c 1997 by Jay H.

Lee, Jin Hoon Choi, and Kwang Soon Lee

Part II
OVERVIEW OF INDUSTRIAL MPC
TECHNIQUES

1
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

Contents
1 INTRODUCTION TO MODEL PREDICTIVE CONTROL 5
1.1 BACKGROUND FOR MPC DEVELOPMENT . . . . . . . . . . . . . . . . 5
1.2 WHAT'S MPC . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.3 WHY MPC? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.3.1 SOME EXAMPLES . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.3.2 SUMMARY . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
1.4 INDUSTRIAL USE OF MPC: OVERVIEW . . . . . . . . . . . . . . . . . . 25
1.4.1 MOTIVATION . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
1.4.2 SURVEY OF MPC USE . . . . . . . . . . . . . . . . . . . . . . . . . 31
1.5 HISTORICAL PERSPECTIVE . . . . . . . . . . . . . . . . . . . . . . . . . 32
1.6 CHALLENGES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
1.6.1 MODELING & IDENTIFICATION . . . . . . . . . . . . . . . . . . 34
1.6.2 INCORPORATION OF STATISTICAL CONCEPTS . . . . . . . . . 41
1.6.3 NONLINEAR CONTROL . . . . . . . . . . . . . . . . . . . . . . . . 48
1.6.4 OTHER ISSUES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49

2 DYNAMIC MATRIX CONTROL 50

2.1 FINITE IMPULSE AND STEP RESPONSE MODEL . . . . . . . . . . . . 50

2
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

2.1.1 OVERVIEW OF COMPUTER CONTROL . . . . . . . . . . . . . . 50

2.1.2 IMPULSE RESPONSE AND IMPULSE RESPONSE MODEL . . . . 52
2.1.3 STEP RESPONSE AND STEP RESPONSE MODEL . . . . . . . . 54
2.2 MULTI-STEP PREDICTION . . . . . . . . . . . . . . . . . . . . . . . . . . 57
2.2.1 OVERVIEW . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
2.2.2 RECURSIVE MULTI-STEP PREDICTION FOR AN FIR SYSTEM 58
2.2.3 RECURSIVE MULTI-STEP PREDICTION FOR AN FIR SYSTEM
WITH DIFFERENCED INPUT . . . . . . . . . . . . . . . . . . . . . 62
2.2.4 MULTIVARIABLE GENERALIZATION . . . . . . . . . . . . . . . . 66
2.3 DYNAMIC MATRIX CONTROL ALGORITHM . . . . . . . . . . . . . . . 67
2.3.1 MAJOR CONSTITUENTS . . . . . . . . . . . . . . . . . . . . . . . 67
2.3.2 BASIC PROBLEM SETUP . . . . . . . . . . . . . . . . . . . . . . . 68
2.3.3 DEFINITION AND UPDATE OF MEMORY . . . . . . . . . . . . . 69
2.3.4 PREDICTION EQUATION . . . . . . . . . . . . . . . . . . . . . . . 70
2.3.5 QUADRATIC CRITERION . . . . . . . . . . . . . . . . . . . . . . . 73
2.3.6 CONSTRAINTS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
2.3.7 QUADRATIC PROGRAMMING . . . . . . . . . . . . . . . . . . . . 79
2.3.8 SUMMARY OF REAL-TIME IMPLMENTATION . . . . . . . . . . 83
2.4 ADDITIONAL ISSUES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
2.4.1 FEASIBILITY ISSUE AND CONSTRAINT RELAXATION . . . . . 84
2.4.2 GUIDELINES FOR CHOOSING THE HORIZON SIZE . . . . . . . 85
2.4.3 BI-LEVEL FORMULATION . . . . . . . . . . . . . . . . . . . . . . 86
2.4.4 PROPERTY ESTIMATION . . . . . . . . . . . . . . . . . . . . . . . 89
2.4.5 SYSTEM DECOMPOSITION . . . . . . . . . . . . . . . . . . . . . . 91

3
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

2.4.6 MODEL CONDITIONING . . . . . . . . . . . . . . . . . . . . . . . 98

2.4.7 BLOCKING . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102

3 SYSTEM IDENTIFICATION 107

3.1 DYNAMIC MATRIX IDENTIFICATION . . . . . . . . . . . . . . . . . . . 107
3.1.1 STEP TESTING . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
3.1.2 PULSE TESTING . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
3.1.3 RANDOM INPUT TESTING . . . . . . . . . . . . . . . . . . . . . . 112
3.1.4 DATA PRETREATMENT . . . . . . . . . . . . . . . . . . . . . . . . 118
3.2 BASIC CONCEPTS OF IDENTIFICATION . . . . . . . . . . . . . . . . . . 120
3.3 MODEL DESCRIPTION . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
3.3.1 NONPARAMETRIC MODEL . . . . . . . . . . . . . . . . . . . . . . 124
3.3.2 PARAMETRIC METHOD . . . . . . . . . . . . . . . . . . . . . . . . 125
3.4 EXPERIMENTAL CONDITIONS . . . . . . . . . . . . . . . . . . . . . . . . 128
3.4.1 SAMPLING INTERVAL . . . . . . . . . . . . . . . . . . . . . . . . . 128
3.4.2 OPEN-LOOP VS. CLOSED-LOOP EXPERIMENTS . . . . . . . . . 129
3.4.3 INPUT DESIGN . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
3.5 IIDENTIFICATION METHODS . . . . . . . . . . . . . . . . . . . . . . . . 132
3.5.1 PREDICTION ERROR METHOD . . . . . . . . . . . . . . . . . . . 132
3.5.2 SUBSPACE IDENTIFICATION . . . . . . . . . . . . . . . . . . . . . 137
3.6 IDENTIFICATION OF A PROCESS WITH STRONG DIRECTIONALITY 138

4
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

Chapter 1
INTRODUCTION TO MODEL
PREDICTIVE CONTROL
1.1 BACKGROUND FOR MPC DEVELOPMENT
Two main driving forces for a new process control paradigm in the late 70's
 early 80's:
 Energy crisis + global competition + environmental reg.
+
{ process integration
{ reduced design / safety margin
{ real-time optimization
{ tighter quality control
+
higher demand on process control.
 (Remarkable) advances in microprocessor technology.
5
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

{ cheap, fast and reliable medium for implementation.

{ network environment(e.g., DCS) conductive to hierarchical approach.
Industry's response ) MPC

1.2 WHAT'S MPC

It's a computer control system.

PLANT

HOLD SAMPLER
CLOCK

COMPUTER
1 2 3 4 5 6 CONTROL 1 2 3 4 . . .

ALGORITHM

On - Line
memory model
Optimizer

6
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

It's a computer control system consisting of an observer & an optimizer.

The optimization is based on prediction of future behavior of y.

MPC(software packages) is sold under dierent names:

 DMC (Dynamic Matrix Control, now AspenTech)

 IDCOM (Setpoint, now AspenTech)
 SMCA (Setpoint, now AspenTech)
 RMPCT (Honeywell)
 PCT (Promatics)

7
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

 HEICON (Adersa)
 OPC (Treiber)
 MAC
 IMC
 GPC
 GMC
 UPC
...

It's major features are

 model based
 explicit prediction of future system behavior
 explicit consideration of constraints
 use of on-line mathematical programming
 receding horizon control : repeated computation of open-loop optimal
trajectory with feedback update ) implicit feedback control.

1.3 WHY MPC?

Dicult elements for process control:

 delay, inverse response

 interaction
8
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

 constraints
 competing optimization requirements
MPC provides a systematic, unied solution to problems with these charac-
teristics.

1.3.1 SOME EXAMPLES

Example I : Blending systems (input constraints)

Valve
Positions Additive A
Stock stock
Additive B
Additive A stock
Blending System
Additive B Model total blend flow

 control rA & rB (rst priority).

 control q if possible (second priority).
 possibility of valve saturation must be taken into account.

9
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

Classical Solution :

Selector
<
Valve-position
controller

95% VPC FC Setpoint

Setpoint

Feedback

FT
FT

Total
Stock
blended
flow

Ratio
Setpoint Setpoint
FC
> Blend of
A and B
High
FT
Selector

Additive A

X Setpoint
FC

Ratio
FT
Setpoint

Additive B

MPC Solution :
At t=k, solve 2 3 2 3 2
p (rA )k+ijk
X 64 75 , 64 (rA)ref 75 + kq
k+ijk , qref kR
min 2
ui
i=1 (rB )k+ijk (rB )ref
Q
QR
2 3 2 3 2 3
66 (u1)min 77 66 (u1)j 77 66 (u1)max 77
66 (u2)min 77  66 (u2)j 77  66 (u3)max 77 ; j = 0;


; p , 1
4 5 4 5 4 5
(u3)min (u3)j (u2)max
10
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

Example II : Two-point control in a distillation column (input con-

straints, interaction)

0.5

- 0.5
0.5

- 0.5

 strong interaction
 \wind-up" during saturation
 saturation of an input requires recoordination of the other input
11
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

Clasical Solution: Two single-loop controllers with anti-windup scheme

(decouplers not shown)

PID w/ anti-windup

C1 + (T )
1 ref

0.5 +
T1
-0.5 L +
D1

G
D2 T2
0.5 V +

-0.5 +

-
C2 (T 2) ref
+

PID w/ anti-windup

 T1 controller does not know that V has saturated and vice versa )
coordination of the other input during the saturation of one input is
impossible.
 mode-switching logic is dicult to design / debug (can you do it?) and
causes "bumps", etc.

12
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

MPC Solution:

F ZF

0.5 L +
T1
-0.5 +
D1

G
D2 T2
0.5 V +
-0.5 +

MPC

(T 1 ) ref
MIMO Constraints
Model -0.5 L 0.5 (T 2 ) ref
-0.5 V 0.5

At t = k, solve
2 3 2 3 2 2 3 2
p (T1)k+ijk 1
Lk+ijk
min
X 64 75 , 64 (T1)ref
75 + m
X , 64 75

Uk i=1 (T2)k+ijk (T2)ref i=0

Vk+ijk
Q R
with 2 3 2 3 2 3
L L L
64 min 75  64 k+ijk 75  64 max 75 for i = 0;


; m , 1
Vmin Vk+ijk Vmax
 easy to design / debug / recongure.
 anti-windup is automatic.
 optimal coordination of the inputs is automatic.

