Copyright © 2002 IFAC

15th Triennial World Congress, Barcelona, Spain

www.elsevier.com/locate/ifac

TUNING DECENTRALIZED PID CONTROLLERS

FOR MIMO SYSTEMS WITH DECOUPLERS

Francisco Vázquez Fernando Morilla

[email protected]. Dpto de Informática y Análisis Numérico.

Universidad de Córdoba. Córdoba (SPAIN)
[email protected]. Dpto. de Informática y Automática.
Facultad de Ciencias. UNED. Madrid (SPAIN)

Abstract: In this paper, a method for controlling multivariable processes is presented. The
controller design is divided into two parts: firstly, a decoupling matrix is designed in order
to minimize the interaction effects. Then, the controller design is obtained for the process
+ decoupler block. For this purpose, an iterative numeric algorithm, proposed by same
authors, is used. The aim is to meet the design specifications for each loop independently.
This sequential design method for multivariable decoupling and multiloop PID controller
is applied to several examples from literature. Decentralized PID controller design,
specifications analysis and time response simulations has been made using the TITO tool,
a set of m functions written in Matlab. It can be obtained in web page
http://www.uco.es/~in2vasef. Copyrigth  2002 IFAC

Keywords: PID controllers, multivariable control systems, decoupling.

1. INTRODUCTION
First group bases their design on some SISO method.
Most of industrial processes are multivariable in These formulas indicate the direction in which the
essence (Shinskey, 1995). For such systems, loop PID parameters have to be detuned to compensate the
interactions can arise and cause difficulties in interaction effects when all loops are closed. Method
feedback controller design. To solve this problem of Shinskey (1995) and BLT method (Luyben, 1992),
there exist several alternatives that can be centralized one of most cited in literature, are included in this
multivariable control (LQ, LQR and LQG, robust group. However, with these methods it is difficult to
control, predictive multivariable control, fuzzy establish a priori design specifications in all loops
control,...) or decentralized control. This last is the and they can be rather considered of trial and error.
most broadly used in the industrial environment.
With decentralized techniques, a multivariable A second group includes works that look for critical
system with n inputs and n output variables is treated gains of the system in order to tuning the PID
as n monovariable systems. The sacrifice that controllers. This gain can be obtained by means of
supposes the invariable performance deterioration of proportional controllers (Niederlinski, 1971), or by
a decentralized control structure when it is compared means of relay method (Zhuang and Atherton, 1994),
with a full multivariable control strategy is (Halevi et al, 1997), (Toh and Devanathan, 1993),
compensated with certain advantages as design and (Shiu and Hwang, 1998). Not all these methods
hardware simplicity or easiness of use. require a complete model of the systems.

Some of decentralized methods found in literature Lastly, methods included in a third group use the
could be classified under the following topics, whole transfer function matrix taking into account
showing the interest that it has raised in the last the interaction effects, and the controllers are
years: obtained by means of analytic, numeric or graphics

