SPEEDAM 2010

International Symposium on Power Electronics,

Electrical Drives, Automation and Motion

Advanced PID controllers in MIMO systems

A. Balestrino*, E. Crisostomi*,**
* Department of Electrical Systems and Automation, University of Pisa, Largo L. Lazzarino 1, 56126, Pisa, (Italy)
** Interdepartmental Research Center "E.Piaggio", University of Pisa, Largo L. Lazzarino 1, 56126, Pisa, (Italy)

solved trivially by the knowledge of the physics

Abstract-- This paper revisits the problem of underlying the dynamical system, this can still be solved
decentralised control of Multiple Input Multiple Output according to the classic simple Relative Gain Array
(MIMO) systems. In industrial process control this problem method proposed by Bristol [8] in 1966 or to more
is typically solved through the use of Proportional Integral sophisticated techniques [9]. Both methods require
and Derivative (PID) regulators that are used to close single however the availability of the matrix transfer function of
loops, according to the implicit assumption that interactions
the system.
between input and output variables can be neglected.
However special care and conservatism are required for the
choice of the regulator gains so that stability can be Applications of decentralised control of MIMO systems
preserved also in the presence of the interactions. The cover all areas of process/industrial control, and the
decentralised MIMO control problem is solved here by following references are limited to some papers regarding
extending the Variable Structure - Proportional Controller different aspects of interest within power system control:
(VS-PI) proposed in [1] to the MIMO case, and considering [10] for optimal decentralised load frequency control,
a priority-based approach for the sequential control of the [11] for the design of a decentralised Power System
single loops. In this paper the general idea is presented and Stabilizer (PSS) in a multi-machine power system and
preliminary simulation results including a robot control
using brushless DC motor drives are included
[12] where interactions are compensated to enlarge the
operative range of a power plant. In this paper each
Index Terms—MIMO Systems, PID Control, Variable decentralised controller is a slight modification of a
Structure System. simple Proportional/Integral (PI) regulator, as this is the
solution more welcomed by process engineers, both for
I. INTRODUCTION its simplicity and because it does not involve accurate
modeling (or identification) of the system plant. The
The history of PIDs for process control goes back to the choice of PI controllers rather than full PID is not
pioneering works of Sperry [2] and Minorski [3], and it constraining, as the derivative action can still be added
has been improved by the well known tuning rules of according to standard techniques. Here it is not
Ziegler and Nichols [4], and the more recent updated considered mainly because proposed improvements only
rules by Astrom and Hagglund [5] and Luyben and involve Proportional and Integral action, and considering
Eskinat [6]. The use of PIDs has also been widely applied the derivative action could be misleading. Besides, it is
to MIMO control, although it raises a much more very common that process engineers prefer not to use the
challenging problem due to loop interactions. derivative action in industrial PIDs.

A popular empirical tuning method for the MIMO case, The contribution of this paper is twofold: first the single
at least in the field of chemical process control, is the so PID controllers obtained by applying the Luyben method
called Biggest Log modulus Tuning (BLT) method, are improved according to the form proposed in [1].
introduced by Luyben in 1986 [7]. According to this Secondly, priority driven set-point weighting strategies
procedure, which will be also used in this work both as a are employed to reduce overshoots for set-point responses
starting base and as a comparison term, single PID and the interactions between the single loops.
controllers are obtained using Ziegler-Nichols tuning
rules, and they are detuned (i.e. some conservatism in the The proposed approach is tested only in simulation on a
regulator gains is introduced) to preserve stability also in classic challenging chemical process problem [13] and on
the presence of interactions which had not been taken into the Robot Axis Control problem using Brushless DC
account previously. It is important to remark that this Motor Drive modeled in Matlab/Simulink [14]. This
approach, as in general all decentralised control methods, paper is organised as follows: next section describes the
requires a prior knowledge of which input should be used MIMO control problem and the proposed solution.
to control a certain output, which is known as the pairing Section 3 shows the performances of the proposed
problem. Whenever the pairing problem can not be controller while compared with classic regulators.
Finally, in the last section the achieved results are
summarised and future lines of research are outlined.

