Professional Documents
Culture Documents
Abstract
Keywords: HVAC systems; Multi-zone building; Temperature control; Robust decentralized controllers;
H 1 -performance specications; Constrained optimization
0016-0032/$30.00 r 2004 The Franklin Institute. Published by Elsevier Ltd. All rights reserved.
doi:10.1016/j.jfranklin.2004.06.001
ARTICLE IN PRESS
544 S.A.K. Al-Assadi et al. / Journal of the Franklin Institute 341 (2004) 543567
1. Introduction
Fig. 2. (ac) Comparison of the open-loop responses for the linear and bilinear system to 1 C step
change in the operating point and DT e 1 C, DT P 4 C in the external disturbances.
ARTICLE IN PRESS
548 S.A.K. Al-Assadi et al. / Journal of the Franklin Institute 341 (2004) 543567
stations in the decentralized system, and for Station i, Dui t 2 Rmi is the control
input vector, Dyi t 2 Rl i is the measurable output vector and DyRi t 2 Rl i is the
reference signal (set-point) vector. h i h i
Let B B1 . . . BN , C T C T1 . . . C TN , Du1 Dn3 , Du2 Dn
Dn4
1 , Du Dn2
3 Dn5
and
let DyR tT DyR1 tT . . . DyRN tT 0, i.e., we consider a regulation problem for
the linearized model.
In order to reject a class of disturbances Dwt and track reference signals DyRi t
acting on the system (3.1), we use a general servocompensator [7,8] given by
Dx_ i t Oi Dxi t Yi Dei t; 3:2
where
Oi 2 Rql i ql i block diagOi ; Oi ; . . . ; Oi
and
Yi 2 Rql i l i block diagYi ; Yi ; . . . ; Yi
with q, Yi and Oi completely determined by the disturbances and reference signals
acting on the system [9] e.g., for constant disturbances and reference signals, q 1,
Oi 0, Yi 1. Thus the augmented decentralized system is
X
N
_^ AD
Dxt ^ xt
^ B^ i Dui t EDwt
^ F^ DyR t; 3:3a
i1
Dyi t
D^yi t C^ i Dxt
^ 3:3b
Dxi t
and the output to be regulated is
D^zi t C i Dxt Dyi t; i 1; . . . ; N; 3:3c
where
" # " # " #
Dxt A 0 Bi
^
Dxt ; A^ ; B^ i ;
Dxt Y C O 0
" # " # " #
E 0 Ci 0
E^ ; F^ and C^ i
0 Y 0 Ip
with O block diagO1 ; . . . ; ON
and Y block diagY1 ; . . . ; YN .
For the robust decentralized servomechanism problem given by (3.3), the task is to
nd N local controllers with inputs D^yi t, and outputs Dui t; i 1; . . . ; N, so that
(i) the resulting closed-loop system is asymptotically stable, (ii) the outputs D^zi t
asymptotically track the reference signals for all constant disturbances and set-
points, and (iii) for all perturbations in the parameters of the system which do not
cause instability in the closed-loop system, property (ii) still holds.
If we dene a set K of block diagonal matrices:
K fK j K blockdiagK 1 ; . . . ; K N ; K i 2 Rmi l i g 3:4
ARTICLE IN PRESS
550 S.A.K. Al-Assadi et al. / Journal of the Franklin Institute 341 (2004) 543567
then the set of decentralized xed modes (d.f.m.s) of (3.1) with respect to K is
dened [6,10] as
\
LA; Bi ; C i ; K sA BKC; 3:5
K2K
where Ki K 1
i K 2
i .
The resulting decentralized control inputs, ui t ui0 t Dui t; i 1; 2; 3, are
the actual control inputs nk t, to the plant, k 1; . . . ; 5. The control inputs nk t are
normalized with respect to their maximum capacities, and vary between 0 and 1.
