01a 104
01a 104
01a 104
MULTILOOP CONTROL APPLIED TO INTEGRATOR MIMO
PROCESSES. A Preliminary Study
Eduardo J. Adam1,2*, Carlos J. Valsecchi2
1
Instituto de Desarrollo Tecnológico para la Industria Química (INTEC)
(Universidad Nacional del Litoral – CONICET)
Güemes 3450, S3000GLN, Santa Fe – Argentina
2
Facultad de Ingeniería Química, Univ. Nac. del Litoral, Santa Fe – Argentina
Santiago del Estero 2654, S3000AOJ, Santa Fe – Argentina
Email: [email protected]
Abstract. This work presents a multiloop control study applied to integrator
controller design techniques for MIMO systems are studied BLT method
textbook (Skogestad and Postlethewaite, 1996; Zhou and Doyle, 1998) and,
methods (BLT)) when a proportionalintegral multiloop control is designed.
A discussion together with some recommendations about these problems is
included. Also, an industrial application example (Tan et al., 2002) is
design. This design is showed in detail so that the reader can see some
complications present in these particular MIMO processes.
Keywords: Integrator system, MIMO, Multiloop, PID controllers, H∞ design.
* To whom all correspondence should be addressed
AAIQ Asociación Argentina de Ingenieros Químicos
IACCHE Interamerican Confederation of Chemical Engineering
XXII IACChE (CIIQ) 2006 / V CAIQ
1. Introduction
Multivariable and multiloop control have received much attention in the academic
community since several industrial applications have multipleinput and multipleoutput
variables, such as distillation columns, heat transfer processes and chemical reactors,
among others. Thus, multiloop PID controllers are being widely used in the industry
because the engineers can control MIMO processes with only a few parameters to tune.
Three types of tuning methods for multiloop control systems are available in the
literature; (1) detuning method, (2) sequential closing method and (3) independent
design method.
In the detuning method, each controller of the multiloop control system is first
designed ignoring process interactions from the other loops. Then, the interactions are
taken into account and each controller is detuned until some performance criterion is
reached. In this context, the most popular technique of the detuning methods is
probably the BLT method (Luyben, 1986 and Monica et al., 1988). However, for the
sequential closing method, each controller is designed sequentially. That is, an input
output pair is designed and this loop is closed. Then a second pair is considered while
the first controller is closed and so on (Mayne, 1973; Loh et al., 1993). Finally, in the
independent design method, each controller is designed based on the paired transfer
functions while satisfying some constraints due to the process interactions (Skogestad
and Morari, 1989).
In this paper, two tuning methods for multiloop control are studied to control
integrator MIMO processes. For this purpose, an attractive and interesting example that
shows several problems in the multiloop PI control design is included.
AAIQ Asociación Argentina de Ingenieros Químicos
IACCHE Interamerican Confederation of Chemical Engineering
XXII IACChE (CIIQ) 2006 / V CAIQ
In order to give a systematic exposition of this work, in Section 2 the semiempirical
method known as BLT technique is applied to the integrative MIMO system and a brief
integrative MIMO system based on H∞ design. Then, an industrial application example
is analyzed in Section 4. Finally, in Section 5 the conclusions are presented.
2. Semiempircal Methods
2.1. Biggest log modulus Tuning (BLT)
Detuning techniques for MIMO process control are based on disadjusting the
parameters in the controller matrix by using some semiempirical criteria. BLT method
multiloop PI control), uses a detuning factor F for all gains and integrative times of the
control following similar ideas to the original method. Both techniques use semi
empirical functions and begin an iterative sequence with PI controllers designed by
Ziegler and Nichols' (ZN) settings (continuous cycling technique). For this reason, the
following remark is enunciated:
Remark 1. If at least one independent control loop in the multiloop system presents
infinity gain margin then, BLT methods cannot be applied.
PROOF
Since BLT methods compute ultimate gain (Ku) and period (Pu), if at least one
AAIQ Asociación Argentina de Ingenieros Químicos
IACCHE Interamerican Confederation of Chemical Engineering
XXII IACChE (CIIQ) 2006 / V CAIQ
proportional controller for all 0 < Kc < ∞ then, it is not possible to compute Ku and Pu
and BLT methods cannot be applied.
