BDK Pi Thoi Gian Thuc

INTERNATIONAL JOURNAL OF ENERGY RESEARCH
Int. J. Energy Res. 2004; 28:1313–1327

Published online 18 August 2004 in Wiley InterScience (www.interscience.wiley.com). DOI: 10.1002/er.1030
Real-time tuning of PI controllers in HVAC systems
Guang Qu and M. Zaheeruddinn,y

Building, Civil and Environmental Engineering, Concordia University, Montreal, Quebec, Canada H3G 1M8
SUMMARY
The problem of tuning single-loop controllers in heating, ventilating and air conditioning (HVAC) systems
is explored. The HVAC process was described by a first-order-plus-dead-time (FOPDT) model. By using
recursive least squares method, the model parameters were updated while the system remained in closed-
loop. The H1 loop-shaping tuning rules published in the literature were transformed to discrete-time
tuning rules and were implemented in an adaptive PI control strategy. The methodology was applied to a
discharge air temperature (DAT) control system. The output responses of adaptive PI controller were
compared with a LQR optimal adaptive controller. Simulation results show that the adaptively tuned PI
controller is able to track setpoint changes very well in the presence of changes in plant parameters,
disturbances and external noise acting on the system. Copyright # 2004 John Wiley & Sons, Ltd.
KEY WORDS: HVAC systems; DAT system; temperature control; online tuning; PI control
1. INTRODUCTION
The local control loops in heating, ventilating and air conditioning (HVAC) systems must be
designed and operated to maintain dynamic equilibrium between the zone thermal loads and
plant heat extraction rates. Because of its impact on occupants’ comfort, equipment
performance and operating costs, the HVAC control problem is a challenging research problem
with the potential for significant economic benefits.
From a practical standpoint, the HVAC controller should be (i) simple to implement on the
available hardware, (ii) adaptive to load changes, (iii) stable, and (iv) give near optimal
temperature control. To this end, the principal objective of this paper is to develop a real-time
tuning method for local loop control in HVAC systems. Comparisons with other methods will
be made to evaluate the performance. As an example, the tuning and control of a discharge air
temperature (DAT) system will be presented.
n
Correspondence to: M. Zaheeruddin, Building, Civil and Environmental Engineering, Concordia University, Montreal,
Quebec, Canada H3G 1M8.
y
E-mail: zaheer@cbs-engr.concordia.ca
Contract/grant sponsor: Natural Sciences and Engineering Research Council of Canada; contract/grant number: OGP
0036380.
Received 24 November 2003

Copyright # 2004 John Wiley & Sons, Ltd. Accepted 19 January 2004
1314 G. QU AND M. ZAHEERUDDIN
1.1. PID control in HVAC systems