Since any violation of this constraint is not realistic, we seek a set of Ki by optimizing
an H 1 performance criterion. By incorporating performance specications and
robust stability within an H 1 optimization criterion, robust performance of the
HVAC system can be achieved. The specications of the H 1 -based performance
measures are discussed next.
considered is
supfkW 1 Sdk2 : kdk2 p1g kW 1 Sk1 : 3:7
It is desired to minimize kW 1 Sk1 over the set of all constant decentralized
stabilizing controllers Ki . This is also related to shaping the frequency response
characteristic of the decentralized sensitivity Ss via an appropriate choice of the
weighting function W 1 s. For example, a specication of the form
jSjojojW 1 joj1 for all o 2 R
is equivalent to the H 1 norm specication kW 1 Sk1 o1.
Modeling error (as discussed in Section 2) in the system may be represented in the
~ I dsW 2 sPs. Here, Ps represents the transfer function matrix
form Ps
of the nominal HVAC linearized model, W 2 s and ds are the parametrization
transfer function matrices of the uncertainty in the model. ds is a variable, stable
transfer function matrix satisfying kdk1 p1. In this study the uncertainty weighting
function W 2 s is assumed to be a known diagonal matrix with equal constant gains
s2 in the diagonal entries. This uncertainty modeling accounts for the phase
uncertainty and acts as a scaling factor on the magnitude of the perturbations.
Robust stability is achieved by putting an H 1 -norm constraint on the transfer
function T from the disturbance Dwt to the control input Dui t, i.e.
kW 2 Tk1 o1; 3:8
where, T represents the decentralized complementary sensitivity function to be
minimized. This minimization corresponds to reducing the energy in the control
input and also to improving the robustness of the system against any multiplicative
perturbations in the HVAC model.
In order to ensure that the controller Ki designed based on the linear model (3.1)
results in a stable bilinear closed-loop system and the input energy of the system does
not exceed its maximum capacity, we introduce additional constraints
0pnk tp1; k 1; . . . ; 5 3:9c
and
x0j t 1:5oxj tpx0j t 1:5; j 1; . . . ; 7 3:9d
where x0 t T b0 T l10 T h10 T z10 T l20 T h20 T z20 T are the temperatures at the
operating point, nk t; k 1; . . . ; 5, are the closed-loop control inputs and xj t are
the actual states of HVAC system (A.1A.9). Note that the nk s are normalized with
respect to their maximum capacities so that they vary between 0 and 1. The
constraint (3.9d) ensures that the actual system states remain bounded and close to
the operating point x0 t.
In order to normalize the H 1 -norm criterion given in Eq. (3.9a) to be less than 1,
we need to choose suitable weighting functions W 1 s and W 2 s.
Remark 1. In the standard H 1 optimization problem, the weighting functions are
used to capture the structure of the disturbances and plant uncertainty. Thus, the
weighting functions inuence the structure of the controller. In this paper, we imbed
the problem of constraint optimization of an H 1 criterion in robust servomechan-
ism formulation. The servocompensator captures the characteristics of the
disturbance and reference signals [5] and is embedded in the decentralized control
law (3.6) as a xed part. The remaining part consists of a constant gain Ki designed
via H 1 optimization problem with additional constraints. For this optimization
problem, a choice of constant weighting functions W 1 s and W 2 s is considered to
be suitable.
Remark 2. Another difference from the standard H 1 optimization problem is that
there is no direct transmission from Dwt to Dui t and D^zi t in the augmented
decentralized system (3.3). Consequently, many of the currently available computing
methods for H 1 design (such as the MATLAB mAnalysis and Synthesis Toolbox
given by Balas et al. [13]) cannot be used directly for solving this constrained H 1
optimization problem.
Remark 3. It should be mentioned that the technique used in this paper for designing
controllers for HVAC systems differs from the conventional H 1 methods, and is
considered advantageous for the following reasons: (i) Use of conventional H 1
design for this problem would give rise to a feedback controller of order equal to the
order of the HVAC system plus the order of the servocompensator; our technique,
on the other hand employs a servocompensator and a constant gain feedback
controller. This feature is considered to be of signicance from the viewpoint of
practical applications. (ii) The well posedness requirements for our problem, which
guarantee the existence of a solution, are signicantly weaker than those required in
the so called standard H 1 design methods , e.g. see [13]. This makes it easier for us
to solve the optimization problem. No additional weighting functions are needed to
ARTICLE IN PRESS
S.A.K. Al-Assadi et al. / Journal of the Franklin Institute 341 (2004) 543567 553
4. Simulation results
from Fig. 3a that the minimization of the H 1 criterion results in faster response.
