You are on page 1of 18

Copyright © IFACI L\C In strumentation

trumentation and Automation


Pa per. Rubber,
in the Paper Rub ber. Pia
Plastics Pol\'lnerisation
tic and Polymeri ation
Industries.
Indu Ant wer p. Belgium
trie Antwerp, Belgiulll 19
19833

INTERNAL MODEL CONTROL - THEORY


AND APPLICATIONS
M. Morari
Califo rnia In
California Institute Pasadena , California,
tilute of Technology, Pasadena, Califon/ia, USA

Abstract . Process control is characterized hy severe


Abstract. s eve re modelling nroblems.
nroblems .
Therefore , robustness
Therefore, robustn ess of the control system, that is, stabilitystabilit y and acc en t-
acceryt-
pe rformanc e in the event of plant narameter
able performance naramet e r changes
chan8es and sensor and/or
and /o r
~aramount imnortance.
actuator failure is of oaramount Durin~ the nast few years a
imnortance . Durin~
multivariable control system design method, Internal Model (I~~C),
Mod e l Control (P~C),
hee n developed
has been specifically these issues. The I~1C
de ve loped which addresses sgecifically !~C
trans p are nt, can be easily ad4usted
scheme is transparent, on- line and is there~ore
adiusted on-line th e reFore
r eadily accepted by the
readily th e ooerating nersonnel.
n e rsonnel.

nrincioles behind IJ!C


The basic theoretical nrinci9les DIC are described and narallels are
drm-m to other design
dravln desi gn schemes ('10del Al go rithmic
C10del Algori 1)ynamic ~1atrix
r;ontrol, !)ynamic
thmic r:ontrol 'latrix
Control, Linea r 0uadratic Ootimal Control, S~ith
Control , Linear S~ith nredictor, etc.).
e tc.). Exten-
nonlin e ar systems are indicated. Annlications of I~C both
sions of IMC to nonlinear bo th
in simulations and on nilot
oilot ulants
olants are discussed.

~:eywords .
~:eywords. control , robustness,
Process control, robustness , time lag systems,
svstems, samn1ed
samn1.ed data
nonlin e ar control systems.
systems, nonlinear

INTRODUCTIOn
INTRODUCTIon

11any
l1any of the th e modern controller
controll e r design d es i gn techni-
t echni - exnlicitly
take the constraints into account exnlicitlv
"ues have not
0ues n ot found
fou nd their way into the nro- ke en the nlant safely within the
in order to keen
desnite ~roven
industri es desryite
cess industries n roven success in bounds .
nrescribed bounds.
aerosoace
aerospace applications. reas ons are the
The reasons
different underlying sys tems on one hand and
systems The dissatisfaction with the abilit y o~
th e abilitv OF the
th e
different rerfor~ance
the dif:erent r e~ uir eme nts on the
perfo rman ce requirements available control system design methods to
other. Snace
Sryace st ructur es are notorious for
structures deal effectively with these th ese issues and the
th ei r lar
their ge number of modes most of which are
large incre as e d power
increased powe r of readily availabl e co~nu
readil y available ter
co~nuter
on l y slightly
only s li ghtly damped.
damped . On the contrary, co ntrary , most hav e led a number
hardware have ~rouos
numb e r of research ~rouus
sys t ems found in the ~rocess
systems n ro cess industries
indu s tri es are in industry
industr y and academia to search for new
slugg ish, overdamoed
slu?gish, overdamped and their dynamic alternativ es . The prominent
alternatives. prominen t ones have become
characteristics can ca n generally
ge n e rally be he approximated
aporoximated known as Model Algorithmic Control (Richalet,(F.ichale t,
well by a first or o r second order lag combined 1978), Dynamic Hatrix
Ilatrix Control
Contr o l (Cutler,
(Cutler ,
wit h a dead time. A second
with seco nd im~ortant
imnortant dis dis-- Ramak e r, 1980), Inferential Control (Brosilow,
Ramaker,
tin gu i shi n g feature
tinguishing featu r e is that nany many chemical 1979) and Internal Model Control (Garcia,
pprocesses
r ocesses are stronglys tr on g l y nonlinear
n onlin ea r and can ca n only Horari, 1982). Though this was clearly not
110rari,
be poorly
poo rl y modeled,
mode l ed , while the mathematical
ma themati ca l des- des - recognized by b y most of the develooers the
c ri ntio n s of sattelite
crintions satte lit e motions
mo tions are usually usual I v ~rinciple
principle features which give these th e se methods
quite accurate.. ~he
qu it e accurate The requirements
reauirements on o n the th e ir power, are identical and will he
their eluci -
be eluci-
closed loop l oop transient res~onse r esno n se of a chemical dated next.
next . The key issue is the caoabilitv
processing system s ys tem are generally
ge n e rally ouite
nuite lose;los e ; technioues to
of the new techniaues t o combine the advan-
advan-
the s t eady state ~erformance
th e steady pe r fo r mance is of maior tages of open-looo (feed forward) and feed- feed -
importance
impo rtance (no offset). offset) . On the othe otherr hand,
hand , back control and to eliminate
elimina t e their
th e ir disadvan-
in s9ace
space applications
aoplications the t he prohlen
p robl em is most tages.
often
of t en 0: of the servotype and there is no n o steady
stead y
state to worry worrv about.
abou t. Another
An othe r imnortant
imoortant The advantage
advanta ge of the onen loon
l oon scheme (Fig.
requirement is that for fo r controllers td to be lA) is that the stability
s tability question
ques ti on is trivial
acce~ted by the orocess
accented o r ocess industries they th ey have (the system is stable when both the th e control-
con tr ol-
t o be easily
to easi l y adjustable on-lin on-linee by oneratin~
oneratin~ ler and the system are stable)
s tabl e) and that the
personnel
pe r sonne l without university level l eve l trainin~.
trainin ~ . (gC=g- I) . The
contro ller is easy to design (gc=g-l).
controller
f ina l imnortant issue is that the o~erating
A final operatin g disadvantages are the sensitivity of the
region
r egion of a chemicalchemica l nlantn l an t is usuallY hi ~h l v
usuall v hiqhlv performance
pe rformance to modelling
mode lling errors and the
constrained and a controller cont r o ll e r must be able to to inabilitv
inability to cone
cooe with unmeasured disturbances.
2 M.. Horari
10rari

With the feedback arrangement (Fig


(Fig.. IB)
lB) the Property P2 (Perfect Control):
Control) : Assump.
Assume that
reversed . Modelling errors and
situation is reversed. the controller is equal to the model inverse
unmeasured disturbances can be dealt with (gc=g- l) and that the closed loop system in
(gc=g-l)
effectively but the tuning is complicated bv by Fig . lE is stable.
Fig. stable . Then yet)
_ (t) = ys( t) for all
s (t)
loo p stability problem.
the closed loop problem . t > 0 and all disturbances d(t)
d(t)..

We can now augment the open-loop


\e open- loop and closed-
closed- Pr oper t y P3 (Zero Offset):
Property Offset) : Assume
ssume that the
loop systems as indicated in Fig. Fig . 1 C & D steady state gain of the contoller
con toIler is equal
without effecting their performance
performance:: In Fig.Fig . to the inverse of the ~ain
th e model roain
lC , d=O,
lC, d=O , and therefore
th erefo re the ssystemstem is still (gc(O)=g(O)-I)
(gc(O)=g(O)-l) and that the closed 1000
open- loop , in Fig.
open-loop, Fig . ID the two blocks g (g system in FigFig.. lE is stable.
stable . Then for an
indicates a model of the plant g) cancel each asymptotically constant setpoint
other . Comparing Fig.
other. Fig . lC and D and using (lim ys(t)=ys) and asymptotically constant
th e appropria
the appropriate t e definitions we arrive at the t -+«>
disturbances there will be no offset
general structure in Fig Fig.. lE which has all
) = ys) •'
( I im yy(( tt)=ys)
(Hm
t he advantages of both the open-
the loop and the
open-loop tt-KO
-+«>
closed-loop structures:
s truc tures: 1{henl.Jhen the model of
PI simply expresses the fact that unless
plan t is perfect (g=g) and there are no
the plant
there are modelling errors and as long as the
(d=O)~ feedback is not needed
disturbances (d=O).
open loop system is stable, the stability
and structure E behaves identically as
trivial . P2 reasserts that the
issue is trivial.
structure A. When l.Jhen there are modelling errors
open- loop controller leads to nerfect
ideal open-loop oerfect
and/or disturbances feedback is needed ne eded and
closed - loop performance when the IMC struc-
closed-loop struc -
structure E behaves identically as structure
employed . P3 states that integral-
ture is employed. integral -
B. Because the plant p lant model g appears expli-
expli-
type control action can be easily achieved
citely in E, thisthi s structure is referred to
without the need of introducing additional
as the Internal Model structur~
10del Control (IMC) structur~
parameters .
tuning parameters.
As a first approximation we can say that the
controller gc in E can be designed with the
Superficially these properties seem too good
simplicity of an open-loop controller but
to be true.
true . However,
However , it should be emphasized
that the structure E has all the nice perfor- perfor-
that structures Band E are eouivalent (Fig(Fig..
mance characteristics of a feedback system. system .
1) as is apparent f~om
from the
th e following trans-
Obviously
Obviously,, the situation is not quite as
formation equations
straightfonyard, but this argument should
straightforward,
provide sufficient motivation to explore exp lore the c (3)
prope rties of the IMC
theoretical properties IHC structure gc =
l+cg
in more depth
depth.. Striving for clarity rather
than ge ne ralit y in this expository paper we
generality gc
c (4)
will start with a discussion of continuous l - gg c
l-gg
single- input-single- output (SISO) systems
single-input-single-output systems..
resu l ts will be extended to
After that the results and therefore the properties can be easily
multi- input - multi- output (MIMO) and sampled
multi-input-multi-output explained . Whatever is possible with struc
explained. -
struc-
data systems
systems.. The paper will conclude with ture B is possible with structure E and vice
a number of comparative simulation and ex- ex- versa . We know intuitively that P2 re~uires
versa. re~uires
studies .
perimental studies. an infinite controller gain and this is con-
con-
firmed by substitutin~
SUbstitutin~ gc = g-l
g- l in (4).
(4) . By
setting gc(O) = g(O)-
g(O) - as postulated for P3
SYSTEI1S
SISO SYSTE'1S we find c(O) = 00 which implies integral con-
con-
trol action as expected.
From the block diagram for the IMC structure
(Fig.e lE) follow
(Fig fo llmy the relationships The advantage of the IMC structure is two- two-
fold . In simplified terms we
fold. lye can say the
gc
u = PQ (ys-d)
(Ys-d) (1) " ga in" the better the performance.
larger the "gain"
l+gc(g - g)
l+gc(g-g)
co nventio nal structure B the objective
In the conventional
" gain " as large as possible
is to make the "gain"
without causing instabilit_.
instability . Simultaneously
Simul taneously
ggc
yy (Ys-d)+d (2) attention is to be paid to t o other cri teria
criteria
l+gc(g- g)
r obustness to modelling errors and to
like robustness
constraints like input saturation
saturation.. Accord-
ccord-
The advantages of the IiCIMC structure discussed ing to P2~2,, with IiCIIIC we
lye can start with a
qualita tively in the introduction can be
qualitativel_
stable closed loop system with perfect con- con-
stated more
mo re precisely in the
th e form of three
trol . Thus the first design problem is
trol.
properties which
Iyhich can be proved easily from
eliminated altogether and full attention can
(1 ) and (2)
(1) (2)..
be devoted to the additional criteria c riteria and
constraints
constraints.. The second advantage of
Prope rt y PI (Dual Stability): Assume the
Property
HIC
I1C is that the design philosophphilosophy lends
(g=g) . Then the closed loop
model is perfect (g=g).
itself much better to t o be extended to multi-
multi -
system in Fig.
Fig . lE is stable if the controller
variable and nonlinear systems as we will
th e plant g are stable.
gc and the stable .
see later.
later .
Internal Model Control - Theory and Applications 3