2.2 An Illustrative Example
It is important to highlight that the BLT methods were not developed for integrator
or unstable MIMO processes because the authors never considered these systems
(Luyben, 1986 and Monica et al., 1988). One of the purposes of this work is to study
this alternative.
Consider a 2x2 system with a transfer function matrix given by
e 2e
−0.5s −0.1s
G s=
[ s s1
3e
−0.3s
2s1
4e
3s1
−0.2s
s s3
] . (1)
Following the BLT method, two semiempirical functions are computed. One of
them is defined as
W(s) := 1 + det(I + G(s)C(s)) , (2)
and the other one is a semiempirical measurement (Lcn) designated by the author as
multivariable closed loop log modulus (similar to complementary sensitivity function
for SISO systems) defined as
W j
L cn j :=20log
∣ 1W j ∣ . (3)
AAIQ Asociación Argentina de Ingenieros Químicos
IACCHE Interamerican Confederation of Chemical Engineering
XXII IACChE (CIIQ) 2006 / V CAIQ
Luyben (1986) suggests finding the parameters of the PIcontroller diagonal matrix
by using a detune parameter F such as
where n is the MIMO system dimension. Thus, the PI controller actions result
Kci = KciZN/F , (5)
and
TIi = FTIiZN , (6)
with i = 1, ..., n.
due to G(s) is 2x2 system. Figure 1 shows Lcn(jω) computed in the frequency interval
result is shown in Fig. 2. According to this figure there is not only one solution to this
particular problem since there are three values for the detuning factor F such that
∥L cn∥∞=4 DB. Furthermore, there exits an instability region for 0 < F < 2.1. And,
means that for integrative MIMO systems, the design problem can be nonconvex, and
a stability condition should be included. While, Luyben and coworkers state that the
integrator elements.
10
Lcn = 4
Modulus de L cn (jω ) (DB)
∞
0
5
10
15
20
2 1 0 1
10 10 10 10
Frequency ω (rad/sec)
Fig. 1. Lcn(jω) vs frequency ω with F = 8.2345.
25
20 Unstable Stable Region
Region
15
L cn (DB)
∞
10
Lcn = 4
∞
0
0 1 2 3 4 5 6 7 8 9 10
Detune Factor F
AAIQ Asociación Argentina de Ingenieros Químicos
IACCHE Interamerican Confederation of Chemical Engineering
XXII IACChE (CIIQ) 2006 / V CAIQ
Table 1 summarizes the settings of the controller diagonal matrix and the stability
and instability situation for this example.
Table
1 . PI controller settings for integrative MIMO system (1).
According to the BLT method (Luyben, 1986), if the designer choices F = 2 in other
to initializes the detuning procedure, an unstable solution can be found.
3. Mixed Sensitivity Function H∞ Design
and secondly, a diagonal controller K is designed in frequency domain. Invariably this
twostep procedure results in a suboptimal design.
With the purpose of quantifying the interaction degree in the MIMO system, the
relative gain array (RGA) evaluated in s = 0 is traditionally used. Thus, RGA is defined
as
AAIQ Asociación Argentina de Ingenieros Químicos
IACCHE Interamerican Confederation of Chemical Engineering
XXII IACChE (CIIQ) 2006 / V CAIQ
Λ := G(0)x[G(0)1]T , (7)
where x denotes elementbyelement multiplication.
Remark 2. If the transfer function matrix G(s) has at least one integrator element
traditional RGA are not possible.
Therefore, when the interaction is strong, the decoupling control must be considered.
Thus, decoupling control results when W(s) is chosen such that Gs(s) = G(s)W(s) is a
dynamic decoupling is reached. Usually, G1(s) results in a nonrealizable matrix and in
consequence, one alternative for this problem is the steady state decoupling. This
Gs(0) in a diagonal matrix. Note that the following remark for integrator MIMO
processes:
state decoupling is not possible.