Kasahara et al. (1999) have presented a stability limit analysis and a tuning method for PID
controllers in bilinear systems with time-delayed feedback. The bilinear system they considered
is a variable air volume (VAV) system. The main limitation for the online application of the
tuning method is that it requires extensive computations to find an important gain reduction
factor needed for implementation.
Wang et al. (1999) have proposed a PID controller design method that achieves high
performance for a wide range of linear self-regulating processes. Also, real-time experiments
were carried out in the laboratory. The experimental results show that improved performance is
achieved by using the PID tuning method. The main limitation of the tuning method is that
robustness to uncertainties in the models is not considered and uncertainties in the models
certainly exist and have an effect on performance since the method uses an exact cancellation
technique.
Seem (1997) presented a method for implementing a new pattern recognition adaptive
controller (PRAC) developed through optimization, for automatically adjusting the parameters
of PI controllers while under closed loop control. Depending on patterns of the closed loop
response, PRAC will determine the parameters of the digital PI controller used in a HVAC
system. The limitation of this study is that PRAC was developed for systems that can be
characterized by a first-order plus dead-time model with the ratio of dead time to time constant
between 0.25 and 1, and the ratio of sampling time to time constant is between 0.1 and 1 for
‘good control’.
Nesler (1986) has reported the implementation of three automated controller-tuning methods.
The author noted that unmodelled load disturbances, drifting parameters and actuator
nonlinearity are limitations of self-tuning controllers.
Chen et al. (1990) have presented an adaptive robust control scheme applied to a single-zone
HVAC system. It is a nonlinear system, and the uncertainty is assumed bounded but the bound
is unknown. Then, a class of adaptive robust controls which was originally designed by Corless
and Leitmann (1984) is used to achieve the control objective to drive the room’s state (which is
related to its temperature and humidity) into a comfort region. A comparison between the use of
on-off control and the use of the adaptive robust control from simulation results is given. Chen
et al. conclude that simulation results depict a satisfactory transient performance, in the sense of
maintaining small overshoot, under significant deviation of the initial state from the comfort
region. A drawback reported in Chen et al. (1990) is that the steady state performance has
certain oscillations.
There are a number of apparent advantages in using adaptive controls in HVAC applications.
Commissioning costs would be less since the commissioning period would be shorter. Since
HVAC field engineers are familiar with direct digital control (DDC) PID controllers, it is of
interest to develop an online adaptive tuning method to update the PI parameters. Motivated by
these considerations, we are interested, in exploring the design of adaptive PI controllers for
HVAC systems. To illustrate the design methodology, we consider a discharge air temperature
(DAT) system.
Several authors have studied the modelling aspects of DAT systems. For example, the most
detailed models based on the principle of energy and mass balances are described in Zheng and
Zaheeruddin (1999). Other simplified models such as first-order models have been reported in
Shavit and Brandt (1982). These latter models have been used in the design of PID controllers
Copyright # 2004 John Wiley & Sons, Ltd. Int. J. Energy Res. 2004; 28:1313–1327
REAL-TIME TUNING OF PI CONTROLLERS 1315
for DAT systems. The controller design methods include both classical design methods (Shavit
and Brandt, 1982) and more recent advanced design methods (Seem, 1996).
2. DAT SYSTEM: PHYSICAL MODEL
Figure 1a shows a schematic diagram of a DAT system. Mixed air enters the cooling coil at
temperature Ta0 and flow rate Q. The air is cooled in the cooling coil using chilled water. The
temperature of the air leaving the cooling coil Ta is controlled by modulating the chilled water
flow rate in the coil as shown by the feedback control loop.
The DAT process was described as a sensible cooling process. The input-output
representation of the DAT model is shown in Figure 1b. In Figure 1b, u; the chilled water
flow rate, is the input and Ta, the discharge air temperature, is the output. The entering air
temperature Ta0 is considered as disturbance on the system.
2.1. First-order model

From the point of view of online control, it is necessary to reduce the computational effort
required in the identification and control of the DAT system. To this end, we model the
DAT system as a first-order plus dead-time (FOPDT) system. The model was represented as in
Figure 2a. In Figure 2a, Ta0 is the output of the system when u ¼ 0; Ta is the output of the
system and u is the input of the system. Assuming T is the sampling time and
t
l¼ ð1Þ
T
is an integer, the discrete time model of the DAT system was derived. The discrete time model of
the system is shown in Figure 2b.
where; a ¼ eT =Ts ð2Þ
and
b ¼ ks ð1 aÞ ð3Þ
Figure 1. (a) Schematic diagram of DAT system, (b) block diagram of DAT system.
u ks y_
e −τs Ta
Ts s + 1 +
(a) Ta0
u bz − l − 1 y _
Ta
1 − az − 1 +
(b) Ta0
Figure 2. (a) First-order plus dead-time model, (b) First-order plus dead-time model in discrete system.
3. ONLINE IDENTIFICATION: RLS METHOD FOR THE FOPDT MODEL
The recursive least squares (RLS) method is an effective approach in online identification of
model parameters. In order to implement the RLS algorithm for online identification of the
DAT system, we follow the method described in Franklin and Powell (1980) and shown in
Figure 3.
Considering the model shown in Figure 2b, we note that, if l is fixed, the system identification
reduces to the estimation of parameters a and b of the plant
Y ðzÞ bzl1
GðzÞ ¼ ¼ ð4Þ
U ðzÞ 1 az1
Using measured values for Ta0, Ta(k) and uðkÞ for some k; we have
yðkÞ ¼ Ta0 TaðkÞ ð5Þ
and
yðkÞ ayðk 1Þ buðk l 1Þ ¼ eðk; yÞ ð6Þ
Let
y ¼ ½a bT ð7Þ
and
cðkÞ ¼ ½yðk 1Þ uðk l 1ÞT ð8Þ
then, through analysis, we can write the error equation as
Y ðN Þ ¼ CðN Þy þ eðN ; yÞ ð9Þ
where
T
CðN Þ ¼ cðl þ 1Þ cðl þ 2Þ cðN Þ ð10Þ
T
eðN ; yÞ ¼ eðl þ 1; yÞ eðN ; yÞ ð11Þ
T
Y ðN Þ ¼ yðl þ 1Þ yðN Þ ð12Þ
u(k) y(k)
Plant
0 + e(k ; )
_
Prediction yˆ (k )

