c

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

Part I
OVERVIEW OF PROCESS
CONTROL

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

Contents
I OVERVIEW OF PROCESS CONTROL

1 BAISIC LOW-LEVEL CONTROL

1.1 BASIC IDEA OF FEEDBACK/FEEDFORWARD CONTROL . . . . . . . .

1.2 MOTIVATION- WHY(NEGATIVE) FEEDBACK CONTROL ? . . . . . . .

1.3 ELEMENTS OF A FEEDBACK SYSTEM . . . . . . . . . . . . . . . . . . . 13

1.4 PID FORMULATIONS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
1.4.1 INTEGRAL(I) CONTROL . . . . . . . . . . . . . . . . . . . . . . . 20
1.4.2 PROPORTIONAL(P) CONTROL . . . . . . . . . . . . . . . . . . . 21
1.4.3 PROPORTIONAL INTEGRAL(PI) CONTROL . . . . . . . . . . . . 24
1.4.4 PROPORTIONAL INTEGRAL DERIVATIVE(PID) CONTROL . . 25
1.4.5 DIGITAL IMPLEMENTATION . . . . . . . . . . . . . . . . . . . . . 27
1.5 QUANTITIVE PID TUNING METHODS . . . . . . . . . . . . . . . . . . . 29
1.5.1 CONTINUOUS CYCLING METHOD . . . . . . . . . . . . . . . . . 29
1.5.2 REACTION-CURVE-BASED METHOD . . . . . . . . . . . . . . . . 38
1.5.3 FOPDT-BASED TUNING RULES . . . . . . . . . . . . . . . . . . . 41
1.5.4 DIRECT SYNTHESIS METHOD - IMC TUNING . . . . . . . . . . 46
1.6 PRACTICAL CONSIDERATIONS . . . . . . . . . . . . . . . . . . . . . . . 50
2

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

2 MULTI-LOOP CONTROL AND FURTHER PRACTICAL ISSUES

55

2.1 FEEDFORWARD-FEEDBACK CONTROL . . . . . . . . . . . . . . . . . . 55

2.2 CASCADE CONTROL . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
2.3 OVERRIDE CONTROL . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61

3 CONTROL OF MULTI-INPUT MULTI-OUTPUT(MIMO) PROCESSES 62

3.1 MIMO PROCESS ? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
3.2 INTERACTION AND I/O PAIRING . . . . . . . . . . . . . . . . . . . . . . 64
3.2.1 INTERACTION . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
3.2.2 I/O PAIRING . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
3.3 DECOUPLING . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
3.4 CURRENT TRENDS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
3.4.1 COMPUTER INTEGRATED PROCESS MANAGEMENT . . . . . 74
3.4.2 MODEL-BASED APPROACHES . . . . . . . . . . . . . . . . . . . . 75
3.4.3 COMPUTING ENVIRONMENT . . . . . . . . . . . . . . . . . . . . 77
3.4.4 COMPUTER CONTROL SYSTEM . . . . . . . . . . . . . . . . . . 78
3.4.5 SMART INSTRUMENT AND FIELD BUS . . . . . . . . . . . . . . 79

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

Chapter 1
BAISIC LOW-LEVEL CONTROL
1.1 BASIC IDEA OF FEEDBACK/FEEDFORWARD
CONTROL
Example Process

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

Control Objective keep the water outlet temp(T ) at a desired set point(T ).
Major Disturbance hot water demand(m )
Manipulated Input steam ow rate(m )
wo

sp

st

 APPROACH 1 : FEEDFORWARD CONTROL

 Measure m
 Calculate m needed to maintain T for given m . For example,
w

st

wo

stmst = mw Cpw (Two , Twi) ! mst = mw Cpw (Tsp , Twi)

st

 Apply m .
st

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

No guarantee that Two ! Tsp because of various uncertainties which come

from the model description, measurements, and signal implemetation, etc.

 APPROACH 2 : FEEDBACK CONTROL

 Measure T
 Compare with T
 Take approriate action to elimiate the observed error.
wo

sp

Two can be steered to Tsp. But Two may undergo a long transient period,
frequently with oscillation, due to the trial and error nature of the feedback
action. If the feedback controller is designed based on a process model, however, the transient can be adjusted as desired (with some limitations).
6

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

 APPROACH 3 : FEEDFOWARD-FEEDBACK CONTROL

 Apply m from the feedforward block to the process.
 Provide additional corrective signal through feedback control when there
st

is control error.