Consequently, the boiler temperature reaches the setpoint more quickly. Likewise
Figs. 3b and c show that the output responses obtained with the H 1 criterion are in
general better than those obtained with the quadratic performance criterion.
The control input proles corresponding to the outputs (Figs. 3ac) are shown in
Figs. 4ae. Note that the boiler input energy n3 remains the same (Fig. 4c), but
better temperature responses (Fig. 3a) are achieved. Furthermore, heat pump-1
energy is minimized (Fig. 4d) at the expense of slightly higher air ow rate to zone-1
n1 which is intuitively a smart strategy because running the heat pump with
higher capacity is more expensive than running the fan at higher speed to deliver
higher air ow rate. Since, both options theoretically yield the same heat ow rate to
offset the effect of the disturbance on zone-1, it is indeed worth noting that the
H 1 -based performance measure chose the cost effective alternative. This result
is of signicant economic benet towards achieving energy conservation through
better controller design. Overall, it appears that the output regulation with
the H 1 criterion is better than that obtained with the quadratic performance
criterion.
Next, we compare the results obtained by minimization of the H 1 -norm in Eq.
(3.7), i.e. the sensitivity function only, with the H 1 -norm in Eq. (3.9a), i.e.
combination of the sensitivity and complementary sensitivity functions. The output
responses are shown in Figs. 5ac. For the problem considered here, it appears that
the use of Eq. (3.9a) results in a signicant improvement in the boiler temperature
response (Fig. 5a) and a slight improvement in the zone temperature responses (Figs.
5b and c). The corresponding control input proles are shown in Figs. 6ae. Note
that the addition of energy input in minimizing the H 1 -norm in fact saves some
energy in running the fan n1 (Fig. 6a). While holding the other energy inputs at the
same level and giving better output temperature regulation.
The closed-loop responses presented thus far show the minimization of the H 1 -
norm sensitivity function not only achieves faster response but minimizes control
input energy as well. On the other hand, a slight further improvement in output
responses is also achieved at about the same level of energy input by minimizing
the H 1 -norm of the mixed sensitivity function. For the HVAC system considered
in this paper, the above results corroborate the fact that external disturbances
affect the system output through certain states which indirectly inuence the
control energy inputs. Thus, by minimizing the disturbance-to-output sensitivity
function, a reduction in input energy is achieved as well. The complementary
sensitivity function plays a secondary role in this problem because of the fact that
there is no direct transmission from the disturbance-to-input energy in the bilinear
model.
4.2. Robustness
Fig. 3. (ac) Comparison of the closed-loop responses (quadratic and H 1 solutions) of the decentralized
system acted upon by the same magnitude of step changes and disturbances as in Figs. 2ac.
ARTICLE IN PRESS
556 S.A.K. Al-Assadi et al. / Journal of the Franklin Institute 341 (2004) 543567
Fig. 4. (ae) Control input proles corresponding to the outputs shown in Figs. 3ac.
these changes is to slow down the output responses. The responses of the closed-loop
system are depicted in Figs. 7ac for the decentralized feedback controller designs
based on optimization of the H 1 -norm criterion of the sensitivity function, and
those obtained by optimization of the quadratic criterion [1], subject to the same
initial step changes and disturbances as those used earlier. The responses obtained
ARTICLE IN PRESS
S.A.K. Al-Assadi et al. / Journal of the Franklin Institute 341 (2004) 543567 557
Fig. 5. (ac) Comparison of the output responses (H 1 solutions) of the decentralized system.
ARTICLE IN PRESS
558 S.A.K. Al-Assadi et al. / Journal of the Franklin Institute 341 (2004) 543567
Fig. 6. (ae) Control input proles corresponding to the outputs shown in Figs. 5ac.