Figure 3. Block diagram showing the generation of output prediction error.
Assume that we observe the set of inputs and outputs
fuð0Þ; uð1Þ; . . . ; uðN Þ; yð0Þ; yð1Þ; . . . ; yðN Þg
and that we need to estimate the parameter y0 of the plant and the prediction is y: The least-
squares method can be expressed as
min J ðyÞ ð13Þ
y
where
X
N
J ðyÞ ¼ wðkÞe2 ðk; yÞ ¼ eT W e ð14Þ
k¼n
and the weighting function wðkÞ is positive. By using the methodology given in Franklin and
Powell (1980), set of equations were solved to update the parameters a and b:
3.1. Online identification and simulation

The plant model with the following parameters was simulated.
Y ðzÞ bz16
GðzÞ ¼ ¼ Ta0 ¼ 66:028F;
U ðzÞ 1 az1
and sampling time T=4 s, where

( (
0:9558 t51000 s 0:1326 t51000 s
a¼ and b¼
0:9797 t51000 s 0:1658 t51000 s
The implementation scheme for online identification of parameters a and b in the closed loop
mode in simulink is shown in Figure 4.
Figures 5 and 6 show two sets of results one without delay (Figure 5) and the other with input
delay (Figure 6). It is evident that consideration of input delay improves the model response
(Tam ) very close to plant response (Ta ). The identified parameters converge fast and reach final
values as shown in the figures. From Figure 6, we note that the model response has almost the
same shape as the plant response. This fact means that the online identification becomes
accurate after a time equal to 2 l sampling intervals counted from the start of the next update
of the model parameters.
Figure 4. Online identification in closed loop mode.
70 6
65
4
u (gpm )
Ta (F)
60
2
55
50 0
0 500 1000 1500 2000 0 500 1000 1500 2000
time (Sec) time (Sec)
70 1
a
0.8
65
Tam (F)
0.6
a, b
60
0.4
55
0.2
b
50 0
0 500 1000 1500 2000 0 500 1000 1500 2000
Figure 5. Closed loop mode identification results (without delay).
70 6
65
4
u (gpm )
Ta (F )
60
2
55
50 0
0 500 1000 1500 2000 0 500 1000 1500 2000
70 1
a
0.8
65
Tam (F )
0.6
a, b
60
0.4
55
0.2
b
50 0
0 500 1000 1500 2000 0 500 1000 1500 2000
Figure 6. Simulation result with 30 samples input delay.
4. H1 ADAPTIVE PI CONTROL
In this section, we explore a PI tuning rule for first-order plus dead-time (FOPDT) models in
discrete-time systems. To this end, we have adopted the PID tuning rules (Tan et al., 1998) based
on loop-shaping H1 control for FOPDT models in continuous-time systems. We refer to this
method as H1 PI tuning rules. We also study an adaptive control structure, in which H1 PI
tuning rules will be used. Finally, comparisons between H1 adaptive control and LQR optimal
control, will be made.
4.1. PID tuning rules based on loop-shaping H1 control

PID tuning rules based on loop-shaping H1 control for first-order plus dead-time (FOPDT)
models in continuous-time systems were proposed by Tan et al. (1998). We first briefly describe
these rules.
Given a plant pðsÞ that can be described by an FOPDT model
ks
pðsÞ ¼ ets ð15Þ
Ts s þ 1
the PID structure based on loop-shaping H1 control becomes

1 1
KðsÞ ¼ Kp 1 þ þ Td s ð16Þ
Ti s 1 þ Tf s
and the PID tuning rules are given by