Better control performance can be expected than with feedback or feedforwardonly control.

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

Block Diagram Representations

Feedforward Control

Feedback Control

Feedforward-Feedback Control

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

1.2 MOTIVATION- WHY(NEGATIVE) FEEDBACK

CONTROL ?
Some Comments on Negative Feedback

 In (a), as y deviates from zero, it is pulled back toward zero. Hence,

negative feedback has self-stabilizing tendency.

 If an external r is put as in (b), y will tend to r.

 When the positive signal is fed back (positive feedback), it adds to itself
and will tend to diverge.

What Can We Gain Through Feedback Control in Process Control

Problems ?
1. To steer the process variables to desired steady states

 Even when there is no disturbance, it is hard to manulally drive PVs

to desired states.
 With the aid of integral action, the feedback controller continues the
corrective action until PVs reach their respective SPs.
2. Disturbance Rejection

 Try to keep the PVs at their SPs against various disturbances.

9

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

3. Stabilization

 Some processes are intrinsically unstable.

Autothermal Reactor
(a) When Ti is perturbed to increase, TRi is increased.
(b) This accelerates reaction rate and induces more heat of reaction.
(c) This again increases TRi and boosts up (b).
Positive feedback path exists between TR and TRi.

 There are processes with integrating dynamics.

Surge Tank - Level is not self-stabilizable.

10

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

4. Linearization

 Usually, m (steam ow rate) changes nonlinearly with vp(valve post

sition).
 When TC is congured to directly manipluate vp, the process seen
by TC includes a nonlinear control valve block.
 In the cascade conguration, mst  mspst if the slave controller is
tightly tuned. Thus, the nonlinearity by the control valve block can
be removed.
5. Improving Dynamics

 The dynamics of the slave loop in the cascade conguration can be

adjusted to have a faster response than the control valve block has.
 Suppose Two is to be changed to another value. If the change in mst
is made manually, the settling time is set by the intrinsic process
11

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

time constant. On the other hand, putting a feedback loop (TC)

can speed up the response time.
 Thus, TC in the cascade conguration controls a faster process than
the one in the direct conguration.

12

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

1.3 ELEMENTS OF A FEEDBACK SYSTEM

Process seen by the Controller

13

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

Transmitter
 Conditioning the sensor signal
{ Ohm, mv, mA


to voltage
{ linearization
{ conversion of the voltage signal to 4-20 mA(0-100 %) standard signal.

 To transmit a signal over a long distance while minimizing corruption

with electro-magnetic noise, the signal needs to be converted into a current form.

 4 mA(0%) bias is for detection of a sensor failure.

I/P (Current to Pneumatic Pressure) Transducer

 Linearly convert 4-20 mA to 3-15 psig(0.2-1.0 Kgf/cm2) air pressure
 A separate device or embedded in a valve positioner.

Control Valve
 Mostly pneumatic for safety reason
14

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

 Composed of an actuator and a valve body

 Actuator
{ ATO(air to open) and ATC types. Sometimes called NC(normally
close)/NO or FC(fail close)/FO
{ Selection is based on what the fail-safe position is.
{ Depending the actuator type, sign of the process gain is reversed.

15

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

 Valve Body
{ According to intrinsic ow characteristics (Linear/Equal Percent-

age), leakage(single seated/double seated), critical ow characteristics, noise, etc, many dierent types of valves are designed.
{ Flow characteristics
v
u
u
q(gal=min) = Cv  f (vp)  t
Pv (psig)

sp:gr

8
>> vp ) linear
<
f (vp) = > pvp ) quick opening
>: vp,1

) equal percentage

Intrinsic ow characteristic

16

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

Installed ow characteristic

As q increases,
p decreases.
A properly sized EQ-valve can show approximately linearized ow
characteristics.

{ Valve Size: C = [gallon water]/[min][psig]

v

Choose it to cover the needed ow range (with some additional

room). A larger valve covers a larger range and reduces pumping
cost, but tends to be more nonlinear and sensitive to pressure drop
changes.

