308 Ind. Eng. Chem. Process Des. Dev.

1982, 2 1 , 308-323

process phosphoric acid solutions. The uranium extraction OPAP was reduced slightly by increasing the CP04con-
coefficients and the organic phase iron loading were not centration, and was not affected appreciably by changing
affected by the MOPPA/DOPPA ratio within the range the concentration of other aqueous phase components.
of 0.7 to 1.5 or by differences in the coficentrations of Rased on these results, the most important impurities
impurities in the two OPAP batches tested. Extraction in wet-process phosphoric acid with regard to uranium
of uranium from simulated wet-process phosphoric acid extraction with OPAP are Fe"' and C F . Iron(II1) is
solutions was slightly lower than OPAP solutions prepared particularly important because it is extracted by OPAP
from commercial OPAP than with solutions prepared from and competes with uranium for the extractant. Uranium
purified MOPPA and DOPPA fractions. This effect was extraction could be improved by minimizing the concen-
due to the presence in the commercial OPAP of impurities trations of these impurities in the acid. Both of these
that depressed uranium extraction. Octyl phenol was impurities would have a smaller effect on uranium ex-
possibly involved but other impurities were more impor- traction from hemihydrate process acids (high CPO,) than
tant, since the effect of octyl phenol on uranium extraction from dihydrate process acids (low CPO,).
was shown to be small in earlier studies (Arnold et al., Literature Cited
1980). The increased uranium extraction power with pu- Arnold, W. D.; McKamey, D. R.; Baes, C. F. "Progress Report on Uranium
rified fractions was not large enough to justify separation Recovery from Wet-Process Phosphoric Acid with Octyiphenyl Acid
and purification of the fractions for process use. In ad- Phosphate"; ORNL/TM7182 (Jan 1980).
Goers. W. E. Proceedings of 28th Annual Meeting, Fertilizer Industry Round-
dition, the combination of purified fractions also extracted table, Atlanta, GA, 1978, p 99.
iron more strongly and was more susceptible to precipi- Hurst, F. J.; Crouse. D. J. Ind. Eng. Chem. Process D e s . Dev. 1974, 13,
286.
tation as a complex with ferric iron. McGinley, F. E. Paper presented at Uranium Industry Seminar, Grand Junc-
Uranium extraction was more strongly affected by tion, CO, 1972.
changes in aqueous phase composition. The extraction Or6, F. Proceedings of 28th Annual Meeting, Fertilizer Industry Roundtable,
Atlanta, GA. 1978, p 58.
coefficient was strongly depressed by increasing the con- Schwer, E. W.; Crozier, 8. T. Proceedings of 28th Annual Meeting, Fertilizer
centrations of C P 0 4 , FeIII, or HF, and moderately de- Industry Roundtable, Atlanta, GA, 1978, p 69.
pressed by increasing the concentration of CS04. The
uranium extraction coefficient increased when the con- Received for review March 19, 1981
Accepted December 28, 1981
centrations of Si, Al, or Mg were increased. The increases
were probably due to complexing of part of the fluoride This research was sponsored by the Division of Chemical Sciences,
by Si or Al, and to reduction of the aqueous phase acidity Office of Basic Energy Sciences,U S . Department of Energy, under
by the addition of A1 or Mg. The extraction of Fel" with Contract W-7405-eng-26 with the Union Carbibe Corporation.

Internal Model Control. 1. A Unifying Review and Some New

Results

Carlos E. Garcia and Manfred Morari'

Chemical Engineering Department, University of Wisconsin, Madison, Wisconsin 53706

Internal Model Control (IMC) is defined for single input-single output, discretstime systems and its relationships
with other control schemes (Optimal Control, Smith Predictor, Inferential Control, Model Algorithmic Control, Dynamic
Matrix Control) are established. Several new stability theorems for IMC are proven which provide practical tuning
guidelines. I t is concluded that the IMC structure allows a rational controller design procedure where control quality
and robustness can be influenced in a direct manner. Finally, the transparency and intuitive appeal of IMC make
it attractive for industrial applications.