349
 methods. Some methods (Wang et al, 2000) use interactions brought about by cross coupling of the
optimization algorithms to obtain controllers, other process variables.
(Zhang et al, 2000) uses pole placement techniques,
and some (Ho et al, 1995) provides on-line tuning In literature there are different decoupling methods:
formulas. Method described in section 3 (Vázquez et Lineal decoupling (Desphande, 1989) is most
al, 1999) also belongs to this group. extended method. In this, the decoupling matrix try
to eliminate interactions from all loops, obtaining
But sometimes, interaction between variables or following elements for a 2 x 2 system.
system dynamics can impede the application of one g (s ) g (s )
of previous decentralized methods, and design d11=1, d 12 (s) = − 12 , d 21(s) = − 21 , d22=1 (1)
g11 (s) g 22 (s)
specifications are not met. In this case, and before
using some centralized method, there is an hybrid
alternative consisting of decomposing the design Implementation of this decoupling matrix has some
problem into two parts: firstly, decoupling the system problems: What happens if numerator has bigger
in order to minimize interaction or to make the order than denominator?, what if delays exist?, and
system diagonal dominant; then, design the what if these delays only appear in the denominator?.
controllers using some decentralized method. The The different works propose some solutions like Pade
final control system will be the product of the approximations for delays or steady state decoupling
decoupling and the controller matrixes. The as a first option to prevent this problems. Other
decouplers, together with single-loop controllers, solutions consist of applying partial decoupling,
constitute the multivariable controller. This setting d12 or d21 null. This fact avoid problems
alternative could be used as another design option. originated in one loop reach the other.
But not all decentralized design methodologies are
prepared to do that. Section 2 described how There are some other methods: one of them looks for
decoupling can be done, and section 3 how a diagonal dominance in the system as ALIGN
decentralized design technique (Vázquez et al, 1999) algorithm designed by McFarlane and Kouvaritakis
fulfill this purpose. in 1974 and described in (Maciejowski, 1989). Other
design by means of singular value decomposition
(Desphande, 1989) or by means of inverse
2. DECOUPLING DESIGN decoupling (Wade, 1997) that produces the process
input signals by combining one controller output with
The design of a decentralized control system with a the other process input signals.
decoupling matrix can be done combining a diagonal
controller Kd(s) with a block compensator D(s), so Even if decouplers are incorporated, the interaction
that the controller manipulates the variable u’i instead effects cannot be completely eliminated because of
of the ui, as can be appreciated in figure 1, for the 2 x model mismatch. Then, single-loop controllers
2 case. With this configuration the controller see the cannot be tuned independently, and a sequential
process as a set of n completely independent tuning algorithm that takes interactions into account
processes or with the interaction minimized. should be used. Next section describes one of these
algorithms.

u’1
k1 d11 + u1 g11 y1
r1 + + 3. MIMO DESIGN METHODOLOGY
+
d21 g21
Kd(s) D(s) Analysis of decentralized methods described in
section 1 show several interesting points:
g12
d12
y2 - Obtaining different design specifications for
+
r2 k2
u’2
d22
+ g22 + each loop implies solution methods based on
+ u2
some kind of iteration. If designs are carried out
loop-by-loop, tuning one of them can detune the
others.
Fig. 1 : General 2 x 2 system with decouplers and - In spite of abundance of methods, there are no
single-loop controllers simple and general solutions for the
multivariable tuning problem. Or methods are
The essence of decoupling is the imposition of a used by optimization (Wang et al, 2000) that do
calculation net that cancels the existent process not guarantee an appropriate solution and they
interaction, allowing the independent control of the require several iteration processes, or the
loops. In decentralized design, the question is not to solutions are too particular, for concrete models
eliminate interaction, but to take it into account. A (Ho et al, 1996) or too simple (Shinskey, 1996),
multivariable system may still experience (Luyben, 1992) that only get a first approach to
interactions and responds poorly. The objective in the problem.
decoupling is to compensate for the effect of