17

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

Controller
 Appearance

 A/M : Auto/Manual mode selection for CO adjustment

 D/R : Direct/Reverse mode selection

Direct (Reverse) - CO increases(decreases) when PV increases. Negative

(Positive) gain control.

ATO valve, select DIRECT.

ATC valve, select REVERSE.
18

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

 L/R : Local SP/Remote SP selection

 PB, TI, TD : PID parameters
 In computer control, engineering units can be used for SP and PV instead
of 0-100%.

19

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

1.4 PID FORMULATIONS

1.4.1 INTEGRAL(I) CONTROL
Zt
Zt
1
1
u(t) = ubias + T 0 e( )d = ubias + T 0 (r( ) , y( )) d
I
I

 u is the manually set CO value before switching to AUTO mode.

 t = 0 represents the instant the AUTO mode is started.
 Corrective action based on accumulated control error.
 If control error persists, u(t) continues to increase (decrease) until e(t)
bias

is settled at zero.

 The I-mode ultimately nds a new u

for exact correction. In this

sense, the I-mode is often called the RESET mode.
bias

 When switching AUTO to MANUAL, u(t) replaces u . (Bumpless

bias

transfer)

20

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

PB(%) = 100
K ;
c

u(t) = 100
PB e(t)

PB represents the percent change of control error which induces

100% movement of CO.
{ Note that 0%  u(t); y(t)  100%.
Depending on r and ubias settings, the controllable range of y(t)
which induces the maximum travel of CO may be dierent from PB.
The controllable range of y(t) is not necessarily centered with respect to r.

Example. 1 u(t) = 20 + (100=PB)(80 , y(t)); PB = 50%

At y(t) = 90%; u(t) = 0%
At y(t) = 40%; u(t) = 100%

22

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

Example. 2 u(t) = 80 + (100=PB)(80 , y(t)); PB = 50%

At y(t) = 70%; u(t) = 100%
At y(t) = 100%; u(t) = 40%

Example. 3 u(t) = 80 + (100=PB)(20 , y(t)); PB = 200%

At y(t) = 0%; u(t) = 90%
At y(t) = 100%; u(t) = 40%

23

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

 The process is in control state only when PV is within the true PB.
Outside the PB, CO is staturated (or PV itself is saturated.)

1.4.3 PROPORTIONAL INTEGRAL(PI) CONTROL

"

#
Zt
1
u(t) = ubias + Kc e(t) + T e( )d
I
0

 Immediate action by P-mode PLUS oset elimination by I-mode

 TI : Integral time (min/repeat)

Consider the situation where e(t) = 1 persists.

uP (t) = Kce(t) = Kc
t
cZt
e
(

)
d
=
K
uI (t) = K
c
TI
TI
0

24

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

uI (TI ) = Kc = uP ; uI (2TI ) = 2Kc = 2uP ;

I-mode repeats uP every TI minutes.

 Sometimes, Reset Rate RI dened by RI (repeat/min) = 1=TI is used in

industrial controllers.

1.4.4 PROPORTIONAL INTEGRAL DERIVATIVE(PID) CONTROL

2
3
Zt
1
de
(
t
)
5
u(t) = ubias + Kc 4e(t) + 0 e( )d + TD

TI

dt

 de(t)=dt detects how fast the error increases(or decreases).

 D-mode can much imporve the control performance by anticipating the

future trend of control error (in a linear fashion) and taking a compensating action in advance.
25

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

 TD : Derivative time (min)

Consider the situation where e(t) = t (linearly increases).

uP (t) = Kce(t) = Kct

uD (t) = KcTD dedt(t) = KcTD

) uD (t) = uP (TD)

or uPD (t) acts TD (min) ahead of uP (t).

In this sense, D-mode is sometimes called a preact mode and TD is called

the preact time.

 D-mode can also serve to reduce oscillation, thereby allowing use of

higher gains.

 D-mode is sensitive to high frequency noise in the control error. Because

of this trouble, not employed for ow rate, pressure, and level control
26

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

where the measurements usually have signicant process noise. Mostly
used for temperature control.

