PART - 4

KNU/EECS/ELEC 835001

Multivariable Control
for MIMO processes

Multivariable Control Dr. Kalyana C. Veluvolu

Outline - Module 5.4

Decoupler Design for MIMO processes
– Ideal Decoupler
» Simplified Decoupler
» Generalized Decoupler
– Limitation of Decoupler
– Simpler Decoupling
» Partial Decoupling
» Steady-state Decoupling
– Effects of MV Constraints
– Ill Conditioned Process
» Degeneracy
» Singular Value Decomposition
» Decoupling Based on SVD

Multivariable Control Dr. Kalyana C. Veluvolu 2

Multi-loop vs. Multivariable Control
• multi-loop - use of several single-loop controllers (e.g., PID) on
pairs of manipulated/controlled variables

• multivariable - make control adjustments decisions jointly

considering all outputs simultaneously

• Multi-loop control configurations are typically used as a base

control configuration and reside in the Distributed Control
System (DCS).
» e.g., flow control, temperature control, pressure control

Multivariable control configurations typically require additional

computational capability, and sit over a base multi-loop control
configuration, sending setpoints to the multi-loop controllers.

Multivariable Control Dr. Kalyana C. Veluvolu 3

Multi-loop vs. Multivariable Control

Under the multi-loop control strategy, each controller gci operates according to:

u i = g ci ( y di − y i ) = g ci ε i
Multivariable controller must decide on ui, not using only εi, but using the
entire set, ε1, ε2,, ...,εn,;. Thus, the controller actions are obtained from

u1 = f 1 (ε 1 , ε 2 ,
ε n ) u1
++

G11(s)
y1
u 2 = f 2 (ε 1 , ε 2 ,
ε n )
y1,sp G21(s)
u 3 = f 3 (ε 1 , ε 2 ,
ε n ) y2,sp
Multivariable
Controller

=
G12(s)

y2
u n = f n (ε 1 , ε 2 ,
ε n )
+

u2
+

G22(s)

Multivariable Control Dr. Kalyana C. Veluvolu 4

Principles of Decoupling
Main loop y1 — u1 , y2 — u2,…, yn — un, couplings

desirable for control
Cross-couplings, yi — uj (i ≠ j)

undesirable; loop interactions
Eliminates the effect of the undesired cross-couplings

improve control performance.
Objective is to compensate for interactions by cross-couplings

not to “eliminate” the cross-couplings; impossibility, require
altering the physical nature of the system.

Multivariable Control Dr. Kalyana C. Veluvolu 5

Simplified Decoupling
Two compensator blocks gI1 and gI2.

Controller outputs v1 and v2, actual control on the process u1 and u2.

Without the compensator,

u1 = v1 and u2 = v2, and
the process model
y1 = g11u1 + g12 u 2
y1 = g 21u1 + g 22 u 2

Compensator, Loop 2
“informed” of changes in
v1 by g12, u2 is adjusted.
The same for Loop 1

Multivariable Control Dr. Kalyana C. Veluvolu 6

 Design Simplified Decoupler
y1 = g 11u1 + g 12 u 2 u1 = v1 + g I 1v 2
⇒ ⇒
y 2 = g 21u1 + g 22 u 2 u 2 = v 2 + g I 2 v1
y1 = ( g 11 + g 12 g I 2 )v1 + ( g 12 + g 11 g I 1 )v 2
⇒ y 2 = ( g 21 + g 22 g I 2 )v1 + ( g 22 + g 21 g I 1 )v 2

g 12 g 21
g I1 = − gI2 = −
g 11 g 22
g 12 g 21
y1 = ( g11 − )v1
g 22
g12 g 21
y 2 = ( g 22 − )v 2
g 11

Multivariable Control Dr. Kalyana C. Veluvolu 7

Difficulties for Simplified Decoupler

3x3 , six compensator.
larger than 2 x 2, decoupling become tedious.
NxN: (N2-N) compensators.

Multivariable Control Dr. Kalyana C. Veluvolu 8

Generalized Decoupling
MIMO process
y =Gu u = GI v ⇒ y = GG I v
To eliminate interactions, y to v : a diagonal matrix; GR(s).