1. Introduction McAvoy, 1976; Palmor and Shinnar, 1979). Though these

Spawned at least partly by the introduction of the state are quite straightforward applications of optimal control
space approach, we witnessed a surge in process control theory which unfolds its power more impressively in the
research during the 60's. This period of activity was fol- multivariable case, it is nevertheless an issue of significant
lowed by a period of reflexion during the 70's (Foss, 1973; practical interest.
Lee and Weekman, 1976; Kestenbaum et al., 1976) which Before proceeding with any more detail we would like
was precipitated by the realization that the wealth of new to emphasize that in this paper we are not concerned with
theory had only a minimal impact on the process control the 90+% of the control problems encountered in industry
practice. where essentially any control scheme will work satisfac-
One repeatedly addressed issue was the performance of torily but with the other situations where high performance
optimal controllers for single input-single output (SISO) loops are of critical importance for safety or economic
systems in comparison with that of standard controllers reasons but are difficult to design. In these cases we can
of the PID type (Kestenbaum et al., 1976; Bohl and assume that increased manpower for the design and in-
1121-0308$01.25/0 0 1982 American Chemical Society
0 196-4305/82/
 Ind. Eng. Chem. Process Des. Dev., Vol. 21, No. 2, 1982 309

creased hardware expenditures for the implementation are clusions are not surprising. Though the quadratic per-
economically justifiable. formance index is often criticized for just being a mathe-
For the purpose of comparison we postulate that with matical convenience, it yields a controller where the gain
regard to chemical processes the quality of a controller can be increased arbitrarily and thus a “perfect” response
should be judged according to the following criteria: (1) can be approached asymptotically. On the other hand, it
Regulatory behavior. The output variable is to be kept has been stated repeatedly (Horowitz and Shaked, 1975)
at its setpoint despite unmeasured disturbances affecting that the state space description which forms the framework
the process. (2) Servo behavior. Changes in the setpoint for most of today’s optimal control theory is unsuitable
should be tracked fast and smoothly. (3) Robustness. for robustness investigations. Sensitivity theory based on
Stability and acceptable control performance should be the assumption of infinitesimal perturbations is no viable
maintained in the face of structural and parametric substitute here.
changes in the underlying process model. (Note that this We have to add that all methods, modern as well as
is entirely equivalent to requiring that a controller should conventional, are equally incapable of handling constraints.
be designable with a minimum of information about the We disregard here indirect tools like penalty functions
process.) (4) Ability to deal with constraints on the inputs added to the objective or ad hoc solutions like switching
and states. the controller to “manual”.
It is a well-accepted fact (e.g., Lee and Weekman, 1976) While the academic discussions about the relative merits
that the major economic return from process control arises of the different schemes were in progress, apparently new
from the optimization of the operating conditions, i.e., techniques were developed in industry independently in
determination of the optimal setpoints, rather than from France (Richalet et al., 1978) and in the US. by Shell
regulation. A well-functioningregulatory layer is necessary, (Cutler and Ramaker, 1979). They are not the result of
however, to implement the actions dictated by the optim- a new theory but have a heuristic basis. As an attractive
izing layer. Optimal operating points generally lie at the feature the four criteria stated above are addressed quite
intersection of constraints (Arkun and Stephanopoulos, directly in the design. Moreover, simulations and indus-
1980). Therefore the ability of the regulatory controller trial applications showed the performance of these heu-