350
In (Vázquez et al, 1999), a method of tuning PID  K1 0 
controllers for systems with decentralized control is K =   (2)
 0 K 2 
presented. Its fundamental characteristics can be
summarized in following points:
- It is a generic method for n x n systems with The process element gii has its feedback loop open.
decentralized control. The rest of elements of its same row are called G12.
- It is a method based on successive SISO tuning The rest of elements of its same column are called
that does not suppose any additional constraint G21. And the rest of elements of G are called G22.
to obtained controllers (except for their Then, the process matrix G can be written as:
decentralized structure) neither to the transfer  g ii (s) G 12 (s) 
functions matrix. This matrix can include the G(s) =   (3)
transfer functions of only the process or also  G 21 (s) G 22 (s) 
including the decoupling net. Present work
exploits this characteristic. The process seen from input i to output i when the
- The SISO methodology integrated in the other loops has been closed is called ~ g i and it is
algorithm has certain imposed limitations: it obtained from
should allow the controller be designed from a ~
g i = g ii − G 12 K 2 (I + G 22 K 2 )−1 G 21 (4)
frequential description of the process (their
frequency response). And it should quantify the where, using the above notation, gii is an element of
specifications achievement by means of a quality G, that is, a SISO transfer function, G12 a transfer
index, J. This index controls the iteration function row vector of dimension 1 x (n-1), G21 a
evolution of the tuning algorithm. transfer function column vector of dimension (n-1) x
- The method allows a decoupling matrix to be 1, K2 is a diagonal square matrix of dimension (n-1)
included between the plant and the controllers. and I is the identity matrix of order n-1. In the case of
- The design methodology is divided into two 2 x 2 processes, elements G12 and G21 are also
phases that are carried out in a sequential and individual SISO transfer function elements of G, and
combined way: the structural decomposition and the calculation of ~ g i do not imply matrix operations.
the controller design. These two phases are
described next. Opposite to standard transfer functions, the ~
g i are not
an intrinsic property of the system, and they depend
3.1. System structural decomposition on the designed controllers, as can be observed in
equation (4). Possibly, they have not a representation
This phase consists of the decomposition of a n x n by means of Laplace transformation.
multivariable system into n SISO systems. To solve
this problem the structural decomposition, introduced The structural decomposition implies some important
by Zhu (1996), is used. Let be a n input system considerations for the decentralized control: some
controlled by means of a decentralized control global system properties, as interaction, stability,
strategy. Now, n-1 loops have been closed by means integrity, etc, could be deduced from the properties
of n-1 controllers. The process ‘seen’ from the free of the obtained n SISO subsystems.
input to the free output is needed. The scheme
corresponds to the figure 2, where loop i has not yet Also, no supposition about G has been made, and
been closed (the pairing problem has been already then, structural decomposition could be applied to
solved and input-output pairs are in the diagonal). every transfer functions matrix. Applying Nyquist
theorem to each individual loop obtained with
ri ~
gi
ui structural decomposition, stability criterion is more
K1 1 y
interesting from the point of view of applicability to
1 G decentralized multivariable systems. Then, the n
SISO systems stability implies MIMO stability.
n-1 n-1
1 Theorem 1: Supposing that individual elements of
K2 G(s) and their SISO independent subsystems do not
1 have poles in the right hand plane, the system with
decentralized control is stable if and only if the
Fig. 2: Structural Decomposition Nyquist contour of the equivalent open loop transfer
function k i ~
g i does not encircle point (-1,0), i.e.
In the structural decomposition scheme, K1 is the N (− 1, k ~
g ) = 0 ∀i
i i (5)
controller that closes the loop between input i and
output i. The set of n-1 controllers, whose loops are
already been closed, is K2. This way, the controller The structural direct Nyquist arrays (SDNA) are the
matrix is: representation of the n direct Nyquist diagrams of
N (− 1, k i ~
g i ) = 0 ∀i .

