Multivariable Control

Chapter 81
81.1 Case Study
81.2 Compensator Concept
81.3 Control System Model
81.4 Compensator Design
81.5 Worked Example No 1
81.6 Decouplers
81.7 Sampled Data Model
81.8 Impulse Compensator Design
81.9 Worked Example No 2
81.10 Sampled Data Decoupler
81.11 Comments

Some of the techniques of state-space were intro-

duced in Chapter 80. The emphasis there was prin- CC
cipally on the analysis of the behaviour of multi-
variable systems. This chapter introduces the use c0
f2
of matrix techniques, and diagonalisation in par- f0
ticular, for design purposes. A case study based #
c
upon a blending system is used as a realistic, but f1
simple,example to demonstrate the principles.The
approach used is known as internal model control FC
because the inverse of the model of the plant to be
controlled is embedded in the design of the con- Fig. 81.1 An in line blending system
troller. This gives an insight to the problems of
the design of large multivariable control systems
dynamics are fairly fast because the blending sys-
which provides a basis for considering related tech-
tem is in-line. Also it is a highly interactive system.
niques such as sensitivity analysis in Chapter 111,
Any change in either flow f 1 or f 2 will affect both f 0
state feedback in Chapter 112 and model predictive
and c# . Note that changes in flow are propagated
control in Chapter 117.
throughout the system instantaneously, whereas
changes in concentration have an associated time
delay.
81.1 Case Study A model of this system in matrix form is devel-
oped in Chapter 86 using deviation variables:
An in-line blending system is depicted in Fig-     
f0 (s) 1 1 f1(s)
ure 81.1. = (81.1)
In summary, two streams f 1 and f 2 are blended c# (s) −K1 .e−Ls K2 .e−Ls f2(s)
to produce a third stream f 0 of concentration c# . Its which may be denoted:
 674 81 Multivariable Control

fm
x(s) = P(s).f (s) M1 (s)

fr +
- e1 u1 f0
f1
C1 (s) V1 (s)
+
This is depicted in block diagram form in Fig- +

ure 81.2.
− K1
f1 f0
+ c0
+ cr e2 u2 + c#
+ C 2 (s) V2 (s) K2 e − Ls
+ f2
-
M 2 (s)

− K1 Fig. 81.3 Model of blending system with simple feedback loops

+ However, to minimise interaction, it is necessary to

f2
K2 + c0 c#
e − Ls detune the controllers, e.g. by using low gains and
large reset times.This makes for poor dynamic per-
Fig. 81.2 Model of in line blending process formance in terms of speed of response and poor
disturbance rejection in terms of steady state off-
From a control point of view, this is a MIMO sys- set. Much better control can be realised by means
tem with two inputs and two outputs. Suppose of a 2 × 2 multivariable controller, as depicted in
that both the flow rate and concentration of the block diagram form in Figure 81.4.
dilute product stream are to be controlled simulta-
neously. A scheme consisting of two conventional
feedback control loops as depicted in Figure 81.1
will work. The block diagram for this is depicted
in Figure 81.3.

fm
M1 (s)

- u1
fr + e1 f0
C11 (s) V1 (s)
+ +
+ +
f1
C21(s)

C12 (s) − K1
f2
+ +
cr + u2 + c0 c#
e2
C22(s) V2 (s) K2 e − Ls
+
-
M 2 (s)
cm

Fig. 81.4 Model of blending system with multivariable controller

 81.2 Compensator Concept 675

81.2 Compensator Concept trol and C21 and C12 counteract the process inter-
actions:
The multivariable controller may be considered to     
consist of four compensators. In principle, com- u1 (s) C11 (s) C12(s) e1 (s)
= (81.4)
pensators C11 (s) and C22 (s) handle the flow and u2 (s) C21 (s) C22(s) e2 (s)
concentration control loops, in a conventional way,
which may be denoted:
and compensators C21(s) and C12(s) handle the in-
teractions.Thus,for example,following an increase u(s) = C(s).e(s)
in flow set point f r , compensator C11 (s) will in-
crease flow f 1 and hence f0 to its desired value. The compensator outputs are applied to the control
However, in doing so, it will cause concentration valves via I/P converters:
c0 to decrease via the gain K1 . This interaction is     
f1 (s) V1 (s) 0 u1 (s)
compensated for by C21 (s) which increases f 2 to = (81.5)
f2 (s) 0 V2(s) u2 (s)
produce an equal but opposite effect on c0 via K2 .
If the compensators are designed correctly, other which may be denoted:
signals in the concentration control loop, such as
e2 and c# should be unaffected by the change in f 0 . f (s) = V(s).u(s)
Similarly, C12 (s) protects f 0 against disturbances Equations 81.1–81.5 are a complete, but generic,
from the concentration control loop. description of the system.