to deal with constraints on both the inputs and the states ristic methods to be vastly superior in particular with
is very important. respect to robustness.
All other criteria mentioned in the literature (Kesten- Because of the empirical origin of these techniques,
baum et al., 1976; Palmor and Shinnar, 1979) are either reservations about its general qualities and limitations are
implied by the four above or considered to be of minor only appropriate. In this paper we put them in the proper
importance. In particular, closed-loop stability of the place in the overall control theory framework. A rigorous
system in the absence of plant variations (compare formulation allowed us to discover similarities and dif-
“robustness” above) is almost never an issue because the ferences between them and many other control schemes,
majority of chemical processes is open-loop stable. among them the optimal controller and the Smith pre-
Sometimes a trade-off situation is imagined between sta- dictor. This comparison provided new insight into the new
bility and control quality: For example, for a system of methods as well as the old ones. This insight yielded not
order higher than 2 or a system with time delay the gain only a set of new theoretical results which significantly
of a P-controller and thus the achievable control quality expand the scope of the techniques but also pointed out
is limited by stability considerations. This, however, is only a large number of new promising research areas.
an artifact of the assumed simple control structure. With 2. Model Formulation
the proper dynamic compensator any gain is possible for The strategies considered in this paper provide discrete
any system. time control for the single input, single output continuous
Controller complexity is an important issue in the sense process
that the control structure and the effects of the tuning
parameters should be transparent to the operator. This x = A x + Bm(t - 0 ) +w (1)
is not to be misunderstood as a shortsighted endorsement y =ex
of controllers of the PID type. We have to realize that the
classic three-mode controller is a consequence of hardware where x E R“ is the state, 0 is a delay on the input, and
limitations. Any kind of dynamic compensation can easily w takes into account all unmeasurable disturbances af-
be added when computer control is used. Extremely fecting the process. For a given sampling time T the
speaking we could sprinkle poles and zeroes like salt and controller prescribes constant input values m(k) to be
pepper over the complex plane if this should turn out to applied during the intervals kT < t I ( k + 1)T, k = 0,1,
be desirable and as long as it does not impair on the overall ... . These inputs are based on output measurements at
transparency. .
times kT, k = 0, 1, ... Making the simplifying but not
essential assumption that 0 is an integer multiple of T (0
The response of the controller to specific disturbances = TT),the description of system 1at the discrete i n t e ~ a l s
(e.g., use of stochastic noise model for controller design, is
avoiding resonance frequency in certain range, etc.) is of
secondary importance. A filter inserted in the completed + +
x(k + 1) = (Px(k) I ” ( k - T ) w(k) (2)
control system can correct any undesirable features to the
maximum possible extent. y ( k ) = Cx(k)
The results of the extensive comparisons between op- where
timal control and conventional control can be summarized
as follows. Optimal control yields improved servo- and @ = BAT; I’ = ( L T e A Y dq)B
regulator behavior but the crucial robustness issue cannot
be addressed directly. Weighting matrices and/or noise and w ( k ) is the discretized disturbance vector w(t).
models have to be varied in a roundabout obscure fashion Analogous to the Laplace domain representation for
in the hope of achieving some robustness which then has continuous systems, we find it convenient to express sys-
to be checked through simulation. Actually these con- tem 2 in the z domain as follows (Jury, 1964)
310 Ind. Eng. Chem. Process Des. Dev., Vol. 21, No. 2, 1982