0:265l þ 0:307 Ts
Kp ¼ þ 0:5 ð17Þ
ks t
t
Ti ¼ Ts þ ð18Þ
2
Ts t
Td ¼ ð19Þ
2Ts þ t
t
Tf ¼ ð20Þ
5:314l þ 0:951
where l is a design parameter.
4.2. H1 PI tuning rules for discrete-time systems

In this section, we transfer the PID tuning rules in continuous-time systems to PI tuning rules in
discrete-time systems by using the equivalent transform method. First, we assume that we have
already designed a PID controller for a continuous-time system based on loop-shaping H1
control. For the plant shown in Figure 2a, the controller becomes

1 1
KðsÞ ¼ Kp 1 þ þ Td s ð21Þ
Ti s 1 þ Tf s
and the parameters can be tuned by using Equations (16)–(20). Secondly, using the
transformation
2z1
s¼ ð22Þ
T zþ1
the H1 PI tuning rules for discrete-time systems were derived. These are
Kp
Kid ¼ ð23Þ
Ti
and
Td
Kpd ¼ Kp ð24Þ
Tf
where

ð0:265l þ 0:307Þð1 aÞ 1
Kp ¼ þ 0:5 ð25Þ
b l ln a
T lT
Ti ¼ þ ð26Þ
ln a 2
Tl
Td ¼ ð27Þ
2 l ln a
lT
Tf ¼ ð28Þ
5:314l þ 0:951
The design parameter l can be chosen as l ¼ 0:1974 for all l; or l ¼ ðl ln a 0:951Þ=5:314 for
2=ln a4l4 8:922=ln a in order to get faster response of the closed loop system.
4.3. H1 adaptive PI control implementation

The proposed H1 adaptive PI control structure is shown in Figure 7. The function of the
identifier is to estimate the parameters of the FOPDT discrete time model. The function of
the PI tuner is to update the parameters of the PI controller using the updated parameters of the
plant while the controller remains in closed loop.
4.4. Simulation and comparison

We consider the same plant model as the one used in Section 3.1. The implementation scheme in
Simulink is shown in Figure 8. Figure 9 shows the H1 adaptive PI control response for a
PI tuner Identifier
Tset e[k ] u[k] Ta[k ]

PI controller Plant
+
_
Figure 7. H1 adaptive PI control structure.
Figure 8. Implementation of H1 adaptive PI control in Simulink.
68 6
66
4
64
u (gpm)
Ta (F)
62
2
60
58 0
0 500 1000 1500 2000 0 500 1000 1500 2000
Time (Sec) Time (Sec)
-3
x 10
0 0
-0 . 1
-0 . 2 -2
Kpd
-0 . 3
Kid
-0 . 4 -4
-0 . 5
-6
0 500 1000 1500 2000 0 500 1000 1500 2000
Figure 9. H1 adaptive PI control response for setpoint change.
setpoint change. The PI parameters were updated by the H1 PI tuning rules. Figure 10 shows
the responses using an LQR optimal adaptive PI controller. (The design details of the LQR
optimal PI controller are omitted.) The LQR results correspond to the following values:
" #
0:001 0
N ¼ 200; Q ¼ and R ¼ 0:1
0 80 000
In these simulations, the initial values of the PI parameters are Kpd ½0 ¼ 0:35 and
Kid ½0 ¼ 0:0035; and
(
608F 04t51000 s
Tset ¼
628F t51000 s
The simulation results in Figures 9 and 10 show that the H1 adaptive PI control response for
setpoint change (we focus on the response from 1000 to 2000 s) is smoother compared with the
LQR optimal control case. The response has a smaller overshoot and a shorter settling time. It
should be noted that the PI parameter tuning in the LQR optimal control is very difficult to
implement in real time because it requires significant computational time to update the PI
parameters. Also, several trial runs are needed to choose the appropriate weighting factors in
matrices Q and R in order to get good response. On the other hand, the H1 PI tuning rules in the
H1 adaptive PI control are very easy to implement.
68 6
66
4
u (gpm)
64
Ta (F)
62
2
60
58 0
0 500 1000 1500 2000 0 500 1000 1500 2000
-3
x 10
0 0
-0 . 1
-0 . 2 -2
Kpd
Kid
-0 . 3
-0 . 4 -4
-0 . 5
-6
0 500 1000 1500 2000 0 500 1000 1500 2000
Figure 10. LQR adaptive PI control response for setpoint change.
68
Case 1
66
Case 2
64
Case 3
62
Ta (F)
60
58
56
54
0 200 400 600 800 1000 1200 1400 1600 1800 2000
Time (Sec)
Figure 11. H1 adaptive PI control responses to a setpoint change (Case 1, 2 and 3).
Figure 11 shows the H1 adaptive PI control responses to a setpoint change. In this simulation
three different plant models were used. These are