81.3 Control System Model 81.4 Compensator Design

Two feedback loops are established such that both
The starting point is to establish the transfer ma-
f 0 and c# may be controlled against set points. The
trix of the closed loop system from Equations 81.1–
measurement functions may be articulated in ma-
81.5. Using block diagram algebra gives:
trix form:
     x(s) = P(s).f (s)
fm (s) M1(s) 0 f0(s)
= (81.2) = P(s).V(s).C(s). (r(s) − M(s).x(s))
cm (s) 0 M2(s) c# (s)
Matrix manipulation gives:
which may be denoted:
x(s) = (I + P(s).V(s).C(s).M(s))−1
m(s) = M(s).x(s) . P(s).V(s).C(s).r(s)

The comparators are handled simply by vector ad- Let G(s) be some user defined transfer matrix that
dition: specifies the desired closed loop response:
     
e1 (s) f (s) f (s) x(s) = G(s).r(s)
= r − m (81.3)
e2 (s) cr (s) cm (s)
The objective is to minimise interaction between
the loops.The ideal is zero interaction which corre-
which may be denoted:
sponds to making G(s) diagonal such that changing
set point f r only affects f 0 and changing set point
e(s) = r(s) − m(s)
cr only affects c# :
    
Four compensators are required. In essence, C11 f0 (s) G11 (s) 0 fr (s)
and C22 handle the feedback requirements for con- =
c# (s) 0 G22(s) cr (s)
 676 81 Multivariable Control

Whence: Choose a sensible closed loop transfer matrix:

⎡ ⎤
G(s) = (I + P(s).V(s).C(s).M(s))−1 ) 1
0
⎢ ⎥
. P(s.V(s).C(s) G(s) = ⎣ 4s + 1 e−2s ⎦
0
Solving for C(s) gives the desired multivariable 4s + 1
controller design: In the case of the flow loop, which has first order
lags of 2 s and 3 s, a closed loop response equiva-
C(s) = V(s)−1 .P(s)−1 .G(s) (81.6) lent to a first order lag of 4 s is reasonable. Likewise
. (I − M(s).G(s))−1 for the concentration loop except that a 2 s delay in
the closed loop response has been allowed for the
C(s) is totally dependent upon the accuracy of the measurement delay.
model and the sensible choice of G(s). If some G(s) Substitution of these values into Equation 81.6
is chosen that is unrealistic, perhaps looking for a and manipulation yields
faster response than the system is physically capa- ⎡ ⎤
(3s + 1) (2s + 1) − (3s + 1) (2s + 1)
ble of achieving, then C(s) will be unrealisable.
⎢ 4s (4s + 3) 2 (4s + 1) (2s + 1) − e−2s ⎥
⎢ ⎥
C(s) = ⎢ ⎥
⎣ (3s + 1) (2s + 1) (3s + 1) (2s + 1) ⎦
4s (4s + 3) 2 (4s + 1) (2s + 1) − e−2s
81.5 Worked Example No 1
All four compensators consist of combinations
The dynamics have been chosen as typical of pro-
of lead, lag, integrator and delay terms. This is
cess systems. If the measurements of flow and con-
quite normal for multivariable control, the com-
centration are made by means of an electromag-
pensators having been designed to give a specified
netic flow meter and conductivity cell respectively,
response for a particular plant.
time constants of 2 s are reasonable. Likewise, for
the pneumatically actuated control valves, time
constants of 3 s are assumed. The time delay for
the blending process, which only affects the con- 81.6 Decouplers
centration measurement, is 2 s:
⎡ ⎤ The principal disadvantage of multivariable con-
1 trollers is that the compensators C11 (s) to C22(s)
⎢ 0 ⎥
M(s) = ⎣ 2s + 1 do not have the structure of a PID controller. Their
1 ⎦ form is not intuitive which is undesirable for oper-
0
2s + 1 ational reasons. An alternative approach is to use
⎡ ⎤
1 decouplers, as depicted in Figure 81.5.
⎢ 0 ⎥ In this case the two compensators, C1 (s) and
V(s) = ⎣ 3s + 1 1 ⎦
0 C2 (s), may be realised by conventional PID con-
3s + 1 trollers:
$ %  
1 1 1
P(s) = C(s) = KC 1 + + TD s
−e−2s e−2s TR s
The decouplers,D1 (s) and D2 (s),and compensators
Note that a value of unity is assumed for all the are related according to:
steady state gains. This implies that either all the     
signals have the same range, such as 4–20 mA, or u1(s) C1 (s) C2 (s).D2 (s) e1 (s)
=
that they have been scaled on a common basis,such u2(s) C1 (s).D1 (s) C2(s) e2 (s)
as percentage of range. (81.7)
 81.7 Sampled Data Model 677

fm
M1 (s)