zx(z) = @x(z) + z-'I"(z) + w(z) (3)

342) = CXb)
where Z ( x ( k + a)) = z a x ( z )or in the input/output form
as a difference equation
y ( z ) = C(z1- @)-lI'z-"~(z) + d(z) (4)
where d(z) = C(z1 - @)-lw(z).For a stable system (i.e.,
I(@(I < l),one can further expand the transfer function as
an infinite series in z
m m
Figure 1. The discrete-time feedback controller with the model
~ ( - @)-IF = z-lChlz-(l-l)=
~ 1 = Z-lC(C@-ly)z-(l-l) modifications (dashed lines).
1=1 1=1
z-'H(z) (5)
finally obtaining the "impulse response model"
y(z) = ~ - ~ z - ~ H ( z ) m+(d(z)
z) (6)
Note that the coefficients hl in the convolution sum
represent the discrete unit impulse response of the system
and satisfy
lim h1= 0 (7)
1-c-

Consequently, for any t > 0, there exists an N such that Figure 2. Basic IMC structure.
lc@l-lrl
ct for 1 > N impulse response, the impulse response model is very
convenient for a theoretical investigation of robustness.
such that (6) can be approximated to any desired degree
of accuracy by 3. Structural Analogies
N Motivated by Brosilow's (1979) discussion, we develop
y(z) = z - 7 ~ - ~ C h ~ z - ( " ~+) md ((zz)) (8) the structure of Internal Model Control (IMC) and es-
1=1 tablish analogies with conventional control schemes.
In addition, one can obtain the equivalent "step response 3.1. The IMC Structure. 3.1.1. Evolution of the
model" from (8) by making the substitution IMC Structure. Consider the familiar feedback control
arrangement shown in Figure 1, where the process is rep-
hl = a1 - q - 1 (9) resented by a discrete transfer function G(z)with setpoint
s(z) and disturbance d(z). Denoting the model transfer
where function by G ( z ) one can subtract and add the effect of
1 the input m on the measurement signal (y&)) (dashed
= Chi lines in Figure 1)yielding an entirely equivalent setup. If
i=l
we consider the control block C(z) with model feedback
are the discrete unit step response coefficients. Therefore as our new controller we obtain the basic IMC structure
(8) becomes shown in Figure 2. Hence any conventional feedback
N controller can be restructured to yield IMC. Furthermore,
- ~ ) ) ( ~+ d ( z )
( ~z-l)m(z)
y(z)= z - ~ { z - ~ ~ u ~ z -- (10) any IMC can be put into the conventional feedback form
1=1 by defining
where now the model relates y ( k ) to changes in the input Gdz)
C(Z) =
Am(k) = m(k) - m(k - 1) 1 - G,(z)G(z)
Most of the results derived in this paper are independent If the two structures are interchangeable, what is the ad-
of the particular process representation. Very often in vantage of using IMC? We claim that the controller G,(z)
process control a "natural" fundamental state space pro- is much easier to design than C(z) and that the IMC
cess model is not available or the order is too high for it structure allows us to include robustness as a design ob-
to be useful for control system design. Then the impulse jective in a very explicit manner. This_is partly due to the
(and equivalently the step) response model offers several special form of the feedback signal d ( z )
advantages: no assumptions have to be made about the d ( z ) = (1 + ( G ( z )- G ( Z ) ) G ~ ( Z ) ) - ' ~ ( Z )(11)
number of poles and zeroes present in the system; therefore
we call the model structure free. As an apparent drawback We note that for a perfect model the feedback signal is just
many more parameters have to be determined than in a the disturbance d(z). Thus the system is effectively
structured model, but because of this nonminimal de- open-loop and stability is not an issue. When there are
scription the robustness is improved. The use of a state differences between the plant and the model, d@)contains
space model would require in addition the selection of a some information on them and by modifying d(z) appro-
coordinate system. The closed-loop robustness properties priately we can obtain robustness. We will elaborate on
are influenced by the selected coordinate system in a these points in more detail in the following.
nontrivial way and therefore this choice is very difficult. 3.1.2. General Properties of IMC and the Perfect
Finally, we repeat that the frequency domain is the ap- Controller. After some manipulations the following
propriate framework for a robustness analysis. Because transfer functions are obtained from the block diagram in
the transfer function is simply the 2-transform of the Figure 2.
 Ind. Eng. Chem. Process Des. Dev., Vol. 21, No. 2, 1982 311

In terms of the sum of the squared errors the best

performance achievable by any control system can be ob-
tained through the “perfect controller”.
Property 2: Perfect Controller. Under the assump-
tion that G(z) = G(z) and that G(z) is stable, the sum of
the squared errors is minimized for both the regulator and
For stability it is necessary and sufficient that both of the the servo-controller when
following characteristic equations have their roots strictly
inside the unit circle. G,(z) = l/G-(z) (19)
This property can be derived from optimal control theory.
Rather than doing that we will discuss the consequences
of (19). For the perfect controller (13) becomes
~ ( 2 =) G+(zMz) + (1- G+(z))d(z) (20)
We note that G+(z)contains the part of the process which
Property 1: Dual Stability Criterion. When the limits the achievable control quality. These limitations
model is exact, stability of both controller and plant is are inherent in the physical system and cannot be removed
sufficient for overall system stability. by any control system. In the absence of time delays and
Proof. When G(z) = G(z) the characteristic equations nonminimum phase behavior truly perfect control would
become be possible. It is well known that the system is very sen-
1 sitive to modelling errors when the perfect controller is
-=o used.
G,k) Property 3:-Zero Offset. A controller G, which sat-
isfies G,(1) = G(l)-l yields zero offset.
Proof. From (13) we have for a unit step change in s(z)
This implies that both the system and the controller poles lim y ( t ) =
t-m

