You are on page 1of 5

PROCESS DESI GN AND CONTROL

Improved Filter Design in Internal Model Control


Ian G. Horn, J effery R. Arulandu, Christopher J . Gombas,
J eremy G. VanAntwerp, and Richard D. Braatz*
Department of Chemical Engineering, University of I llinois at UrbanasChampaign, 600 South Mathews
Avenue, Box C-3, Urbana, I llinois 61801-3792
The wi del y publ i shed i nternal model control (I MC) proporti onal -i ntegral -deri vati ve (PI D) tuni ng
rul es provi de poor l oad di sturbance suppressi on for processes i n whi ch the desi red cl osed-l oop
dynami cs i s si gni fi cantl y faster than the open-l oop dynami cs. The I MC fi l ter i s modi fi ed to
deri ve l ow-order control l ers that provi de effecti ve di sturbance suppressi on i rrespecti ve of the
l ocati on at whi ch the di sturbances enter the cl osed-l oop system.
1. Introduction
For decades engi neers have worked to devel op i m-
proved tuni ng rul es for proporti onal -i ntegral -deri vati ve
(PI D) control l ers (Chi en and Fruehauf, 1990; Cohen and
Coon, 1953; Hang et al ., 1991; Smi th and Corri pi o, 1985;
Zi egl er and Ni chol s, 1943). The wel l -known i nternal
model control (I MC) PI D tuni ng rul es have the advan-
tage that a cl ear tradeoff between cl osed-l oop perfor-
mance and robustness to model i naccuraci es i s achi eved
wi th a si ngl e tuni ng parameter (Ri vera et al ., 1986). The
I MC-PI D tuni ng rul es are taught to most undergradu-
ate chemi cal engi neers (Ogunnai ke and Ray, 1994;
Seborg et al ., 1989) and are wi del y appl i ed i n i ndustry
(Chi en and Fruehauf, 1990).
However, several academi c and i ndustri al process
control engi neers (A strom and Haggl und, 1988; A strom
et al ., 1993; Bergh and MacGregor, 1987; Chi en and
Fruehauf, 1990; Ho et al ., 1994; Ho and Zhou, 1995)
have noted that the wi del y publ i shed I MC tuni ng rul es
(Ogunnai ke and Ray, 1994; Ri vera et al ., 1986; Seborg
et al ., 1989), whi l e provi di ng adequate suppressi on of
output di sturbances, do a poor job suppressi ng l oad
di sturbances when the process dynami cs are si gni fi -
cantl y sl ower than the desi red cl osed-l oop dynami cs.
Morari and co-workers proposed to address thi s probl em
by i ncl udi ng an addi ti onal i ntegrator i n the output
di sturbance whi l e performi ng the I MC desi gn procedure
(Scal i et al ., 1992). Thi s method was found to provi de
adequate l oad di sturbance suppressi on for many pro-
cesses and has been appl i ed to model predi cti ve control
(Morari et al ., 1996). However, the resul ti ng control l ers
do not have PI D structure.
Here we devel op a tabl e of I MC tuni ng rul es whi ch
provi de adequate di sturbance suppressi on i rrespecti ve
of the l ocati on at whi ch the di sturbance enters the
cl osed-l oop system. The tuni ng rul es provi de the pa-
rameters for a PI D control l er i n seri es wi th a fi l ter and
are deri ved by i ncorporati ng phase l ead i n the I MC
fi l ter. Such control l ers are easi l y i mpl emented usi ng
modern control hardware. The basi c approach was fi rst
proposed and appl i ed by Brosi l ow and Markal e (Brosi l ow
and Markal e, 1992) to cascade control l er desi gn and
provi des performance si mi l ar to that obtai ned by the
method of Morari and co-workers (Scal i et al ., 1992).
2. Controller Design
The cl assi cal control structure used for the feedback
control of si ngl e-l oop processes i s shown i n Fi gure 1. I n
the di agram, p refers to the transfer functi on of the
process; d and l refer to the output and l oad di stur-
bances, respecti vel y; yrefers to the control l ed vari abl e;
n refers to measurement noi se; r refers to the setpoi nt;
and u refers to the mani pul ated vari abl e speci fi ed by
the control l er k. The control l ed vari abl es are rel ated
to the setpoi nt, measurement noi se, and unmeasured
di sturbances by
where
are the sensitivity and complementary sensitivity func-
ti ons, respecti vel y. I n I MC, the process model p i s
factored i nto an al l -pass porti on p
A
and a mi ni mum
phase porti on p
M
The al l -pass porti on p
A
i ncl udes al l the open ri ght-hal f-
pl ane zeros and del ays of p and has the form
where > 0 i s the ti me del ay, z
i
i s a ri ght-hal f-pl ane
zero (Re{z
i
} > 0) i n the process model , and zj
i
i s the
compl ex conjugate of z
i
.
I n I MC, the compl ementary sensi ti vi ty i s equal to
(Garci a and Morari , 1982; Ri vera et al ., 1986)
where the conventi onal I MC fi l ter f i s sel ected to have
one of the fol l owi ng forms:
* To whom correspondence shoul d be addressed. Tel e-
phone: 1-217-333-5073. Fax: 1-217-333-5052. e-mail: braatz@
ui uc.edu.
y ) Sd + pSl + T(r - n) (1)
S )
1
1 + pk
; T ) 1 - S )
pk
1 + pk
(2)
p ) p
A
p
M
(3)
p
A
) e
-s