There are several reasons why the "perfe ct


"perfect of the error for a specific input (vs - d),
(vs-d),
controller" implied by P2 cannot be realized for example
in practice.

1) Right half plane (RHP) zeros


zeros:: If the Jf oo 2
e dt (8)
model has a RHP zero, the controller gc == g_
g_l1 o
has a RHP pole and if g = g the closed- loop
closed-loop
system will be unstable according to PI. Frank (1974) proposes a gene ral procedure
general
for this op timization which is valid for
optimization
2) Time dela y : If the model/plant contains
delay: arbitrary inputs
in~uts. . In summary, the integral
a time delay, the controller gc = gg_l
_1 is square error op timal factorization for steD
optimal ste~
predictive and cannot be realized by a physi- inputs is as follows.
cal system
system..
Theorem 1 (Frank
(Frank,, 1974): Let
3) Constraints on the manipulated variables
variables::
If the model is strictly proper then the
perfect controller gc == g-g-lI is improper which
(s-zj)
(S-Zl)(S-(S- Z2) (s-Ztu '~
(s-zfn
g =
= (9)
(S - PI) (S- P2)
(S-P1)(S-P2) (s-Pn)
(s-P n )
implies lim IIgcl
gc I = 00
00.. in£initesi~~lly
Thus infinitesi'.
" llly
(,;.,)-KC
w-><'"
small high frequency disturbances would give where zZl, ••• zm < o.
••• zi > 0 and zi+l' ••.
I , •.• O.
rise to infinitely large excursions of the
manipulated variables which is physically orytimal g+ minimizing (R) is
Then the oT)timal
impossible..
impossible
(-S+ZI)
(-s+z 1 ) (-s+Zj)
(-s+z j )
4) error : If g # gg,, PI does not
Modelling error:
g+ = (S+ZI)
(10)
8+ (s+z 1) (s+zi)
hold and the closed loop system will general-
6 ~. f
l y be unstable for the controlle
ly controllerr gc = g-l.
g- I . . 1 ISE . 1
IS E 1S 2
The optimal
Th e opt1ma is ~.
j=l J
In order to deal with these four issues the
ideal of perfect control has to be abandoned.
Theorem 3 (Frank, 1974): Let
steps .
This is done in two steps.

1) The model g is factored n (s)


n(s) - (-) s
-(-)s
g = des) e (11)
(ll)
(5)
polynomials .
where n(s) and des) are polynomials. Then
g- l is stable and causal and
such that g:l the o~timal
ontimal g+ minimizing (8) is
8+(0) == 1. -
g+(O)
- 0s
g+ == e -0s (12)
2) The controller is

(6) Fil ter Design


Filter

where f is an adjustable low-pass filter re~uired to make the controlled


The filter is required
which guarantees that gc is proper and the gc proper and thus realizable and to make
robust . By definition
closed loop system is robust. the closed loo~loon system robust to modelling
modelling
realizable .
of the factorization (5) gc is realizable. errors . For realizability the order differ-
errors.
ence between numerator and denominator
The design
desi gn of the IMC controller involves as polynomial of the filter should be at least
aa first step the factorization of gg in some n- m (c.f.
n-m (c . f . (9)).
(9» . In the absence of modelling
th eselec -
suitable manner and subsequently theselec- errors the closed loop response is
tuning of the filter f.
tion and tuning f . In the
design
desi gn procedure we have developed, a (13)
perfect model is assumed for the first step
and the factorization is performed to opti- minimize ,
and the filter could be selected to minimize,
measure . In the
mize some performance measure. for example
second step the filter is selected to t o make
model -
the closed loop system robust against model-
ling errors.
errors . (14)
o

Factorization
Factor ization of gg Frank (1974) provides aa table of optimal
ad;ustable parameter which
filters with one adlustable
I n the
In the absence of modelling
mo dellin g error (g=g) and has aa direct effect on the speed but not the
has the
with
with gc g=l
g=1 the
the control error ee is shape of the closed loop
shape loop response.
response . The main
main
obiective of the
obiective the filter
filter however
hOlvever is to
to
(7)
(7) reasonably good
guarantee reasonably good and
and at least stable
stable
presence of
behavior in the presence of plant/model
plant/model
g+
g+ can be
be selected
selected to
to minimize
minimize some function
function mismatch .
mismatch.
4 M. Morari

The model uncertainty is commonly assumed to


to (llorari, lS83a): T~ere
Theorem 4 (Uorari, T~ere exists a
be of the multiplicative type stabilizing filter with f(O) == 1 for the
closed loop system in Fig. lE with the con-
g(s) = g(s) (l+ £ (s»
(l+£(s» (15)
(15 ) troller (6) if and only if g(O)~(n) - l > 0 or
g(O)~(O)-l
o ther words if the steady state gain of
with other
£ (s) is constrai.ned
where £(s) constrained by a real non- the model and the svstem ha ve the same sign.
system have
negative function

!£(iw) ~£
!£ (iw) !I ~ :[ (w) (16) Predic t o r
IHC and the Smith Predictor

or A look at IMC for systems with a time delay


(Fig. 3A) shows the completee~uivalenc
completee0uivalencee with
g(iw) ! ~
Ig(iw) - 8(iw)!
!g(iw) £ (w) (17)
Pred ictor (Fig. 3B) where
the Smith Predictor
IIgg(iw) 1I
1 f
This implies that the Nyquist
Nyqu ist plot of the c = ~_ l+f
(22)
plant g(s) can lie within a band around the
Nyquist plot of the model g(s). This band t o a number of important con-
This leads to
is described by a set of circles centered at clusions:
£
! g(iw) I 9: (w) (Fig. 2).
g(iw) with radius !g(iw)
At high frequencies the model is essentially • IHC
HIC includes time delay comnensators in
always of lower order than the plant. There- a natural manner.
fore • The factorization (5) implied impli ed by the
Predic tor is only optimal
Smith Predictor op timal in the
liro
lim £9: (w) 1 (13) sense of ISE for step inputs.
w+<'"
0J"+C0 • Robustness studies via the IMC structure
l e d to
have led t o very simple filter design rules
For the controller (6) and the uncertainty
uncertaint y for the caBecase that the modelling error is
(16,17) the closed loop expression (2) only in the time delay (Brosilow, 1979;
becomes Clinch, 1982). Let Le t the possible time
delay error be ±E and let a first order
filter l/(TS+l)
l /(Ts+l) be sufficient to make the
y 8± f (1+£) (ys-d) + d (19) controller gc realizable. If T is select-
l+g+ f£
ed equal to t o (1.4
(1 .4 times) E,E , then the maxi-
mum closed loop amplitude ratio peak will
Theorem 3: The closed loop system is stable not exceed 2 (1.4). With these filter
£ (s) fatisfying
for all uncertainties £(s) satisfying (16,17) settings it can be shown that the comhina- combina-
if and only if !If(iw) ~ ~
f(iw) !I ~ ~( o ~ a ~:
t ion o~
tion ~~ controller with a Smith Pre-
£ (w)
w) ou tperforms a PID controller
dictor always outperforms
alone. Therefore all the reports about
Proof: Follows the same outline as in Garcia
the impracticality and a nd poor performance of
and Morari (1984). Thm. 3 provides a design
Smith Predictors because of their sensiti-
rule for the filter given a specific uncer-
vity are myths generated by incorrect
tainty range for the plant. It guarantees
tunin g procedures.
tuning
stability but for adequate performance it is
desirable to limit the maximum peak of the
closed loop transfer function
Summary