1.4.5 DIGITAL IMPLEMENTATION

Digital control is run according to the following sequence:

Position Form Digital PID Algorithm

At the instance t = kh,

e(t) = ek
Zt
k
X
e( )d  eih
i
de(t)  ek , ek,
dt
h
0

=1

Using the above approximations, we can derive a digital PID control

algorithm
1
0
k
X
T
h
ei + D [ek , ek,1]A
uk = ubias + Kc @ek +

TI i

=1

27

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

 The above is called the position form because the positional

value of uk is directly computed.
 Required to store ek and Pki ei in the memory.
 Inconvenient to handle the reset windup problem.
=1

Velocity Form Digital PID Algorithm

The position from algorithm at t = (k , 1)h is

1
0
kX
,1
h
T
D
ei + [ek,1 , ek,2]A
uk,1 = ubias + Kc @ek,1 +

TI i

=1

Subtracting this from uk yields

uk = uk,

!
h
T
D
+ Kc [ek , ek, ] + T ek + h [ek , 2ek, + ek, ]
I
1

 It is called the velocity form algorithm since only
uk = uk ,

uk, is computed at each sampling instant.

 Can eectively and conveniently handle the reset windup problem.
1

 As h ! 0, digital PID ! continuous PID.

In single loop controllers or local control units of DCS, 100-500 msec of
h is usually used (  continuous controller)
In computer implementations, h  (time constant)/(1020) is sucient
for most applications.
28

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

1.5 QUANTITIVE PID TUNING METHODS

 Tuning PID parameters is not a trivial task in general.
 Various tuning methods have been proposed for dierent model descriptions and performance criteria.

1.5.1 CONTINUOUS CYCLING METHOD

Frequently called Ziegler-Nichols method since it was rst proposed by Ziegler
and Nichols (1942). Also referred to as loop tuning or the ultimate sensitivity
method.

Procedure:
step1 Under P-control, set Kc at a low value. Be sure to choose the right
(direct/reverse) mode.

step 2 Increase Kc slowly and monitor y(t) whether it shows oscillating re-

sponse.
If y(t) does not respond to Kc change, apply a short period of small
pulse input on r(t).

step 3 Increase Kc until y(t) shows continuous cycling. Let Ku be Kc at this

condition. Also let Tu be the period of oscillation under this condition.
29

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

Kc < Ku

Kc = Ku

Kc > Ku

step 4 Calculate and implement PID parameters using the the Ziegler-Nichols
tuning tables:

30

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

Ziegler-Nichols Controller Settings

Kc
0:5Ku
0:45Ku
0:6Ku

Controller
P
PI
PID

TI

,,

Tu=1:2
Tu=2

TD

,,
,,

Tu=8

Remarks :
 We call
{ Ku ultimate gain
{ Tu ultimate period (!co = 2=Tu critical frequency)
 Ziegler-Nichols tuning is based on the process characteristic at a single
point where the closed-loop under P-control shows continuous cycling.

At !co,

jGcKcj! = 1
co

y(t) is 180o (phase lag) behind u(t):

31

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

 Ku is the largest Kc for closed-loop stability under P-control.

When Kc = 0:5Ku under P-control, the closed-loop approximately shows

1/4 (Quarter) decay ratio response. This roughly gives 50% overshoot.

32

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

 An alternative way to emperically nd Ku and Tu is to use relay feedack

control (sometimes called bang-bang control).

Under relay feedback, the following repsonse is obtained:

, umin) ; T : from the repsonse

Ku = 4(umaxA
u

33

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

Typical Responses of Z-N Tuning

Process: G(s) = (s +1 1)

34

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

 Since 50% overshoot is considered too oscillatory in chemical process

control, the following modied Ziegler-Nichols settings have been proposed for PID contollers:

Modied Ziegler-Nichols Settings

Kc
0:6Ku
0:33Ku
0:2Tu

Original(1/4 decay ratio)

Some Overshoot
No Overshoot

Process: G(s) = (s +1 1)

35

TI
Tu=2
Tu=2
Tu

TD
Tu=8
Tu=3
Tu=3

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

Eects of Tuning Parameters: G(s) = 1=(s + 1)

36

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

37

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

1.5.2 REACTION-CURVE-BASED METHOD

Not all but many SISO(single-input single-output) chemical processes show step responses which can be well approximated by that
of the First-Order Plus Dead Time(FOPDT) process.