GG I = G R (s ) ⇒ y = G R ( s) v

Choose GI such that

G I = G −1GR ( s)

Selected to provide desired decoupled behavior with the simplest form

A commonly employed choice

GR ( s) = Diag[G( s)]

Multivariable Control Dr. Kalyana C. Veluvolu 9

Relation Between the Two Schemes

Simplified decoupling y = GG I v

2 x 2 and 3 x 3 system, the compensator transfer function matrix:

 1 g I1   1 g I 12 g I 13 
GI =  G I =  g I 21 g I 23 
1 
1
g I 2  g I 31 g I 32 1 

For the desired GI , task is to find g Iij to make GGI diagonal

General decoupling
Final diagonal form GGI specified as GR, then GI can be derived.

Multivariable Control Dr. Kalyana C. Veluvolu 10

Example Distillation Column
 12.8e − s − 18.9e −3s 
 
G ( s) = 16.7 s−+7 s1 21.0s + 1 
 6.6e − 19.4e −3 s 
10.9s + 1 14.4s + 1 
simplified decoupler
(16.7 s + 1)e −2 s (14.4s + 1)e −4 s
g I 1 = 1.48 gI2 = 0.34
21.0s + 1 10.9s + 1

actual implementation
(16.7 s + 1)e −2 s
u1 = v1 + 1.48 v2
21.0 s + 1
(14.4 s + 1)e −4 s
u 2 = v 2 + 0.34 v1
10.9 s + 1
Multivariable Control Dr. Kalyana C. Veluvolu 11

Example Distillation Column

Generalized decoupling:
 12.8e − s   − 19.4e −3 s 18.9e −3 s 
 0  1  g g I 12 
G −1 ( s ) =  14.4 s +−17 s 21.0 s + 1 G I =  I 11
G R ( s ) = 16.7 s + 1

−3 s 
− 19.4e  ∆  − 6.67e 12.8e −s
 g I 21 g I 22 
0  10.9 s + 1
 14.4 s + 1  16.7 s + 1 

− 248.32(21.0s + 1)(10.9s + 1)e −4 s + 124.74(16.7 s + 1)(14.4s + 1)e −10 s

∆=
(21.0s + 1)(10.9s + 1)(16.7 s + 1)(14.4s + 1)
− 248.32( 21.0 s + 1)(10.9 s + 1)
g I 11 =
124.74(16.7 s + 1)(14.4s + 1)e −6 s − 248.32(21.0s + 1)(10.9s + 1)
− 366.66(16.7 s + 1)(10.9s + 1)e − 2 s
g I 12 =
124.74(16.7 s + 1)(14.4s + 1)e −6 s − 248.32(21.0s + 1)(10.9 s + 1)
84.48(21.0s + 1)(14.4s + 1) g I 22 = g I 11
g I 21 =
124.74(16.7 s + 1)(14.4 s + 1)e −6 s − 248.32(21.0s + 1)(10.9s + 1)

The actual implementation: u1 = g I 11v1 + g I 12 v 2

u 2 = g I 21 v1 + g I 22 v 2
Multivariable Control Dr. Kalyana C. Veluvolu 12
Comparison of the Two Methods
Simplified decoupling: “equivalent” open-loop decoupled system

 g12 g 21   12.8e − s 18.9 × 6.6(14.4 s + 1)e −7 s 

y1 =  g11 − v1 =  − v1
 g 22   (16 . 7 s + 1) 19 . 4 ( 21 . 0 s + 1)(10 . 9 s + 1) 
 g12 g 21   − 19.4e −3 s 18.9 × 6.6(16.7 s + 1)e −9 s 
y 2 =  g 22 − v 2 =  − v 2
 g 11   (14 . 4 s + 1) 12 . 8( 21 . 0 s + 1)(10 . 9 s + 1) 

much more complicated than GR specified in the Generalized decoupling

Difficult to tune controller

Generalized decoupling:
tuning and performance better than for Simplified decoupling

complicated decoupler

Multivariable Control Dr. Kalyana C. Veluvolu 13

Limitations in Application
Perfect decouple if model perfect - impossible in practice.
The simplified decoupling similar to feedforward controllers

realization problems, time delay elements
Perfect dynamic decouplers based on model inverses.

can only be implemented if inverses causal and stable.