(20) The IMC design procedure consists of two


steps.
where a ~ 0.4 would be a reasonable require-
~
~:: Fac t or the
Factor th e model transfer function
ment for process control applications.
invertible Dart
into an invertihle part g_g _ and a noninverti-
Except for very simple analytic expressions
ble part g+. If the factorization is per-
£ , the search for the filter f to
for £, t o satis-
accordingg to Thm. I1 & 2 and
formed accordin a nd if f is
fy (20) has to be performed
pe rformed numerically.
a low pass filter chosen
c hosen to t o make gc proper
con tr olle r gc = g:lf
the controller g: l f minimizes the ISE
According to P3 the requirement for integral
for step changes in the inputs.
control is that

g- l (0) ~
~:: In the presence of model uncertain-
(21)
ty the filter time contants have tot o be
increased to satisfy the conditi on ot Thm.
condition Ihm.
and because of g+(O) 1, f(O) == 1. IItt then
3.
follows from Thm. 3 that integral control is
impossible if the steady state gain error
The simplicitv
simplicity of the design procedure should
can exceed 100% 0: (0» 1). This is expressed
100% (£(0»1).
more precisely in Thm. 4. be apparent. In the absence of modelling
Internal Model Control - Theory and Applications 5

errors there is only ~one adjustable parameter, be pointed out at this point
ooint that zeros of
response~termined
the speed of response determined by the transfer matrices can be defined in a number
fil ter, which can be selected by the designer
filter, of ways (1acFarlane,
(HacFarlane, 1974) but that in
at will. If modelling errors are present general they
the y bear no connection to the zeros
t o be addressed
the problem of robustness has to of the individual transfer matrix elements.
elements .
by adjusting the filter.
filter . For stable systems the RHP zeros can be
determined from the determinant of the trans-
trans -
The tuning is so transparent because the fer matrix.
matrix . The factorization of time
designer selects the closed loop transfer delays is complicated by the fact that in
function g+ f directly (see (13)).
(13)) . A "fast" general the time delays in the different
filter pushes the system hard and increases matrix elements are different.
the possibility of an instability if the
model is inaccurate. If not much is demanded
(conse rvative filter) it
from the system (conservative Factorization of G
will be stable even when the modelling errors
are severe.
severe . The same could be accomplished
accomolished In princi~le
princiole G+ could be determined again by
with the classic feedback structure (Fig. lB) minimizing a scalar function of the error
IHC uses the inverse of g_ exnlicitely
but THe exnlicitE'ly
approximatin g it inGirectl~T
instead of approximating in~irectl~ bv -e
-e== y - ys (25)
Als~ i~
selecting a high controller gain . Also ir. the
classic structure a series of parameters in e . g . the ISE.
e.g. ~rocedure has
ISE . Though such a nrocedure
c would have to be adjusted simultaneously been develoned
develooed (Frank, 1974) it is not recom-
recom-
to have the effect of the single IMC filter mended in oractice because it is extremely
parameter .
parameter. cumbersome and also requires a relative
weighting of the different outout errors
Several questions of theoretical
the ore tical and practi-
practi- arbitrary . The
which is usually quite arbitrary.
cal interest are currently the focus of our simole results have emerged from
following simule
research efforts: the investigations by Holt and ~10rari (1983 ,
Jorari (1983,
1984) . Note that without modelling errors,
1984). errors ,
1) The uncertainty description (17) destroys and for
phase information and can therefore lead to
very conservative control systems. G
G = G- 11 F
= G- (26)
c -

2) No simple, practically effective filter reduces to


design methods to satisfy criteria like (20)
have been proposed yet.
yet . (27)

3) In deciding on the optimal factorization Thus , the type of factorization determines


Thus, de t ermines
(5) attention should be paid to modelling response .
directly the closed loop response.
ISE . This can
errors and not only to the ISE.
have a profound effect on the performance Theorem 5 (Holt and Morari,
Morari , 1984): Let the
(Brosilow, 1983) but only demostrative case HI10 svs tem G(s)
HPm system G (s) have RHP zeros at S=Zl,
s=z 1 ,
studies and no fundamental analyses are avail- Z2,
Z2 , ••• zi '
• • • zi. Then, in general, the "bad"
able to date. effect of the RHP zeros can be localized to
outout,
any particular output,

1.
HIMO SYSTEMS
11Hm o
0
1 ••• 1 (-s+z j) ••• (- s+z . ) ]
properties , relation-
The basic structure, properties, relation - = [
x···xC-S+Z1)···C-s+Zj)
X ''' x ~ X ''' x (28)
ships and design philosophy carry over to the (s+zl)
(s+z 1) "
• • '• (s+zi) 1.
multivariable case and will not be elaborated
on in detail. Transfer functions
funct ions are replac- o '1
ed by transfer matrices
mat ri ces which will be denoted
by canital letters.
letters . Again all systems
s y stems will where all the off-diagonal
off - diagonal elements are zero
be assumed to be strictly open-loop
open- loop stable RFP
exceot in the row which contains the RE~
and to have the same number of inputs and zeros.
zeros .
outputs.
ou tputs. Then for the IMC structure multi-
mul ti-
variable equivalent of Fig. lE we find For example,
example , consider the system

y G(I+Gc(G-G))-lG
G(I+ Gc (G-G )) - IG c ((yss-d)
-d) + d (23)
c

u = (I+Gc(G-~)) - IG c(Ys - d)
= (I+Gc(G-G))-lGc(ys-d) (24) G(s)
1
8+1
s+l
[ 1
1+2s :J
Properties Pl- P3 carr~
Proper ti es Pl-P3 carry over simply by substi-
suhsti- which has a zero at s =
= 1/2.
1 /2 . Three possible
oossible
tuting matrices for scalars.
scalars . The factoriza-
factoriza - factorizations are shown below together with
tion of G into an invertible and a noninver-
noninver - the ISE resulting from a unit steo
step change
tible part and the design of the robustness in both set points
poi nts
filter need special attention.
attention . It should
6 M. Horari

[_25+1
-2~+1l
is given to the second output. Holt & Morari
1
G+(S)
G!(s)
l2"+l -2:+r}
2s+l
22 1
G - 8s
+
[ 1 _2:+1]
[
_2:+1] ;
;
(1983) have shown that a diagonal G+ which
renders G- 1l causal is "optimal" if and only
0
2s+l 2s+l 2s+l ro~s and columns of G can be rearrang-
if the ro;s
delav of each
ed such that the smallest time delay
TSE
ISE = 8 ISE
lSE 4 exampl~, the
row is on the diagonal. For examnl~,
Wood & Berry (1973) distillation column has

G33
+
[[~'+l
~5+1
2s+l
2s+1

0 1
]
252S25'"l+1
+l ]
the transfer matrix

G(s) =
.
12.8 e
[ l6.7s+l
-s -s
--18.9
--18.9 e-
2ls+1
2ls+l
-3s
e 3S ]

(29)
-3s
lSE
ISE 1
6.6 ee--7s
7s -19.4 ee- 3s
10.9s+l l4.4s+l
The optimal G+ can be found using the men-
tioned matrix factorization procedure Here the smallest time delays are on the
(Frank, 1974): diagonal and therefore G+ =~
=- diag(e- SS , e- 3s )
"oTJtimal"
is "oDtimal" upfler bounds on
The lower and upper
the settling time coincide.
1
5(1+2s)
[
5-6S

[ 8s
8s
5+6s
J ;. ISE
"
!!.
= !!..
55 Contrarv to the results obtained for systems
Contrary
involving RHP zeros, the effects of time
delays are structured, i.e. they
th e y are general~
g eneral~

For a different set of inputs or a different 1y


ly associated with a particular outnut and
weighting of the outputs the lSE-optima}
lSE-optimal cannot be shifted around.
factor G+(s) would be different. Thus
lSE optimality does not aopear
striving for ISE appear
a very practical proposition. Factori~ations
Factori~ations Filter Design
1 2 3
of the type G+ ' ,, G+ 2 & G+ 3 are much easier
to obtain and allow the designer to clearly Again the function of the filter is twofold:
indicate his preference. If a decoupled It serves to make the controller Gc (26)
response is sought G+1(s)
G+ 1 (s) is the answer. If relizab1e and to provide
proper and thus relizable
output 1 is more important G G+ 22 should be robustness against modelling errors for the
selected, if output 2 is critical G 3 is loop system. The realizability issue
closed 1000
the best candidate. + can be resolved trivially simply by providing
enough poles in the filter. The main obiec-
Similarly, in the case of time delays a tive of the filter, namely to guarantee
trade-off between the speed of the closed reasonably good, but at least stable behavior
loop response and decoupling is possible. is the presence of plant-model mismatch is
For example three possible factorizations for more difficult to accomplish.