13
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

Performance of classical solution vs. MPC

SISO loops w/ anti-windup & decoupler (no mode switching):

14
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

Example III : Override control in compressor(output constraint)

 control the owrate
 but maintain P  Pmax
Classical Solution :

LS

<
Set
Press
Feed
back FC PC

SC Set
Discharge Flow

Motor Time
Compressor

15
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

MPC Solution:

Compressor flowrate
motor speed Model
pressure

At t = k, solve p 2 mX
,1 2
X q
min

U k+ijk , q
ref Q +
u k+ijk R
k i=1 i=0
with
Pk+ijk  Pmax for i = 1;


; p

16
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

Example IV : Override control in surge tank(output constraints)

 control the outlet owrate
 but maintain L  Lmin
Classical Solution :

Hot
Staurated
Liquid

L
PI
LC FC
1 LS 1

vp q

17
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

MPC Solution:

Surge tank flowrate

Pump speed
Model
liquid level

At t = k, solve p 2 mX
,1 2
X
min qk+ijk , qref Q +


Uk i=1

uk+ijk R
i=0
with
Lk+ijk  Lmin for i = 1;


; p

18
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

Example V : Valve position control in air distribution network (op-

timization requirement)
 control the owrates of individual channels
 minimize the air compression
Classical Solution :

Valve-position
controller
HS
VPC
>
Feedback

Signals
from
individual
PC process
control
loops

Header
pressure

PT Process
SC
demands
for air

AIR
compressor

Header

19
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

MPC Solution :

PRIMARY
CONTROL VARIABLES
Valve
Positions
air flow rates

Air Distribution
Network
Model
Valve positions

Header
Pressure
SECONDARY
CONTROL VARIABLES

At t = k, solve
2 3 2 3 2
6 (q1)k+ijk 77 66 (q1)ref 77 mX
X 66
p
.. 77 , 66 .. 77 + ,1 P 2
min

Uk i=1 4
66 75 64 75 k+ijk , P min R
(qn)k+ijk (qn)ref Q i=1
with Q  R and
2 3 2 3 2 3
66 min 77 66 Pk+ijk
P P
77 66 max 77
66 (u1)min 77 66 (u1)k+ijk 77 66 (u1)max 77
66 . 77  66 ... 77  66 . 77 for i = 0;


; m , 1
66 .. 77 66 77 66 .. 77
4 5 4 5 4 5
(un)min (un)k+ijk (un)max

20
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

Example VI : Heavy oil fractionator (all of the above)

 y7 must be kept above Tmin.
 y1 and y2 is to be kept at setpoint(measurements delayed).
 BRD must be minimized to maximize the heat recovery.

PC

reflux drum
T

LC
FC
Upper reflux duty

A Top draw

Intermediate reflux duty

stri-
T pper
FC

T
Bottoms reflux
duty Side draw
A
FC

LC
T

Feed Bottoms

21
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

Classical Solution:
Not clear

 how to use temperature measurements to ght the eect of delays, un-

reliability, etc. of analyzers.
 how to accommodate the optimization requirement.
MPC Solution :

w1
w2

y1
u1 comps.
y2
Heavy-Oil y3
u2
Fractionator
y6 temps.
u3
y7

2 3 2 3 2 2 3 2

u1
Xp 66 y1 77 66 y1 77 Xm 66 77
min 6 y 77 6 7
, 64 y2 75 + 664
u2 77

Uk l=1 64 2 5 i=1 5
u3 u
u3
k+ljk 3 ref Q k+ijk R
y7  Tmin
plus other input constraints.

22
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

Example VII : Tennessee Eastman process(supervisory control re-

quirements)

LOOP #1 PI

PI

COMPRESSOR
PI CWS XA
LOOP #2 A
SC N XB
CONDENSER
PI A
LT PI XC
LT PT CWR L
TI Y XD
Z
XE
CWS VAP/LIQ E
TI
SEPARATOR R XF

LOOP #4 XH
XA
A
TI TI
XB N
A CWR
XC STRIPPER
L PI A
XD
XD Y PI
N
REACTOR
Z PI A XE
LT
XE L
E
STM XF
R Y
XF
Z XH
COND
E
R
PI LOOP #3

PI

Tier Loop # Controlled Variables Manipulated Variables

1 Reactor Level Compressor recycle valve
I 2 Separator Level Separator liquid ow
3 Stripper Level Stripper liquid ow
4 Reactor Pressure Reactor cooling water ow
2 3 2 3 2
p
X 6 Q 7 r mX ,1 2
min 4 5 6 17
4 5 + ,
uk+ijk R

Uk l=1 G=H r 2 k +l jk i=0
k+ljk Q
Pr  (Pr )max
(Hr )min  Hr  (Hr )max
where
Pr : reactor pressure, (Pr )s: setpoint to reactor pressure loop
Hr : reactor level, (Hr )s: setpoint to reactor level loop
Q: total product ow G=H : mass ratio between products G and H
FD : D feed ow FE : E feed ow

23
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

1.3.2 SUMMARY
Advantages of MPC over Traditional APC

 control of processes with complex dynamics

 decoupling and feedforward control are \built in" (traditional approaches
are dicult for systems larger than 2  2).
 constraint handling
 utilizing degrees of freedom
 consistent methodology
 realized benets: higher on-line times and cheaper implementation /
maintenance

24
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

1.4 INDUSTRIAL USE OF MPC: OVERVIEW

1.4.1 MOTIVATION
Protability potential with multivariable control
 reduce variability of key variables by 50 % or more.
 increase yield of more valuable products,

25
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

Benets of control and optimization

26
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

MPC within plant automation hierarchy

27
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

100
Traditional
80 APO
Investment, %

Real-time
60 Optimization
Constrained
Multivariable
40 Control

20
DCS
0
0 20 40 60 80 100

Potential,%

28
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

Benets of control and optimization

1.4.2 SURVEY OF MPC USE

Current status of MPC application(North America/Europe)
Applications by ve major MPC vendors (Badgewell, 1996)
Area DMC Coop. Setpoint Inc.Honeywell Adersa Treiber Total
Promatics Controls
Rening 360 320 290 280 250 1500
Petrochemicals 210 40 40 - - 290
Chemicals 10 20 10 3 150 193
Pulp and Paper 10 - 30 - 5 45
Gas - - 5 - - 5
Utility - - 2 - - 2
Air Separation - - - - 5 5
Mining/Metallurgy - 2 - 7 6 15
Food Processing - - - 41 - 41
Furnaces - - - 42 - 42
Aerospace/Defence - - - 13 - 13
Automotive - - - 7 - 7
Other 10 20 - 45 - 75
Total 600 402 377 438 416 2233
First App DMC:1985 IDCOM-M:1987 PCT:1984 IDCOM:1973 OPC:1987
SMCA:1993 RMPCT:1991 HIECON:1986
Largest App 603283 3528 28 20 - 2419

31
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

Current status of MPC application(Japan):

1.5 HISTORICAL PERSPECTIVE

 The idea of using a model for prediction and optimal control computation
has been around for long time.

32
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

Note that research on optimal control was most vigorous in the 50s and
60s. Most of the results during this period were for open-loop optimal
control. For feedback implementation, they hinted the idea of receding
horizon control. However, most of the results were impractical due to
the lack of implementation hardware.
Some remarkable results of this period include
{ Pontryagin's maximum principle.
{ Hamilton-Jacobi-Bellman equation for optimal feedback control.
{ Feldbaum's dual control concept.
 Due to the lack of sophisticated hardware, it was highly desirable to de-
rive a closed-form control law that could be implemented with compu-
tational equipments available at reasonable costs. The work of Kalman
represents a major achievement in this regard.
Kalman derived analytical solutions to
{ linear quadratic optimal control problem for deterministic systems
) (1-horizon) LQR
{ the same problem for Gaussian stochastic systems ) LQG = LQR
+ Kalman lter
These solutions were important because they represented very few ana-
lytial solutions to optimal feedback control problems.
However his work (based on the idea of stage-wise solution using dynamic
programming) could not be extended to constrained systems or nonlinear
systems.
 In the 70's, Kwon and Pearson discussed the idea of receding horizon
control (a cornerstone of MPC) in the context of LQ optimal control
and how to achieve stability with such a control law.
However, they did not consider constrained / nonlinear problems and
failed to motivate the need for such an approximation.
33
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

 In the late 70s and early 80s, there were several reports of successful use
of optimal control concepts in oil industries. For instance, Charlie Cut-
ler reported the success of implementing the so-called Dynamic Matrix
Control in Shell Oil rening units.
This started an avalanch of similar algorithms and industrial projects.
From here on, process control would never be the same.
 The industrial success renewed the academics' enthusiasm for optimal
control. Prototypical algorithms were analyzed and improved. Also,
connections to the Kalman's work and other optimal control theories
were brought forth.
 Now, MPC is an essential tool-of-trade for process control engineers.
There are more than a dozen vendors oering commercial software pack-
ages and engineering services. There is probably not a single major oil
industry where MPC is not employed in its new instaillations or revamps.
MPC is also very well understood from a theoretical standpoint.

1.6 CHALLENGES
1.6.1 MODELING & IDENTIFICATION

34
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

 most models for MPC design came from identication rater than funda-
mental modeling.
 system ID takes up to 80-90% of the cost and time in a typical imple-
mentation of a model based controller.

35
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

Current Practice

36
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

 Example illustrating the diculty in multivariable system identication

True plant(Model A)
2 3
1 64 0:878 ,0:864 75
10s + 1 1:086 ,1:092

Model B
2 3
1 64 0:867 ,0:875 75
10s + 1 1:095 ,1:083

Model C
2 3
1 ,0:804 75
64 0:808
10s + 1 1:006 ,1:025

37
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

1.6.2 INCORPORATION OF STATISTICAL CONCEPTS

 Improved Disturbance Rejection:
One can capture the temporal / spatial correlations of disturbance eects
in the form of statistical models, using historical data, and use them for
better prediction / control.
 Inferential Control of Composition / Quality Variables
Many quality variables (e.q., variables directly related to end-use prop-
erty) and compositions are not on-line measurable or dicult to measure
on-line. Delays and large sample time involved in the laboratory anal-
ysis can make tight control impossible. In this case, correlations with
other more easily measurable variables can be captured and utilized for
inferencing.
 Control System Performance Monitoring / Failure Diagnosis
The concept of Statistical Process Control (SPC) can be used to detect
unusual behavior and also diagnose the cause of performance deteriora-
tion of MPC.

41
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

Motivational Example: Batch Pulp Digester

On-line On-line
Model
measurements prediction

y q

input
process controller
u

42
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

Key Aspects of the Problem

 Frequent Batch-to-Batch Variations in Operating Conditions
- Feed conditions
- Process parameters
! heat transfer coecients
! reaction rate parameters
 Lack of On-Line Quality Measurements
- Prohibits real-time feedback control

43
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

Traditional Control Method: Statistical Quality Control

 SPC Based on On-Line Measurements

44
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

Inferential Prediction and Control of Product Quality

On-line On-line
Model
measurements prediction

y q

input
process controller
u

47
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

1.6.3 NONLINEAR CONTROL

Linear model-based control can be inadequate for
 highly nonlinear processes (reactors, high-purity distillation column, batch
processes, etc)
 process with large operating windows
MPC is a promising approach, but diculties are in
 obtaining models (esp. through identication)
 computational complexity (NLP must be solved on-line)
 lack of theoretical understanding on the stability and robustness.

48
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

1.6.4 OTHER ISSUES

 control system maintenance
 integration with on-line optimization
 discrete-event systems, hybrid systems(e.q., start up)

49
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

Chapter 2
DYNAMIC MATRIX CONTROL
Dynamic Matrix Control
 Proposed by C. Cutler at Shell (later became the President of DMCC).
 Based on a system representation using step response coecients.
 Currently being marketed by AspenTech (in the name of DMC-Plus).
 Prototypical of commercial MPC algorithms used in the process
industries.
We will discuss the core features of the algorithm. There may be some
dierences in details.

2.1 FINITE IMPULSE AND STEP RESPONSE

MODEL
2.1.1 OVERVIEW OF COMPUTER CONTROL
Computer Control System

50
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

PLANT

HOLD SAMPLER

Input Sequence Output Sequence

CONTROL
ALGORITHM

1 2 3 4 5 6 1 2 3 4 . . .