3.1 Design Procedure
The µiteration measure by Grosdidier and Morari (1986) is very convenient for a
multiloop control system design. However, this measure does not ensure a stable
multiloop control system when integrator elements are present in the controllers
(integrator modes) or in the plant (one or more integrator transfer functions). And this
AAIQ Asociación Argentina de Ingenieros Químicos
IACCHE Interamerican Confederation of Chemical Engineering
XXII IACChE (CIIQ) 2006 / V CAIQ
is the central point of the investigation problem since the integrator elements are present
in the plant matrix and in the PI/PID controllers.
applicable to multivariable controllers (Skogestad and Postlethewaite, 1996; Zouh and
Doyle, 1998):
Step 1. Adopt an objective function as N = [WpS] or N = [WpS WuKS]T, where S is
the sensitivity transfer function and, Wp and Wu are weights function matrix.
Step 2. A reasonable initial choice is to set the weights Wp = diag{wpi} with i = 1, 2, ...,
n, where n is the number of channel and wpi = (s/Mi+ωBi)/(s+ωBiAi) where Ai << 1,
Step 3. Solve the following optimization problem:
W p j S j ,
min ∥N K ∥∞ =min max ]
K K [ K ,
(8)
with K ∈ C and C the stabilizing controller set. In the particular case of this paper, the
attention is centered on the PI/PID controller multiloop design. Finally, while the
nominal stability is guaranteed in the design of H2/H∞ optimal controllers by the Youla
parameterization procedure (Youla, 1976), here, for a controller with a fixed structure
such as PI or PID controllers the nominal stability must be checked.
AAIQ Asociación Argentina de Ingenieros Químicos
IACCHE Interamerican Confederation of Chemical Engineering
XXII IACChE (CIIQ) 2006 / V CAIQ
4. An Industrial Application Example
input and output variables are
input: output:
u1 feedwater flow rate (Kg/s), y1 drum level (m),
u2 fuel flow rate (Kg/s), y2 drum pressure (Mpa),
(Kg/s),
Tan et al. (2002) identifies the following LTI model using experimental inputoutput
data and the MATLAB System Identification Toolbox:
2 −3 −3
3.1s−0.032 10
g11 s g12 s g 13 s u 0
−0.16s 0.052s0.001410
y1 u1
][ ] [ ][ ]
s 2 0.0168s s 20.0215s
[][
y3 g31 s g32 s g 33 s u
1
y 2 = g21 s g22 s g 23 s u2 =
3
−0.039510−3
s0.018
−0.00118s0.000139
s 2 0.01852s0.000091
2.5110− 3
s0.0157
0.448s0.0011
s20.0127s0.000095
0.588s 20.2015s0.000910− 3
2
s 0.0352s0.000142
0.582s−0.0243
s20.1076s0.00104
u2
u3
. (9)
The normal setpoints for this operation are
y 1sp 1.00
y[ ][ ]
y sp = y = 6.45 ,
2sp
466.7
3sp
(10)
and, the model constraints of the manipulated and controlled variables are not taken into
account in this paper.
Note three important observations in this example.
AAIQ Asociación Argentina de Ingenieros Químicos
IACCHE Interamerican Confederation of Chemical Engineering
XXII IACChE (CIIQ) 2006 / V CAIQ
reason, it is impossible to calculate a Ku and Pu with a proportional controller because
it presents infinity gain margin. In consequence, BLT methods cannot be applied
(remark 1).
2. Clearly, the system is coupled, but the traditional RGA evaluated in s = 0 is not
possible to apply due to the integrator elements in the plant matrix (9) (remark 2).
3. According to the previous topic, the steady state decoupling is not applicable in
this example (remark 3).
For the reasons stated above, three PI control feedback loops (Fig. 3) were
implemented without precompensator for decoupling the system, and the control
technique enunciated in Section 3.1 was chosen for the design.
Drum
Level Drum
Setpoint Level
Feedwater Feedwater
Controller Valve
Drum
Presure Drum
Setpoint Presure
AirFuel AirFuel
Controller Valves
Boiler
Steam
Steam
Temperature
Temperature
Setpoint
Steam Temp. Steam
Controller Valve
Fig. 3. Classical PI feedback control loop for the utility boiler.