Case 1: a b ¼ 0:9358 0:1026 and l ¼ 15:

Case 2: a b ¼ 0:9558 0:1326 and l ¼ 15:

Case 3: a b ¼ 0:9758 0:1626 and l ¼ 15:
Figure 12. Implementation scheme for testing the effects of disturbances and noise.
68 4
66
3
64
u (gpm)
Ta (F)
2
62
1
60
58 0
0 1000 2000 3000 0 1000 2000 3000
-3
x 10
0. 1
0 0
-0 .1
-0 .2 -2
Kpd
Kid
-0 .3
-0 .4 -4
-0 .5
-6
0 1000 2000 3000 0 1000 2000 3000
Figure 13. Response without disturbances and noise.
In all three cases, the initial values of the PI parameters were Kpd ½0 ¼ 0:35 and Kid
½0 ¼ 0:0035; and
(
608F 04t51000 s
Tset ¼
628F t51000 s
As shown in Figure 11 after 1000 s, the responses for the three cases are the same. This means
that we can obtain the same setpoint tracking response irrespective of changes in model
parameters (a and b) using H1 adaptive PI control.
4.5. The effect of noise and disturbance

In order to study the performance of H1 adaptive PI controller subject to disturbances and
noise, several simulations runs were made. Figure 12 shows the implementation scheme. Using
(
608F 04t51000 s
Tset ¼
618F t51000 s
the output responses subjected to cyclic disturbances and band-limited white noise with different
levels are presented in Figures 14–16; and the output response without disturbances and noise is
shown in Figure 13. It is noted that the H1 adaptive PI controller remains stable and is able to
68 4
66
3
u (gpm )
64
Ta (F )
2
62
1
60
58 0
0 1000 2000 3000 0 1000 2000 3000
-3
x 10
0.1
0 0
-0.1
-0.2 -2
K pd
K id
-0.3
-0.4 -4
-0.5
-6
0 1000 2000 3000 0 1000 2000 3000
Figure 14. Response under noise level between 0 and 1.0.
68 4
66
3
u (gpm )
Ta (F ) 64
2
62
1
60
58 0
0 1000 2000 3000 0 1000 2000 3000
-3
x 10
0.1
0 0
-0.1
-0.2 -2
K pd
K id
-0.3
-0.4 -4
-0.5
-6
0 1000 2000 3000 0 1000 2000 3000
Figure 15. Response under noise level between 0 and 0.01.
68 4
66
3
u (gpm )
64
Ta (F )
2
62
1
60
58 0
0 1000 2000 3000 0 1000 2000 3000
-3
x 10
0.1
0 0
-0.1
-0.2 -2
K pd
K id
-0.3
-0.4 -4
-0.5
-6
0 1000 2000 3000 0 1000 2000 3000
Figure 16. Responses with: amplitude 1, period 500 s, duty cycle 2%.
reject the effects of disturbances very well. From the figures, we see that the H1 adaptive PI
controller has good ability to reject the effects of disturbances and is comparable to the pattern
recognition adaptive controller (Seem, 1997).
5. CONCLUSIONS
An adaptive PI control strategy for single loop HVAC processes is developed. The tuning
methodology updates the PI parameters in real-time while the control system remains in the
closed-loop. Simulation results show that the adaptively tuned PI controller is stable under the
effects of variable disturbances and external noise. The setpoint tracking performance is not
affected and remains smooth as the plant parameters are continuously updated in the closed-
loop operation of the DAT system. Being simple and robust, the adaptive PI controller is well
suited for control of HVAC thermal processes in buildings.
ACKNOWLEDGEMENT
This work was funded by a research grant (OGP 0036380) from the Natural Sciences and Engineering
Research Council (NSERC) of Canada.
REFERENCES
Chen YH, Lee KM, Wepfer WJ. 1990. Adaptive Robust Control Scheme Applied to a Single-zone HVAC System.
Proceedings of the 1990 American Control Conference, San Diego, vol. 2. 1076–1081.
Corless M, Leitmann G. 1984. Adaptive control for uncertain dynamical systems. In Dynamical Systems and
Microphysics: Control Theory and Mechanics, Blaquiere A, Leitmann G (eds). Academic Press: Orlando.
Franklin GF, Powell JD. 1980. Digital Control of Dynamic Systems. Addison-Wesley: Reading, MA.
Kasahara M, Matsuba T, Kuzuu Y, Yamazaki T, Hashimoto Y, Kamimura K, Kurosu S. 1999. Design and tuning of
robust PID controller for HVAC systems. ASHRAE Transactions Part 2; 105:154–166.
Nesler CG. 1986. Automated controller tuning for HVAC applications. ASHRAE Transactions Part 2B; 92:189–201.
Seem JE. 1996. A pattern recognition adaptive controller. IFAC 13th Triennial World Congress, San Francisco, 121–126.
Seem JE. 1997. Implementation of a new pattern recognition adaptive controller developed through optimization.
ASHRAE Transactions Part 1; 103:494–506.
Shavit G. 1994. Tuning of PID DDC controllers. American Control Conference, Baltimore, vol. 3. 3065–3069.
Shavit G, Brandt SG. 1982. The dynamic performance of discharge air temperature system with a PI controller.
ASHRAE Transactions Part 2; 88:826–828.
Tan W, Liu L, Tam PKS. 1998. PID tuning based on loop-shaping H1 control. IEE Proceedings on Control Theory
Applications 145(6):485–490.
Wang QG, Lee TH, Fung HW, Bi Q, Zhang Y. 1999. PID tuning for improved performance. IEEE Transactions on
Control Systems Technology 7(4):457–465.
Zheng GR, Zaheeruddin M. 1999. Discharge air system: modelling and optimal control. International Journal of Energy
Research 23:727–738.