Model for Computer Control

Should provide the following relation:
fv(0); v(1); v(2);


; v(1)g model
! fy(1); y(2);


; y(1)g

51
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

We will concentrate on linear models. v and y are deviation variables, i.e.,

steady state is dened as
v (k) = 0 8k
0
! y (0) = 0 8k
0

2.1.2 IMPULSE RESPONSE AND IMPULSE RESPONSE

MODEL
Impulse Response

PLANT

Assumptions:
- H0 = 0: no immediate eect of impulse response
- 9 n s.t.Hn+1 = Hn+2 =


= 0: \Finite Impulse Response"
(reasonable for stable processes).
Examples:

FC

52
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

Finte Impulse Response Model

Superposition means =) \Response adds and scales."

PLANT ?

PLANT

PLANT

Using the superposition described above,

y(k) = H1v(k , 1) + H2v(k , 2) +





+ Hnv(k , n)

53
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

v(k-1)

y(k)
v(k-2)

=
k-n k
v(k-1) δ(k-1)

H 1 V (K -1)

+ k +

+ +

H N V (K - N )

v(k-n) δ(k-n)
k-n k

NOTE: need to have n-past inputs (v(k , 1);


; v(k , n)) in the memory.

2.1.3 STEP RESPONSE AND STEP RESPONSE MODEL

Step Response

PLANT

Assumptions:
- S0 = 0: no immediate eect of step input
- Sn+1 = Sn+2 =


= S1: equivalent \Finite Impulse Response"
54
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

(reasonable for stable processes)

Relation between Impulse Response and Step Response:
k
X
Sk = Hiv(k , i)
i=1
where v(k , i) = 1 for 1  i  k. Hence,
Sk = Pki=1 Hi
Hk = Sk , Sk,1
Truncated Step Response Model

υ
2

1 1 1
1
∆υ(0) ∆υ(1)
1 2 1 2

1
∆υ(2)
3
−
2

As shown above, any z.o.h. signal v(t) can be represented as a sum of steps:
1
X
v(t) =
v(i)S (t , i)
i=0
where
v(i) = v(i) , v(i , 1) and S (t , i) is a unit step starting at the ith
time step.
Using this and superposition,
y(k) = S1
v(k , 1) + S2
v(k , 2) +






| v(k , n) +
v(k ,{zn , 1) +


+
v(0))}
+Sn (

v(k,n)

55
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

More compactly,
,1
nX
y(k) = Si
v(k , i) + Snv(k , n)
i=1

v(k-1)

y(k)
v(k-2)
k-n =

=
k

∆v(k-1)
S 1 ∆v(k-1)
k-1
+ +

+ +

∆v(k-n+1)

k-n+1

+ +
k-n
S n v(k-n)
v(k-n)

Note:
1.
v(k , i) instead of v(k , i) appears in the model.
2. v(k , n);
v(k , n + 1); : : : ;
v(k , 2);
v(k , 1) must be stored in the
memory (Same storage requirement as in the FIR model).

56
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

2.2 MULTI-STEP PREDICTION

2.2.1 OVERVIEW
 In control, we are often interested in describing the future output
behavior with a model.
 For a dynamic system, future output behavior depends on both past
and future inputs.

past input trajectory

"Dynamic
states"
X(k) :
memory of
past

Predicted future output

past future

hypothesized future input

trajectory PREDICTOR

past future

Hence, past inputs must be remembered in some form for prediction.

Dynamic states (in an input / output description) are dened as
memory about the past inputs needed for prediction of the
future output behavior
For a same system, states can be dened in many dierent ways, i.e., there
are many ways to remember the past for the purpose of future prediction).
- For instance, states can consist of the entire past input trajectory:
x(k) = [v(k , 1); v(k , 2); : : : : : : ; v(0)]T
57
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

This choice is not practical since the memory size keeps growing with
time.
- For an FIR system, one only has to keep n past inputs (Why?) :
x(k) = [v(k , 1); v(k , 2); : : : : : : ; v(k , n)]T
With this choice of x(k), we can certainly build the prediction of the
future output behavior.
- Since the ultimate purpose of the memory is to predict future output, the
past may be more conveniently tracked in terms of its eect on the
future rather than the past itself. This is discussed next.

2.2.2 RECURSIVE MULTI-STEP PREDICTION FOR AN FIR

SYSTEM
 Separating Past and Future Input Eects
For linear systems, due to the separation principle, the eect of past
and (hypothesized) future inputs can be computed separately and
added:

past input trajectory

past future past future

Y 0(k)

+ +
hypothesized future input trajectory
=
=

past future

58
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

 Past Eects As Memory

Dene Y 0(k) as future output deviation due to past input deviation:
Y 0(k) = [y0(k=k); y0(k + 1=k); : : : : : : ; y0(1=k)]T
where
y0(i=k) =
y(i) assuming v(k + j ) = 0 for j  0
Note that
y0(k=k) = y(k)
since the assumption of v(k + j ) = 0; j  0 does not aect the output
at time k.
Although Y 0(k) is innite dimensional, for FIR system, we only have
to keep n terms (why?):
Y 0(k) = [y0(k=k); y0(k + 1=k); : : : : : : ; y0(n=k)]T
This vector can be chosen as states since it describes the eect of past
input deviation on future output deviation.

59
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

Future output can be written as

2 3 2 3
66 y ( k + 1) 7 66 y 0 (k + 1=k )
7
66 y(k + 2) 777 66 y0(k + 2=k) 777
66 .. 7777 = 66 .. 77
66 66 77
66 ... 777 66 ... 77
66 75 66 77
4 4 0 5
y(k + p) y (k + p=k)
| {z }
Eect of2Past3Inputs From 2 Y 03(k) 2 3
66 H1 77 66 0 77 66 0 77
66 H2 77 66 H1 77 66 0 77
66 . 77 66 . 77 6 77
+ 666 . 777 v(k) + 666 . 777 v(k + 1) + : : : : : : + 6666 ... 77 v(k + p , 1)
77
66 ... 77 66 ... 77 66 ... 77
64 75 64 75 64 5
H Hp,1 H1
| p {z }
Eect of Hypothesized Future Inputs
We can see that such a detion of states can be very convenient for
predictive control.
 Recursive Update of Memory
Memory should be updated from one time step to next. For computer
implementation, the update should occur in a recursive manner.

Y 0 (k)

current memory
update Y 0 (k+1)

υ(k)
new input

Y 0(k) can be updated recursively as follows:

60
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

2
Y 0 (k )
3
! M 0 Y 0 (k ) + Hv (k ) = Y 0 (k + 1)

66 y 0 (k=k )
7 2 0 3 2 3 2 0 3
66 y0(k + 1=k) 777
66 . 77 66 y (k + 1=k) 77 66 H1 77 66 y (k + 1=k + 1) 77
66 .. 77 66 y0(k + 2=k) 77 66 H2 77 66 y0(k + 2=k + 1) 77
66 .. 7
77 6
66 .. 7 6
77 66 .. 77 7 66 .. 77
66 77 66 77 + 66 77 v(k) = 66 6 77
66 0 66 ... 77 66 ... 77 66 ... 77
64 y (k + n , 2=k) 775 66 77 66 77 66 77
y0(k + n , 1=k) 66 y0(k + n , 1=k) 77 66 Hn,1 77 66 y0(k + n , 1=k + 1) 777
4 0 5 4 5 4 0 5
0 y (k + n=k) Hn y (k + n=k + 1)
..

v y Y0 (k)

k-1 k k+1 k+2 k k+1 k+2 k+n-1 k+n k+n+1

v y M 0 Y(k)

k-1 k k+1 k+2 k k+1 k+2 k+n-1 k+n

υ(k)
y H υ(k)
v
k-1 k k+1 k+2 k k+1 k+2 k+n-1 k+n k+n+1

61
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

Mathematically, the above can be represented as

2 3 2 3
66 0 1 0


0 77 66 H1 77
66 0 0 1


0 777 66 H2 77
66 .
Y (k + 1) = 666 .
0 .. .. . . . .. 777 Y 0(k) + 6666 .. 77
77 v(k)
66 0 77 66 77
64 0 0


1 75 7 66 Hn,1
4
77
5
|
0 0 0
{z



0
}
Hn
M0
Note that multiplication by M 0 in the above represents the shift
operation (which can be eciently implemented on a computer).

2.2.3 RECURSIVE MULTI-STEP PREDICTION FOR AN FIR

SYSTEM WITH DIFFERENCED INPUT
Multi-step prediction model can be developed in terms of step response
coecients as well.
 Separating Past and Future Input Eects
Apply the superposition as before, but in a slightly dierent manner:

62
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

past input trajectory

past future past future

~
Y (k )

k+n-1

+ +
hypothesized future input trajectory

=
 Past Eects As Memory
Dene Y~ (k) as future output deviation due to past input deviation plus
current bias:
Y~ (k) = [~y(k=k); y~(k + 1=k); : : : : : : ; y~(k + n , 1=k)]T
where
y~(i=k) =
y(i) assuming
v(k + j ) = 0 for j  0
Note that y~(k + n , 1=k) = y~(k + n=k) = : : : = y~(1=k), thus allowing
the nite-dimensional representation of future output trajectory. This
vector can be chosen as states.

63
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

Future output can be written as

2 3 2 3
66 y ( k + 1) 7 66 y~( k + 1 =k ) 7
66 y(k + 2) 777 66 y~(k + 2=k) 777
66 .. 7777 = 66 .. 77
66 66 77
66 ... 777 66 ... 77
66 75 66 77
4 4 5
y(k + p) y~(k + p=k)
| {z }
Eect
2 of3 Past Inputs
2 + Current 3 Bias (from Y~2 (k))3
66 S1 77 66 0 77 66 0 77
66 S2 77 66 S1 77 66 0 77
66 . 77 66 . 77 66 . 77
+ 66 . 77
v(k) + 66 . 77
v(k + 1) + : : : + 666 .. 777
v(k + p , 1)
6 . 7 6 . 7
66 ... 77 66 ... 77 66 .. 77
64 75 64 75 64 75
S Sp,1 S1
| p {z }
Eect of Hypothesized Future Input Changes
We can see that such a detion of states can be very convenient for
predictive control.
 Recursive Update of Memory
Y~ (k) can be updated recursively as follows:
2
Y~ (k) 3 M Y~ (k) + S
v(k) = Y~ (k + 1)
66 y~(k=k) 7 2 3 2 3 2 3
66 y~(k + 1=k) 777 y
~ (k + 1 =k ) S y
~ ( k + 1 =k + 1)
66 77 66 77 66 1 77 66 77
66 .
.
. 77 6
6 y
~ (k + 2 =k ) 7
7 6
6 S 7 6 y
~ ( k + 2 =k + 1) 77
6 7 6 2 77 6
6 77
66 . 77 6 .
. 7 6 .
. 7 6 .
.
66 . 77 6
6 . 7
7 6
6 . 77 7 6
6 . 77
66 77 6
6 .
. 7
7 + 6
6 . 7
v (k ) = 6
6 . 77
64 y~( k + n , 2 =k ) 75 6
66 . 7 6
77 66 . 7
77 6
66 . 77
y~(k + n , 1=k) 66 y~(k + n , 1=k) 77 66 Sn,1 77 66 y~(k + n , 1=k + 1) 777
4 5 4 5 4 5
y~(k + n=k) y~(k + n=k) Sn y~(k + n=k + 1)
...