The FOPDT process is represented by three parameters

 Kp steady state gain dened by
yss=
uss

 d dead time (min), no response during this period
38

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

  time constant, represents the speed of the process dynamics.

,ds
K
pe
G(s) = s + 1

Procedure
step 1 Wait until the process is settled at the desired set point.
step 2 Switch the A/M toggle to the manual position and increase
the CO (u(t)) by
uss stepwise.

step 3 Record the output reponse and nd an approximate FOPDT

model using one of the following methods:

39

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

 Drawing a tangent is apt to include signicant error, espe-

cially when the measurement is noisy. To avoid this trouble,

the following method is recommended:

1. Obtain Kp =
yss=
uss.
40

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

2. Estimate the area A .

0

3. Let  + d = A =Kp and estimate the area A .

0

4. Then  = 2:782A =Kp and d = A =Kp , 

1

step 4 Once a FOPDT model is obtained, PID setting can be done

based on a tuning rule in the next subsection.

Popular tuning rules are Quater-decay ratio setting and Integral

error criterion-based setting.

1.5.3 FOPDT-BASED TUNING RULES

 The following PID tuning rules are applicable for FOPDT pro-

cesses with 0:1 < d= < 1.

1/4 Decay Ratio Settings

 Z-N tuning for the FOPDT model.
Controller
P
PI
PID

Kc
(=Kpd)
0:9(=Kpd)
1:2(=Kpd)

41

TI

,,

3:33d
2:0d

TD

,,
,,

0:5d

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

Integral Error Criteria-based Settings

 PID parameters which minimizes one of the following errorintegration criteria:

IAE =
ISE =
ITAE =

Z1
0

Z1
0

Z1
0

je(t)jdt
e (t)dt
2

tje(t)jdt

IAE Tuning Relations

Type of Input
Load

Controller
PI

Load

PID

Set Point

PI

Set Point

PID

42

Mode
P
I
P
I
D
P
I
P
I
D

A
0.984
0.608
1.435
0.878
0.482
0.758
1:02b
1.086
0:740b
0.348

B
-0.986
-0.707
-0.921
-0.749
1.137
-0.861
-0.323
-0.869
,0:130b
0.914

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

ISE Tuning Relations

Type of Input
Load

Controller
PI

Load

PID

Set Point

PI

Set Point

PID

Mode
P
I
P
I
D
P
I
P
I
D

A
1.305
0.492
1.495
1.101
0.56
-

B
-0.959
-0.739
-0.945
-0.771
1.006
-

ITAE Tuning Relations

Mode
A
B
P
0.859
-0.977
I
0.674
-0.680
Load
PID
P
1.357
-0.947
I
0.842
-0.738
D
0.381
0.995
Set Point
PI
P
0.586
-0.916
I
1:03b
-0.165
Set Point
PID
P
0.965
-0.855
I
0:796b ,0:147b
D
0.308
0.929
Design relation: Y = A(d= )B where Y = KKc for P-mode,
=I for I-mode, and D = for D-mode.
b For set-point change, the design relation for I-mode is
=I = A + B (d= ).
Type of Input
Load

Controller
PI

43

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

Performance of 1/4-Decay Ratio Tuning

,s
e
G(s) = 3s + 1

44

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

Performance of Integral Error Criteria-based Tuning

,s
e
G(s) = 3s + 1

set-point change

load change

45

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

1.5.4 DIRECT SYNTHESIS METHOD - IMC TUNING

 In the direct synthesis method, a desired closed-loop response is specied

rst for a given process model, and then the controller which satises
the specication is determined.

 IMC(internal model control) deals with virtually the same problem but
from somewhat dierent point of view.
Consider a closed-loop system.

Let the unit step response of the process look like

,ds

pe
y(s) = G(s)u(s) = K
s + 1 u(s)
46

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

We want to design a controller which gives the following response

to unit step change in set point.

,ds
y(s) = Gcl (s)r(s) =  es + 1 r(s)
c

 Since the process has time delay of d(min), the closed-loop re-

sponse should have at least d(min) of time delay.

 For zero oset error, the steady state gain of the closed-loop
system is specied as unity.

Gc which satises the above objective ?