978-1-4244-4987-3/10/$25.00 ©2010 IEEE

693
 II. DECENTRALISED MIMO CONTROL between the corresponding Proportional controller and
the corresponding PI, it always requires a smaller control
The problem of closed-loop control is usually described
effort and it has the beneficial side-effect property of
in the classic block diagram of Fig. 1.
decreasing undesired overshoots. Here, the same
approach is extended to the MIMO case to design single
loop regulators, where a second modification is added
through the term γ (i ) which provides a set-point
weighting action. Benefits of set-point weighting are
well-known, see for instance the contribution of Chen and
Seborg [14] where it is reminded that set-point weighting
Fig. 1. Closed-loop control system
both reduces overshoots and loop interaction. We propose
to change the value of the set-point weighting factor γ (i )
In the general MIMO case, assuming that the system P of according to a priority based method which assigns more
Fig. 1 is linear and it has n inputs and n outputs, then it is importance to the loops which are desired to behave
generally described by an n× n full matrix transfer faster. On the other hand, the other loops are detuned so
function. We suppose that the inputs are re-numbered so to reduce their interference with the most important
that input i corresponds to output i according to a pairing loops. Preliminary evaluation of the proposed MIMO
procedure (for instance Relative Gain Array (RGA) [8] or controllers is shown in the next section for some simple
Absolute RGA (ARGA) [9]). Therefore the objective is to simulation examples.
design a diagonal matrix transfer function C where the
diagonal terms are PI controllers. It is reminded here that III. EXAMPLES
while a full matrix C is preferable as it provides more
degrees of freedom to shape the desired control of the The first example taken from [14] concerns an industrial-
system, a diagonal (decentralised) control can be more scale polymerization reactor control problem. After a
convenient as it requires the design of a smaller number model identification procedure, [14] proposes the
of regulators and eventual failures only affect the following matrix transfer function, where here
involved loop. disturbances are not considered:

According to the classic notation, in Fig. 1 the error e(t )

ª 22.89 −0.2 s 11.64 º
is a vector corresponding to the point-wise difference « 4.572s + 1 e − e −0.4 s »
ª y (1) (s )º 1.807 s + 1
» ª u (1) (s )º (2).
between the vector reference signal r (t ) and the output « »=
«
» «u (2 ) (s )»
¬ y (2 ) (s )¼
«
¬ ¼
vector y (t ) . The controller provides the vector signal « 4.689 e −0.2 s 5.80
e −0.4 s »
[
u (t ) = u (1)(t ), u (2) (t ), ..., u (n ) (t ) , where ] each
¬« 2.174s + 1 1.801s + 1 ¼»

component u(i ) (t ), i = 1,..., n is computed as The time scales are in hours, so the process dynamics are
slow, and the interaction between the two components is
u (i ) (t ) = [ ]
k (pi ) γ (i )r(i ) (t ) − y (i ) (t ) + quite strong as there is no diagonal dominance.
e (τ )2 º
Both RGA and ARGA methods suggest a diagonal
ª − i2 2
t« β σ » (1) pairing, therefore the first input controls the first output,
+ ³ «k i(i )e(i ) (τ )exp (i ) (i ) »dτ
2
and the same does the second input with the second
0
« » output. According to Luyben’s tuning procedure and the
¬ ¼
notation of (1), we find controller gains k (p1) = 0.213 ,
where k (pi ) and ki(i ) are the proportional and integral
gains of the i_th PI; γ (i ) and β (2i )σ (2i ) are two extra ki(1) = 0.096 , k (p2 ) = 0.178 and ki(2 ) = 0.043 . Extra
parameters which are used to make the priority based set- parameters β (i ) and σ (i ) are chosen equal to 5 and 1
point weighting and to delay the integral action until respectively for both control loops. These values are
settling stage is reached; again, the index (i ) has the known to provide a good trade-off between the
objective of reminding the input-output pairing. More corresponding proportional and PI response [1]. Finally,
details about the tuning of the parameter β (2i )σ (2i ) can be it is (arbitrarily) decided to give more importance to the
first channel rather than to the second one, and the set-
found in [1]. Here, we remark that the exponential term in point weighting factors are chose as γ (1) = 0.5 and
(1) is always constrained between 0 (no integral term)
when the error is large, and 1 (full contribution of the γ (2 ) = 0.25 . As a consequence the response is much
integral term) when the error tends to zero. As a slower, but overshoots and interactions are minimal, as it
consequence the integral action is reserved for is shown in Fig. 2 in the case of the response to a constant
accomplishing fine tracking tasks while it is reduced unit signal on both the channels.
during the transient stage when the proportional action is
enough to get close to the desired reference. As proved in As can be seen from Fig. 2, the 'important' component has
[1] the response of the modified PI is always included the same settling time of the original case, without