G
=[[0
= 0
-2s
2S
-2S]
ee- ]
As shown in Fig. 4 the multivariable multi-
plicative uncertainties can act either on
the imputs (L I ) or the outputs (L
(LO)
O)
-e 1
are G(s) c:
~(s)(I+LI(s))
(s) (I+L r (s» (30A)

~2'] e~2" ]
[ -2, G(s) (I+L (s))(;(s)
(I+LO(s»C:(s) (30B)
Gl -f:4" G2
O
+ e + (1:e- 2S ) 1I~-1
II C:- 1 (G-(;)
(G-C:) 1111 < 9"I(W)
£r( W) (3IA)
(3lA)

e-2, :'-0-2,) ]
G_1
(I;-G ),1;-1 (3IB)
(3lB)
[ e-4, 1I11 (C;-G) 11 < £O(w)
9, O(W)
G3 =
+ 0 where £1'
9, r' £0 are scalar functions defined on
the positive reals. These functions do not
allow to distinguish between uncertainty
uncertaint v
G indicates that output 1 can react only
localized in one element and uncertainty
uncertaint y
after two time invervals, output 2 can ~eact
react
"soread"
"spread" over all elements. This might-
might or
immediately. These figures area~e a lower
depending on
might not be disadvantageous denending
bound on the response time but they are not
how much uncertainty information is available.
an indication of the actual settling time.
Also let us define
If both outputs are equally important and
decoupling is chosen, G+ 11 provides an upper
(32)
bound on the settling time. This is veri-
fied by G+ 22 , where preference is given to
the first output which settles in minimum
time (cf. G), at the cost of decoupling and What norms whould be used will denend
deTJend on
a maximum settling time for the second out- the application. Here we will use the
put (cf. G+ 1l ). Analogously, in G+jpreference spectral norm
Internal Model Control - Theory and Applications 7
1
II GII
IIGII max ;\A.~
Y, (G*G) (33) Theorem 7 (Mo rari,1 983a) : There exists no
(Morari,1983a):
ii i
1
stabilizing filter with F(O) = I for the MIMO
cclosed
losed loop system in Fig. lE with
wi th the con-
con -
which is compatible with the Euclidean vector det(G(O)~(O)-l) ~ O.
tr o ller (26) if det(G(O)C(O)-I)
troller
norm. We will employ the following notation
no tation
for the singular values , 1983a) : There exists a
Theorem 8 (Morari, 1983a):
stabilizing filter with F(O) == II for the
k~ MIMO closed
MIMG c l osed loop system in Fig.
Fig . lE with the
A 22
;\ (G*G)
max controller (26) if all the eigenvalues of the
G(O)~(O)-l
matrix product G(O)C(O) - l are in the RHP.
RHP .

;\~. (G*G) Comparing Thm ' s . 4 and 7 we note that in


Thrn's.
mln
MIMO system the eigenvalues of the steady
state gain matrix play a similar role as the
It can be shown that ga in of SISO system
gain system.. In SISO system the sign
of the gain is usually known from physical
(34) arguments and thus the condition postulated
in Thm. 4 can be satisfied easily.easily . In ill-
ill -
Thus the maximum singluar value is a natural conditioned MIMO systems the accuracy of G G
definition of gain for multivariable systems.
systems . Thm . 8 can often
required by Thm. of ten be excessive.
Substituting the uncertainty description Thm . 8 is only
Also, Thm. onl y sufficient while Thm.
Thm . 4
(30- 32) and the controller (26) into (23)
(30-32)
necessary . ~~en
was also necessary. T.Jhen the number of
eigenvalues of G(O)G(O)
G(O)G(O)-l- l in the LHP is odd
the following guidelines emerge for the
there clearly exists no stabilizing filter
design of the multivatiable filter.
filter .
(Thm . 7), when it is even there could exist
(Thm.
Theorem 6 (Grossmann & Morari,
Morari , 1983)
1983):: The
one .
one.
uncer-
closed loop system is stable for all uncer-
tainties (30A) or
or (30B) satisfying (3lA) or ll1C and Multivariable Time Delay
Hll10 ll·1C Delav
(3lB) and (32) if Compensation

After many attempts in the literature


IIFII .;;
,;;; y(C)l£Cw)
y (C)\(W) (35)
(35) ~levisakis & Seborg, 1973;
~levisakis 197 3; Ogunnaike &
Ray , 1979) of varying degree of success and
Ray,
where restrictiveness the IMC structure points out
a new way of multivariable time delay com- com-
pensation . For the case of equal time delays
pensation.
in all transfer matrix elements the IMC pro- pro-
posed compensator structure (Fig(Fig.. 3B) reduces
of~.
is the condition number of C. y is a measure
to that of Alevisakis & Seborg (1973) or
of singularity
singularity..
Ogunnaike & Ray (1979).
(1979) . Otherwise it
generall
generallyy does not
no t remove all time delays
Thm . 6 with Thm.
Comparing Thm. Thm . 3 we note that the
selec-
from the transfer matrix but does so selec-
ga in is not only inverse-
bound on the filter gain
tivel y (see the definition of C_)
tively ~_).. This
lly
y proportional to the uncertainty (£) ( ~ ) but
always leads to
t o significant performance im- im-
number . Because y ~ 1
also to the condition number.
wil l be demonstrated in the
provements as will
this implies that ill- conditioned systems can
ill-conditioned
example section
section.. More details are available
amplify modelling errors in a manner unknown
systems . This is probably the main
in SISO systems. from Holt and coworkers (1984).
(1984) . To date the
systems .
source of design difficulties in MIMO systems. r obus tness results (Thm.
robustness (Thm . 6) have not
no t been
translated into simple filter design rules
Thm
Thm.. 6 guarantees stable but not necessarily t o guarantee
to guaran te e satisfactory performance in the
performance . In our experience (35) is
good performance. i~ errors .
presence of time delay modelling errors.
generally too conservative because the un- un-
certainty description (30),
(30) , (31) is inherently
conservative . It is likely that some recent
conservative. Summary
results by Doyle (1983) lead ttoo a more prac-
prac-
result .
tical result. The IMC design procedure for MIMO systems
sy stems
steps :
consists of two steps:
According to P3 the requirement for Mll10
MlllO
integral control is that ~:
~: tran sfer matrix has to be
The model transfer
factored into an invertible part C_ ~_ and a
(37) noninvertible part ~+..
par t C+ fac t oriza ti on
In this factorization
the designer has some flexibility to localize
~+(O) = I,
and if we define C+(O) I , F(O) = I.
I . Thus the detrimental effect of the nonminimum
when y(~(O)£(O)
y (C(O)) ~ (O) > 1 the existence of a phase elements on one or the other
o ther output
outpu t
filter with unity steady state gain (F(O) = 1) I) t o choose a decoupled response or to
and to
more . Two tighter
is not guaranteed any more. ti gh ter allow full or partial interactions.
interactions .
theorems provide more information.
8 M. Morari

~:
~: In the presence of model uncertainty where A.(A) denotes the jth eigenvalue of A.
a low pass filter has to be introduced, for J
example of the form Because of the discrete nature a first order
k ii filter is sufficient for stability as long as
F = diag(l/(Tis+l) ) (38) (40) is satisfied. Depending
Deoending on the tvpe
type of
uncertaint y a higher order filter can be
uncertainty
By choosing the filter time constants and the required for continuous systems. Also, the
filter order sufficiently large, Thm. 6 condition (40) is necessary
necessar y and sufficient
guarantees that the system can be stabilized for the existence of a range ran ge of aui's
i's
without sacrificing integral control action (a * ~ a i < 1) for which the system is closed
(a* c losed
as long as the condition of Thm. 8 is satis- i
loop stable. Note, however, that some
fied. specific set of ai'sa i's (instead of the open
interval extending to 1) might exist which
The tuning procedure is inherently simple and stabilizes the closed loop system even when
transparent. If the filter (38) is used then (40) is not satisfied. In this case the
for each output there is a single tuning system will become unstable when the ai's a. 's
parameter Ti which affects directly the speed are increased. This conditional stabilitystabilitv
of response of the particular output. If a makes the on-line tunintuningg much more difficult
fast response is demanded, a good model is hi ghl y undesirable.
and is highly
required.
The role of the filter for the robustness of
The technique bears some resemblance to the SISO control systems is illustrated by
bv Reid
method of "decoupling" prominent in process and coworkers (1979) in a spec ific case
specific
control applications. The analysis here has study, but the general theoretical explana-
shed light onto the old question when com- tion offered by Thm. 9 is not provided.
plete decoupling might be detrimental to
performance. This is the case when time The THC
IHC structure starts to display
displa y its full
delays or RHP zeros are present in the trans- power when instead of simply translating
translatin g
insight has
fer matrix. Furthermore some ins.ight from continuous to discrete time, specific
been gained into the question of robustness
robustness. e use is made of the discrete formulation in
the computation of the control law. The
In terms of open research questions the same controller (26) is composed of the inverse of
types of problems as listed for 5150
51S0 systems the invertible part of the model and the
await solution. Some help should be avail- filter. This combination can be interpreted
able from the works of Zames (1981) and as an approximate inverse of the model con-
Doyle (1982). structed to be stable and to avoid excessive
actions of the manipulated variable or at
\Ve can find approxi-
least to be realizable. He
TIME SYSTEHS
DISCRETE TINE mate inverses in an alternate manner which
offers increased flexibility.
flexibilit y .
Most modern control systems ace
a£e micropro-
cessor or minicomputer based and unless the The process model can be employed to predict
oredict
sampling rate is very fast a discrete time the outputs resulting from a series of in-
domain analysis and synthesis is more appro- puts. Or alternatively, desired outputs can
priate. All the results derived for SISO
5150 and co uld be ca
be prescribed and the inputs could lcu-
calcu-
MlMO
MIMO continuous systems in the preceding lated such that the predicted ooutputs
utouts follow
sections can be easily rederived
reder1ved for discrete the presctibed outputs in some "optimal"
" oo timal"
time systems. In most cases equivalent manner. If one requires the predicted values
properties and theorems are found -- now to agree with the prescribed ones exactly
exactl y
formulated in terms of z-transforms instead the system inputs resultin
resultingg from the solution
of Laplace Transforms. However, it turns of this matching problem will be the same as
out that Thm. 8 can be considerahly
considerably strength- would be obtained by an inversion of the
ened for sampled data systems. process model. If one requires the predicted
values only to be close to the desired ones
Theorem 9 (Garcia & Horari, 1984): Assume
& 110rari. Asswne in the least square sense, for example, the
that the robustness filter F(z) is diagonal solution of the optimization problem will
and of the exponential type provide an approximate inverse of the process
model. The characteristics of the approxi-
F(y)
F (y ) (39) mate inverse can then be affected by the
choice of weighting matrices in the least
squares objective function. This method of
C:
and that Gc = ~:l1 F. There exists an a* a* computing the control law is referred to in
(O~a*< 1) such that the system
(O';;;a*< svstem is closed the literature as "model-predictive
"model-predic ti ve control
loop stable for all ai a i in the open interval law formulation".
Ci*~ a
cx*';;; a.1. < 1 iif only iiff G and ~ satisfy
f and onlv ~
l.
We can pose the following problem to be
(40) K subject to
solved at time k t o the model equa-
tions relating u and y
Internal Hodel
Model Control - Theory and Applications 9

