standards we consider are inspired by those used by
the American Society of Heating, Refrigeration and Air-Conditioning Engineers (ASHRAE). For typical summer conditions, assuming
that the room occupants are wearing light clothing, the ASHRAE
comfort standards (ASHRAE 55-2004) entail maintaining the zone
temperature between 22.3 and 24.7 1C and for this work, a zone
temperature set-point, Tzone,SP , of 24 1C is selected. Another
important ASHRAE comfort standard is that the zone temperature
not drift, which is dened as the temperature violating a band
around the set-point for longer than 15 consecutive minutes. In
this work, a 70.5 1C band around the set-point is considered.
The proposed control structure for meeting the temperature
control objectives is shown in Fig. 9. The cascade control structure
was motivated by the time-scale of the VCC dynamics compared
to the zone temperature dynamics. Step tests in the VCC inputs
(compressor RPM and valve opening) revealed the supply air
and superheat temperatures evolve roughly in the same time
scale (110 min) whereas the zone air temperature dynamics
were signicantly slower (nearly 50 min). Varying internal gains
and ambient conditions act as disturbances to the zone air
temperature; however, by using a cascade control structure, the
relatively faster dynamics of the inner loop are exploited to
eliminate these disturbances (using the VCC inputs) before they
signicantly affect the zone temperature.
In this cascade control structure, the inner loop consists of a
stand-alone VCC controller (either the predictive controller or PI
controllers designed in Section 3.2.1). The outer loop is used to

56

M. Wallace et al. / Chemical Engineering Science 69 (2012) 4558

o
Ts,e ,T a,e

Disturbances

VCC
k
Tzone,SP +

Compressor

T oa,e,SP
PI

MPC/PI

vo

EnergyPlus
Model

Evaporator

Condenser

Tzone

Valve

Fig. 9. Proposed cascade control structure for energy efcient temperature control.

Remark 9. One natural extension of the proposed control structure is to replace the outer PI loop with a model predictive
controller. The main requirement for this extension is to identify a
model between the supply air temperature and the zone air

24.6
24.4
Zone air temperature (C)

regulate the zone temperature and is connected to the inner loop


via the supply air temperature set-point. Based on the error
between the zone temperature and its set-point (Tzone,SP
o
24 1C), the outer loop prescribes a supply air set-point, Ta,e,SP
, to
the inner loop controller. The zone temperature is sampled every
15 min, which also corresponds to the frequency of the supply air
set-point updates. Faster sampling times led to excessive uctuations in the prescribed supply air set-point, resulting in poor
tracking performance by the inner loop controller. The outer loop
was tuned iteratively such that it yielded trackable supply air setpoints by the VCC. The outer loop tunings were kept consistent for
both control strategies in the inner loop.
The zone temperature response for each control strategy is
shown in Fig. 10, which compares the efcacy of each control
strategy in meeting the primary control objective. The MPC-based
strategy is able to satisfy the comfort standards for the entire test
period (any zone set-point violations lasted less than 15 min)
while the comfort standards are violated for approximately the
last 40 min of the test period when using the PI-based controller.
From Fig. 3, after 16:00 or 4:00 P.M., there is a signicant decrease
in the internal gains owing to a decrease in the zone occupancy
and also a decrease in the ambient temperature. The zone
temperature response after 4:00 P.M. indicates that the MPCbased design is able to respond to these disturbances more
effectively than the PI-based control strategy.
With regard to the secondary control objectives, Fig. 11 displays the closed-loop VCC input and output proles for the two
inner loop control strategies. The performance metrics of the
inner loops are shown in Table 5. Similar to the results in the
stand-alone VCC case, for the MPC-based design, the superheat is
allowed to oat between its bounds and ended up evolving
closer to its lower bound, allowing for better supply air temperature tracking using considerably less compressor power. As
shown in the supply air proles in Fig. 11, in contrast to the PIbased design, the MPC-based controller provides an offset free
supply air temperature prole for the majority of the feasible setpoint values prescribed by the outer loop controller. To achieve
this offset free performance (in addition to improved zone air
temperature regulation), aggressive control action is prescribed.
We note, however, that the key idea in the results with the
interfaced system is not so much to demonstrate improved
control of the zone conditions (which depends on several factors,
including the outer loop) but more to evaluate the performance
of the VCC control structure subject to realistic disturbances.