i
-s + z
i
s + zj
i
(4)
T ) p
A
f (5)
3437 I nd. Eng. Chem. Res. 1996, 35, 3437-3441
S0888-5885(96)00287-4 CCC: $12.00 1996 Ameri can Chemi cal Soci ety
The type I fi l ter ensures perfect setpoi nt tracki ng of
steps, whi l e the type I I fi l ter ensures perfect setpoi nt
tracki ng of ramps. The fi l ter order n i s sel ected l arge
enough to make f/p
M
proper, and the adjustabl e fi l ter
parameter provi des the tradeoff between performance
and robustness. From (5) we see that i s the desi red
cl osed-l oop ti me constant for setpoi nt tracki ng.
Whi l e the conventi onal fi l ters provi de good perfor-
mance for setpoi nt tracki ng and output di sturbance
suppressi on, they provi de poor performance for l oad
di sturbance suppressi on when the process contai ns a
pol e sl ower than the desi red cl osed-l oop speed of re-
sponse. Thi s can be seen from (1), where the sl ow
process pol e appears i n pS i rrespecti ve of the sel ecti on
of .
To fi x thi s probl em, we propose to use the al ternati ve
fi l ter forms:
where i s chosen so that the sl ow pol e of p i s cancel ed
by a zero i n the sensi ti vi ty S.
The expressi on for T i n (5) can be substi tuted i nto (2)
and rearranged to arri ve at the control l er
Because most model s for chemi cal processes are l ow
order, I MC control l ers based on these model s are of l ow
order and can be wri tten i n the form of a proportional-
integral-derivative (PI D) control l er i n seri es wi th a
second-order fi l ter
where k
c
i s the gai n,
I
i s the i ntegral ti me,
D
i s the
deri vati ve ti me constant, and a, b, c, and d are fi l ter
parameters. The second-order fi l ter ensures that the
nomi nal PI D control l er i s proper (as must be requi red
by any practi cal control l er) and i s easi l y i mpl emented
usi ng modern control hardware. The form of the
control l er i n (11) i s an extensi on to the modi fi ed PI D
control l er structure proposed by Ri vera et al . (1986).
3. Example: Tuning Rules for First Order with
Time Delay
Consi der a fi rst-order process wi th ti me del ay:
where k
p
i s the steady-state gai n, i s the ti me constant,
i s the ti me del ay, and the desi red cl osed-l oop ti me
constant i s l ess than . The process model pi s factored
i nto mi ni mum-phase and al l -pass porti ons
where the ti me del ay has been model ed wi th a fi rst-
order Padeapproxi mati on. Usi ng the conventional type
I fi l ter (6) wi th n ) 1 and substi tuti ng i nto (10) gi ves
whi ch can be rearranged to be i n the form of (11), wi th
Appl yi ng thi s control l er (15) to the model (13) yi el ds
the nomi nal cl osed-l oop rel ati onshi p between the di s-
turbances and the control l ed output
The dynami cs between the l oad di sturbance and the
control l ed output contai ns the sl ow process pol e at s )
-1/. The effect of thi s sl ow process pol e can be seen
the l ong tai l i n Fi gure 2 for ) 20 and ) 40 (the
control l er k from (14) and the process transfer functi on
p i n (12) were substi tuted i nto (1) for the si mul ati ons).
Note that sel ecti ng a hi gher order (n > 1) conventi onal
fi l ter (7) cannot remove the sl ow pol e i n (16) and hence
wi l l al so resul t i n I MC control l ers wi th sl uggi sh l oad
di sturbance suppressi on.
Figure 1. Cl assi cal control structure.
Type I : f(s) )
1
(s + 1)
n
(6)
Type I I : f(s) )
ns + 1
(s + 1)
n
(7)
Type I : f(s) )
s + 1
(s + 1)
n
(8)
Type I I : f(s) )
s
2
+ ns + 1
(s + 1)
n
(9)
k )
T
p(1 - T)
)
p
A
f(s)
p
A
p
M
(1 - p
A
f(s))
)
f(s)
p
M
(1 - p
A
f(s))
(10)
k ) k
c
(
1 +
D
s +
1