min and should be reserved for the trained


engineer
engineer..
m. (k) . • • ,m
(it),, ••• ,m.. (k+P)
1 1
i = 1,
l,rr Model Algorithmic Control (Richalet and
P Hehra, 1982)
coworkers, 1978, Rouhani and Mehra,
L {lly
{lI y (k+£)
(kH ) - y(k+ k) lI \2
y(kHlIk)11 (41) includes the filter but thet he parameter choices
£=1 s T in (41) are so restrictive that tha t systems
sys t ems with
wi t h
£=1 s f£f£
£ £ ze r os outside the unit circle can only be
zeros
22 difficult y (Mehra and
treated with greatest difficulty
lI u(k+£-1) 1I
+ Ilu(k+£-1)11 }
T 1930) . The IHC
Rouhani, 1930). al go rithm developed
IMC algorithm
B£ B£
B£B£ (1982 , 1984a) offers
by Garcia and Morari (1982,
flexibilit y of Dynamic Matrix Con-
all the flexibility Con-
T
where Ilxll
IIx ll Q x Qx trol in terms of the objective (41) but
Q also includes the filter which is a trade- t rade-
P controller horizon (P>l)
(P >l) mark of Model Algorithmic Control and thus
y s (k)
ys(k) traject or y (set points)
desired trajectory features .
combines all the best features.
f£ , B£
f£,B£ weighting matrices
y(kH lk)
y(k+£lk) ou tput
model predicted system output Formulation (41) together
to ge ther with the impulse
u (k) system input response model as equality constraints can
be augmented by inequality constraints
representin g bounds on the manipulated vari-
representing vari-
The formulation (41) is analyzed in grea greatt ables, process states oorr ooutputs.
utputs . If these
depth by Garcia and Horari
Morari (1982, 1984b) and inequality constraints are linear a quad- quad-
the effect of the different tuning tunin g and weight- ratic program (QP) results which has to be
ing parameters in (40) on the properties of solved at each time step.
step . This was suggested
the approximate inverse is elucidated.
elucidated . In by Mehra and coworkers (1982) and is dis- dis -
particular , it is shown how the adjustable
particular, cussed in detail by Ricker (1984) (1984).. When the
parameters in (41) have to be selected to algorithms for solving the QP are tailor made
yield the sum- of - squared- error optimal time
sum-of-squared-error the on- line computational requirements are
on-line
delay compensator and how in the p~esencep~esence of very modest.
modest . This is probably
probabl y the first
firs t
RHP zeros the same optimal controller can practical approach to include constraints
be obtained as by the factorization method computation .
explicitely in the control law computation.
(5) . The horizon
(5). ho rizon length and weighting It is also the first time that more use is
matrices can be cchosen
hosen to suppress excessive on - line computing
made of the available on-line
input action or even to yield "state-dead-
"state-dead- power and not just algorithms developed for
beat" control action (Reid and coworkers,
coworkers , analog systems are implemented on digital
1979) . The rules and the theoretical analy-
1979). analy- hardware .
hardware. An alternative to (41) is to
sis are sufficiently refined to reduce trial penalize the absolute value of the deviations
and error search for the parameters to a instead of the square.
square . Zhao and Brosilow
minimum and in most casesca ses eliminate it all (1933) have developed a very efficient
to ge ther.
together. s o lve the Linear Program
algorithm to solve Pro gram which
formulation .
results from this formulation.

Control Law
Law Computation and Implementation
Summary
(41) describes a standard linear quadratic
open - loop op
open-loop timal control problem and could
optimal All desirable properties of continuous time
be solved in the state space employing any Il1C are
ll1C a re preserved oorr strengthened in discrete
of the standard methods.methods . It turns out to t o be IMC . The major advanta
time IMC. ge of discrete
advantage
simpler to use a truncated
trun ca ted impulse response
resp onse IHC is that constraints
time IMC con s traints on the inputs,
inputs ,
o r discrete convolution
or convolut i on model.
model . Then the ou tputs and states can be included explicitly
outputs
resultin g linear
resulting linea r least squares problem can be co ntrol law.
in the control law . Thanks to t o the impulse
ic:ally with simple matrix opera-
solved analyt ically opera - response model description, the on - line
on-line
y ieldin g a control
tions yielding contr ol law with constant computational effort even for fo r systems
s ys tems with
coefficients. This is basically what is fo ur inputs and four ooutputs
four utputs is easily with- with-
done in Dynamic Matrix Ma trix Control (DiC)
(DlIC) deve-
deve- in the capability of current process control
(C ulter & Ramaker, 1980).
loped by Shell (Culter 1980) . In systems .
computer systems.
DMC no ffilter
ilter is used and all the tuning tunin g is
done viav ia the parameters
pa r ame ters in (41). We believe
that the filter should be present in all NONLINEAR
ONLI EAR SYSTEMS
implementations because the filter parameters
have a much more direct effect on the closed c l osed All physical ssystems
ys tems are nonlinear and some-
some-
lloop
oop response
resp onse than the parameters in (41). (41) . times the linear models employed for control
The function of the ffilter ilter parameters is more system design are only very poor approxima-
approxima-
easily understood by the operating personnel tions of the real behavior.
behavior . While we can
and they
the y are therefore ideal for oon-line n- line deal with mild nonlinearities just by detun-
detun-
tunin g . Shaping the response
tuning. resp onse via the para-
para- ing linear controllers,
controllers , it is likely that
meters in (41) requires a higher skill leve levell in the presence of stron g nnonlinearities,
strong onlinearities,

IAA-B
10 H. Horari
Norari

nonl~near controllers
nonlinear offer distinct multiple steady states in open-loop. Under
advantages. The design of these assumptions and if a good model of the
open-loop nonlinear controllers is a well Proper tv 2 prescribes
plant is available, Property
established field. Csing
Using variational methods exactly the structure and parameters of the
conce i vable problem has been
virtually every conceivable controller
con tr oller which will bring about "perfect
t ackled. On the other hand a general theory
tackled. control", i.e. exact set point following
for the design of nonlinear feed-back con- despite unmeasured disturbances. Moreover
trollers does not exist at present for all Property I1 guarantees the stability of the
practical purposes. One of the very few closed loop nonlinear control system. The
exceptions is the work by Frank (1974) who trick is, that as far as the design is
was probably the first one to use extensive- concerned, IHC transforms the problem into
ly the advantageous features of I~C for the a feedforward control problem, which can be
design of linear and nonlinear control solved easily even for nonlinear systems.
systems. But on the other hand IMC preserves all the
important characteristics of feedback con-
Most of the available nonlinear feedback trol, in particular the suppression of
control literature concentrates on stability unmeasured disturbances. This is evident
analysis. The recent results of Safonov from Properties 2 and 3.
(1980)
(1930) should be mentioned here, who extended
the pioneering work of Zames (1966). He In terms of stability
s ta bility and robustness, theorems
again put the work of Popov (1962) in a Ihm. 6 can be proven for nonlinear
similar to Thm.
general context. The practical applications IMC now involving appropriately defined
TMC
of all these results are mostly limited to gains of nonlinear operators (Morari, 1983b).
feedback systems consisting of a linear Except in the case of linear dynamic opera-
dynamic part and a nonlinear memory less ele-
memoryless tors with nonlinear static elements there
ment whose characteristics can be bounded ga ins
is no simple way of evaluating these gains
by a conic sector. numerically. Therefore these theorems are
only useful in a qualitative sense for the
Morari (1983b) made a first step toward a
Horari design of the robustness filter.
practical approach for the design of non-
linear feedback controllers. He adopted the A further question involves the existence,
A
general operator formalism introduced by con-
uniqueness, stability and numerical con-
Zames (1966) to
t o state the nonlinear IMC struction of the inverse of the nonlinear
structure and to derive its properties. operator used in properties P2 and P3.
Conditions for the existence and uniqueness
The nonlinear li1C
DIC has the same block diagram of the left inverse are given by Hirschorn
as shown in Fig. lE but now g,g and gc are (1979) and Rebhuhn (1980). AA possible
nonlinear operators representing the plant, iterative numerical procedure for the con-
model and controller respectively. The struction of the inverse has been suggested
following relationships can be deduced by Morari (1983b).
directly from the diagram