64
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

v ∆υ (k ) y ~
Y (k + 1 )

k-1 k k+1 k+2 k k+1 k+2 k+3 k+n-1 k+n

v y ~
M Y (k )

k-1 k k+1 k+2 k k+1 k+2 k+3 k+n-1 k+n

y S ∆υ (k )
v
∆υ (k )
k-1 k k+1 k+2 k k+1 k+2 k+3 k+n-1 k+n

Hence,
2 3 2 3
66 0 1 0


0 77 66 S1 77
66 0 0 1


0 777 66 S2 77
6 ... ... . . . ... 777 Y~ (k) + 6666 ... 77
~Y (k + 1) = 666 ... 77
v(k)
66 7 66 77
66 0 0 0


1 777 66 Sn,1 77
4 5 4 5
0 0 0


1 Sn
| {z }
M
Note that multiplication by M in the above represents a shift
operation of dierent kind (the last element is repeated).

65
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

2.2.4 MULTIVARIABLE GENERALIZATION

Si =
i2th step response3 coecient matrix

S S
= 64 1;1;i 1;2;i 75
S2;1;i S2;2;i
In general 2 3
66 S1;1;i S1;2;i





S1;nv ;i 77
66 S ... . . . . . . ... 77
66 . 2 ; 1 ;i 77
Si = 666 .. ... . . . . . . ... 77
66 ... ... . . . . . . ... 77
64 77
5
Sny;1;i Sny;2;i





Sny ;nv ;i

66
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

Again, dene Y~k+1 and Y~k in the same manner as before (now they are
(n
ny )-dimensional vectors). Then,
Y~ (k + 1) = M Y~ (k) + S
v(k)
where 2 3
66 0 I 0


0 77
66 0 0 I


0 777
66 . .
M = 666 . . .. . . . .. 777
66 0 0 7
64 0


I 7775
20 0 30


I
66 S1 77
66 S2 77
66 . 77
S = 666 . 77
77
66 S 77
64 n,1 5
Sn
where I is an ny  ny indentity matrix. Again, it merely represents the
shift-operation; such a matrix does not need to be created in reality.

2.3 DYNAMIC MATRIX CONTROL ALGORITHM

2.3.1 MAJOR CONSTITUENTS

67
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

past future
target
projected
outputs

inputs

k k+1 k+m k+p

horizon
 Memory: stores the eect of past inputs on future outputs.
 Predictor: combines information stored in the memory with model
information to predict the eect of hypothesized future input
adjustments on future outputs.
(y(k+1jk); y(k+2jk); : : :; y(k+pjk)) = f (I (k); (
u(k); : : :;
u(k + m , 1)))
where I (k) denotes the all the available information at k (stored in the
memory).
 Objective function and constraints
 Optimization program
User-chosen parameters are the prediction horizon, control horizon, and
parameters in the objective function and constraints.

2.3.2 BASIC PROBLEM SETUP

The basic system description we assume is

68
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

u : manpulated variable
d : measured / modelled disturbance
wy : unmeasured disturbance + model / bias error eect

2.3.3 DEFINITION AND UPDATE OF MEMORY

Dene the memory (state vector) as the eect of past deviation + current
bias of known inputs (u and d) on the future output behavior:
2 3
y(k) 66 77
6
u(k) =
u(k + 1) =





= 0
77
y(k + 1)
~Y (k) = 666
... 66
d(k) =
d(k + 1) =





= 0
77 with
77
4 wy (k) = wy (k + 1) =





= 0
5
y(k + n , 1)
The memory update simply consists of:
z2
v(}|k,1) 3{
  6
u(k , 1) 7
~
M Y (k , 1) + S S 4
u d 5 ! Y~ (k)

d(k , 1)

69
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

2.3.4 PREDICTION EQUATION

We can develop the prediction based on Y~ (k) in the following manner
(y(k + `jk) denotes y(k + `) predicted at t = k):
2 3 2 3
66 y ( k + 1 j k ) 7 66 y~( k + 1 =k ) 7
66 y(k + 2jk) 777 66 y~(k + 2=k) 777
66 ... 77 66 ... 77
66 77 = 66 77
66 ... 77 66 ... 77
66 77 66 77
4 5 4 5
y(k + pjk) y~(k + p=k)
| {z }

u(k) =
u(k + 1) =





= 0
y(k+1);


;y(k+p) with
d(k ) =
d(k + 1) =





= 0

2 3 2 wy (k3) = wy (k + 1) =




2
= 03
66 S1u 77 66 0 77 66 0 77
66 S u 77 66 S u 77 66 0 77
66 .2 77 66 1 77 66 . 77
+66 . 77
u(kjk) + 66 S2 77
u(k + 1jk) +


+ 666 . 777
u(k + p , 1jk)
6 7 6 u 7
66 ... 77 66 .. 77 66 0 77
64 75 64 75 64 75
2 Sp 3 2 Sp,31 2 S31
u u u

66 S1d 77 66 0 77 66 0 77
66 S d 77 66 S d 77 66 0 77
66 .2 77 66 1 77 66 . 77
.
+66 . 77
d(k) + 66 S2 77
d(k + 1jk) +


+ 666 . 777
d(k + p , 1jk)
6 7 6 d 7
66 ... 77 66 ... 77 66 0 77
64 75 64 75 64 75
2 Sp 3 Sp,1 S1d
d d

66 w(k + 1jk) 77
66 w(k + 2jk) 77
66 . 77
+66 6 .
. 77
66 ... 77
64 77
5
w(k + pjk)
There are more than a few terms (marked with (
jk)) on the right-hand side

70
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

that are unavailabe at time k.

 Assume piece-wise constant disturbances, i.e.,

d(k + 1jk) =
d(k + 2jk) =





= 0
 Assume the eect of unmeasured disturbance, model error, etc. are
described as a piece-wise constant signal.
wy (k +1jk) = wy (k +2jk) =





= wy (kjk)  y| ({zk)} , y|~(k=k
{z )}
real meas: model prediction

 For exibility in adjusting the computational load, consider only m

( p) input moves (
u(kjk);
u(k + 1jk);


;
u(k + m , 1jk)). This
means, in your prediction, assume

u(k + mjk) =
u(k + m , 1jk) =


=


=
u(k + p , 1jk) = 0

71
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

In summary,
2 3 2 3
66 y~(k + 1=k ) 77 66 S1d 77
66 y~(k + 2=k) 77 66 S d 77
66 . 77 66 .2 77
Yk+1jk = 66 6 . 77 + 66 .
77 66 .
77
d(k)
77
66 ... 77 66 .. 77
64 5 4 d 5
y~(k + p=k) S
| {z } | n {z }
MpY~ (k) S d
d(k)
from feedforward
memory term
2 u 3
2 3 6 S 0





0 7 2 3
66 y(k) , y~(k=k) 77
u(kjk)
1
66 u 77 6 7
66 y(k) , y~(k=k) 77 6
66 . S 2 S u
1 0


0 7
7 66
u(k + 1jk) 777
6
66 77 66 . .. . . . . . . .. 77 66 77
+ 666 .
.. 77 + 6 77 6 .
. 77
7 66 S u S u





S u 77 66 77
66 .. 77 66 .m m,1 1 7 66 .. 77
64 75 66 . . . . . . . . . . .. 7 64
7
5
7
y(k) , y~(k=k) 4 u u 5
u( k + m , 1 j k )
| {z } Sp Sp,1





Spu,m+1 | {z }
| {z }
Ip(y(k) , y~(k=k)) S u
U (k)
feedback future
dynamic
term input
matrix
moves
NOTE: More complex (dynamic) extrapolation of the feedback errors is
possible. For instance, for ramp disturbances, use

d(k) =
d(k + 1jk) =





=
d(k + p , 1jk)
wy (k + `jk) = wy (kjk) + `(wy (kjk) , wy (k , 1jk , 1))

72
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

w(k)-w(k-1)

k-1 k k+1 k+2

∆d(k+2)
∆d(k+1)
∆d(k)

k-1 k k+1 k+2

2.3.5 QUADRATIC CRITERION

min
u(j jk) [V (k) =
Ppj=1(r(k + ijk) , y(k + ijk))T Q(r(ki + ijk) , y(k + ijk))
+ Pm`=0,1
uT (k + `jk)R
u(k + `jk)
Q and R are weighting matrices; they are typically chosen as diagonal
matrices.

73
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

Note that the objective function can be rewritten as

2 3T 2 32 3
66 r ( k + 1j k ) , y (k + 1 jk ) 7 6 Q 77 66 r ( k + 1 j k ) , y ( k + 1 jk ) 7
66 r(k + 2jk) , y(k + 2jk) 777 666 Q 77 66 r(k + 2jk) , y(k + 2jk) 777
V (k) = 666 .. 77 66
7 6 . .
77 66
7 6 . 77
77
64 75 64 . 76 .
54 5
r(k + pjk) , y(k + pjk) Q r(k + pjk) , y(k + pjk)
2 3T 2 32 3
66
u(kjk) 77 66 R 77 66
u(kjk) 7
66
u(k + 1jk) 77 66 R 77 66
u(k + 1jk) 777
+ 666 ... 77 66
7 6 . .
77 66
7 6 ... 77
77
64 75 64 . 75 64 5

u(k + m , 1jk) R
u(k + m , 1jk)
+
V (k) = (R(k +1jk),Y (k +1jk))T Q (R(k +1jk),Y (k +1jk))+
U T (k)R
U (k)
where
2 3 2 3
66 r (k + 1 j k ) 7 66 y ( k + 1 j k ) 7
66 r(k + 2jk) 777 66 y(k + 2jk) 777
66 . 77 66 . 77
R(k + 1jk) = 66 6 .
. 77 ; Y (k + 1jk) = 66
7 6 .
. 77
66 ... 77 66 ... 77
64 75 64 77
5
r(k + pjk) y(k + pjk)
and
Q = blockdiag[Q; Q; : : : ; : : : ; Q]; R = blockdiag[R; R; : : : ; : : : ; R]
Note that
Y (k + 1jk) = |MpY~ (k) + S d
d(k){z+ Ip (y(k) , y~(k=k))} +S U
U (k)
known term
Hence, V (k) is a quadratic function of
U (k).

74
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

2.3.6 CONSTRAINTS
Constraints include
 Input magnitude constraints
 Input rate constraints
 Output magnitude constraints
At t = k, one has

past future y
target max

projected output u
max

inputs

k k+1 k+m-1 k+p

horizon

umin  u(k + `jk)  umax

j
u(k + `jk)j 
umax; ` = 0;


; m , 1
ymin  y(k + j jk)  ymax; j = 1;


; p
We want to express the above constraints as a linear inequalty in the form of
C U
U (k)  C (k)

75
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

Manipulated Input Magnitude Constraints

umin  u(k + `jk)  umax; ` = 0;


; m , 1

+
u(k+`jk)
z }| {
X̀
u(k , 1) +
u(k + ijk)  umin
i=0
,u(k , 1) , X̀
u(k + ijk)  ,umax
| i{z=0 }
,u(k+`jk)
+
2 2 33 2 2 33
66 66 I 0


0 7 77
7 66 66 umin , u(k , 1) 77 77
66 66 I I 0 .. 77 77 66 66 u , u(k , 1) 77 77
66 66 77 77 2 3 66 66 min 77 77
66 66 .. .. ... 7
0 75 77 6 7
u(kjk) 6 6 .
. 77 77
66 4 6 77 66 64 57
66
66 2
I
I


I 3 7777 666
u(k + 1jk) 777  6666 2 umin , u(k , 1) 3 7777
66 6 I 0


0 7 777 6664 .. 77 6
66 6 umax , u(k , 1) 7 777
7
66 66 ... 777 77
u(k + m , 1jk) 5 66 666 umax , u(k , 1) 777 77
66 , 66 I
I 0 77 777 66 , 6 77 777
66 66 .... ... 0 775 77 6 6
66 66 .
. 77 77
64 64 5 4 4 55
I I


I umax , u(k , 1)
+
2 2 33
66 66 umin , u(k , 1) 77 77
66 66 ... 77 77
2 3 6
66 4 6 75 77
7
64 IL 75
U (k)  66 2 umin , u(k , 1) 3 77
66 7
,IL 66 66 umax , u(k , 1) 77 777
66 , 66 ... 77 77
64 64 75 75
umax , u(k , 1)

76
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

Manipulated Variable Rate Cosntraints

j
u(k + `jk)j 
umax; ` = 0;


; m , 1
+
,
umax 
u(k + `jk) 
umax
+

u(k + `jk)  ,
umax
,
u(k + `jk)  ,
umax
+
2 2 3 3 2 2 33
66 66 I 0


0 77 77 66 66
umax 77 77
66 66 0 I 0 . 77 77 . 66 66
u 77 77
66 66 7 7
7 2 3 66 , 66 max 77 77
. .
66 66 .