I
s
)
1 + cs + ds
2
1 + as + bs
2
(11)
p )
k
p
e
-s
s + 1
(12)
p )
k
p
s + 1
(
1 -
(

2
)
s
1 +
(

2
)
s
)
; p
M
)
k
p
s + 1
;
p
A
)
1 -
(

2
)
s
1 +
(

2
)
s
(13)
k )
f
p
M
(1 - p
A
f)
)
s + 1
k
p
1
1 + s
1
1 -
1 -

2
s
(
1 +

2
s
)
(s + 1)
)
1
k
p
1 +
(
+

2
)
s +
(

2
)
s
2
( + )s +
(

2
)
s
2
(14)
k
c
)
2 +
2( + )k
p
;
I
)

2
+ ;
D
)

2 +
;
a )

2( + )
; b ) c ) d ) 0 (15)
y )
s
(
+ +

2
s
)
(
1 +

2
s
)
(s + 1)
d +
k
p(
1 -

2
s
)
(1 + s)
(
1 +

2
s
)
s
(
+ +

2
s
)
(
1 +

2
s
)
(s + 1)
l (16)
3438 I nd. Eng. Chem. Res., Vol . 35, No. 10, 1996
The alternativetype I fi l ter (8) wi th n ) 2 gi ves the
nomi nal cl osed-l oop rel ati onshi p
The extra degree of freedom i s sel ected to cancel the
open-l oop pol e at s ) -1/ that causes the sl uggi sh
response to l oad di sturbances:
Note that the assumpti on that < i mpl i es that > 0;
thus, the fi l ter wi l l not i ntroduce any undesi rabl e RHP
zeros i nto the cl osed-l oop system.
Substi tuti ng i nto (17) gi ves the nomi nal cl osed-l oop
rel ati onshi p wi th the al ternati ve fi l ter
The control l er wi th the al ternati ve fi l ter i s computed
from (10) and can be rearranged i nto the modi fi ed PI D
structure (11)
The fi rst-order Pade approxi mati on of the ti me del ay
begi ns to deteri orate for frequenci es greater than 1/,
whi ch i mpl i es that the desi red cl osed-l oop ti me con-
strant shoul d be sel ected greater than for the
control l er to be robust.
The response to a l oad di sturbance wi th the al terna-
ti ve fi l ter does not contai n the l ong tai l that resul ted
when usi ng the conventi onal fi l ter (see Fi gure 2). The
expl anati on for thi s i s reveal ed i n the output responses
(16) and (19). The l oad di sturbance transfer functi on
i n (16) contai ns the sl ow pol e of the process, whereas
cancel s thi s sl ow pol e i n (19). Wi th the al ternati ve fi l ter,
the settl i ng ti me i s approxi matel y equal to + 5 )
210 s, whi ch corresponds to the desi red cl osed-l oop ti me
constant ) 40. For the conventi onal fi l ter, the
response i s not settl ed by 210 s, even wi th reduced to
20. Wi th the conventi onal fi l ter, the cl osed-l oop speed
of response to a l oad di sturbance i s a stronger functi on
of the sl ow process pol e than of the desi red cl osed-l oop
ti me constant. The fi l ter parameter must be very nearl y
zero for the l ong tai l to di sappear, whi ch woul d l ead to
very poor robustness. Al though the l oad di sturbance
responses are qui te di ssi mi l ar for the di fferent I MC
fi l ter desi gns, the output di sturbance responses are al l
acceptabl e (see Fi gure 3).
RobustnessvsPerformanceTradeoff. The al ter-
nati ve I MC fi l ter i ncorporates phase l ead i n the control -
l er to offset the sl ow pol e (phase l ag) i n the process. The
ti me domai n responses i n Fi gures 2 and 3 show that
the control l er i s robust to the pl ant/model mi smatch
caused by the Pade approxi mati on.
The magni tude of T quanti fi es the robustness of
si ngl e-l oop systems. The Bode pl ots of T for di fferent
control l er desi gns are shown i n Fi gure 4. The magni -
tude of T for the conventi onal fi l ter desi gn wi th ) 20
i s very si mi l ar to that of the al ternati ve fi l ter desi gn
Figure2. Output response to step l oad di sturbance for fi rst-order
pl us ti me del ay process wi th kp ) 1, ) 10, and ) 100:
conventi onal fi l ter desi gn ( )20, - - -; )40, ); al ternati ve fi l ter
desi gn ( ) 40, s). A sol i d hori zontal l i ne at y ) 0 i s i ncl uded to
make the compari son cl earer.
y )
(
1 +

2
s
)
(s + 1)
2
-
(
1 -

2
s
)
(s + 1)
(
1 +

2
s
)
(s + 1)
2
d +
k
p(
1 -

2
s
)
(s + 1)
(
1 +

2
s
)
(
1 +

2
s
)
(s + 1)
2
-
(
1 -

2
s
)
(s + 1)
(
1 +

2
s
)
(s + 1)
2
l
(17)
)

2
+ 2(( - ) + (2 - ))
( + 2)
(18)
y )
(
1 +

2
s
)
(s + 1)
2
-
(
1 -

2
s
)
(s + 1)
(
1 +

2
s
)
(s + 1)
2
d +
k
p(
1 -

2
s
)
(
1 +

2
s
)

2
s +
1

2
+ +

2
-

2
2
)
(
1 +

2
s
)
(s + 1)
2
l (19)
k
c
)
2 +
2(2 + - )k
p
;
I
) +

2
;

D
)

2 +
; a )
2 + 2
2
+
2(2 + - )
;
b )