2 x 2 compensators, gI1 and gI2 must be causal (no e+αs terms) and stable

time delays in g11 smaller than time delays in g12

time delays in g22 smaller than time delays in g21

g11 and g22 no RHP zeros

g12 and g21 must no RHP poles

Multivariable Control Dr. Kalyana C. Veluvolu 14

 Implementation
Adding delays to the inputs u1, u2, ..., un, by define: G m = GD

e − d11s 0 
 
e − d 22 s
D( s) =  
  
 
 0 e − d nn s 
Simplified decoupling: requiring the smallest delay in each row on the diagonal,
designed by using Gm.

Generalized decoupling: use modified process Gm so that GI=(GD)-1GR are

causal which requiring that GR-1(GD) have the smallest delay in each row on the
diagonal.

Multivariable Control Dr. Kalyana C. Veluvolu 15

Example: Distillation Column

add a time delay of 3 minutes to the input u1:

 12.8e −4 s − 18.9e −3s  Smallest delay in each row is not on diagonal,

  simplified decoupling compensator becomes:
G ( s ) =  16.7 s −+101s 21.0 s + 1 
 6.67e − 19.4e −3 s 
(16.7 s + 1)e s
 10.9 s + 1 14.4 s + 1  g I 1 = 1.48
21.0 s + 1

Design D(s) to add a time delay of 1 minute to the input u2, i.e.:

1 0 
D( s) =  −s 
0 e  (16.7 s + 1)
g I 1 = 1.48
 12.8e −4 s
− 18.9e  −4 s
21.0s + 1
 
Gm = GD =  16.7 s −+101s 21.0s + 1  (14.4s + 1)e −6 s
− 19.4e − 4 s  gI2 = 0.34
 6.67e 10.9s + 1
 10.9 s + 1 14.4 s + 1 
Multivariable Control Dr. Kalyana C. Veluvolu 16
Example: Distillation Column
 12.8e −4 s − 18.9e −3s 
  (16.7 s + 1)e s
G ( s ) =  16.7 s −+101s 21.0s + 1  g I 1 = 1.48
21.0 s + 1
 6.67e − 19.4e −3s 
 10.9s + 1 14.4s + 1 
As time prediction term much small than time constant, drop prediction
(16.7 s + 1)
g I 1 = 1.48
21.0 s + 1
Effective time constant of g12 and g11 are similar 16.7 + 4 ⇔ 21 + 3
Steady-state decoupling
g I 1 = 1.48

Multivariable Control Dr. Kalyana C. Veluvolu 17

Partial Decoupling
Consider partial decoupling if

some of the loop interactions are weak

some of the loops need not have high performance

Partial decoupling focused on a subset of control loops

interactions are important, and/or

high performance control is required.

Consider partial decoupling for 3x3 or higher systems

main advantage: reduction of dimensionality.

Multivariable Control Dr. Kalyana C. Veluvolu 18

 Partial Decoupling Example
Grinding circuit analysis  119 153 − 21 
 10 s + 1  u1 
 y1   217 s + 1 337 s + 1

Least sensitive variables y2  y  =  0.00037 0.000767 − 0.00005   
 u2
 2   500 s + 1 33s + 1 10 s + 1   

Most interaction: Loops 1 and 3,  y3   930 − 667e −320s − 1033  u 3 

Decouplers: loops 1 and 3,  500 s + 1 166 s + 1 47 s + 1 

Loop 2 without decoupling.
 119 − 21 
 y1   217 s + 1 10 s + 1  u1 
the transfer function matrix for the subsystem  y  =  930
 3  − 1033  u 3 

 500 s + 1 47 s + 1
using the simplified decoupling approach
21 930
0.176(217 s + 1) 0.0(47 s + 1)
g I 1 = 10s + 1 = ; g I 3 = 500s + 1 =
119 10s + 1 1033 500s + 1
217 s + 1 47 s + 1

Multivariable Control Dr. Kalyana C. Veluvolu 19

Steady-State Decoupling
Steady-state decoupling: uses steady-state gain of transfer function
2 x 2 system
K12 K
Simplified steady-state decoupling g I1 = − , g I 2 = − 21
K11 K 22

Generalized steady-state decoupling GI = K −1 K R

Very easy to design and implement, first technique to try;

ideal decoupler only if dynamic interactions persistent

big performance improvements with very little work or cost

most often applied in practice.