. . 0 77 77 7

u ( k j k ) 66 66 .. 77 77
66 4 5 76 6 7
7 6 4 5 77
66 0 0


I 777 66
u(k + 1jk) 77 666
umax 7
66 2 3 77 66 .. 7
7  6 2 3 777
66 6 I 0


0 7 77 66 77 66 6
umax 77 777
66 66 . 77 77 4 5 666 66 77 77
66 , 66 0 I 0 . 77 77
u(k + m , 1jk) 66 , 66
umax 77 77
66 6 ..


. . . 0 7 77
6 7 66 66 .. 77 77
64 46 7
5 75 64 64 55
0 0


I
umax
+
2 2 33
66 66
umax 77 77
66 , 66 ... 77 77
2 3 66 64 75 77
6 7
64 I 75
U (k)  66 2
umax 3 777
66
,I 66 66
umax 77 777
66 , 66 ... 77 77
64 64 75 75

umax

77
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

Output Magnitude Constraints

ymin  y(k + j jk)  ymax; j = 1;


; p

+
y(k + j jk)  ymin
,y(k + j jk)  ,ymax
+
2 3 2 3
M
64 p ~
Y ( k ) + S d
d(k ) + I (y (k ) , y~(k=k )) + S U
U (k )
p Y
75  64 min 75
,MpY~ (k) , S d
d(k) , Ip(y(k) , y~(k=k)) , S U
U (k) ,Ymax
where 2 3 2 3
66 ymin 77 66 ymax 77
66 ymin 77 66 ymax 77
Ymin = 666 . 77
77 Ymax = 666 . 77
77
64 . 5 64 . 5
ymin ymax
since
2 3
6 y(k + 1jk) 77

666 ... 77 = Mp Y~ (k)+S d
d(k)+Ip(y(k),y~(k=k))+S U
U (k)
Y (k+1jk) = 6 75
4
y(k + pjk)
+
2 3 2 3
64 S U 75
U (k)  64 Ymin , MpY~ (k) , S d
d(k) , Ip(y(k) , y~(k=k)) 75
,S U ,Ymax + MpY~ (k) + S d
d(k) + Ip(y(k) , y~(k=k))

78
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

In summary, we have
2 2 33
66 66 umin , u(k , 1) 77 77
66 66 .. 77 77
66 64 75 77
2 umin , u(k , 1) 3 777
66 7
66
2 3
66
66 66 umax , u(k , 1) 77 777
, 666 ... 77 77
66 IL 77 66
66 4 75 77
66 ,IL
66
77
77 66 umax , u2(k , 1) 3 777
66 I 77 66 7
66 77
U (k)  666 66
umax 77 777
66 ,I 77 66 , 666 ... 777 777
66 U 77 66 4 5 77
64 S 75 66 2
umax 3 777
,S U 66
66 66
umax 77 777
66
66 , 666 ... 777 777
66 4 57
66 2
umax 3 777
66 6
44
Ymin , MpY~ (k) , S d
d(k) , Ip(y(k) , y~(k=k)) 75 7775
~
,Ymax + MpY (k) + S
d(k) + Ip(y(k) , y~(k=k))
d

The above is in the form of linear equality,

C U
U (k)  C (k)
Note that C U is a constant matrix while C (k) must be updated at each time
step.
Although not treated here, time-varying constraints can easily be
incorporated into the formulation.

2.3.7 QUADRATIC PROGRAMMING

Problem:

79
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

At each sample time, we have a minimization with an objective function

V (k) = (R(k +1jk),Y (k +1jk))T Q (R(k +1jk),Y (k +1jk))+
U T (k)R
U (k)
with the prediction equation constraint
Y (k + 1jk) = MpY~ (k) + S d
d(k) + Ip (y(k) , y~(k=k)) + S U
U (k)
and the inequality constraint
C U
U (k)  C (k)
Putting Into the Standard QP Form:
Substituting the prediction equation constraint into the objective gives
 
V (k) = E T (k)Q E (k) , 2E T (k)Q S U
U (k) +
U T (k) S U T Q S U + R
U (k)
| {z } | {z }
G T (k ) H
E (k) = R(k + 1jk) , MpY~ (k) , S d
d(k) , Ip (y(k) , y~(k=k))
Note that E (k) can be computed with information given to us at time k.
Hence, V (k) is a quadratic function of
U (k) with hessian matrix H and
gradient vector G (k).
Since we have a minimization of a quadratic objective with a linear
inequality constraint, we have a quadratic programming (QP). The
standard form of quadratic programming is
minx xT HxT , gT x
Cx  c
The parameters that should be supplied to the QP solver are H; g; C and c.

80
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

In our case, at t = k,
x :
U (k)
H : H =
(S U )T Q S U + R
g : G (k) =
2 2(S U )T3 Q (MpY~ (k) + S d
d(k) + Ip(y(k) , y~(k=k)) , Rk+1jk )
66 IL 77
66 ,IL 77
66 77
C : C u =
666
6 I 77
7
66 ,I 777
66
64 S U 7775
2,S
U
2 33
66 66 umin , u(k , 1) 77 77
66 66 .. 77 77
66 64 75 77
2 umin , u(k , 1) 3 777
66 7
66
66
66 66 umax , u(k , 1) 77 777
66 , 666 .. 77 77
66 4 75 77
66 umax , u2(k , 1) 3 777
6 77

666 6
u
66 . 77 777 7
c : C (k ) = 6 max
66 66 . 77 77
66 4 5 77
66 2
u max 3 777
66
66 66
umax 77 777
66
66 , 666 .. 777 777
66 4 5 77
66 2
u max 3 77
66 6 Ymin , MpY~ (k) , S
d(k) , Ip(y(k) , y~(k=k)) 7 777
d
44 ~ 55
,Ymax + MpY (k) + S
d(k) + Ip(y(k) , y~(k=k))
d

Following are some comments on quadratic programming.

 QP is convex and therefore fundamentally tractable.
 The solution doesn't necessarily lie at the vertices of feasible
region(unlike LPs). One may have any number of active constraints
81
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

(up to the QP dimension).

 The size of QP is m  nu where m is the control input horizon and nu
is the number of input. Computational time of QP depends on many
things (e.g., Hessian size, its structure, number of constraints, the
proximity of the solution to the active constraints) and is dicult to
predict.
 O-the-shelf QP solver is available, but is often not the best choice in
terms of computational eciency. Because the hessian and gradient for
QP tends to be highly structured (sparse), an algorithm tailored to
take advantage of this is recommended.
 QP solver requires inversion of the hessian. Since the hessian is a
constant matrix (given xed input / output weights and model
parameters), it only needs to be inverted once o-line. This eliminates
the time-consuming step of inverting the hessian at each QP run. Only
when the weighting matrices are model parameters are changed,
hessian needs to be recomputed and inverted in the background.
 Since most QPs are feasible-path algorithms, the number of inequality
constraints also aect the computational time. One should use the
constraints sparingly.
 The most well-known solution strategy is the active set strategy. In
this method, rst a feasible solution is found. Then, the least squares
problem is solved with the active constraints as equality constraints.
The optimality of the solution is checked through Kuhn-Tucker
conditions. If they are not satised, the active constraint set is
updated and the procedure is repeated.
 Another emerging method is the interior point (IP) method. In the IP
method, a barrier function is used to trap the solution within the
feasible region. Newton interation is used to converge to the optimum.
This method has many attractive features like quick convergence (most
82
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

problems converge with 5-50 iterations regardless of the problem size)

and ability exploit the problem structure.

2.3.8 SUMMARY OF REAL-TIME IMPLMENTATION

1. Initialization: Initialize the memory vector Y~ (0) and the reference
vector R(1j0). Set k = 1.
2. Memory Update:
Y~ (k , 1) ! M Y~ (k , 1)+ S u
u(k , 1)+ S d
d(k , 1) ! Y~ (k)
2 3
66 y
~ ( k , 1 =k , 1) 77 2 3 2 3 2 3
66 y~(k=k , 1) 777 y k=k ,
66 66 ~ ( 1) 77 66 S1 77 66 y
~ ( k=k ) 7
66 .. 77
77 66 y~(k + 1=k , 1) 77 66 S2 77 66 y~(k + 1=k) 777
66 . 77 66 ... 77 66 . 77 66 ... 77
66 . 66 77 66 .. 77 66 77
66 77 66 .. 77+66 .. 77
v(k,1) = 66 .. 77
64 y~(k + n , 3=k , 1) 775 66 77 66 77 66 77
66 7 6 77 66 7
y~(k + n , 2=k , 1) 64 y~(k + n , 2=k , 1) 775 664 Sn,1 75 64 y~(k + n , 2=k) 775
y~(k + n , 1=k , 1) y~(k + n , 1=k , 1) Sn y~(k + n , 1=k)
..
3. Reference Vector Update: Update R(k + 1jk) by shifting R(kjk , 1)
and entering a new reference value.
2 3 2 3
66 r (k jk , 1) 7 66 r ( k + 1 j k ) 7
66 r(k + 1jk , 1) 777 66 r(k + 2jk) 777
66 ... 77 66 ... 77
66 77 66 77
66 .. 77 66 .. 77
66 77 66 77
66 7 66 7
64 r(k + p , 2jk , 1) 775 64 r(k + p , 1jk) 775
r(k + p , 1jk , 1) r(k + pjk)
4. Measurement Intake: Take in new measurement y(k) and
d(k).

83
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

5. Calculation of the Gradient Vector and Constraint Vector:

G~ (k) = 2(S U )T Q (MpY~ (k) + S d
d(k) + Ip(y(k) , y~(k=k)) , R(k + 1jk))
Update the constraint vector C (k) similarly.
6. Solve QP: Call the QP subroutine with pre-inverted H,C U and
computed G (k); C (k).
7. Implementation of input: Implement
u(kjk):
u(k) = u(k , 1) +
u(kjk)
8. Go back to Step 2 after setting k = k + 1.

2.4 ADDITIONAL ISSUES

2.4.1 FEASIBILITY ISSUE AND CONSTRAINT
RELAXATION
 Output constraints can become infeasible (impossible to satisfy). For
example, if we require ,  y(k + `jk)   for all `, as  ! 0, the
constraint becomes infeasible.
 When the QP is declared infeasible, one must relax the output
constraints. Various ways to relax the constraints exist:
{ Relax the constraint starting from the initial time one by one until
it is feasible.