2

2(2 + - )
; c ) ; d ) 0 (20)
Figure3. Output responses to step output di sturbance for fi rst-
order pl us ti me del ay process wi th kp ) 1, ) 10, and ) 100:
conventi onal fi l ter desi gn ( )20, - - -; )40, ); al ternati ve fi l ter
desi gn ( ) 40, s).
Figure4. Bode magni tude and phase pl ots of T for process model
wi th Pade approxi mati on and kp ) 1, ) 10, and ) 100:
conventi onal fi l ter desi gn ( )20, - - -; )40, ); al ternati ve fi l ter
desi gn ( ) 40, s).
I nd. Eng. Chem. Res., Vol . 35, No. 10, 1996 3439
wi th ) 40, i ndi cati ng that the two control l ers provi de
very si mi l ar robustness. On the other hand, the con-
venti onal fi l ter resul ts i n poor l oad di sturbance sup-
pressi on for any not nearl y zero. I n other words, the
al ternati ve fi l ter provi des a tradeoff between robustness
and good performance, whereas the conventi onal fi l ter
provi des a tradeoff between robustness and poor per-
formance (i n terms of l oad di sturbance suppressi on).
4. Table of Improved Tuning Rules
The tuni ng parameters for common l ow-order process
model s have been cal cul ated usi ng the al ternati ve fi l ter
forms (see Tabl e 1). Strai ghtforward mani pul ati ons of
the al gebrai c expressi ons l i sted i n Tabl e 1 i ndi cate that
al l the control l er parameters are posi ti ve and wel l -posed
(e.g., do not i ntroduce ri ght-hal f-pl ane zeros i n the
control l er) when the desi red cl osed-l oop speed of re-
sponse i s faster than the open-l oop speed of response
(that i s, < ). For < , extensi ve si mul ati ons have
confi rmed that the control l ers provi de the desi red
performance for setpoi nt tracki ng and both l oad and
output di sturbance suppressi on. The conventi onal I MC
tuni ng rul es of Ri vera et al . (1986) that are reported i n
undergraduate textbooks (Ogunnai ke and Ray, 1994;
Seborg et al ., 1989) shoul d be used for > .
5. Extension to Higher Order Models
The versi on of I MC summari zed i n secti on 2 i s that
of (Garci a and Morari , 1982; Ri vera et al ., 1986), i n
whi ch the I MC control l er (commonl y wri tten as q) i s set
equal to the i nverse of the mi ni mum-phase porti on of
the process model (see references for detai l s). Thi s i s
equi val ent to the di rect synthesi s method descri bed i n
undergraduate textbooks (Ogunnai ke and Ray, 1994;
Seborg et al ., 1989). A more general expressi on for the