-e y + y ( 42)
(42) Summary

(43) IHC is very much


The research on nonlinear IMC
y
at its infancy. Several simulation examples
(Morari, 1983b)
1933b) have demonstrated the simpli-
The three characteristic properties Pl-P3
Pl - P3
of the IHC city of the design procedure and show that
HIC structure can be derived directly
TMC-works
HIC works well even in cases where no linear
from (42,43)
(42 ,43) adapting the definitions to to
controller can vield stable behavio~
behavior. These
nonlinear systems.
promising results are an incentive to resolve
some of the theoretical questions which are
Some discussion of the significance of these
still unanswered and to test the scheme
results for nonlinear systems is in order:
The design of the "classic" feedback control- experimentally.
ler involves the choice
c hoice of a controller type t ype
(i.e. the functional relationship) and of the
APPLICATIONS
APPLICATIO:;S
controller parameters, such that good perfor-
mance and at the same time sufficient robust-
The reported applications can be grouped
ness against modelling errors result. For
into three categories depending on the
nonlinear systems general
g eneral guidelines are not
~Iodel Algorithmic Control,
algorithm used: Model
available on how to design a feedback con-
Dynamic
Dvnamic :la tr ix Contro
clatrix Control1 or :mc.
IHC. The maj ority
majority
troller for which the closed loop system
svstem is
u~ed Model
~odel Algorithmic Control. Mehra ~ehra and
stable and even less for which it has some
coworkers (1982)
(1932) provide an extensive list
desired performance characteristics. I~C
The IMC
and references. The systems include air-
formalism is aimed ata t alleviating this
crafts,
crafts , a superheater, a steam generator,
problem at least for systems which are open
a wind tunnel, jet engines, a utility boiler
loop stable (or easily stabilized by output
connected to a distillation column, a glass
feedback) and systems which
,,,hich do not exhibit
furnace and a deminer boat. Shell claims
Internal Model Control - Theory and Applications 11

more than a dozen successful applications of t o a step disturbance in the feed


Responses to
Dynamic Matrix Contro l, the results of an im-
Control, im- flowrate are shown in Figure 6.
6 . Again
Aga in the
plementation on a fluid catalytic
cataly ti c cracking performance of IMC is superior, the input
unit were described by Prett and Gilette magnitudes are comparable to those needed
(1980) .
(1980). ORC .
with the ORC.

The large number of industrial applications


is very encouraging and a clear indication Fixed-bed Reactor
that the methods are appealing not only to
the theoretician but also to the user. user . seventh- order linear model has been publish-
A seventh-order
On the other hand the
theyy cannot be taken as a ed by Foss and coworkers (1980) describing
proof for the performance superiority of the dynamics of an oxygen- hydrogen experimen-
oxygen-hydrogen experimen-
these te chniques, because none of the experi-
techniques, experi- s y stem. Reactor outlet tempera-
tal reactor system. tempera-
ments were carried out in a controlled ture and oxygen concentration are controlled
environment . Also in the reports the com-
environment. com- by introducing a hydrogen quench stream
parison with more convent i ona l methods is
conventional between the reactor beds and manipulating
usually incomplete at best or lacking com- com- its flow and temperature.
Finally , at the time these appli-
pletely. Finally,
cations were carried out, a theoretical Comparisons of the performance of IMC with
understanding of the new techniques which desi gned using
that of controllers designed usin g the
is essential for the design of representative Characteristic Loci method of MacFarlane and
experiments was lacking. Kouvaritakis(1977) are reported by Garcia
(1984a) . Here we will show some
and Morari (1984a).
Recently , the author and coworkers have
Recently, results demonstrating the robustness features.features .
experi-
completed the first set of controlled experi- When a modelling error of two sampling inter- inter-
mental runs comparing IMC with PID and vals deadtime
dead time is introduced deliberately in
steady state decoupling on a mixing tank the transfer matrix elements the control
with temperature and level control. Current-
Current- results are shown in Fig Fig.. 7.
7.
ly IMC is tested on a system of two thermally When no filtering action is used the control-
coup led distillation columns.
coupled columns . ler tries to t o reach the setpoint at the
sampling intervals but the modelling errors
We will start by describing the simulation eventually propagate and destabilize the
epxeriences on a binary distillation column system.
system~ The superb handling of this struc-struc -
reactor .
and a packed bed reactor. tural model/plant mismatch by IMC with an
exponential filter with a == 00.5 . 5 is shown,
where we note the th e smooth
smoo th approach to the
Me thanol /Wa ter Distillation Column
Methanol/Water setpoint and the almost complete absence of
interactions . From Thm.
interactions. Thm . 7 and 8 we know that
Wood and Berry (1973) have reported transfer for errors in the steady state gain ga in a stabi-
function models of an experimental methanol/ lizing a might not exist. The reactor gain
column . This model has been used
water column. matrix
rna trix is
extensively in the literature for the com-
com-
schemes .
parison of multivariable control schemes.
For example, Wood and Berry (1973) studied G(l)
__ [- 2 . 27
[-2.27 O.73J
0.73]
con trollers
the performance of single loop PI controllers 1. 85
1.85 - 0 .65
-0.65
~l979)
with decoupling and Ogunnaike and Ray (1979)
used the model to test their multivariable Assume that a reactor model with slight
compensator . (29) shows the model
deadtime compensator. errors in the steady
stead y state gains has been
relating overhead and bottom compositions to IMC :
used to tune IMC:
respectively .
the reflux and steam flow rates respectively.

The performance of IMC with a == 00.8 . 8 is com- ~(l)


C(l) == [
- 2 .25 3
-2.253 O
0 .77

77]J
pared with the results of Ogunnaike and Ray [ 1.85 - 0 . 63
-0.63
(1979) in Figure 5. 5 . Using SISO PI loops
significant oscillations and interactions Computing the eigenvalues of G(l)C(l)
G(l)C(l)-l - l we
observed . When the Ogunnaike-Ray-Com-
are observed. Ogunnaike- Ray- Com- the y are 1.0025
find that they 1 . 0025 and -24.398.
- 24 . 398 .
pensator (ORC) is used with the same PI These slight modelling errors were enough
controllers
controllers,, a smoother approach to the for the condition of ThmThm.. 7 not to be satis-
satis-
obtained . However, interactions
setpoints is obtained. fied ! Simulation results confirmed that no
fied:
remain severe and the settling time is quite stabilizing a (0 ~ a < 1) exists.
exists .
long. Noote
te that with a = 0.8
0 . 8 in the filter,
filter ,
HfC input amplitudes are comparable to
the n·1C This simple example shows that it is quite
those exhibited by the PI controllers with easy to demonstrate goo d control system
good s ys tem
and without the ORC but the IHC Il1C response in throu gh simulations.
performance through simulations . However
particular with regard to interactions is for inherently sensitive systems like this
superior .
clearly superior. reactor good control will be difficult to
12 M. Morari

achieve in practice regardless of what con- model uncertainties (and only under these
tr oller design method is used.
troller circumstances!) closed loop performance might
circumstances:)
have to be sacrificed by subjecting it to a
decoupling constraint. Though in particular
tlixing Tank
Hixing the role of zeros has been mentioned previous-
l y (Rosenbrock, 1974; Bristol, 1980) a com-
ly
mixin g tank experi-
The following educational mlxlng prehensive view has not been offered before
f irst at the Unviersity
ment was designed first Unviersit y of
of and is also not reflected in any of the
Wisconsin (Morari anda nd Ray, 1982
1982)) and has available interaction measures.
since been duplicated in about half a dozen
schools in the U.S. The flow rates of hot 2) Hultivariable Deadtime Compensator: Using
and co ld water entering a mixing tank are
cold standard tuning techniques the time delays
manipulated in order to control
co ntrol the tempera- omnipresent in chemical engineering systems
ture and level
leve l in the tank, which has a fixed can lead to stability problems and sluggish
orifice in the bottom. A long hose has been response. Within the IMC framework we dis-
inserted between the valves and the tank covered in a natural manner a new multivari-
introduc ing a deadtime into the transfer
introducing able deadtime compensator (Holt and coworker~
coworker~
relatin g the tank temperature to
functions relating 1984). For some specific cases our trivial
the manipulated variables. The deadtime
dead time is derivation led to the same result as that
about 35
35%% of the tank time constant • obtained previously using optimal control
theory (Soliman and Ray, 1972), where the
In our experiments the performance of IMC enormous complexity has prohibited both under-
was compared with that of a steady state standing and application. The performance
decoupler with two PI controllers. The inte- throu gh IMC over earlier desi
improvements through gns
designs
gral time was set equal to the system time involvin g different extensions of the SISO
involving 5150
constant, the gain was set "optimally"
" op timally" by Smith Predictor to MIMO systems are unques-
trial and error. In this case the steady tionable.
le ads to an almost perfectly
state decoupler leads
dynamicall y decoupled system (Ray, 1981).
dynamically 3) Robustness: With the exception of the
academic environment where several graduate
Fig. 8 shows the response to a cold water students are sometimes kept busy
bus y developing
disturbance steam turned on after 900 sec a detailed model for an essentially trivial
and turned off after 1500 sec. For the level piece of lab equipment, good dynamic models
IHC is clearly superior, for the
control liIC rarel y exist in practice. The route via
rarely
temperature it is about equivalent to PI. fundamental models, response matching by
The response to a setpoint change in tempera- parameter adjustment, order reduction, etc.
ture (Fig. 9) again shows the superiority of often does not result in a more accurate
IMC. The level control
con trol is essentially per- system description than that ggiven
iven by an
fect with both IMC and PI. Most of what is operator sketch of a typical step response.
seen in Fig. 9 is measurement noise. IMe viewpoint allowed to develop new
The !MC
theoretica
theoreticall insight for the identification
of systems whose performance is inherently
CONCLUSIONS sensitive to modelling errors. IMC provides
a unique control structure within which
c lassic control theory
Most classic theor y is centered robustness can be adjusted easily on-line by
around the question of closed-loop stability. the operator. As a key result theorems were
The discovery of simple graphical and numeri- derived which quantity the maximum model/
ca l techniques
cal te c hniques like the Nyquist criterion plant mismatch tolerable by any control
and the Routh array were major accomplish- system involving integral action. If the
ments which allow stability assessment in an model error violates the stated criterion
c riterion the
manne r. For both linear and non-
efficient manner. control loops must be opened to maintain
linear open-loop stable systems (i.e. the stabilit y.
stability.
vast majority) the IHC IMC structure disposes of
the closed-loop stability issue altogether Let us emphasize again that for linear
and thus gives the designer the opportunity systems IMC is just a different way of look-
to address the central issues of control in
ingg at feedback control, that all theoreti-
s ystem performance and robustness directly.
system cal results could have been derived using
This uncluttered
unc luttered viewpoint
viewpo int has allowed to different methods and that the th e same perfor-
resolve a number of important issues which qualit y co
mance quality uld have been obtained
could ob tained
reapp earing in the control litera-
have kept reappearing desi gn techniques. The fact
employing other design
ture for yea rs.
f or year s. that the theoretical results were not no t found
previously and that the performance obtainedob tained
a re:
Among the most important are: previously for typical test examples does
no
nott match up to the standards set by DIC
1) Decoupling: It seems desirable not to is regarded tha t I~IC is able to
re garded as a proof that
introduce disturbances in all the variables impo rtant issues with unparalled
reveal the important
when changing
cha nging one of the setpoints. We\-Ie have simplicity.
shown, however, that in the presence of
fo r structured
nonminimum phase elements and for
Internal Model Control - Theory and Applications 13