351
3.2 Controllers design between reflux and vapor flow and the temperatures
of plates 4 and 17:
The proposed is a frequency-based method and  2.2 −s 1.3 −0.3s 
stability measures as phase and gain margins could  R   7s + 1 e 7s + 1
e  T 
be used. But most of existent methods requires a   =    4  (6)
V
   2 . 8 −1 . 8s 4. 3 − 0 . 35s   T17 
model for obtaining the design parameters. Carrying e e
 9s + 1 9.2s + 1 
out the structural decomposition, the resulting
transfer function ~g i could be not a rational function,
In order to analyze the effects of designing with
possibly due to the inclusion of delays in model decouplers, expressions (1) are used. In this case,
systems. In order to solve this problem a frequential dynamic and steady state decoupling coincide:
representation of this transfer function ~g i it is used, d11=1, d12=-0.59, d21=-0.65 and d22=1 (7)
i.e. a frequency response array. But this take implicit
another problem: what SISO design method to use? Design specifications are phase margin (PM) of 45º
The particular solution has been the monovariable and gain margin (GM) of 4 for both loops, and
methodology described in (Morilla and Dormido, looking for a PID with a limit of the relationship
2000). This methodology supposes, and in fact between the derivative and integral constant (α)
requires, a model with the description of the system equal to 0.01.
dynamics. However it leaves open the possibility to
be adapted to solve the problems outlined previously: The design using iterative algorithm without
it can be modified allowing the controllers to be decouplers gets following controllers:
designed from an array with the frequential  1 
representation of the frequency response and also, the k1 = 0.88 1 + + 0.035s  and
 1.82s 
obtained results can be compared with the design
specifications to present an indicative of quality,  1 
k 2 = 2.70 1 + + 0.073s  (8)
indispensable for the proposed solution search  1. 79s 
method. These two phases cannot be considered
isolated, since the design process consists of an Desing specifications are met with combined (PM
iteration of both. This is because, as equation (4) and GM) tuning design in three iterations. Interaction
shows, the ~ g i transfer functions, obtained with impedes using other methods based on Gershgorin
structural decomposition, incorporate the controller bands, as Ho’s method.
transfer functions in their definitions. Thus, every
time a design is done for one of these transfer Next, a complete decoupling is chosen. Design
functions, it will be needed to recalculate the others. specifications are the same as before (PM=45º and
From the new ~ g i , a new controller could be GM=4 for both loops). The algorithm converges in
same number of iterations as before, obtaining
designed, and then, definitely, the algorithm needs an
following controllers
alternation between the two phases.
 1 
k1 = 0.964 1 + + 0.057s  and
 2.81 
4. EXAMPLES  1 
k 2 = 0.967 1 + + 0.014s  (9)
 0. 766 s 
One of advantages that presents this methodology of
multivariable controller design is to be independent
With this design, specifications are also met, but
of the process model. This has been shown in
now, controlled system was decoupler + process
numerous examples in (Vázquez et al, 1999), which
block. Time responses are shown in figure 3.
can be repeated with TITO tool. Also, the model does
Response without decouplers is also superimposed.
not have reason to be a rational expression in the
1 .5
Laplace operator, but rather it can be a frequency
response array. These characteristics allow the 1
R (% )

immediate extension of this method, using it 0 .5

accompanied by some decoupling strategies (one of 0

the described in section 2 or another), so that in a -0 .5

first step the decoupling matrix is obtained and, in a 0 20 40
T i m e (s e c )
60 80 100

second step, the controllers for the decoupler+process 1 .5

block designed. Some of the possibilities are shown 1
in following examples.
V (% )

0 .5

Example 1: Following shows a system model with a -0 .5

0 20 40 60 80 100
RGA next to 1.7. Some decoupling strategy could be T i m e (s e c )

tested. It is a distillation column (Vinante and Fig. 3: Time response of example 1 with decoupler
Luyben, 1972) that describes the existent dynamics and without decouplers (continuous)

352
 A slight modification can be appreciated, mainly in Example 3: This third example studies a process
second loop, and an interaction decrease on the first proposed by Niederlinski (1971) and analyzed with a
one. Although modifications in time response are set of decouplers proposed by Shiu and Hwang
minimum, design is good to show methodology (1998). The process function transfer matrix is:
effectiveness when decouplers are incorporated. 

0.5 −1 

 y1 (s)   (0.1s + 1) 2 (0.2s + 1) 2 (0.1s + 1)(0.2s + 1) 2   u1 (s) 
 y (s)  =  1 2.4   u 2 (s) 
Example 2: Table 1 shows the results of PID  2 
 2 2 
controller design for the system  (0.1s + 1)(0.2s + 1) (0.1s + 1)( 0.2s + 1) (0.5s + 1) 

 12.8 − 18.9 −3s  (11)