24.2
24
23.8
23.6
23.4
MPC
PI

23.2
8

10

11

12

13
14
Time (h)

15

16

17

18

Fig. 10. Air temperature in zone 2 when using a cascade control structure for
temperature control with a PI or predictive controller in the inner loop.

temperature. This can be identied through step-tests or more


desirably, by generating PRBS-like sequences of the supply air.
However, in any case, the resulting model will be dependent on
the closed-loop dynamics of the stand-alone VCC controller. Due
to the large variation in the time scales between the zone and VCC
dynamics, in addition to the zone air being affected through a
single VCC output variable (supply air temperature), it is advantageous to use separate MPC designs for each level of the cascade
rather than using one model predictive controller to regulate the
zone conditions. As discussed in Remark 5, a weather model/
estimator can also be incorporated in the design by including an
ambient temperature component in the model. In this case,
through an economic objective function that considers varying
electricity costs for the outer loop controller, operating costs
can be reduced by pre-cooling as necessary. In addition to
optimality, the benets of using MPC in the outer loop include
explicitly incorporating comfort specications and accounting
for the VCC cooling capacity in computing the supply air setpoints. Finally, while we use temperature as the comfort measure
in this work, other measures of comfort, such as a Predicted Mean
Vote (Federspiel and Asada, 1994; Hanna, 1997; Brager and de
Dear, 1998; Jones, 2002; Ye et al., 2003; Baus et al., 2008) can be
readily incorporated in the objective function in the MPC
formulation.

Supply air temperature, T oa,e (C)

Supply air temperature, T oa,e (C)

M. Wallace et al. / Chemical Engineering Science 69 (2012) 4558

SP
MPC
24
23
22
21
8

10

12
14
Time (h)

16

57

SP
PI

25
24
23
22
21

18

10

12
14
Time (h)

16

18

103
MPC
PI

MPC
PI

1.2

20

RPM, k

Superheat, Ts,e (C)

25

15

10
0.8
5
10

12
14
Time (h)

Valve opening, vo (%)

16

18

10

12
14
Time (h)

16

18

14
12
10
8
MPC
PI

6
8

10

12
14
Time (h)

16

18

Fig. 11. Closed-loop output and input proles for the VCC when interfaced with the EnergyPlus building model under MPC and PI control.

Table 5
Inner loop performance metrics in the cascade control structure.
Metric

ISESA (s 1C2 )
TEC (kJ)

Control strategy
PI control

MPC

70 978

20 987

5080

4284

4. Conclusions
In this work, control strategies were implemented on a
realistic building model interfaced with a detailed VCC model
which provided the cooling load for a specied zone in the
building. The detailed VCC model was subject to realistic disturbances in the ambient and mixed air conditions as a result of
interfacing the two models. The zone air temperature in the
building model was also subject to realistic ambient air conditions (obtained from actual weather data) and typical internal
load variations. A cascade control structure was proposed to
regulate the zone air conditions in the building model. In the

proposed control structure, an outer loop regulates the zone air


temperature by adjusting the set-point of the VCC supply air
temperature using PI control. The inner loop regulates the VCC
supply air and superheat by manipulating the compressor RPM
and valve opening using an ARX-model-based predictive controller. The proposed control strategy demonstrated better disturbance rejection ability in the zone air temperature than a PI-based
cascade structure. Also the predictive control strategy was more
energy efcient. The improved performance of the MPC-based
design stemmed from its incorporation of the coupled nature of
VCC dynamics (through the ARX model) in computing the control
action. The ability in the MPC to trade-off optimality considerations with tracking requirements enabled achieving improved
set-point tracking while operating at lower RPM values (where
possible) resulting in better energy efciency.

Notation
Variable
V

r
Z

Description
volume
density
efciency

58

o
t

M. Wallace et al. / Chemical Engineering Science 69 (2012) 4558

T
x
u
d
vo
cp
L
w
D
M
t

compressor RPM
time constant
specic enthalpy
pressure drop
discharge coefcient
volumetric mean void fraction
mass ow rate
cross-sectional area
pressure
perimeter
heat transfer coefcient
temperature
state variable
input variable
disturbance variable
valve opening
specic heat capacity
length
humidity ratio
diameter
mass
time

Subscript

Description

k
v
h
e
c
f

r
w
a
amb
W
sat
dp
V
s
SA
SP

compressor
valve
heat exchanger
evaporator
condenser
vapor
liquid
refrigerant
wall
air
ambient
water
saturation
dew point
volumetric
superheat
supply air
set-point

Superscript

Description

i
o

inlet
outlet

DP
Cd

g
_
m
A
P
p

Acknowledgments
Financial support from the Natural Sciences and Engineering
Research Council of Canada through the Collaborative Research
and Development Program (in collaboration with Johnson Controls Inc.) is gratefully acknowledged.
References
Albieri, M., Beghi, A., Bodo, C., Cecchinato, L., 2009. Advanced control systems for
single compressor chiller units. Int. J. Refrig. 32 (5), 10681076.