Fig. 7. (ac) Comparison of the output responses (quadratic and H 1 solutions) of the decentralized
system (with 25% changes in the heat exchanger and boiler ue loss coefcients) acted upon by the same
magnitude of step changes and disturbances as in Figs. 2ac.
ARTICLE IN PRESS
560 S.A.K. Al-Assadi et al. / Journal of the Franklin Institute 341 (2004) 543567
Fig. 8. (ae) Control input proles corresponding to the outputs shown in Figs. 7ac.
change in the external disturbance, the quadratic controller reacts slowly i.e. the
control input n3 is smaller (to begin with) compared to that for the H 1 -based
controller. (iv) The H 1 -based performance measure improves the robustness
properties of the controller. (v) All the inputs are smooth and within bounds,
satisfying the capacity constraints of the physical system.
Figs. 9ac show the improvement in the output responses obtained using the H 1 -
based controller design that incorporates a combination of sensitivity and
complementary sensitivity functions (solid-lines) as compared to that which uses
only the sensitivity function (dashed-lines).
A realistic way of testing the HVAC systems is to operate them over a 24-h
period corresponding to a typical day. To this end, we are interested in simulating
the 24-h operation of the bilinear HVAC system using the designed controllers.
For this purpose, we consider an outdoor temperature prole for a typical day in
March in Montreal as shown in Fig. 10. We approximate this prole by step
functions as shown in the gure by assuming that such step changes are known a
priori.
Simulations showing the results of implementing on the bilinear system, the
decentralized controllers obtained by optimizing the H 1 norm of the mixed
sensitivity function (3.9a) and the quadratic cost function given in [1] for this day are
shown in Figs. 11ac. The boiler and zone temperature responses are held close to
the setpoints throughout the day and are closer for the H 1 solution. The maximum
variations in the outputs from their set-points are within 0:2 C for the quadratic
method and 0:1 C for the H 1 method. It is possible to improve the output
regulation by updating the system model at several points along the outdoor
temperature prole (Fig. 10) and by using techniques such as gain scheduling e.g. see
[15]. However, here we have taken an alternate approach in the design of the
decentralized controller using a 24-h time constrained optimization scheme. The
optimization focuses on a 24-h forecasted outdoor temperature prole which is
assumed to be available ahead of time. This appears to be a reasonable approach
because, in practice, HVAC systems are operated based on past history of the load
patterns and future estimates or forecasts of the 24-h load and building occupancy
schedule.
5. Conclusions
Fig. 9. (ac) Comparison of the output responses (H 1 solutions) of the decentralized system (with 25%
changes in the heat exchanger and boiler ue loss coefcients) acted upon by the same magnitude of step
changes and disturbances as in Figs. 2ac.
ARTICLE IN PRESS
S.A.K. Al-Assadi et al. / Journal of the Franklin Institute 341 (2004) 543567 563
Fig. 10. A typical day prole of outdoor air temperature with step function approximation.
Fig. 11. Comparison of the typical daily regulation of boiler and zone temperatures with the decentralized
controller (quadratic and H 1 solutions).
ARTICLE IN PRESS
564 S.A.K. Al-Assadi et al. / Journal of the Franklin Institute 341 (2004) 543567
quadratic cost optimization approach for linear and bilinear models of the HVAC
system given by Zaheer-uddin et al. [1]. We conclude that the H 1 constrained
optimization design is superior not only in terms of achievability of stability and
disturbance attenuation requirements, but also in terms of improvement robustness
against model uncertainties. The goal in Zaheer-uddin et al. [1], as in the present case
was to achieve good regulation subject to robustness constraints on stability and
control input energy capacities. It should be noted that, the approach used in this
paper involves designing a controller from a frequency domain as well as time
domain viewpoint where as in Zaheer-uddin et al. [1] a time-domain viewpoint was
only used.