- e1 u1
fr + f0
C1 (s) V1 (s)
+ +
+ +
f1
D1(s)

D2 (s) − K1
f2
+ +
cr e2 + u2 + c0 c#
+
C2 (s) V2 (s) K2 e − Ls
-
M 2 (s)
cm

Fig. 81.5 Multivariable control of blending system using decouplers

The design technique essentially consists of choos- 81.7 Sampled Data Model
ing appropriate C1 (s) and C2 (s),empirically or oth-
Multivariable controllers are realised by means of
erwise, and then finding D1 (s) and D2 (s) for the
software. The dynamics of the sampling process
known P(s), M(s) and V(s) to give a specified G(s).
can be ignored if they are fast compared with the
For the system of Figure 81.5, the decouplers
process. However, if they are of the same order of
can be determined by inspection. Firstly consider
magnitude,as with this blending system,they must
the decoupler D1 (s). Following a change in error
be taken into account. Either way, it is convenient
e1 , the affect on C0 due to the interaction inherent
to do the design in the Z domain because the sub-
in the process is given by:
sequent compensators are easily realisable.
c0 (s) = −K1 V1(s)C1 (s).e1 (s) The block diagram of the sampled data multi-
variable control system for the blending plant is as
The corresponding affect on C0 due to the decou-
depicted in Figure 81.6.
pler is given by:
Note that both the error and output signals are
c0 (s) = K2 V2(s)D1 (s)C1 (s).e1 (s) sampled, but only the outputs are held by zero or-
der hold devices. Whereas previous values of the
For the decoupler to be effective, these two affects error signals may be stored in memory, the output
must be equal and opposite: that is they cancel each signals must be held between sampling instants to
other. Thus: create pseudo continuous signals. It is a common
−K1 V1(s)C1 (s) e1 (s) = −K2 V2 (s)D1 (s)C1 (s) e1 (s) mistake to assume hold devices for the error sig-
nals.
Hence: Equations 81.1–81.3 still apply. Allowing for
K1V1 (s) K1
D1 (s) = ≈ the samplers, the equation which describes the be-
K2V2 (s) K2
haviour of the compensators becomes:
By a similar argument, for a change in error e2 to     ∗ 
have no net affect on f 0 , the decoupler D2 (s) re- u1 (s) C11 (s) C12 (s) e1 (s)
= (81.8)
quired is simply: u2 (s) C21 (s) C22 (s) e∗2 (s)

D2 (s) = −1 which may be denoted:

 678 81 Multivariable Control

fm
M1 (s)

fr +
- e1 u1 h1 fo
C11(s) H1 (s) V1 (s)
T + T +
+ +
f1
C 21(s)

C12 (s) –K1

f2
+
cr e2 + u2 h2 ++ co c#
C 22 (s) H 2 (s) V2 (s) K2 e −Ls
+
- T T
cm
M 2 (s)

Fig. 81.6 Control of blending system using impulse compensators

u(s) = C(s).e∗ (s) Solving for e(z) gives:

An additional equation must be introduced to ac-
(I + MPVH(z).C(z)) .e(z) = r(z)
commodate the hold devices:
    ∗  e(z) = (I + MPVH(z).C(z))−1 .r(z)
h1 (s) H1(s) 0 u1 (s)
= (81.9)
h2 (s) 0 H2 (s) u2∗(s)
Assuming there are imaginary samplers on the
which may be denoted: controlled variables:
h(s) = H(s).u∗ (s)
x(z) = PVH(z).C(z).e(z)
The outputs of the holds are then applied via I/P = PVH(z).C(z). (I + MPVH(z).C(z))−1 .r(z)
converters to the two valves which manipulate the
process inputs: = G(z).r(z)
    
f1 (s) V1 (s) 0 h1 (s) G(z) is the closed loop transfer matrix which de-
= (81.10)
f2 (s) 0 V2 (s) h2 (s) termines the overall system performance:
which may be denoted:
G(z) = PVH(z).C(z) (81.11)
f (s) = V(s).h(s) −1
. (I + MPVH(z).C(z))