84
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

y max

k k+H c

Relax the constraints between

k+1 and k+H c
-1

{ Soften the constraint and penalize the degree of softening:

min;
U (k)[Usual Objective] + 2

ymin ,   y(k + `jk)  ymax + 

plus other constraints

2.4.2 GUIDELINES FOR CHOOSING THE HORIZON SIZE

In order to obtain good closed-loop properties and consistent tuning eect
from problem to problem, it is recommended to use a very large or
preferably innite prediction horizon (Long-sighted decision making
produces better results in general). 1-horizon DMC can be implemented in
the following way:
 choose m as large as possible (within the computational limit).
 choose
p=m+n
where n is the number of time steps for step responses to settle.
 add constraint
y(k + pjk) = 0

85
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

Note that the above choice of p with the equality constraint amounts to
choosing p = 1. Stability of the closed-loop system is guaranteed under
this cuhoice (regardless of choice of m). Choice of m is not critical for
stability; a larger m should result in better performance at the expense of
increased computational requirement.

k+m-1 k+m+n-1
N time steps

k+m-1

The lesson is
 Use large enough horizon for system responses to settle.
 Try to penalize the endpoint error more (if not constrain to zero).

2.4.3 BI-LEVEL FORMULATION

In the DMC algorithm, control computation at each sample time is done in
two steps:
 Steady State Optimization: Here model prediction at steady state is
used to determine the optimal steady state. The steady-state model is
86
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

in the form of
y(1jk) = Ks (| u(1jk) ,{z u(k , 1))} +b(k)

us (k)
With only m moves considered,

us(k) =
u(kjk) +
u(k + 1jk) + : : : : : : +
u(k + m , 1jk)
and with FIR assumption,
y(1jk) = y(k + m + n , 1jk)
and Ks = Sn. Hence, the steady prediction equation can be easily
extracted from the dynamic prediciton equation we had earlier.
In terms of the optimization criterion, various choices are possible.
{ Most typically, some kind of linear economic criterion is used along
with constraints on the inputs and outputs:
min [`(u(1jk); y(1jk))]

us (k)
In this case, a linear programming (LP) results.
{ Sometimes, the objective is chosen to minimize the input move size
while satisfying various input / output constraints (posed by
control requirements, actuator limits plus those set by the rigorous
plant optimizer):
min [j
us(k)j]

u (k) s

Again, an LP results.
{ In the pure regulation problems where setpoint for the output is
xed, one may use
min [(r , y(1jk))T Q(r , y(1jk))]

us (k)

87
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

This combined with subsequently discussed QP results in

Innite-Horizon MPC.
 Dynamic Optimization: Once the steady-state target is xed, the
following QP is solved to drive the outputs (and sometimes also inputs)
to their chosen targets quickly without violating constraints:
hPm+n,2
i=1 (y (k + ijk ) , y (1jk )) Q(yi(k + ijk ) , y (1jk ))
min
u(j jk)  T 
+ Pm,1
uT (k + j jk)R
u(k + j jk)
j =0
subject to

u(kjk) +
u(k + 1jk) + : : : : : : +
u(k + m , 1jk) =
us (k)
plus various other constraints. This is a QP.
The last constraint forces y(k + m + n , 1jk) to be at the optimal
steady-state value y(k + 1jk).
Note: The above steady-state optimization is to be distinguished from
the rigorous plant-wide optimization. The above is performed at every
sample time of MPC while the rigorous optimization is done much
more infrequently.

88
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

2.4.4 PROPERTY ESTIMATION

 Property data q are usually obtained through on-line analyzer or lab.

analysis.
 Both have signicant delays and limited sampling capabilities (more so
for lab. analysis).
 On-line anayzers are highly unreliable (prone to failures).
 Using more reliable fast process measurments y (and possibly u), we
can estimate product properties at a higher frequency with a minimal
delay.
 The property estimator (sometimes called soft sensor) can be
constructed from a fundamental model or more commonly through
data regression.
 Almost all estimators used in practice today are designed as static
estimators.
 Since process variables exhibit dierent response times, ad hoc dynamic
compensations (e.g., lead / lag elements, delays) are often added to the
static estimator.

89
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

 If the number of process measurements is too large, the dimension can

be reduced through PCA(principal component analysis) or orther
correlation analyses.
 In some cases where nonlinearity is judged to be signicant, Articial
Neural Networks are used for regression.
 Analyzer or lab results can be used to remove bias from the soft sensor.
Suppose the soft sensor takes the form of q^s(p; i) = f (y(p; i)). Then,
q^(p; i) = q^s(p; i) + | (q(p; 0) {z, q^s(p; 0))}; 0    1
bias correction

90
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

2.4.5 SYSTEM DECOMPOSITION

In MIMO processes, some input-output pairs have no or only weak
coupling. Such systems can be decomposed into several subsystems and
separate MPC can be designed for each subsystem.
The decentralized MPC design can reduce computational demand and
improve numerical stability.

 Number of oating point computation in matrix algebra is proportional

to n2 or n3.
 If we can decompose an n-dimensional system into two subsystems
with equal size, the number of computation can be reduced from O(n2)
or O(n3) to O(n2=4) or O(n3=8).
 System decomposition is not a trivial task in general. It is one of the
continuing research issues studied under the title of Control Structure
Synthesis or Design of Decentralized Control.
 Some of the rather obvious cases are as follows:

91
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

Case 1 : Complete Separation

 The system can be decomposed into (U1 U2 U4) , (Y1 Y2 Y4) and
U3 , Y3 disjoint pairs.

92
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

Case 2 : Partial Separation I

 (Y1 Y2 Y4) is not aected by U3. But Y3 is aected by U4

 The system can be decomposed into two subsystems. In this case,
U4 can be treated as a measurable disturbance to the U3 , Y3 loop.

93
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

Case 3 : Partial Separation II

 Y3 is not in uenced by (U1 U2 U4). But, U3 has an in uence on Y4.

 Similarly to above, the problem can be decomposed into two
subproblems. U3 acts as a measurable disturbance to the rst
block.

94
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

Case 4 : Partial Separation III

 If G34 and G43 have slower dynamics and smaller steady state
gains than the other transfer functions, we may decompose the
system as shown in the gure.

95
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

Example: Extrusion Process This example shows how the feedforward

control can be constructed in a real process situation.

96
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

According to the input-output map, uj , yj pair is decoupled from

others while uj ,1 plays a measurable disturbance to yj .
Instead of treating uj ,1 as a measured disturbance, however, it is
better to take yj ,1 as the measured disturbance and compensate its
eect through the feedforward loop.

Decentralization Options
 Decentralization for both model update and optimization.
 Full model update, but decentralized optimization.
 Full model update, full steady-state optimization (LP), but
decentralized dynamic optimization (QP).

97
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

2.4.6 MODEL CONDITIONING

Ill-Conditioned Process ?
 Consider the following process:
2 3 2 32 3 2 3 2 3 2 3
64 y1 75 = 64 2 4 75 64 u1 75 = 64 2 75 u + 64 4 75 u = 64 2 75 (u + 2u )
1 2 1 2
y2 3 6 u2 3 6 3

Two column vectors of the steady state gain matrix are colinear. As a
result, [y1 y2]T lies on v1 for any u1 and u2.
If the set point is given outside v1, it can never be attained.

 This time, let the steady state gain matrix be

2 3
64 2 4 75
3 6:2
Two column vectors are nearly colinear.
Assume that ysp = (,2; 3)T ? v1.
The input is
2 3 2 3,1 2 3 2 3
64 u1 75 = 64 2 4 75 64 ,2 75 = 64 ,61 75
u2 3 6:2 3 30
On the other hand, for ysp = (2 3)T = v1, the input is
98
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

2 3 2 3
64 u1 75 = 64 1 75
u2 0

{ It is possible to control y along the v2 direction but a large input

possibly beyond the constraints is required.
{ Does it make sense to try to control both y1 and y2 independently ?
The answer will depend on the requirements of the process, but in
many cases would be `Not !"
{ If we give up one of the outputs, say y2, and control only y1 at y1sp,
only a small control energy will be required. In this case, y2 will
stay at around 1:5y1sp.
 Since the above gain matrix has a very large condition number, we say
that the process is ill-conditioned or has a strong directionality.

99
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

Analysis using SVD

 Let
2 32 3
G = [ W1 W2 ] 64 1 0 75 64 V1T 75
T
0 2 V2
 Assume that 1  2 where dim(1) = m < n,
y = Gu ) y  W11V1T u
The output is dominated by the modes with large singular values.
 On the other hand,
2 ,1 32 3
0 75 64 W1 75
T
u = G+y = [ V1 V2 ] 64 1 ) u  V2,2 1W2T y
0 2 ,1 W2 T

where + denotes the pseudo-inverse. The input is mostly determined

by less signicant modes associated with small singular values.
 SVD can be extended to a dynamic gain.
G(j!) = W(j!)(j!)VT (j!)

Model Conditioning in Commercial Packages

step 0 It is assumed that operating regime for output y is given with

priority for each output.
step 1 From the SVD of G(steady state gain matrix), count the number of
signicant modes. Let it be m. Notify that n , m outputs are better to
be removed from the controlled variables (CVs).

100
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

step 2 Using u = G+y, check if the input constraints can be violated for
any y within the dened set. If unacceptable, do the next step.
step 3 The designer takes out some of low propority outputs from CVs.
Let the reduced input-output model be yr = Gr u. Repeat step 2 for
the reduced model until the estimated input is acceptable for all
possible yr .

This idea can be slightly generalized to include quantitative weighting for

each output (rather than strict priority).
Model conditioning is needed not only to prevent input constraint violation
(which would be automatically handled by the constrained MPC), but
because low-gain directions are very dicult to identify and gains typically
carry large multiplicative errors (sometimes more than 100%).

101
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

2.4.7 BLOCKING

 Consider an MPC problem with m = 30, nu = 4. At every sampling

instance, MPC has to decide 120 variables through QP. Are all these
values truly impotant in the prediction of major output modes ?
 The technique to reduce the number of input decision variables while
minimizing the degradation of the intended control performance is
called blocking.
 Blocking can enhance robustness of MPC, too.
Concept of Blocking
 Blocking is an approximation of the future input values by a linear
combination of appropriately chosen small number of basis vectors.

Uk  b1
u1k + bb
ubk = B
Uk b  m  nu
where m  nu is the dimension of Uk .
B is called a blocking matrix.
 QP determines
Uk instead of
Uk .
 The most important step in blocking is to choose an appropriate
blocking matrix.

102
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

Time-Domain Blocking - Signal Approximation

Divide the control horizon into several subintervals and decide
piecewise constant input value for each subinterval.

2 3 2 3
66
uk 77 66 1 0 07
66
uk+1 77 66 0 0 0 777
66 ... 77 66 ... ... ... 77
66 77 66 77
66
uk+m1,1 77 66 0 0 0 777
66
u 77 66 0 1 0 777 2 3
66 k+m1 77 66
u
66
uk+m1+1 77 66 0 0 0 777 666 k;1 777
66
66
.. 77 = 66 .
77 66 .
.. .. 77 4
uk;3 5
7
u
66
uk+m2,1
66
77 66 0
77 66 0 0 777 | {zk; 3 }
66
uk+m2 77 66 0 0 1 777
Uk
66
uk+m2+1 77 66 0 0 0 777
66 .. 77 66 . .. .. 77
64 75 64 . 5
|

uk{z+m,1 } |
0 0
{z
0}

Uk B

Rather heuristic.
Many industrial algorithms employ the above technique.