American Society of Heating, Refrigeration and Air-Conditioning Engineers, Inc.,


ASHRAE Handbook-Heating, Ventillating and Air-Conditioning Systems and
Equipment, ASHRAE.
Aumi, S., Mhaskar, P. Integrating data-based modeling and nonlinear control tools
for batch process control. AIChE J., in press. doi:10.1002/aic.12720.
Aumi, S., Corbett, B., Mhaskar, P., Clarke-Pringle, T. Data-based modeling and
control of nylon-6,6 batch polymerization. IEEE Trans. Control Syst. Technol.,
submitted for publication.
Baus, Z.L., Nikolovski, S.N., Maric, P.Z., 2008. Process control for thermal comfort
maintenance using fuzzy logic. J. Elec. Eng. 59 (1), 3439.
Behidj, N., Brugger, M., Demers, D., Kowal, A., Liu, Y., Warbanski, M., Yamada, F.,
2009. Energy Efciency Trends in Canada, Technical Report, Natural Resources
Canada (09 2009). URL /http://oee.nrcan.gc.ca/publications/statistics/
trends09/S.
Brager, G.S., de Dear, R.J., 1998. Thermal adaptation in the built environment: a
literature review. Energy Build. 27 (1), 8396.
Dincer, I., Rosen, M., 2007. Exergy: Energy, Environment and Sustainable Development. Elsevier, Oxford, UK.
Federspiel, C.C., Asada, H., 1994. User-adaptable comfort control for HVAC
systems. J. Dyn. Syst. Meas. Control Trans. ASME 116 (3), 474486.
Hanna, R., 1997. Relationship between thermal comfort and user satisfaction in
hot dry climates. Renew Energy 10 (4), 559568.
Huang, G., Wang, S., Xu, X., 2009. A robust model predictive control strategy for
improving the control performance of air-conditioning systems. Energy Convers. Manage. 50, 26502658.
Jain, N., Li, B., Keir, M., Hencey, B., Alleyne, A., 2010. Decentralized feedback
structures of a vapor compression cycle system. IEEE Trans. Control Syst.
Technol. 18, 185193.
Jones, B.W., 2002. Capabilities and limitations of thermal models for use in
thermal comfort standards. Energy Build. 34, 653659.
Keir, M., Alleyne, A., 2007. Feedback structures for vapor compression cycle
systems. In: Proceedings of the American Control Conference (ACC),
pp. 50525058.
Leducq, D., Guilpart, J., Trystram, G., 2006. Non-linear predictive control of a
vapour compression cycle. Int. J. Refrig. 29, 761772.
Lin, J., Yeh, T., 2007. Modeling, identication and control of air-conditioning
systems. Int. J. Refrig. 30, 209220.
Ma, J., Qin, J.S., Salsbury, T., 2010. Real-time model predictive control for energy
and demand optimization of multi-zone buildings. In: Proceedings of the
AIChE Annual Conference.
Ma, Y., Borrelli, F., Hencey, B., Coffey, B., Bengea, S., Haves, P., 2010. Model
predictive control for the operation of building cooling systems. In: Proceedings of the American Control Conference (ACC), pp. 51065111.
Mhaskar, P., 2006. Robust model predictive control design for fault-tolerant
control of process systems. Ind. Eng. Chem. Res. 45, 85658574.
Mhaskar, P., El-Farra, N.H., Christodes, P.D., 2005. Predictive control of switched
nonlinear systems with scheduled mode transitions. IEEE Trans. Automat.
Control 50, 16701680.
Mhaskar, P., El-Farra, N.H., Christodes, P.D., 2006. Stabilization of nonlinear
systems with state and control constraints using Lyapunov-based predictive
control. Syst. Contr. Lett. 55, 650659.
Morosan, P., Bourdais, R., Dumur, D., Buisson, J., 2010. Distributed model
predictive control for building temperature regulation. In: Proceedings of the
American Control Conference (ACC), pp. 31743179.
Rasmussen, B., 2005. Dynamic Modeling and Advanced Control of Air Conditioning
and Refrigeration System. Ph.D. Thesis, University of Illinois at Urbana-Champaign.
Sandipan, M., Alleyne, A., Chandan, V., 2010. Predictive control of complex
hydronic systems. In: Proceedings of the American Control Conference (ACC),
pp. 51125117.
Sarabia, D., Capraro, F., Larsen, L., Prada, C., 2007. Hybrid control of a supermarket
refrigeration system. In: Proceedings of the American Control Conference
(ACC), pp. 41784185.
Schurt, L., Hermes, C., Neto, A., 2009. A model-driven multivariable controller for
vapor compression refrigeration systems. Int. J. Refrig. 32, 16721682.
U.S. Department of Energy, EnergyPlus Building Software, Building Technologies
Program. URL /http://apps1.eere.energy.gov/buildings/energyplus/S.
Wetter, M., Haves, P., 2008. A modular building controls virtual test bed for the
integration of heterogeneous systems. In: Third National Conference of IBPSAUSA, pp. 6976.
Xi, X., Poo, A., Chou, S., 2007. Support vector regression model predictive control
on a HVAC plant. Control Eng. Pract. 15, 897908.
Ye, G.D., Yang, C.Z., Chen, Y.M., Li, Y.G., 2003. A new approach for measuring
predicted mean vote ccPMVcc and standard effective temperature. Build.
Environ. 38 (1), 3344.
Zhu, G., Henson, M., Megan, L., 2001. Dynamic modeling and linear model
predictive control of gas pipeline networks. J. Process Control 11, 129148.

You might also like