AAIQ Asociación Argentina de Ingenieros Químicos
IACCHE Interamerican Confederation of Chemical Engineering
XXII IACChE (CIIQ) 2006 / V CAIQ
Following the procedure detailed in Section 3.1,
Step 1. It was adopted N = [WpS], where S = [I + GK]1.
Step 2. By simplicity in a preliminary design, it was chosen Wp = diag {wpi} where
wpi 1/Sii with Sii = 1/(1 + giiki) and ki a PI controller given by ki(s) = Kci(1 + 1/TIis) tuned
(Skogestad and Postlethewaite, 1996; Zhou and Doyle, 1998; among others).
Step 3. The optimization problem (8) was solved with K = diag{ki}.
According to the IMC technique, each controller was tuned as,
1
q i= f i , (11)
0 MP
g ii
element of the diagonal matrix (9), where the superscript 0 denotes the nominal transfer
function. Then, the PI controllers were computed as
qi
k i= . (12)
1−gii q i
0
Thus, following Morari and Zafiriou recommendations, the adopted PI controller
actions result
Kci = τii/Kpiiλi , (13)
AAIQ Asociación Argentina de Ingenieros Químicos
IACCHE Interamerican Confederation of Chemical Engineering
XXII IACChE (CIIQ) 2006 / V CAIQ
and
TIi = τii , (14)
with i = 1, 2 and 3. The constants Kpii and τii are summarized in Table 2, and λi is the
IMC filter time constant. Note that, it is easy to reach to Eqns. (13) and (14) with the
IMC procedure for the nominal plant g22(s). But, some algebraic considerations must be
taken into account for g11(s) and g33(s).
Table
2 . Parameters involved in the Eqn. (13) and (14) in order
to calculate the PI controller actions.
According to the IMC approach, λi is a tuning parameter in each canal. Thus, the
problem basically consists of finding the set of parameters Λ = {λ1, λ2, λ3} such that
∥N∥∞ is minimized. Using computational techniques in the frequency domain for this
particular example with the set of parameters Λ = {10000, 500 ,30}, the ∥N∥∞ =
18.46911 at = 0.0100. It is important to notice that the plant matrix (9) is ill
conditioned due to the nature of the variables involved; in consequence the inputoutput
scaling is not good. For this reason, the computational technique convergence is not
easy to reach.
1 A similar value was reported by Tan et al. (2002).
AAIQ Asociación Argentina de Ingenieros Químicos
IACCHE Interamerican Confederation of Chemical Engineering
XXII IACChE (CIIQ) 2006 / V CAIQ
2
10
N = 18.4691
1
10
Modulus
0
10
1
10
4 3 2 1 0 1 2
10 10 10 10 10 10 10
Frequency (rad/sec)
Fig. 4. Modulus of N vs. frequency.
Table 3 summarizes the PI controller settings implemented in the multiloop control
system suggested in Fig. 3.
Table
3 . PI controller settings for multiloop control of Fig. 3.
Figures 5, 6 and 7 show the step responses of the boiler with setpoint changes in
drum level, drum pressure and steam temperature using PI controller diagonal matrix.
According to Figs. 5 and 6 it is possible to observe that the drum level and the pressure
are linked and this behavior is due to a mixture of liquid and gas in the boiler drum.
AAIQ Asociación Argentina de Ingenieros Químicos
IACCHE Interamerican Confederation of Chemical Engineering
XXII IACChE (CIIQ) 2006 / V CAIQ
Figure 5 shows that a level step change has serious impact on the controlled
temperature and a little effect on the drum pressure. While, according to Fig. 6 a step
change in the pressure has an effect on the liquidgas level inside the boiler drum and a
strong impact on the steam temperature.
Figure 7 shows the effect on the drum level and the pressure when a steam
temperature step change is introduced. In this case, a small interaction is observed.
Also, according to Figs. 5 and 7 the drum level and the steam temperature setpoint
changes have little effect on the drum pressure.
1.2
Drum Level (m)
1.1
0.9
0 5 10 15 20 25 30
6.5
Drum Preasure (Mpa)
6.45
6.4
0 5 10 15 20 25 30
470
Steam Temp. (ºC)
468
466
464
0 5 10 15 20 25 30
Time (min.)