103
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

SVD-based Blocking
SVD of the pulse reponse matrix informs us which input
directions excite the signicant output directions.
Let the SVD of the truncated pulse response matrix over the
control and prediction horizons be
2 32 3
H = [ W1 W2 ] 64 1 0 75 64 V1T 75
T
0 2 V2
where
2 3
66 h1 0


0 77
66 h2 h1


0 77
66 . .. .. .. 77
66 . 77
H = 66 77
66 hm hm,1


h1 77
66 . .. .. .. 77
64 . 75
hp hp,1


hp,m+1
If 1  2, we can choose
B = V1
) Approximate
Uk as a linear combination of dominant
input principal vectors.
 considered to be better than the time-domain blocking in that
it provides structural approximation of MIMO systems, too.

104
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

[Ex. 1] No Model Error Case

2 17 5
3
model 6
G(s) = G (s) = 4 60 :48 s2 +15:6s+1 19:36s2+7:04s+1 75
3 10
10:89s2+4:62s+1 36s2+12s+1

Q = I; R = 0:01I; p = m = 50

,1  Uk  1

Regular MPC(m  nu = 100) SVD of H (b=20)

105
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

[Ex. 2] Model Error Case

2 17 2:3
3
G(s) = 64 (10s+1)(10s2+s+1)
1:3
30s+1
2:8
75
20s+1 (10s+1)(5s2+s+1)

2 1:5 2:2
3
Gmodel (s) = 64 101s:+1
2
30s+1
2:6
75
20s+1 10s+1

Q = I; R = I; p = m = 90
No constraints

Regular MPC(m  nu = 180) SVD of H (b=20)

106
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

Chapter 3
SYSTEM IDENTIFICATION
The quality of model-based control absolutely relies on the quality
of the model.

3.1 DYNAMIC MATRIX IDENTIFICATION

3.1.1 STEP TESTING

107
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

Procedure
1. Assume operation at steady-state with
controlled var.(CV) : y(t) = y0 for t < t0
manipulated var.(MV) : u(t) = u0 for t < t0
2. Make a step change in u of a specied magnitude,
u for
u(t) = u0 +
u for t  t0
3. Measure y(t) at regular intervals:
yk = y(t0 + kh) for k = 1; 2; : : : ; N
where
h is the sampling interval
Nh is approximate time required to reach steady state.
4. Calculate the step response coecients from the data
sk = yk
,uy0 for k = 1; : : : ; N

Discussions
1. Choice of sampling period
 For modeling, best h is one such that N = 30  40.
Ex : If g(s) = Ke,ds=(s + 1),
then settling time  4 + d
Therefore, h  4N+d = 440+d = 0:1 + 0:025d

108
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

 May be adjusted depending upon control objectives.

2. Choice of step size (
u)
 too small :
May not produce enough output change
Low signal to noise ratio
 too big :
Shift the process to an undesirable condition
Nonlinearity may be induced.

109
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

 Trial and error is needed to determine the optimum step size.

3. Choice of number of experiments

110
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

 Averaging results of multiple experiments reduces impact of

disturbances on calculated sk 's
 Multiple experiments can be used to check model accuracy by
cross-validation.
Data sets for Identication $ Data set for Validation
4. An appropriate method to detect steady state is requried.
5. While the steady state (low frequency) charateristics are accurately
identied, high frequency dynamics may be inaccurately charactierized.

3.1.2 PULSE TESTING

Procedure
1. Steady operation at y0 and u0.
2. Send a pulse of size u lasting for 1 sampling period.
3. Calculate pulse response coecients
, y0 for k = 1; : : : ; N
hk = yk u
4. Calculate the step response coecients as a cumulative sum of hk .
k
X
sk = hi for k = 1; 2; : : : ; N
i=1

111
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

Discussions
1. Select h and N as for the step testing.
2. Usually need u 
u for adequate S/N ratio.
3. Multiple experiments are recommended for the same reason as in the
step testing.
4. An appropriate method to detect steady state is requried.
5. Theoretically, pulse is a perfect (unbiased) excitation for linear systems.

3.1.3 RANDOM INPUT TESTING

Concept

fhk g or fA; B; C g or G(s)

112
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

Type of Inputs
1. Pseudo-Random Binary Signal(PRBS)

In MATLAB,  u=u0+del*2*sign(rand(100,1))-0.5;
or  u=mlbs(12);
2. Random Noise

In MATLAB,  u=u0+del*2*(rand(100,1)-0.5);
113
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

Data Analysis - Least Squares Fit

Given fu1; u2; : : : ; uM g and fy1; y2; : : : ; yM g, determine the best t
FIR(nite impulse response) model fh1; h2; : : : ; hN g.

Consider
yk = h1uk,1 + h2uk,2 + : : : + hN uk,N + dk
Assume the eects of initial condition are negligible.
2 3 2 32 3 2 3
66 y1 77 66 u0
u,1 : : : u1,N 7 6 h1 77 66 d1 77
66 77 66 u1 76
66 y2 u0 : : : u2,N 777 666 h2
77 = 66 .
77 66
77 + 66 d2 77
77
66 ... ... ...
77 66 .. ... 77 66 ... 77 66 ... 77
4 5 4 75 64 5 4 5
yM
uM ,1 uM ,2 : : : uM ,N hN dM
y = Uh + d
The least squares solution which minimizes
M
0 N
12
X X
(y , Uh)T (y , Uh) = @ yi , hj ui,j A
i=1 j =1
is  
h^ = UT U ,1 UT y
In MATLAB,  hhat=y\U;

Discussions
1. Random input testing, if appropriately designed, gives better models
than the step or pulse testing does since it can equally excite low to
high frequency dynamics of the process.
2. If UT U is singular, the inverse doesn't exist and identication fails.
! persistent excitation condition.

114
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

3. When the number of coecients is large, UT U can be easily singular

(or nearly singular). To avoid the numerical, a regularization term is
addded the the cost function. ! ridge regression
h i
min
h
(y , Uh)T (y , Uh) + hT h

 
! h^ = UT U + I ,1 UT y
4. Unbiasedness: If d(
) and/or u(
) is zero-mean and u(i) is uncorrelated
with d(j ) for all (i; j ) pairs (these conditions are easily satised.), the
estimate is unbiased.
   
h^ = UT U ,1 UT (Uh + d) = h + UT U ,1 UT d
Since  
E f UT U ,1 UT dg = 0
we have
E fh^ g = h

5. Consistency: In addition to the unbiasedness,

 
^h ! h (or, equivalently E (h^ , h)(h^ , h)T ! 0) as M ! 1:

6. Extension to MIMO identication is straightforward.

The above properties are inherited to the MIMO case, too.

115
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

Example
 Process : h = [h1; h2; h3; h4; h5;


] = [1:5 2:0 5:5 0:1 0


]

 Input : PRBS with

N = 200 u = 0

116
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

 The resulting output response corrupt with measurement noise with

n2 = 0:252 is

 Estimates of fhj g appear as

117
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

3.1.4 DATA PRETREATMENT

The data need to be processed before they are used in
identication.

(a) Spike/Outlier Removal

 Check plots of data and remove obvious outliers ( e.g., that are
impossible with respect to surrounding data points). Fill in by
interpolation.
 After modeling, plot of actual vs predicted output (using measured
input and modeling equations) may suggest additional outliers.
Remove and redo modelling, if necessary.
 But don't remove data unless there is a clear justication.
(b) Bias Removal and Normalization

 The input/output data are biaesd by the nonzero steady state and also
by disturbance eects. To remove the bias, dierence is taken for the
input/output data. Then the dierenced data is conditioned by scaling
before using in identication.
9
y(k) = (yproc(k) , yproc(k , 1))=cy >= ! identication
u(k) = (uproc(k) , uproc(k , 1))=cu >;

118
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

(c) Preltering

 If the data contain too much frequency components over an undesired

range and/or if we want to obtain a model that ts well the data over
a certain frequency range, data preltering (via digital lters) can be
done.

119
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

3.2 BASIC CONCEPTS OF IDENTIFICATION

 u(k) is processed to y(k) by the process, i.e., y(k) contains the process
information. By treating fyk g together with fu(k)g, we can extract the
process characteristics.
 A multivariable process has directional as well as frequency-dependent
characteristics.

120
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

The directional gain changes with frequencies.

 To extract all the information, the input should be properly designed
to excite all these charateristics uniformly.
 Furthermore, the process variables (inputs as well as outputs) are
subject to various random disturbances. To remove the disturbance
eects, some kind of averaging is needed. For perfect averaging, innite
number of data should be collected. The excitation input, however, has
limited magnitude and duration.
 In reality, it is neither necessary nor feasible to identify all the facets of
a process. Depending on the purpose of the model, some
characteristics should be accurately identied while the others are not.
 To nd an appropriate model to given purposes, the following three
elements need to be judiciously selected and/or designed.
{ Model description
{ Experimental condition
121
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

{ Identication method

Ex. Accurate t of the step response does not necessarily imply a
good model for control.

(a) open-loop step response

(b) closed-loop step response

122
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

(a) open-loop step response

(b) closed-loop step response

123
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

3.3 MODEL DESCRIPTION

3.3.1 NONPARAMETRIC MODEL
 Models that are not described by a nite number of parameters.
{ Pulse response model : fh0; h1; h2;


g
{ Step response model : fs0; s1; s2;


g
{ Frequency response model : G(j!)
 Pulse or step response models can be directly identied from pulse or
step test.
 The pulse and step tests are very simple to conduct. However, the step
test too much emphasizes low-frequency excitation while the pulse test
gives too-widely-spread excitation over the whole frequency ranges,
hence may not provide an appropriate model adequate to control
purpose.
 In general, a parametric model is identied rst and then converted to
a nonparametric model.

124
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

3.3.2 PARAMETRIC METHOD

ARMAX Model

y(k) = a1y(k , 1) +


+ ana y(k , na) + b1u(k , 1) +


+ bnb u(k , nb)

+ n(k) + c1n(k , 1) +


+ cnc n(k , nc)
or
A ,1 ,1 ,1
| (q {z)y(k)} = B
| (q {z)u(k)} + C
| (q {z)n(k)}
AR X MA

where fn(k)g is a zero-mean i.i.d.(independent and identically distributed)

sequence or, equivalently a white noise sequence.

 (C (q,1)=A(q,1))n(k) represents the disturbance model. Depending on

the nature of the disturbance, a simpler form can be used.
When the disturbance mostly occurs at the input in the form of white
noise,

A(q,1)y(k) = B (q,1)u(k) + n(k)

might be enough.
When the disturbance mostly occurs at the output in the form of white
noise,

125
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

A(q,1)y(k) = B(q,1)u(k) + C(q,1)n(k) !

2 32 3 2 32 3 2 32 3
66 A1


0 77 66 y1(k) 77 66 B11


B1m 77 66 u1(k) 77 66 C1


0 77 66 n1(k) 77
66 .. . . . ... 77 66 ... 77 = 66 .. ... .. 77 66 ... 77 + 66 .. . . . .. 77 66 ... 77
64 75 64 75 64 75 64 75 64 75 64 75
0


An yn(k) Bn1


Bnm um(k) 0


Cn nn(k)

A1(q,1)y1(k) = B11(q,1)u1(k) +


+ B1m(q,1)um(k) +Cn(q,1)n1(k)

.. .. .. ..
An(q,1)yn(k) = Bn1(q,1)u1(k) +


+ Bnm(q,1)um(k) +Cn(q,1)nn(k)

FIR (Finite Impulse Response) Model

y(k) = H1u(k , 1) +


+ Hn u(k , nb) + w(k)

b

where Hi is a impulse response coecient (matrix) and w(k) is a zero-mean

random noise (not necessarily i.i.d.).