Multivariable Control Dr. Kalyana C. Veluvolu 20

Example Distillation Column
 12.8e − s − 18.9e −3s 
12.8 − 18.9  
K=  ⇐ G ( s ) = 16.7 s−+7 s1 21.0 s + 1 
−3s
 6.6 − 19.4  6.6e − 19.4e 
10.9s + 1 14.4s + 1 
Simplified steady-state decoupling
u1 = v1 + 1.48v 2
gI1 = −
−18.9
= 1.48, gI 2 = −
6.6
= 0.34 ⇒
12.8 −19.4 u 2 = v 2 + 0.34v1

Generalized steady-state decoupling

12.8 0   2.01 2.97 u1 = 2.01v1 + 2.97 v 2

KR =  ⇒ GI =   ⇒
 0 − 19.4 0.68 2.01 u 2 = 0.68v1 + 2.01v 2

Multivariable Control Dr. Kalyana C. Veluvolu 21

Effect of Inputs Constraints

Always existing constraints on the process input variables

valves cannot go beyond full open or full shut

heaters cannot go beyond full power or zero power, etc.

Decoupling ok, if controller outputs not reached constraints

Even one input reaches a constraint, (reset windup)

control system no longer function decoupling

extremely poor (or even unstable) responses

Multivariable Control Dr. Kalyana C. Veluvolu 22

Input Constraints
Example Closed-loop response
Simplified steady-state decoupler for the WB of Y1 and Y2.
Distillation Column with Kc1 =0.30, 1/τI1 =
0.307, Kc2 = - 0.05, 1/τI2 = 0.107.
− 18.9 6. 6 Unconstrained
g I1 = − = 1.48, gI2 = − = 0.34
12.8 − 19.4 manipulated
variables, u1, u2,
If u, 0≤ u1 ≤ 0.15, the closed-loop response
is very poor once the reflux valve is full open
and the system becomes unstable.

Response of Y1 Manipulated
and Y2 with variables u1,
constrained u1, 0≤ u2 when 0≤ u1
u1≤0.l5. ≤0.15.

Multivariable Control Dr. Kalyana C. Veluvolu 23

Sensitivity to Model Error

K: system steady-state gain matrix:

y = Ku
Generalized decoupler

u = GI v G I = K −1 K R ⇒ y = KK −1 K R v = K R v

∆K - error in the estimate of the steady-state gain matrix, then

−1
y = (K +∆K)u ⇒ y = (K + ∆K)K K Rv
−1 ⇒
= K R v +∆KK K Rv

∆K Adj(K)K R v
⇒ ∆y = ∆KK −1 K R v ⇒ ∆y =
|K|
Multivariable Control Dr. Kalyana C. Veluvolu 24
 RGA and Model Error
∆K Adj(K)K R v
∆y =
|K|
If |K|very small, its reciprocal will be very large

Small modeling errors will cause very large errors in y

Small changes in controller output v result in large errors in y

Decoupling difficult: input/output variables are paired on very large

RGA values; system will also be very sensitive to modeling errors.

K ij C ij
λij = Cij cofactor of Kij
|K|

Multivariable Control Dr. Kalyana C. Veluvolu 25

Example Heavy Oil Fractionator

Transfer function model

 4.05e −27 s 1.20e −27 s 

  4.05 1.20
G ( s ) =  50 s +−18 s 45s + 1  ⇒ K = 
 4.06e 1.19  4.06 1.19
 13s + 1 19 s + 1 
determinant very close to zero: decoupling very difficult.

| K |= −0.0525 RGA  − 91.8 92.8 

Λ= 
 92.8 − 91.8
Decoupling extremely difficult

small value of gain matrix determinant

large values of RGA elements

Multivariable Control Dr. Kalyana C. Veluvolu 26

 SVD and Condition Number
The matrix is said to be singular, if its determinant is zero,
Near singularity matrix: singular values
σ i = (λi ( A* A))1/ 2 i = 1, 2,..., n
Condition number: The ratio of the largest and smallest singular value
σ max
k=
σ min
Example: Heavy Oil Fractionator continued

4.05 1.20 singular values σ1=5.978, σ2 = 0.00878

K =  and a condition number:
4.06 1.19 κ = 680.778
clearly indicating serious ill-conditioning.
Multivariable Control Dr. Kalyana C. Veluvolu 27