Appendix A
where
P1 1 P1 max 11 T h1 T l1 =DT 1 max A:8
and
P2 1 P2 max 11 T h2 T l2 =DT 2 max : A:9
ARTICLE IN PRESS
S.A.K. Al-Assadi et al. / Journal of the Franklin Institute 341 (2004) 543567 565
2 3
19:525 7:988 0:0 0:0 7:988 0:0 0:0
6 7
6 28:361 29:759 1:134 0:0 0:0 0:0 0:0 7
6 7
6 0:0 1:134 9:529 8:130 0:0 0:0 0:0 7
6 7
6 7
A 6 0:0 0:0 3:635 5:054 0:0 0:0 0:2364 7
6 7
6 28:361 0:0 0:0 0:0 29:759 1:134 0:0 7
6 7
6 7
4 0:0 0:0 0:0 0:0 1:134 10:546 9:147 5
0:0 0:0 0:0 0:2659 0:0 4:090 5:685
2 3 2 3 2 3
80:698 0:0 0:0 0:0 0:0
6 7 6 7 6 7
6 0:0 7 6 0:0 92:442 7 6 0:0 0:0 7
6 7 6 7 6 7
6 0:0 7 6 160:103 133:284 7 6 0:0 0:0 7
6 7 6 7 6 7
6 7 6 7 6 7
B1 6 0:0 7; B2 6 71:586 0:0 7; B3 6 0:0 0:0 7;
6 7 6 7 6 7
6 0:0 7 6 0:0 0:0 7 6 0:0 96:586 7
6 7 6 7 6 7
6 7 6 7 6 7
4 0:0 5 4 0:0 0:0 5 4 148:033 137:428 5
0:0 0:0 0:0 66:189 0:0
Table 1
Design parameters of the HVAC system
Table 2
System operating point
References
[1] M. Zaheer-uddin, R.V. Patel, S.A.K. Al-Assadi, The design of decentralized robust controllers for
multi-zone space heating systems, IEEE Trans. Control Systems Technol. 1 (1993) 246261.
[2] W.M. Haddad, D.S. Bernstein, C.N. Nett, Decentralized H 2 =H 1 controller design, Proceedings of
the IEEE Conference on Decision and Control, 1989, pp. 932933.
[3] A.N. Gundes, C.A. Desoer, Algebraic Theory of Linear Feedback Systems with Full and
Decentralized Compensators, Lecture Notes in Control and Information Sciences, vol. 142, Springer,
Berlin, 1990.
[4] R.A. Date, J.H. Chow, A parametrization approach to optimal H 2 and H 1 decentralized control
problems, Proceedings of the IEEE Conference on American Control, Illinois, 1992, pp. 11531157.
[5] E.J. Davison, The robust decentralized control of a general servomechanism problem, IEEE Trans.
Automat. Control 21 (1976) 1424.
[6] S.H. Wang, E.J. Davison, On stabilization of decentralized control system, IEEE Trans. Automat.
Control 18 (1973) 473478.
[7] E.J. Davison, A. Solomon, Partial decentralized temperature control of multi-zone buildings, IEEE
Conference on Decision and Control, Texas, 1983, pp. 1016.
[8] E.J. Davison, T.N. Chang, Decentralized controller design using parameter optimization methods,
Control Theory Adv. Technol. 2 (1986) 131154.
[9] R.V. Patel, N. Munro, Multivariable System Theory and Design, Pergamon Press, Oxford, 1982.
[10] E.J. Davison, G. Gesing, Sequential stability and optimization of large-scale decentralized systems,
Automatica 15 (1979) 307320.
[11] R.V. Patel, P. Misra, Numerical computation of decentralized xed modes, Automatica 27 (1991)
375382.
ARTICLE IN PRESS
S.A.K. Al-Assadi et al. / Journal of the Franklin Institute 341 (2004) 543567 567
[12] A. Emami-Naeini, P. Van Dooren, Computation of zeros of linear multivariable systems, Automatica
18 (1982) 415430.
[13] G.J. Balas, J.C. Doyle, K. Glover, A. Packard, R. Smith, MATLAB mAnalysis and Synthesis
Toolbox, Math Works, Natick, MA, 1991.
[14] A. Grace, MATLAB Optimization Toolbox, Math Works, Natick, MA, 1990.
[15] J.S. Shamma, M. Athans, Analysis of gain scheduled control for nonlinear plants, IEEE Trans.
Automat. Control 35 (1990) 898907.