BDK Pi Thoi Gian Thuc

Transféré par

Informations du document

Titre original

Copyright

Formats disponibles

Partager ce document

Partager ou intégrer le document

Options de partage

Avez-vous trouvé ce document utile ?

Ce contenu est-il inapproprié ?

Droits d'auteur :

Formats disponibles

BDK Pi Thoi Gian Thuc

Transféré par

Droits d'auteur :

Formats disponibles

INTERNATIONAL JOURNAL OF ENERGY RESEARCH

Int. J. Energy Res. 2004; 28:1313–1327

Real-time tuning of PI controllers in HVAC systems

Guang Qu and M. Zaheeruddinn,y

Received 24 November 2003

1.1. PID control in HVAC systems

2. DAT SYSTEM: PHYSICAL MODEL

2.1. First-order model

3. ONLINE IDENTIFICATION: RLS METHOD FOR THE FOPDT MODEL

Figure 3. Block diagram showing the generation of output prediction error.

Assume that we observe the set of inputs and outputs

fuð0Þ; uð1Þ; . . . ; uðN Þ; yð0Þ; yð1Þ; . . . ; yðN Þg

3.1. Online identiﬁcation and simulation

and sampling time T=4 s, where

Figure 4. Online identiﬁcation in closed loop mode.

Figure 5. Closed loop mode identiﬁcation results (without delay).

Figure 6. Simulation result with 30 samples input delay.

4.1. PID tuning rules based on loop-shaping H1 control

and the PID tuning rules are given by

4.2. H1 PI tuning rules for discrete-time systems

4.3. H1 adaptive PI control implementation

4.4. Simulation and comparison

Tset e[k ] u[k] Ta[k ]

Figure 7. H1 adaptive PI control structure.

Figure 8. Implementation of H1 adaptive PI control in Simulink.

Figure 9. H1 adaptive PI control response for setpoint change.

Figure 10. LQR adaptive PI control response for setpoint change.

4.5. The eﬀect of noise and disturbance

Figure 15. Response under noise level between 0 and 0.01.

Vous aimerez peut-être aussi