I MC control l er (e.g., see eqs 4.1-7 of (Morari and
Zafi ri ou, 1989) or eq 10.165 of (Braatz, 1995)) provi des
i mproved control l ers for some hi gh-order process mod-
el s. For the process model s i n Tabl e 1, the two I MC
methods resul t i n the same control l er, so for brevi ty onl y
the si mpl er versi on of I MC was descri bed i n secti on 2.
6. Conclusions
A tabl e of I MC tuni ng rul es was devel oped whi ch
provi des good l oad di sturbance suppressi on for processes
i n whi ch the desi red cl osed-l oop speed of response i s
faster than the open-l oop speed of response. For l ow-
order process model s, the control l ers have the form of
a PI D control l er i n seri es wi th a fi l ter. The resul ti ng
control l ers recover the most desi rabl e property of I MC,
that a cl ear tradeoff between cl osed-l oop performance
and robustness to model i naccuraci es i s achi eved wi th
a si ngl e tuni ng parameter, the desi red cl osed-l oop ti me
constant.
Literature Cited
A strom, K. J.; Haggl und, T. AutomaticTuningof PI D Controllers;
I nstrument Soci ety of Ameri ca: Research Tri angl e Park, NC,
1988.
A strom, K. J.; Haggl und, T.; Hang, C. C.; Ho, W. K. Automati c
Tuni ng and Adaptati on for PI D Control l erssA Survey. Control
Eng. Practice1993, 1, 669.
Bergh, L. G.; MacGregor, J. F. Constrai ned Mi ni mum Vari ance
Control l erssI nternal Model Structure and Robustness Proper-
ti es. I nd. Eng. Chem. Res. 1987, 26, 1558.
Braatz, R. D. I nternal Model Control . TheControl Handbook; CRC
Press: Boca Raton, FL, 1995.
Brosi l ow, C.; Markal e, N. Model Predi cti ve Cascade Control and
i ts I mpl i cati ons for Cl assi cal and I MC Cascade Control . AI ChE
Annual Meeti ng, Mi ami Beach, FL, 1992.
Chi en, I .-L.; Fruehauf, P. S. Consi der I MC Tuni ng to I mprove
Control l er Performance. Chem. Eng. Prog. 1990, 86, 33.
Cohen, G. H.; Coon, G. A. Theoreti cal Consi derati ons of Retarded
Control . Trans. ASME 1953, 75, 827.
Garci a, C. E.; Morari , M. I nternal Model Control s1. A Uni fyi ng
Revi ew and Some New Resul ts. I nd. Eng. Chem. Process Des.
Dev. 1982, 21, 308.
Hang, C. C.; A strom, K. J.; Ho, W. K. Refi nements of the Zi egl er-
Ni chol s Tuni ng Formul a. I EE Proc., Part D 1991, 138, 111.
Table 1. IMC Controllers for Processes with Open-Loop Dynamics Slower than Closed-Loop Dynamics ( < ):
a
k ) k
c
(
1 +
1