Fig. 5. Boiler time responses for a 10% drum level increase with the designed PI
controllers designed according to the procedure indicated above.
AAIQ Asociación Argentina de Ingenieros Químicos
IACCHE Interamerican Confederation of Chemical Engineering
XXII IACChE (CIIQ) 2006 / V CAIQ
1.1
Drum Level (m)
1
0.9
0 5 10 15 20 25 30
6.8
Drum Preasure (Mpa)
6.6
6.4
0 5 10 15 20 25 30
475
Steam Temp. (ºC)
470
465
0 5 10 15 20 25 30
Time (min.)
Fig. 6. Boiler time responses for a 5% drum pressure increase with the designed PI
controllers designed according to the procedure indicated above.
1.02
Drum Level (m)
0.98
0 5 10 15 20 25 30
6.5
Drum Preasure (Mpa)
6.45
6.4
0 5 10 15 20 25 30
500
Steam Temp. (ºC)
450
400
0 5 10 15 20 25 30
Time (min.)
Fig. 7. Boiler time responses for a 10% steam temperature decrease with the designed
PI controllers designed according to the procedure indicated above.
AAIQ Asociación Argentina de Ingenieros Químicos
IACCHE Interamerican Confederation of Chemical Engineering
XXII IACChE (CIIQ) 2006 / V CAIQ
Clearly, let's remark that according to the simulation results (Figs. 5, 6 and 7) that, i)
the inputoutput pairing, feedwater – drum level is linked with drum pressure and
steam temperature when a drum level change is introduced while, ii) airfuel – drum
pressure is strongly linked with drum level and steam temperature when drum pressure
change is introduced, and iii) small effects are observed when a steam temperature
change is introduced, indicating that there is a slight interaction of variables in this
direction.
5. Conclusions
Several complications with integrator MIMO processes were detected when the
multiloop control with PI/PID controller diagonal matrix is designed. From the analysis
of the results, it is possible to conclude that for integrator MIMO processes
1) the traditional BLT detuning method can fail since an unstable solution can be
found,
2) the traditional RGA and steady state decoupling can not be applied and
3) H∞ controller design technique shows to be the adequate method for these cases.
integrator elements, and ii) the designer wants to include a precompensator.
understand the interaction level more clearly and a precompensator must be taken into
account. Clearly, if a precompensator designed at a particular frequency is introduced
in the multiloop PI/PID controller design, a better performance in time domain can be
reached.
AAIQ Asociación Argentina de Ingenieros Químicos
IACCHE Interamerican Confederation of Chemical Engineering
XXII IACChE (CIIQ) 2006 / V CAIQ
Reference
Decentralized Control, Automatica, 22, 3, 309319.
Luyben, W.L. (1986). Simple Method for Tuning SISO Controllers in Multivariable
Systems, Ind. Eng. Chem. Process Des., 25, 654660.
Loh A.P., C.C. Hang, C.K. Quek and V.U. Vasnani (1993). Autotuning of Multiloop
ProportionalIntegral Controllers Using Relay Feedback. Ind. Engng. Chem. Res., 32,
11021107.
Mayne D. Q. (1973). The Design of Linear Multivariable Systems. Automatica, 9, 201
207.
Monica, T.J., C. Yu, W.L. Luyben (1988). Improved Multiloop SimgleInput/Single
969973.
Morari M. and E. Zafiriou (1989). Robust Process Control, Prentice Hall.
Skogestad S. and M. Morari (1989). Robust Performance of Decentralized Control
Systems by Independent Design. Automatica, 25, 119125.
Wiley.
Tan W., H. J. Marquez and T. Chen (2002). Multivariable Robust Controller Design for
a Boiler System. IEEE Transaction on Control Systems Technology, 10, 5, 735742.
Zhou K. and Doyle J. (1998). Essentials of Robust Control, Prentice Hall.
Youla D.C., H.A. Jabr and J.J. Bongiorno (1976). Modern WienerHopf Design of
Optimal Controllers – Part II. IEEE Trans., AC21, 3, 319338.
AAIQ Asociación Argentina de Ingenieros Químicos
IACCHE Interamerican Confederation of Chemical Engineering