−s
 X D (s)   16.7s + 1 e 21s + 1
e   F (s) 
(10)
  =   R  Shiu proposes the following decouplers, obtained
 X B (s)   6.6 e −7s − 19.4 −3s   FS (s) 
 10.9s + 1 14.4s + 1
e 
 after identifying the ~g i functions applying a relay
for different decoupling nets (dynamic, in steady method:
state, partial and total) and for different controllers 2(0.587s + 1)(0.0676s + 1)
d 12 = and
(PI, PID). All the designs have been carried out with (0.379s + 1)(0.2x0.0676s + 1)
specifications of PM=60º and GM=4 for both loop. It 0.417 (1.06s + 1)( 0.197 s + 1)
is a water-methanol distillation column (Wood and d 21 = − (13)
Berry, 1973) analyzed in numerous later works (Ho (0.48s + 1)( 0.2 x 0.197 s + 1)
et al, 1996), (Toh and Devanathan, 1993), (Vázquez
et al, 1999) (it appears in the practical whole of In order to compare time responses with and without
multivariable control references). decouplers, the tuning algorithm has been used with
the same design specifications (PM = 60º for both
d12 d21 Kp1 Ti1 Td1 Kp2 Ti2 Td2 loops).
0 0 0.51 12.62 0 -0.027 2.46 0
18.9/12.8 0 0.48 8.05 0 -0.030 8.94 0
0 6.6/19.4 0.35 9.82 0 -0.031 2.39 0 Without decouplers, controllers are
18.9/12.8 6.6/19.4 0.34 6.11 0 -0.027 9.19 0
18.9(16.7s + 1)  1  and
12.8(21s + 1)
0 0.49 8.72 0 -0.038 9.90 0 k1 = 1.11 1 + 
 0.67 s
6.6(14.4s + 1)
0 0.33 7.60 0 -0.033 2.925 0
19.4(10.9s + 1)  1 
k 2 = 0.27  1 +  (13)
18.9(16.7s + 1) 6.6(14.4s + 1)
0.32 5.67 0 -0.031 8.08 0  0.88s 
12.8(21s + 1) 19.4(10.9s + 1)
18.9(16.7s + 1) 6.6(14.4s + 1) obtained in six iterations of the algorithm. With
0.34 4.91 0.34 -0.046 9.626 0.759
12.8(21s + 1) 19.4(10.9s + 1) decouplers, the controller equations are
 1  and
k 1 = 0.91 1 + 
Table 1: Comparative of different decouplers  0.23s 
and designs of example 2
 1 
k 2 = 0.26 1 +  (14)
Figure 4 shows superimposed time responses of the  0.37s 
system without decouplers and with total dynamic obtained in only three iterations, because, in general,
decouplers, with a PI and a PID (the two last designs the algorithm converges quicker the more dominant
of table 1). The PID design has been carried out the system is.
limiting α to 0.25.
1 .4 Figure 5 shows time response, with and without
decouplers. Note that interaction decreases. If this
1 .2

fact is quantified by means of some measure as the

X D (% )

0 .8

0 .6

0 .4
IAE test, it can be proven that in the response without
0 .2 decouplers, IAE is 3.42, while when decouplers are
0
0 20 40 60 80 1 00
Tim e (s ec )
1 20 1 40 1 60 1 80 2 00 applied it is 0.98.
1 .4

1 .2
1 .5
1
1
0 .8
y1 (% )

0 .5 0 .6
X B (% )

0 .4
0
0 .2
-0 .5
0
0 1 2 3 4 5 6 7 8 9 10
-1 Tim e (s e c)
0 20 40 60 80 1 00 1 20 1 40 1 60 1 80 2 00
Tim e (s ec )
1 .2

Fig. 4: Time response with decouplers (PI, and PID) 1

and without decouplers (continuous). 0 .8

0 .6
y2 (% )

0 .4

Note that interaction effects decrease not very 0 .2

significantly. This same reduction could be obtained 0

-0 .2

tuning the second loop in a less aggressive way (for 0 1 2 3 4 5

Tim e (s e c)
6 7 8 9 10

example PM=80º), which would reduce its effect on Fig. 5: Time response of the system of example 3,
first loop. with and without decouplers (continuous).

353
 Gerhgorin bands and SDNA of the process with and REFERENCES
without decouplers can also be compared in figures 6
and 7. Note that dynamic interaction effects are Deshpande, B. (1989): “Multivariable process
reduced significantly. control”. Instrument Society of America.
2
D N A : E le m e n t 1 1
2
D N A : E le m e n t 2 2 Halevy, Y, Palmor, Z., Efrati,T. (1997): Automatic
tuning of decentralized PID controllers for MIMO
processes, J. Proc. Cont. Vol. 7 Nº 2, pp 119-128
1 .5 1 .5

1 1 Elsevier Science, Ltd.