 Cannot be used for description of unstable systems.

 Requires (much) more parameters than the corresponding ARMAX or
state space model does.
For description of a SISO stable system, usually 40 or more pulse
response coecients are needed if the sampling interval is
appropriately chosen (not too short and too long).
 Irrespective of the nature of w(k) as far as it is of zero-mean, unbiased
parameter estimates can be obtained using a simple least squares
method.
127
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

State space model

x(k + 1) = Ax(k) + Bu(k) + w(k)
y(k) = Cx(k) + v(k)
where fw(k)g and fv(k)g are white noise sequences.

 Adequate to MIMO system description.

Many useful canonical forms are well developed
 Powerful identication methods called the subspace method which
directly nds a state space model in a balanced form has been recently
developed.
A version of the subspace method was commercialized by SETPOINT.

3.4 EXPERIMENTAL CONDITIONS

3.4.1 SAMPLING INTERVAL
 Too long sampling interval ! too much loss of information
Too short sampling interval ! too much computation
 There are many dierent guidelines. h  =10 is thought to be
adequate for most applications.

128
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

3.4.2 OPEN-LOOP VS. CLOSED-LOOP EXPERIMENTS

Open-loop experiment

u = process input
y = process output
Closed-loop experiment

u = process input, controller output

y = process output, controller input
129
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

) Identied Model  Process or 1/Controller

 For nonparametric models (typically transfer functions),
G^ model(s)  Gprocess(s) when d = 0
G^ model(s)  Gcontrol (s) when v = 0
 For parametric models (FIR, ARMAX, State Space ..),
G^ model (s)  Gprocess(s)
if identiability is satised.
Identiability is in most case satised if
1. a FIR model is used and/or
2. a high-order controller is used and/or
3. independent excitation is given on v.

3.4.3 INPUT DESIGN

 Remember that the excitation input has limited energy with nite
magnitudes over a nite duration. Hence, it is inevitable that the
identied model has biased information of the process.
 Depending on the way how to distribute the input energy over dierent
frequencies and also over dierent input principal directions (for
MIMO cases), the identied model may have dierent characteristics.
 The input should preferrably be designed to suciently excite the
system modes which are associated with the desired closed-loop
performance.
For a SISO process, process information near the crossover frequency is
most important. The Ziegler-Nichols tuning method (i.e., continuous
130
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

cycling method) is justied in this sense.

 In general, the PRBS(Pseudo-Random Binary Sequence) is used as an
excitation signal. By adjusting the minimum step length, we can
change the frequency contents in the PRBS.

131
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

3.5 IIDENTIFICATION METHODS

3.5.1 PREDICTION ERROR METHOD
 Estimate model parameters which minimizes the optimally determined
one-step-ahead output prediction error.
 The output predictor is constructed using the model.
 The existing identication methods such as LSM, OEM, GLSM,
ELSM, PLRM, MLM etc. are special cases of PEM which are derived
for dierent model types.
Hence, the PEM can be considered a kind of generalized framwork for
system identication.

Example 1: Identication of an ARMAX process using an ARX

Model
True process :
y(k) + ay(k , 1) = bu(k , 1) + n(k) + c1n(k , 1)
where n(k) is white noise( (0; n2 )).
Model :
y(k) + ay(k , 1) = bu(k , 1) + e(k) ARX model
where e(k) is assumed to be a zero-mean white noise.

Procedure
1. Given fy(k , 1); y(k , 2);


g and fu(k , 1); u(k , 2);


g,
the best one-step ahead output prediction is

132
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

2 3
y(^k) = ,ay(k,1)+bu(k,1) = [ ,y(k,1) u(k,1) ] 4 a 5 = (k)T 
b

2. The prediction errror at k is

"(k; ) = y(k) , y^(k) = y(k) , (k)T 
3.  which minimizes the sum of squared prediction error can be
found as
N
X
min

"(k; )T "(k; )
k=1
4. The above precedure can be rewritten in the vector form as
2 3 2 3 2 3
66 "(1; ) 77 66 y(1) 77 66 (1)T 77
64 ... 75 = 64 ... 75 , 64 ... 75  ! EN () = YN , N 
"(N; ) y(N ) (N )T
min EN ()T EN ()
^LS = (TN N ),1TN YN

Discussions
 The PEM tries to seek a parameter which minimizes the
prediction error. In case that the model has a dierent
structure from the process, the parameter is determined such
that the PE is minimized under its structural constraints.
This usually leads to unbiased estimate as shown below.
The process output can be written as

133
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

2 3 2 3 2 3
66 y(1) 77 66 (1)T 77 66 n(1) + cn(0) 77
64 .. 75 = 64 .. 75 +64 .. 75 ! YN = N +VN
y(N ) (N )T n(N ) + cn(N , 1)
Hence,
^ = (TN N ),1TN (N  + VN ) =  + (TN N ),1TN VN
Taking expectation gives
  h i
E ^LS =  + |E (TN N{z),1TN VN }
= 0 ?

Now,
2 3
, y (0) ,
2
y (1)


, y ( N , 1)
n(1) + cn(0) 7 3
6
TN VN = u(0) u(1)


u(N , 1) n(Nn(2)
4 + cn(1).. 775
5 66
4
2
) + cn(N , 3
1)
= 4 ,y(0) (n(1) + cn(0)) , y(1) (n(2) + cn(1)) ,


5

Since y(k) and n(k) have correlation (E fy(k)n(k)g = 2),

 
E ^ =
6  ! BIASED !!
If c = 0, unbased estimate !!

Example 2: Revisit of Example 1 with an ARMAX Model

This time, we consider an ARMAX model
y(k) + ay(k , 1) = bu(k , 1) + e(k) + ce(k , 1)
134
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

 Note that e(k) is not directly measurable. However, it aects

y(k). Hence, by treating y(k) we can obtain an estimate of
e(k).

Procedure
1. Let the estimate of e(k) be e^(k).
At k , 1, the best one-step-ahead output prediction is

y^(k) = ,ay(k , 1) + bu(k , 1) + ce^(k , 1)

2 3
66 a7
= [ , y(k , 1) u(k , 1) e^(k , 1) ] 64 b 775 = Tk 
c

2. As was in Ex. 1,
^ = (TN N ),1TN YN
3. In the above, e^(k) can be obtained by inverting the model
equation.
e^(k) = ,c^e^(k , 1) + y(k) + a^y(k , 1) , ^bu(k , 1); e^(0) = 0

Discussions
 To nd e^(k), ^ should known. On the other hand, e^(k) is
needed to nd ^. ! Nonlinear equation. Backsubstitution or
other nonlinear solver is required.
 Due to the structural consistency, unbiased estimate is
obtained.
135
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

General Properties of PEM

1. Need a priori infomation on the model structure (model type and
orders of each term)
2. Structural inconsistency may lead to biased parameter estimates.
The bias is revealed dierently for dierentl input excitation.
3. An ARMAX model with suciently large orders is OK for most
applications.
To nd a parsimonious model, however, trial and error procedure with
dierent orders is usaully necessary.
4. Generally, nonlinear equation should be solved to nd an estimate.
Optimumj solution is not always guaranteed.
5. Recursive (or On-line) PEM algorithms are avaliable, too
6. The PEM can be extended to MIMO identication, too.
However, lack of an appropriate canonical form for the MIMO
ARMAX model leads to an overparameterized model structure.
The industrial practice for MIMO identication is to seperate the
model into ny MISO (multi-input single-output) susbsystems, conduct
MISO identication independently, and combine the results.

A(q,1)y(k) = B(q,1)u(k) + C(q,1)n(k) !

2 32 3 2 32 3 2 32 3
66 A1


0 77 66 y1(k) 77 66 B11


B1m 77 66 u1(k) 77 66 C1


0 77 66 n1(k) 77
66 ... . . . ... 77 66 ... 77 = 66 ... ... ... 77 66 ... 77 + 66 ... . . . ... 77 66 ... 77
64 75 64 75 64 75 64 75 64 75 64 75
0


An yn(k) Bn1


Bnm um(k) 0


Cn nn (k )

136
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

A1(q,1)y1(k) = B11(q,1)u1(k) +


+ B1m(q,1)um(k) +Cn(q,1)n1(k)

.. .. .. ..
An(q,1)yn(k) = Bn1(q,1)u1(k) +


+ Bnm(q,1)um(k) +Cn(q,1)nn(k)
7. MISO identication is in eect same as SISO identication
8. MIMO identication via MISO identication cannot take the
directional characterisitcs of MIMO systems into account.

3.5.2 SUBSPACE IDENTIFICATION

 Subspace identication(SSID) is a very powerful method that has been
emergered from early 90s.
 SSID
1. is an o-line identication method (at least currently),
2. does not require virtually any a priori information on the model
structure,
3. can be seamlessly extended to MIMO identication,
4. provides an optimally balanced stochastic state space model,
(A; B; C ) and noise covariances, of minimum order using input
/output data.
 Moreover, SSID does not solve a nonlinear equation as in the PEM,
but only relies on numerically stable linear operations.
 The ID package, IPCOS, from SETPOINT is based on a version of
SSID.
 More comments will be given later.

137
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

3.6 IDENTIFICATION OF A PROCESS WITH

STRONG DIRECTIONALITY

Suppose that we want to control the following 2  2 system.

2 32 32 3 2 3
4 y1 54 u1a1 0 e1
54 5+4 5
y2 u20 a2 e2
where a1 = 100; a2 = 0:1 and je1 j; je2j  1.
For controller design, a1 and a2 should be known.
 To identify a1 and a2, assume that we apply excitation signals
of magnitude 1 to both inputs u1 and u2
S/N(signal to noise ratio) for y2  0:1 while S/N for y1  100
The consequence is that a^2 is not correctly identied. In the
worst case, the sign may be reversed.
) The y2 control loop performs poorly or may be unstable.
 In order to get over the trouble,
{ we either apply large exciation to u2 in open-loop

138
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

or
{ close the y2 loop and apply an excitation from outside the
loop.

Now consider a more general case.

2 3 2 3 2 3
64 y1 75 = G 64 u1 75 + 64 e1 75
y2 u2 e2

G is decomposed as
2 32 3
G = [u1 u2 ] 64
1 0 75 64 v1T 75 , ,SVD
0 2 v2T
Here, u1 ? u2, v1 ? v2, ku1k = ku2k = kv1k = kv2k = 1 .
If 1  2, the same problem as before arises.
To avoid the problem, it is necessary to apply large input along
the weak direction to the process either in an open-loop or a
closed-loop manner.

139
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

[Example :] High Purity Binary Distillation

140
 c 1997 by Jay H. Lee, Jin Hoon Choi, and Kwang Soon Lee

In high purity separation,

2 3 2 3 2 3 2 3
4
L 5 = 4 5)41
xD 5=4 +10 ,4
5

V 1
xB ,10,4
2 3 2 3 2 3 2 ,1 3
4
L 5 = 4 1 5)4
xD 5 = 4 10 5

V ,1
xB 10,1
2 3 2 32 1 32 3

xD 5  1 0 56 2 , 2
L 5
p p1
4

xB = [u1 u2] 0  4 p1 p1
4 75 4

V
2 2 2

where 1  2

141