 From the block diagram, we have

(s) r(s) = G (s)r(s)
y(s) = 1 +GcG(s)(G
cl
s)G(s)
c

Solving for Gc gives

47

()

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

+ 1)=Kp
Gc(s) = G1(s) 1 ,GclG(s)(s) = (s
cl
c s , e,ds + 1

, , , ()

When d is small,

e,ds  1 , ds
Therefore,

(
)
(
s
+
1)
=K
1
p @ =Kp A
Gc(s)  ( s + d)s =  + d 1 + s
c
c

, , PI-type

Remarks :
 Simple and easy, but requires a model (of fairly good quality).
 c is de facto a tuning parameter. As c gets small, faster speed

closed-loop response (to set point change) but more sensitive to

measurement noise as well as model error. Note that tuning is
reduced to a single parameter. Hence it is somewhere between
random tuning (two or three parameters) and using PID tuning
table (no parameter).
 Using (*), we can design Gc(s) for arbitrary G(s) and Gcl (s)
(but some restriction applies on the choice of Gcl ). In this case,
a more general type of controller other than PID is obtained.
For example, (**) is in fact a Smith predictor equation.
 The following controller settings was derived for dierent process models based on the IMC-tuining rule.

48

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

IMC-based PID Controller Settings

Type of Model

KcKp

I

Kp
s+1


c

1 +2
c

 +


c

2

2
c +

2

s

Kp
s

( 1 +1)( 2 +1)

Kp
 2 s2 +2s+1
Kp (,s+1)
 2 s2 +2 +1 ; 

>1

Kp
s

c

Kp
s(s+1)

c

D

1 2
1 +2





49

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

1.6 PRACTICAL CONSIDERATIONS

In commercial PID controllers, many modications are made from
the standard form. Some of them are

 remove derivative kick, sometimes together with proportional

kick
 suppress noise sensitivity of the D-mode
 anti-reset windup
 bumpless transfer during Auto-Manual switching
 bumpless parameter change
 nonlinear compensation
 etc.

Removing Kicks
 During regulation, SP is kept constant in most time and changed rarely.
When SP is changed, a surge signal is produced by the derivative mode
which acts as a \shock" on the process. We call this phenomenon derivative kick.
To remove this harmful eect, derivative is not taken on SP in most
commercial PID controllers.

de(t) = d (r(t) , y(t)) ! , dy(t)

dt dt
dt
PID controllers with the above modication is denoted as PI-D controllers.
50

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

 A similar phenomenon can be said for P-mode. In order to mitigate

proportional kick, some commercial controllers modies the P-mode as
follows:

uP (t) = Kc[r(t) , y(t)] where 0    1

Controllers with P and D modications are denoted as I-PD controllers.

Filtered D-mode
 For suppression of high frequency noise sensitivity, the D-mode is further
modied to to take derivative on a (low-pass) ltered output signal

F
F
, dydt(t) ! , dydt(t) where a dydt(t) + yF (t) = y(t) a  3D  10D

With such modication, the high frequency gain of the D-mode is bounded not to exceed some limit.

Bumpless A/M Transfer

2

F (t)
K
dy
cZt
4
u(t) = ubias + T e( )d +Kc e(t) , TD dt 5
I{z
}
|
0

ubias (t)

ubias(t) ) ubias at Auto to Manual switching

51

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

Bumpless Parameter Change

Even when e(t) = 0, adjusting Kc and/or TI modies the I-mode
value, which in turn changes ubias(t).
To avoid this trouble, a special algorithm needs to be introduced in
the I-mode calculation.

Anti-Windup

52

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

 Suppose that the actuator has saturated at 100% (0 %). Without

knowing the actuator saturation, the I-mode continues error integration.

! Integral WINDUP !!

 Once windup occurs, the actuator does not return to its normal range
until the windup is removed by opposite-signed control error over a certain period.
Control is lost for this period.
To prevent the I-mode from being wound up, the controller needs
to monitor vp(t). If vp(t) is observed to be stuck at a saturation
limit, the controller stops integration until opposite-signed control
error enters.
This function is called Anti-Windup.