0 .5 0 .5
Ho, W.K., T.H. Lee, O.P. Gan (1996): Tuning of
multiloop PID controllers based on gain and
Im a g i n a ry

Im a g i n a ry

phase margins specifications. 13th IFAC World

0 0

- 0 .5 - 0 .5 Congress, pp 211-216.
-1 -1
Luyben, W.L. (1992): Practical distillation control.
Editor W.L Luyben.
- 1 .5 - 1 .5
Maciejowski, J.M. (1989). In: Multivariable feedback
-2
-2 -1 0 1 2
-2
-2 -1 0 1 2
design. Addison-Wesley
Real Real
McAvoy,T. (1983). In: Interaction analysis:
Fig. 6: SDNA and Gershgorin bands of system of principles and applications. Ins. Soc. of America.
example 3 with decouplers Morilla, F.; Dormido, S. (2000): “Methodologies for
the tuning of PID controllers in the frequency
D N A : E le m e n t 1 1 D N A : E le m e n t 2 2
2 2 domain”. Proceedings of PID’00. Tarrasa,
1 .5 1 .5 pp.155-160
1 1
Niederlinski, A. (1971): “A heuristic Approach to the
design of linear multivariable interacting control
0 .5 0 .5
Im a g i n a r y

Im a g i n a r y

systems”. Automatica, Vol. 7, pp. 691-701.

0 0
Shinskey (1995): “Process Control Systems”.
- 0 .5 - 0 .5
MacGraw-Hill.
-1 -1 Shiu, S.J.; Hwang, S. (1998): ”Sequential design
- 1 .5 - 1 .5
method for multivariable decoupling and
-2 -2
multiloop PID controllers”, Ind. Eng. Chem.
-2 -1 0
Real
1 2 -2 -1 0
Real
1 2
Res., Vol. 37, Nº 1, pp. 107-119.
Toh, K.; Devanathan, R.(1993): “An expert
Fig. 7: SDNA and Gershgorin bands of system of autotuner for multiloop SISO controllers”.
example 3 without decouplers. Control Eng. Practice, Vol. 1, Nº 6 pp. 999-1008.
Vázquez, F.; Morilla, F.; Dormido, S. (1999): ”An
iterative method for tuning decentralized PID
CONCLUSIONS controllers”. Proceeding of the 14th IFAC World
Congress, pp. 491-496.
This work has been developed to design Vinante, Luyben (1972): “Experimental studies of
multivariable controllers including a decoupling net. distillation decoupling”, Control Eng. Practice,
It takes advantage of methodology proposed in Vol. 1, Nº 6 pp. 999-1008
(Vázquez et al, 1999), and it does not need a transfer Wade, H. L.(1997): “Inverted decoupling: a
function matrix with the system model but a neglected technique”, ISA Transactions, Vol. 36,
representation with frequency response. Nº 1, pp 3-10.
Independently of how decouplers had been obtained, Wang, Q.,; Lee, T; Fung, H. (2000): ”Independent
they can be included between process and design of multi-loop controllers taking into
controllers. Then, a decentralized PID controller is account multivariable interactions”, Journal of
designed, using some of the different possibilities of Chemical Engineering of Japan, 33 427-439.
proposed algorithm: design with only phase margin Wood, R. K.; Berry, M. W. (1973): ”Terminal
specifications for both loops, only gain margin, or a composition control of a binary distillation
combination of PM and GM specifications. column” Chem. Eng. Sci., 28, pp. 707-1717.
Zhang, Y; Wang, Q.; Astrom, K. (2000): “Dominant
All designs have been obtained with TITO tool, a set Pole Placement for Multi-Loop Control Systems”.
of MATLAB m-functions that can be obtained in American Control Conference, Chicago, USA,
web page http:// www.uco.es/~in2vasef. June 29, 2000.
Zhuang, M.; Atherton, D. (1994): “PID controllers
design for a TITO system”. IEE Proc. Control
ACKNOWLEDGEMENTS Theory Appl. Vol. 141, Nº-2, pp. 111-120.
Zhu, Z.X. (1996): ”Structural analysis and stability
This work was supported by the CICYT (Comisión conditions of cecentralized control systems”. Ind.
Interministerial de Ciencia y Tecnología) under Eng. Chem. Res., Vol. 35, Nº 3, pp. 736-745
Grant TAP 96-0404.

354