In addition,
addition , the IMC
DIC structure which is also Grossmann I. I . E. and M. Morari (1983) (1983)..
inherent in Model Algorithmic Control and "Operability, resiliency and flexibility
Dynamic Matrix Control allows for the first -- Process Design Objectives for a
time to deal with constraints on the manipu-
manipu- world" . Proc.
changing world". Proc . Sec.
Sec . Int. ConL
Conf.
lated variables,
variables , states and outputs explicit-
explicit - on Foundations of Computer- Aided Process
Computer-Aided
law . Finally IMC's
ly in the control law. IMC ' s poten-
poten- Design, Snowmass,, CO
Design , Snowrnass CO..
tial for the design of controllers for Hirschorn , R. M. (1979).
Hirschorn, (1979) . "Invertibility of
nonlinear systems is very promising
promising.. Multivariabl e Nonlinear
onlinear Control Systems".
Trans.. Autom.
IEEE Trans Autom . Contr.,
Contr. , AC-24,
AC- 24 , 855.
855 .
Acknowledgement: Financial support from the Holt , B. R. and M. Morari (1983)~esign
Holt, (1983)~esign
National
ational Science Foundation (CPE- 8115022)
(CPE-8ll5022) of Resilient Processing Plants V. The
and the Department of Energy (DOE Contract effect of dead time on dynamic resilience. resilience .
DE-AC02-80ERl0645)
DE- AC02 - 80ERl0645) is gratefully acknowledged.
acknowledged . Chem
Chem.. Eng. Sci.,
Sci . , in press.
press .
Holt,
Holt , B. R. and M. Morari (1984). (1984) . "Design
of Resilient Processing Plants. Plants . The
REFERE
REFERENCES
CES effect of right-half-plane
right - half - plane zeros on
dynamic resilience".
resilience" . Chem. Chem . Eng. Sci.,
Sci. ,
Alevisakis,
Alevisakis , G, D. E. Seborg (1973) (1973).. "An ex- ex- submitted .
submitted.
tension of the Smith Predictor method Holt , B., N.• Jerome,
Holt, Jerome , M. Morari and W. H. Ray. Ray .
to Multivariable Linear Systems tontain- tontain- compensation". ,
"Multivariable time delay compensation".,
ing time delays".
delays ". Int.
In t. J. Cont rol, 2,
J . Control 3 preparation .
in preparation.
541. ' -' MacFarlane, A. G. JJ.. (1974) (1974).. "Relationships
Bristol,
Bris t ol, E. H. (1980). "A new process between recent developments in linear
concept : Pinned zeros".
interaction concept: zeros" . control theory and classical design
Report, Foxboro Company, Boston. Boston . techniques".
techniques" . Proc. Proc . of the IFAC Sympo-Sympo-
Brosilow , C. B. (1979).
Brosilow, (1979) . "The structure and sium, "Multivariable Technological
design of Smith Predictors
Pr.edictors from the Systems"
Systems"..
viewpoint of inferential control". control" . MacFarlane , A. G. J. ,and B. Kouvaritakis
MacFarlane,
Joint Automatic Control Conference (1977).. "A design technique for linear
(1977)
Proceedings , Denver,
Proceedings, Denver , CO.
CO . multivariable
multivarHlble feedback systems". systems" . Int.Int .
Brosilow, C. B. (1983) (1983).. Personal communica-
communica- JJ.. Control, 25 25,, 837.
837 .
tion.
tion . Mehra , R. K•.
Mehra, and R:-
K. _and R.'" Rouhani (1980)
(1980).. "Theo-
"Theo -
Clinch , A. (1982).
Clinch, (1982) . "Designing an IMC Con- Con- retical considerations on Model Algorith- Algorith-
troller for a first order system with mic Control for Non-minimum
Non- minimum Phase
variable time delay"
delay".. ChE 770 Term pro- pro - Systems
Systems". ". Joint Automatic Control Con- Con-
ject Report
Report,, University of Wisconsin,
Wiscon sin , ference Proceedings,
Proceedings , San Francisco,
Francisco , CA.CA .
Chemical Eng. Dept.Dept . Mehra , R. K., R. Rouhani, J.
Mehra, J . Eterno, J. J.
Cutler , C. R., B. L. Ramaker (1980).
Cutler, (1980) . "Dyna-
"Dyna- Richalet and A. Rault (1982). (1982) . "Model
mic Matrix Control
Cont r ol -- a Computer Control Control : Review and recent
Algorithmic Control:
Algorithm".
Algorithm" . AIChE 86th National ational Meet-
Meet- developement". Eng. Eng . Foundation Conf. Conf .
ing, April, 1979;1979 ; also in Joint Auto- Auto - on Chemical Process Control 11, 11 , Sea
matic Control Conference Proceedings, Island,
Island , GAGA..
San Francisco, CA. CA . Morari, M. (1983a)
(1983a).. "Robust stability of
Doyle J. J . (1982).
(1982) . "Analysis of feedback systems with Integral Control~ Control~'. . Proc..
Proc
systems with structured uncertainties".
uncertainties" . IEEE Conf
Conf.. on Dec.Dec . and Control, San
lEE Proc.,
Proc ., 129, Pt.Pt . D, 242.
242 . Antonio, TX.
Foss, A. S., J . ~Edmunds,
S. , J. ~Edmunds , B. Kouvaritakis Morari, M. (1983b).
(1983b) . "Internal Model Control: Control :
(1980).
(1980) . "Multivariable Control System A new approach to the control of Non- on-
for Two-Bed Reactors by the Character- Character- linear Systems".
Systems" . Proc. Proc . Symp.
Symp . on Non-on-
istic Locus 11ethod"
Hethod".. Ind.
Ind . Eng.
Eng . Chem
Chem.. linear Problems in Energy Engineering,
Fundam., 19, 109 109.. Argonne National
ational Laboratory, Argonne Argonne,,
Frank~.
Frank~. (1974) (1974).. "Entwurf von Regel- Regel- IL.
kreisen mit vorgeschriebenem Verhalten".Verhalten" . Morari, M. and W W.. H. Ray (1983).
(1983) . ChE 470
G. Braun Verlag, Karlsruhe.
Karlsruhe . Laboratory Manual, Chemical Eng. Dept., Dept .,
Garcia, C. E. and M. Morari (1982) (1982).. "Internal University of Wisconsin-Madison.
IHsconsin- Madison.
10del Control. 11.. A unifying review
Model Ogunnaike,, B. A. and W. H. Ray (1979).
Ogunnaike
and some new results".
results" . Ind.Ind . Eng.
Eng . Chem.
Chem . "Multivariahle
"Multivariable Controller Design for
Process Des. Dev.,
Des . Dev . , 21, 308
308.. Linear Systems having multiple time
Garcia, C. E. and M. Morari (1984a). (1984a) . "Internal delays".. AIChE J.,
delays" J . , 25,
25 , 1043.
1043 .
Model Control.
Control . 2.
2 . Design procedure for Popov, V. M. (1962). "Absolute stability of
(1962) . "Abwlute
Multivariable Systems".
Systems" . Ind. Ind . Eng. Chem..
Eng . Chem onlinear Control Systems of Automatic
Nonlinear
Process Des.
Des . Dev., submitted..
Dev . , submitted Control" Autom . & Remote Control, ~,
Control".. Autom. 22,
Garcia, C. E. and M. Morari (1984b). (1984b) . "Internal 857.
857 . --
odel Control. 33.. Multivatiable control
Model Prett, D. B. and R. D. Gillette (1980). (1980) .
law computation and tuning guidelines".
guidelines" . "Optimization and Constrained Multi- Multi-
Ind.
Ind . Eng.
Eng . Chem.
Chem . Process.
f rocess . Des.
Des . Dev.,
Dev . , variable Control of a Catalytic Catal y tic Crack-
Crack-
submitted. ing Unit".
Unit" . AIChE 86th National ational Meeting,
Meeting ,
14
14 M. Horari