The anti-windup is a very important concept not only in feedback

control with actuator saturation but also in various multi-loop control techniques.
53

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

A Simple but practical windup algorithm is

2

Zt
Zt
1
1
4
u(t) = ubias+ T (vp( ),u( ))d +Kc e(t) + T e( )d + TD dedt(t) 5
t
I
0

The equivalent discrete form is

!
h
T
h
D

uk = T (vpk ,uk, )+Kc [ek , ek, ] + T ek + h [ek , 2ek, + ek, ]
t
I
1

where Tt is called tracking time constant.

54

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

Chapter 2
MULTI-LOOP CONTROL AND
FURTHER PRACTICAL ISSUES
2.1 FEEDFORWARD-FEEDBACK CONTROL
Revisit of the Heating Tank Process

55

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

 In section 1.1, we considered feedforward control based on steady-state

compensation.

stmst = mw Cpw (Two , Twi) ! mst = mw Cpw (Tsp , Twi)

st

 When the time lag between Twi and Two is signicantly dierent from

that of mst and Two, it is necessary to introduce a dynamic compensation

in addition.
The lead-lag compensator is used in industrial controllers for this purpose.

 When appropriately designed, feedforward control can considerably improve the overall control performace. In this case, feedback control plays
only a trimming role.

 Feedforward control can be designed for measurable disturbances only.

56

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

More Examples
Batch Reactor

Since the jacket has a much smaller volume than the reactor, the batch reactor is considered a good target for cascade control from the viewpoint of
dynamics distribution.
59

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

Flow Rate Control with a Control Valve with Positioner

The control valve usually has a longer time constant than the pipeline itself.
Therefore, FC cannot be tightly tuned.
By this reason, the positioner-control valve is not recommended for ow rate
control where the process dynamics is fast compared with the valve dynamics.

60

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

2.3 OVERRIDE CONTROL

 It is assumed that only FC has an I-mode.

 Under normal condition, q is required to be kept constant.
 When L is lowered below a certain level, LC overrides the con-

trol loop and starts level control.

 When L is returned to the normal range, FC takes over the
control.
 To prevent the integral in FC from being wound up when LC
overrides the actuator, reset feedback is needed.
61

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

Chapter 3
CONTROL OF MULTI-INPUT
MULTI-OUTPUT(MIMO)
PROCESSES

3.1 MIMO PROCESS ?

Many chemical processes have MIMO characteristics.
Each manipulated variable in uences two or more process variables
simultaneously.
This trait gives rise to diculties in controller design, which has
not been observed in SISO control.

62

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

The followings are some of the MIMO process examples:

Crude unit
Distialltion columns

:::
63

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

3.2 INTERACTION AND I/O PAIRING

3.2.1 INTERACTION

 When G 1 is open, the process that G 2 controls is G22.

 When G 1 is closed, however, the process controlled by G 2 becomes
c

G 22 = G22 , 1G+21GG 1GG12

c

1 11

If both G12 and G21 are not zero, G 22 varies with G 1.

c

If G 1 is adjusted, G 2 should be retuned, too.

c

 Same thing can be said for G 1.

 The above probelm is caused by the interaction through G21 and G12.
c

64

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

Ex. Consider the following process:

2 3 2
64 y1 75 = 64
y2

32 3 2
32 3
2
1
5
G11 G12 75 64 u1 75 = 64 10 +1 +1 75 64 u1 75
15
2
G21 G22 u2
u2
+1 10 +1
:

Assume that P-control is used for both G 1 and G 2.

c

 When only one of y1 and y2 is under control, the controller gain can

have value (as far as it is positive) without causing stability problem.

 When both control loops are closed, stability is attained for the
controller gains in a limited region shown below. Both gains cannot
be increased simultaneously.

65

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

Ex. This time, we consider the process

2 3 2
64 y1 75 = 64
y2

32 3 2
32 3
2
1
5
G11 G12 75 64 u1 75 = 64 10 +1 +1 75 64 u1 75
G21 G22 u2
, 1+15 10 2+1 u2
:

Again, it is assumed that P-control is used for both G 1 and G 2.

c

 The stability region is changed to

66

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

3.2.2 I/O PAIRING

 Suppose that we want to control a MIMO process using single-loop controllers (in a decentralized fashion).

 Then the pairing will be the primary question. u1 , y1=u2 , y2 or

u1 , y2=u2 , y1 ?
 From the previous considerations, the judgement for proper pairing can
be made based on an interaction measure.