If G(s) is diagonal then any change in either set

point will only affect one of the controlled vari-
81.8 Impulse Compensator ables, i.e. the system is decoupled. For a specified
Design G(s), and for given M(s), P(s), V(s) and H(s), there
must be a unique C(s):
Block diagram algebra yields:
e(z) = r(z) − m(z) G(z). (I + MPVH(z).C(z)) = PVH(z).C(z)
= r(z) − MPVH(z).u(z) G(z) = (PVH(z) − G(z).MPVH(z)) .C(z) (81.12)
= r(z) − MPVH(z).C(z).e(z) C(z) = (PVH(z) − G(z).MPVH(z))−1 .G(z)
 81.9 Worked Example No 2 679

81.9 Worked Example No 2 Extensive manipulation yields

Assume that M(s), V(s) and P(s) are as before, and C(z) =
that the zero order holds have a one second delay: ⎡
0.441 1 − 1.323z−1 + 0.4346z−2
⎢
⎡ ⎤ ⎢ 1 − 1.441z−1 + 0.430z−2
1 − e−s ⎢
0 ⎥ ⎣ 0.441 1 − 1.323z−1 + 0.4346z−2
⎢ s
H(s) = ⎣ 1 − e−s ⎦ 1 − 1.441z−1 + 0.430z−2
0 ⎤
s −0.441 1 − 1.323z−1 + 0.4346z−2
1 − 1.385z−1 + 0.4724z−2 − 0.05597z−3 − 0.04239z−4 ⎥
⎥
The 1 sec delay associated with the zero order holds ⎥
−1
0.441 1 − 1.323z + 0.4346z −2 ⎦
implies a 1 sec sampling period.
1 − 1.385z−1 + 0.4724z−2 − 0.05597z−3 − 0.04239z−4
Choose a closed loop transfer matrix, taking (81.13)
into account the dynamics of the sampling process: which is of the form:
 
⎡ ⎤ ˛(z) −ˇ(z)
e−s C(z) = (81.14)
0 ˛(z) ˇ(z)
⎢ ⎥
G(s) = ⎣ 4s + 1 e−3s ⎦
0 It can be seen by inspection that all four compen-
4s + 1 sators are realisable, i.e. there are no terms in z+1 .
The coefficients of z−1 in successive terms of both
The rationale for choosing G(s) is essentially the the numerator and denominator series are of re-
same as in the continuous case, except that a 1 sec ducing magnitude and, with the exception of the
delay has been included in the flow loop to allow least significant term of the second denominator,
for the hold. Likewise for the concentration loop are of alternate sign. This is necessary for stabil-
where a 3 sec delay has been allowed for both the ity. The compensators may be realised by means
measurement delay and the hold. of software using any of the techniques described
Assume a sampling period of T = 1 s. in Chapter 78. Remember that the variables in the

fm
M1 (s)

f r +- e1 u1 h1 fo
C1 (s) H1 (s) V1 (s)
+ T +
+ +
f1
D1 (s)

D 2 (s) K1
f2
+ h2 - co
cr e2 + u2 + c#
C 2 (s) H 2 (s) V2 (s) K2 e −Ls
+ T
-
cm
M 2 (s)

Fig. 81.7 Control of blending system using sampled data decouplers

 680 81 Multivariable Control

model are all in deviation form so a steady state Given the assumption that K1 = K2 = 1, this re-
bias must be added to the controller outputs for sult is entirely consistent with that obtained in Sec-
implementation purposes. tion 81.6.
Note that it is often sufficient to realise the two
compensators C1 (z) and C2 (z) by means of the dis-
crete form of PID controllers developed in Chap-
81.10 Sampled Data Decoupler ter 78:
The block diagram of a sampled data decoupler a0 + a1 z−1 + a2 z−2
C(z) = (78.3)
type of control system for the blending plant is as 1 − z−1
depicted in Figure 81.7. The approximation may simply be based on the
Apart from the compensators and decouplers first terms of ˛(z) and ˇ(z) which are the dominant
being of an impulse nature, the approach to design terms. Otherwise, the PID compensator is taken as
is essentially the same as for the continuous system a factor out of C1 (z) and C2 (z). The parameters of
of Figure 81.5. Control action is of the form the compensator would be presented to the oper-
ators in terms of the familiar gain, reset and rate
   
 times, whose values may be tuned empirically.
u1(z) C1 (z) C2 D2 (z)
e1 (z)
=
u2(z) C1 D1 (z) C2 (z)
e2 (z)
(81.15)
Comparison of Equations 81.14 and 81.15 reveals 81.11 Comments
that,to meet the same closed loop performance cri-
terion G(s) for the physical example of the blending Although all these compensators have been de-
system, the decouplers must be signed for servo operation as opposed to regulo
control, i.e. the effects of disturbances haven’t been
taken into account explicitly, they can nevertheless
D1 (s) = −D2 (s) = 1 be just as effective at disturbance rejection.