694
overshoots. The second component becomes secondary Absolute value of the control effort, and it is should be
through a smaller weighting factor so to avoid undesired small to have a reduced average control effort; IADU is
interactions with the important component. Weighting the Integral of the Absolute value of the Derivative of the
factors can be modified to stress or to relax the priority control action, and it should be small to avoid stress of
classification of output variables. the control actuator.

The second example concerns a robot axis control

problem using Brushless DC Motor Drive, provided by
Matlab/Simulink [15] and it is available as a demo in the
SIMPowerSystem Toolbox. This example may be more
of interest for the power system community and intends
to prove that even without an accurate model of the
process plant in exam, the proposed small modifications
to classic PI controllers often provide immediate and
evident advantages. In the example a step in the angle
position of two joints is the reference command and the
original conventional decentralised control involving two
PI regulators is modified according to equation (1), where
the free parameters β (i ) and σ (i ) are chosen again equal
Fig. 2. Response of the classic Luyben controller and the proposed one to 5 and 1 respectively for both the links, while the set-
are shown with dashed and solid line respectively. Output y (1) is shown point weighting factors are again γ (1) = 0.5 and
with light colour, while y(2 ) is in dark colour.
γ (2 ) = 0.25 , thus higher with priority given to the first
link. Figure 2 compares the original conventional PI
Results obtained here can be further improved, for
MIMO control with the proposed one for a step response.
instance by inserting a static decoupler to increase
It should be noticed that the evident improvements have
independence among variables [14], however this
been achieved without changing the PI gains (which are
solution is avoided here because the interest is to modify
the same in both controllers), but only thanks to the
the standard PI structure as less as possible so to make it
priority set-points weighting and to the reduced integral
appealing for industrial control applications. Finally, it is
action, which is reserved for the final tuning action. (i.e.
also important to remark that not only the response in
with the same gains, but without priority set-point
case of control with the proposed PI is smoother
weighting and with full integral action during the control
compared to the conventional one, but there is also an
effort).
important reduced control effort which is both caused by
preserving the integral action only for the last stage, but
also thanks to the set-point weighting which avoids the
kicking behaviour of the proportional action. In particular,
we emphasise that there is an important reduction also in
the derivative of the control action, as illustrated in Table
I. This is particularly appealing if the control action
involves the motion of mechanical actuators when a
reduced stress of the actuator is mandatory.

TABLE I. COMPARISON OF THE CONTROL EFFORTS REQUIRED BY

THE CONVENTIONAL PI AND THE PROPOSED VARIABLE STRUCTURE PI.
THE COMPARISON INVOLVES BOTH CONTROL LOOPS.

IAU IAU IADU IADU

Controller /
First Second First Second
Control Index Fig. 3. Response of the classic Luyben controller and the proposed one
Channel Channel Channel Channel
are shown with dashed and solid line respectively. Output y (1) is shown
below with dark colour, while y(2 ) is above in light colour.
PI 9.516 9.547 0.3051 0.2082

VSPI 9.201 9.051 0.0877 0.1162 As can be seen from Figure 2 the proposed controller
again provides a smoother response which reduces
Regarding the control properties summarised in Table I, overshoot in the reference following. Also notice that in
we remind that IAU represents the Integral of the this case the reduced control effort does not slow down
the time required to achieve steady state behaviour.

695
Statistics concerning the control properties are and voltage control with the proposed priority based
summarized in Table II. approach (giving more priority to the frequency control).