 How to measure interaction ? Many dierent ways, but the most widely
used one is the relative gain by Bristol(1968).

Relative Gain

67

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

Denition :

y1 =
u1
11 =


y1 =
u1
gain between u1 and y1 when all other loops are open
= gain
between u1 and y1 when all other loops are closed
 The relative gain is usually dened under steady state conditions.
  = [ ] is called the Relative Gain Array.
op

cl

ij

Interpretations:

 Obviously, 11 = 1 when G12 and/or G21 is zero. ) No

interaction, u1 , y1 pair is decoupled from other loops.
 11 = 0 when G11 = 0 ) No coupling between u1 and y1;
y1 should be paired with u2 for a 2  2 process.
 11 > 1 ) The gain is increased when other loops are

closed.
 11 < 1 ) Sign of the gain is reversed when other loops
are closed.
 11  1 or 11  1 implies that the system has serious interaction. ! SISO pairing has limitations. ! should rely on MIMO
control.

68

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

Properties:

 For an n  n process, P  = P  = 1.
For a 2  2 process,
i

ij

ij

11 = 1 , K K1 =K K ; 22 = 11; 12 = 21 = 1 , 11

12 21 11 22

 The relative gain can be directly computed from the steady state gain matrix of the process.

Let K be the steady state gain matrix of a process.

Then,
=K

O ,
K

where N denotes the element-by-element multiplication.

69

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

3.3 DECOUPLING
 When 11 is far from one for a 2  2 process, decentralized control has
limitations. ) Multivariable Control
 One of the classical multivariable control techniques is decoupling control.

 Decoupled Process

If we neglect the input saturation blocks,

y1 = (G11 + G12D12)m1 + (G12 + G11D21)m2

y2 = (G21 + G22D12)m1 + (G22 + G21D21)m2

70

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

Let
12
D21 = , G
G11
21
D12 = , G
G
22

Then
!
G
G
12
21
m1 = G01m1
y1 = G11 , G
22 !
21 G12 m = G0 m
y2 = G22 , GG
2
2 2
11

The relative gain for the decoupled process, 11 , is one.

d

71

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

 Multivariable Control
u1 = m1 + D21m2
u2 = m2 + D12m1

Remarks :
 A decisive drawback of decoupling control is that it is sensitive to model
error.

{ For processes with 11  1, 11  1.

{ As 11 deviates from one, 11 also deviates from one.
d

Let

D21
D12

!
G
12
= (1 + ) , G
11 !
G
= (1 + ) , G21
22

72

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

 The decoupler can be designed at the output.

 To design decoupling control, Process Model is required !
Why not try other MIMO control techniques ?

73

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

3.4 CURRENT TRENDS

3.4.1 COMPUTER INTEGRATED PROCESS MANAGEMENT

74

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

3.4.2 MODEL-BASED APPROACHES

 At every level of the previous CIM hierachy possibly except low-level
loop control and corporation management, process models play very important roles.

 Multivariable Control

{ In multivariable control, modelless approaches such as decentralized PID has limitations in achievable performance. Model-based
approach is indeed mandatory.

{ In SISO control, too, controllers designed based on a process model

always shows better performance than modelless ones. Especially,

for processes with long time-delay and/or inverse response, performance of PID control appears very poor.

{ In industries, economically optimal operating conditions are fre-

quently given at intersetions of process constraints. For this reason, consitraint handling is an important part of multivariable control. This problem is most eectively handled with model-based
approaches.

{ Model Predictive Control is a convenient and eefctive vehicle to

embrace all the above requirements in a single framework and has
now become the industrial standard for multivariable control.

75

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

 On-line Optimization

{ Dependable on-line process optimization has been a long-cherished

desire in the process industries.

{ Success of on-line optimization denitely relies on the quality of the

process model.

{ Thanks to recent advances in system identication and parameter

estimation techniques, creation of a reliable model tuned to an industrial process is now much easier than before. As a consequence,
on-line optimization is becoming a practical gear which makes existing processes more protable.

76

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

3.4.4

COMPUTER CONTROL SYSTEM

From proprietary systems to open architectured systems

Popularization of PC-PLC systems

With the eld-bus

Wiring cost can be greatly reduced.
{ diaglog between the eld instruments and the computer control system is possible.
{

80