1
s
+
D
s
)
1 + cs + ds
2
1+ as + bs
2
model p
system
type fi l ter f kc I D a b c d
A
k
p
s + 1
I
1 + s
(1 + s)
2
+
k
p
(2 - )
+

2
2 -
0 0 0
2 -
2

B
k
p
(-s + z)
(
1
s + 1)(
2
s + 1)
I
1 + s
(1 + s)
2

1
+
2
k
p
(z(2 - ) + 2)

1 + 2

1
+
2
z
2
+ 2 +
z(2 - ) + 2

2
z(2 - ) + 2
0 0
2
2
z + 2
2
- z
2
- 2 +
2

2
z +
C
k
p
(-s + z)
s + 1
I
1 + s
1 + s
+
k
p
(z - z + 2)
+

+
+
z( - ) + 2
0 0 0
(z + 2) -
z + 1
D
k
p
(
1
s + 1)(
2
s + 1)
I
1 + s
(1 + s)
3

1
+
2
k
p
(3 - )
1 + 2

1
+
2
3
2
3 -

3
3 -
0
3
2
- 3
2
+
3

2
E
k
p
s(s + 1)
I I
1 + 4s + s
2
(1 + s)
4

k
p
(6
2
- )
0
4
3
6
2
-

4
6
2
-
4
6
2

2
- 4
3
+
4

2
a
Al l process parameters are assumed to be posi ti ve and real (z, 2, 1, , kp > 0). The i n the col umn of rows B and D represents the
sl owest pol e i n the process (ei ther 1 or 2). Al l control l er parameters are posi ti ve (Kc, I , D, a, b, c, d > 0) and provi de the desi red
performance for < .
3440 I nd. Eng. Chem. Res., Vol . 35, No. 10, 1996
Ho, W. K.; Zhou, J. H. Performance and Gai n and Phase Margi ns
of Wel l -known PI Tuni ng Formul as. I EEE Trans. Control Syst.
Technol. 1995, 3, 245.
Ho, W. K.; Hang, C. C.; Cao, L. S. A Compari son of Two Desi gn
Methods for PI D Control l ers. I SA Trans. 1994, 33, 147.
Morari , M.; Zafi ri ou, E. Robust Process Control; Prenti ce-Hal l :
Engl ewood Cl i ffs, NJ, 1989.
Morari , M.; Garci a, C. E.; Lee, J. H.; Prett, D. M. Model Predictive
Control; Prenti ce-Hal l : Engl ewood Cl i ffs, NJ, 1996; i n prepara-
ti on.
Ogunnai ke, B. A.; Ray, W. H. Process Dynamics, Modeling, and
Control; Oxford Uni versi ty Press: New York, 1994.
Ri vera, D. E.; Skogestad, S.; Morari , M. I nternal Model Control
4: PI D Control l er Desi gn. I nd. Eng. Chem. Process Des. Dev.
1986, 25, 252.
Scal i , C.; Semi no, D.; Morari , M. Compari son of I nternal Model
Control and Li near Quadrati c Opti mal -control for SI SO Sys-
tems. I nd. Eng. Chem. Res. 1992, 31, 1920.
Seborg, D. E.; Edgar, T. F.; Mel l i champ, D. A. Process Dynamics
and Control; John Wi l ey: New York, 1989.
Smi th, C. A.; Corri pi o, A. B. Principles and Practiceof Automatic
Process Control; Wi l ey: New York, 1985.
Zi egl er, J. G.; Ni chol s, N. B. Opti mum Setti ngs for Automati c
Control l ers. Trans. ASME 1943, 65, 433.
Received for review May 22, 1996
Revised manuscript received June 13, 1996
Accepted June 14, 1996
X
I E9602872
X
Abstract publ i shed i n AdvanceACS Abstracts, August 15,
1996.
I nd. Eng. Chem. Res., Vol . 35, No. 10, 1996 3441

You might also like