April, 1979; also in Joint Automa tic


Contro l Confere nce Procee dings, San
co, CA.
Francisco,
francis
Ray, W. H. (1981). Advanc ed Proces s Contro l,
McGraw Hill.
Rebhuh n, D. (1980). "Inver tibilit y of CM
Hultiv ariable Input-O utput System s".
IEEE Trans. Autom. Contr. , AC-25, AC-2s, 207.
Reid, J. J., R. K. Mehra and E. ~ood Kirkw8~d

(1979). "Robus tness proper ties of Out-


put Predic tive Dead-B eat Contro l: SISO
case". Proc. IEEE Conf. on Dec. and
Fort Lauder date, fL,
Contro l, fort FL, p. 307.
Richal et, J. A., A. Rault, J. D. Testud and
"llodel Predic tive
(1978). "Hodel
J. Papon (1970).
Heuris tic Contro l: Applic ations to
indust rial proces s". Automa tica, 14
413. --
Ricker , N. 1.L. (1984). "The use of quadra tic
program ming for Constr ained Intern al
Model Contro l". Ind. Eng. Chem. Proces s
Design Dev., submit ted.
Rosenb rock, H. H. (1974). "Compu ter-Aid ed
Contro l System Design ". Academ ic Press.
Rouhan i, R. and R. K. Mehra (1982). "Model
Algorit hmic Contro l (MAC); Basic Theo-
retica l Proper ties". Automa tica, 18,
401. --
Safano v, M. B. (1980). "Stabi lity and
Feedba ck
robustn ess of Multiv ariable feedbac k
System s". HIT Press.
Solima n, M. A. and W. H. Ray (1972). "Opti-
mal Feedba ck Contro l for Linear- Quad-
ratic System s having time delays ". Int.
J. Contr. , 15, 609. ---
M:
Wood, R. K. and M: W. Berry (1973). "Termi nal
Compo sition Contro l of a binary distil -
lation column ". Chem. Eng. Sci., ~,
1707.
Zames, G. (1966). "On the input-o utput-
stabil ity of Time-V arying Nonlin ear
Feedba ck System s". IEEE Trans. Autom.
Contr. , AC-ll, 228
-;C;:-(19~ "feedba
Zames-;-C: "Feedb ackck and optima l
semsit ivity: Model referen ce transfo r-
mation s, multip licativ e semino rms, and
inverse
approx imate invers s". IEEE Trans.
es".
Autom. Contr. , AC-26.
Zhao and C. B. Brosilo w (1983). "Const rained
ariable Contro l". Sponso rs Meet-
Multivariable
Hultiv
ing, Contro l Science Center , Case
Western Reserv e Univer sity, Clevela nd,
OH.
Internal Model Control - Theory and Applications 15

Ys
--1 gc
u

A
g yY

y't c
H
B
g pl+
I

+ +
vs +
y
-
f-~_- y H"g-~- y

-....,d -
d
c o

+ e
y

Fig . 1.
Fig. 1. Evolu tion of the TIMe-Structure
Evolution IMC- Structure
)6
16 ~1. 110rari
I!. Horari

lmr-------~--------_.--------,_------__,
Im , - - - - - - . . . - - - - - - , . . . - - - - - - , - - - - - -........

g{iw)
1__ --,\-y
l~y
f-------t1L-.-_G_(s_)
(_s_) G

Re
u-:__
u - ! ,_ _G_<s_)_
G_(S_)_ _ H I+L(j(s)
I+Lo(s) rt-- yy

Fig. 2. Uncertainty band around model Fig. 4. Multiplicative input and output
g(iw) within which the real uncertainties for a multivariable
plant g(iw) lies. system.
(r = £,Q, (w) Igg(iw)
(r= I).•
( iw) I)

'7. , - - - - - - - - - . . . . . . . - - - - . . - - - - - . - - - -

" .• - t - - - . - - - i - - w - - t - - . . . . . - - i - - w - - I - - _ - - I

• 2a ..a ,a 18 188

1.5

+ 0• I., J ..

," .. ,
0
e ./ \
!~" ........ -- . _..
~ '-
.
~

A
I.'
• 2. ... , la 11.

2.2

I
E
f ~
.-"-""-
£-
,F
-.- -
_." -- ..
+ + f----,.--y
y l
2.1
Ys u
X

I .•
28 "I 'I .1.
I.'.
•. 11 -.--.----,r----,-----,---,----,

•l,.7'
I
: 1.7'
..-:::,
A
B
"
1.7.
ZI 4t

lINUT£1
•• NUT£I
" .. .11

Fig. 3. Equivalence of the TI1C


IHC struc
structure
ture Fig. 5. \,ood / Berry column, responses to
Wood/Berry
A and the Smith Predictor B. setpoint change in overheads
composition; - - 1~1Cn1C with :1J == 0.8;
--- S1S0 PI loops; _._.- Ogun- _e_._
naike/Ray compensa tor + PI.
na ike / Ra y compensator
Internal Model Control - Theory and Applications 17

9£i.6-----_---~---~---r_--....,
" .6

o0 ,"\
H
: \
C96 . 4 -+--~~~---+---+----t_--___1
C9£i.

U~ ~ - ~--.:---
o0
"
p
P
.. - ll-----.,r---.,.-----r---r---..,...----,
11
96. 2 -J------+--__--J--.,.--+---r--t---..,r----1
96 . Z
48 68 8B
88 lee
11
28
"
8B
o0
I.~

1.8
,-,
,,-.,
N

:z
.
N5-l---~~-~~~--+---t----t---..,
5
';2:."
o0
~
/'- ~,-
H
C "\. ~
f- aI-J\~J!.l~4~L++I4~f_\:i~_d~ffi~~n~~f'1~
<
,2 8.S
0
I.~ . ...•..
.........__
- .......... H

"" · 1 .8
8.8
0
Q
~


8 28
Z8 41
4' 61 8e
88 188
118 Ej
H
4l
:>
~
4l
~
-5
-S
\f
'i
H
~

Fig . 6.
Fig. 6. col umn, rejection
Wood/Berry column, rejec t ion to -u -l-----+-y-~-,-h..,._T""T+r_T-r-r+...,r_r-r1h'".,...,n-1
-11

disturbance change of + 0.34 0 . 34 • 588


'88 1188
1181 1588 2 •••
2811 2511
2588 J ...
3818

lb/min; wi t h a = 0,
l b/min ; --- IMC with 0,
compensa t or +
Ogunnaike/Ray compensator
PI.

u
o
... 4-4------1~--_+_--_+_---f_--_+_--_f
Z

,
o
H
E-. h
~
H
:>
~
.~-:;A.
~
-2-.-A A
/-~ /.
:\
..-,;;;
1-4.."..;;,;)~~~-I-~_+_-?~~J--__.~~::-or::~~~~~
·V··
~ ".
~

~ -2-+---1---..w.:-1=---+---+----t---~
~
r
T
0
U
T
22
~
~
J
-4 4-----1~--~--_+_---+_--_t_--_t

ffi
E-! -, -+..,.....,~---lI-r..,.....,r-T"_+_.....,..-r-r+_r_r-T'_r_+_r_r_r-T_+_T""T..,...T""1
-1

• 188 288 388 488 ,.8 688


511
'88 I".
1888 1511
1588 2118
2888 2518
2588 JI . .
38••

s[COHDS
SECONDS
5

C
0
U
T
• Fig.
Fig . 8.
8. Level and temperature response
cold- water disturbance
to a cold-water
stream . (PI = dashed line,
stream. line ,
-5 IMC = solid line).
line) .
1S8 28. 388 488 5U 68.

SECONDS

Fig . 7. Fixed bed reac t or, robustness


to modelling errors ; - IMC
wi t h a = 0', --- IMC with a = 0 . 5 .

IAA-B*
18
18 Horari
M. Morari

18
le

~ :s5
zZ
....E-1
0
~
f-;
<X:
<I;
....
H

:>
i;i
~
Cl
Q

~
....:I
w
W
-,-5
>
:>
w
W
~
....:I
-18
-1'

I 588
se8 18e8
188. 1588
Isee 288.
28"

2.8 -r-

U
°
Z
Z 1.5

~ .f-".\
....0
~
~
f-;
'I"'"
<X:
<I;
~
H
>
:>
~
W
1..
1.1 \
\ ! ~
'w
... _·r '.- "'C

Cl
Q
W
t.tJ
i \
~
::>
~
E-1
f-; 1.5
'.5
!
I
'ii
~
w
~

U
..
p..
~

ffi~E-1 •••
1.1
"
f-;

8 588
58. 1t8.
1••8
s(COHOS
SECONDS
Ise8 ...
22•••

Fig. 9. Level and temper ature respons e to a


IDoC setpoi
10°C ntt change in temper ature
setpoin
IMC == solid line).
(PI == dashed line, THC

You might also like