TABLE II. COMPARISON OF THE CONTROL EFFORTS REQUIRED BY V. REFERENCES

THE CONVENTIONAL PI AND THE PROPOSED VARIABLE STRUCTURE PI.
THE COMPARISON INVOLVES BOTH CONTROL LOOPS.
[1] Balestrino, A.; Biagini, V.; Bolognesi, P.; Crisostomi, E.:
Advanced Variable Structure PI Controllers, IEEE
Conference on Emerging Technologies and Factory
Automation, Mallorca, Spain, September 2009.
Settling [2] Sperry, E.: Automatic steering, Society of Naval Architects
IAU IADU Time and Marine Engineering, 1922.
Controller / Control First / First / (ˢ ≤  ) [3] Minorski, N.: Directional stability of automatically steered
index Second Second First / bodies, Journal of American Society of Naval Engineers,
Channel Channel Second 1922.
Channel [4] Ziegler, J.; Nichols, N: Optimum settings for automatic
controller, Trans. ASME, vol. 75, pp. 827-833, 1942.
[5] Astrom, K.; Hagglund, E: Adaptive tuning of simple
15.626 / 178.556 / 1.2720 / regulators with specifications on phase and amplitude
PI
8.901 133.794 1.0387 margins, Automatica, vol. 20, pp. 645-651, 1984.
[6] Luyben, W.L.; Eskinat, E.: Nonlinear autotune
identification, International Journal of Control, vol. 59, pp.
7.353 / 98.768 / 1.1446 / 595-626, 1994.
VSPI
1.996 52.925 0.7752 [7] Luyben, W.L.: Simple method for tuning SISO controllers
in multivariable systems, Ind. Eng. Chem. Des. Dev., vol.
25, pp. 654-660, 1986.
The settling time computed in the last column of Table [8] Bristol, E.H.: On a new measure of interaction for
II refers to the time instant from which the absolute value multivariable process control, IEEE Trans. on Automatic
of the error between the reference signal and the actual Control, vol. 11, 1966.
output remains definitively smaller than 0.001. Settling [9] Balestrino, A.; Crisostomi, E.; Landi, A.; Menicagli, A:
time is expressed in seconds. ARGA loop pairing criteria for multivariable systems,
IEEE Conference on Automation and Control, Cancun,
IV. CONCLUSIONS Mexico, 2008.
[10] Feliachi, A.: Optimal decentralized load frequency control,
IEEE Trans. on Power Systems, vol. PWRS-2, no. 2, pp.
This paper presents preliminary results achieved on a new 379-385, 1987.
way of implementing a MIMO decentralised control. The [11] Yang, T.C.; Yu, H.; Young, R.C.D.; Chatwin, C.R.: Power
main motivation behind this work is that several robust system stabilizer design based on structured singular
and analytically proven methods exist to perform MIMO values, Int. Journal of Power and Energy Systems, vol. 25,
no. 2, 2005.
control, but they usually do not correspond to the well
[12] Garduno-Ramirez, R.; Kwang Y. Lee: Compensation of
established practice of using simple PID controls to control-loop interaction for power plant wide-range
regulate loops independently. At the first stage, the operation, Control Engineering Practice, vol. 13, pp. 1475-
existing interactions are therefore neglected, then the 1487, 2005.
controllers are detuned in the hope that by adding some [13] Chien, I.-L.; Huang, H.-P.; Yang, J.C.: A simple
conservatism it is possible to preserve stability also in the multilooptuning method for PID controllers with no
presence of the interactions. Here, small improvements proportional kick, Ind. Eng. Chem. Res., vol. 38, pp. 1456-
are suggested which only slightly complicate the usual 1468, 1999.
PIDs, while still preserving the industrial philosophy [14] Chen, D; Seborg, D.E.: Multiloop PI/PID controller design
method based on Gershgorin bands, Proc. Of the
behind MIMO control. Some preliminary results are
American Control Conference (ACC), Arlington, 2001.
shown within this paper, which should emphasise that [15] Demos of the Simulink Power Systems, description is
good improvements can be achieved independently from available online from URL
the particular applicative field of interest. http://mathworks.com/access/helpdesk/help/toolbox/physm
od/powersys/index.html
As a future work, properties of the proposed controller
will be tested in more realistic and severe situations,
taking also into account disturbances and plant parameter
variations. A specific case study will be developed within
the field of power system control. In particular, the
interest concerns the modeling of Power System
Stabilizers, which are widely studied in the literature as
Single Input Single Output (SISO) systems (or in a few
cases as Multiple Input Single Output (MISO) systems),
as MIMO systems and perform a frequency (or angular)

696