have to be inside the unit circle. Q.E.D. G(l)d(1)-l

Property 1 confirms our statement above that for an (1 - d(1)) + d(1) = 1
open-loop stable plant closed-loop stability (but not nec- 1 + G(I)-’(G(l) - G(1))
essarily robustness) is obtained automatically for any for any asymptotically constant disturbance d. Q.E.D.
stable control block G,(z). This is also trivially true even 3.1.3. Controller Design Procedure. The properties
when the control block is nonlinear as long as it is in- derived above indicate that IMC allows a two step con-
put-output stable. Thus any constraints on the control troller design. Step 1: select G,(z) = 1/G-(z) or a suitable
action, for example, would not affect the stability. For approximation thereof. Step 2: add a filter to make the
conventional feedback structures with nonlinear controllers system robust. We will discuss Step 2 in the next section.
the stability issue is much more complex (Popov, 1963). As a result of Property 2 it is desirable to perform the
We should also note, however, that in general IMC does factorization (18) and use controller (19). However, if an
not allow the stabilization of an open-loop unstable plant. impulse response model (6) is used this may not be pos-
This could only be achieved when the controller cancels sible. Hence, our contention is that we may conveniently
exactly the unstable poles of the system which is impos- approximate the process inverse and still obtain a con-
sible in practice. An unstable plant has to be stabilized troller with desirable properties. This may be done in
first by conventional feedback before IMC is used. several ways. One may factor out the delays and then
Throughout the remaining part of this paper we assume approximate the remaining part of G(z). Alternatively,the
the plant to be open-loop stable. complete model inverse may be approximated. For ex-
After the resolution of the stability issue the question ample, proportional control (C(z) = K ) in Figure 1yields
of the best choice for G,(z) arises. Equation 13 shows that the IMC control block
“pejfect” control could be achieved by selecting G,(z) =
l/G(z). Because G,(z) has to be stable and realizable we
introduce the following factorization
&) = G+(z)G-(z) (18) which can be interpreted as an approximation to G(z)-l,
where G+(z) contains all the time delays (their inversion not necessarily a very good one.
would require a prediction) and all the zeroes outside the Assuming exact model description (6) and that H ( z ) has
unit circle (their inversion would yield an unstable con- all roots inside the unit circle, the perfect controller is
troller). These system characteristics are referred to as m(z) = H(z)-’(s(z) - d(z)) (21)
“non-minimum phase” (NMP) in classical frequency re-
sponse theory. yielding an overall response
The factorization (18)is non-unique, but the following y ( z ) = Z - ~ - ~ S ( Z+
) (1 - z-‘l)d(z)
form of G+(z) is advantageous (see “Property 2”)
Note that besides the dead times there is an additional
sampling interval inherent in all discrete-time systems
which makes the complete inversion impossible even when
r = 0. The resulting delay in the setpoint can be com-
where vi are the p zeroes of G ( z ) and Oi are their images pensated by specifying the required value r + 1sampling
inside the unit circle times ahead, thus yielding
Oi = vi (Vi1 I 1 y(2) = s ( z ) + (1- z-r-l)d(z) (22)
= l/Vi lvil >1 which means perfect compensation after r + 1 invervals.
312 Ind. Eng. Chem. Process Des. Dev., Vol. 21, No. 2, 1982

mi%]

I
PROCESS

1 i1.l

F 3,
~

Air,

Figure 4. The discrete-time Smith predictor.

model block R ( z ) . The setpoint following characteristics

of perfect control will make the output track the desired
dynamics. The complete IMC setup including all filter
blocks is shown in Figure 3.
In summary, the structure of IMC allows a rational
design procedure where in the first step the controller is
selected to give perfect control. In the second step we back
away from the ideal of perfect control by introducing a
This result stresses the importance of the dual stability filter which makes the system robust to a specified mod-
criterion (Property 1). Hence, since the unstable roots el-plant mismatch. This distinguishes IMC from all other
cannot be factored conveniently, H(z)-' has to be ap- feedback control schemes where a compromise between
proximated. Section 4 is devoted solely to the design of good response, stability, and robustness is accomplished
stable approximations of H(z)-'. in a very indirect fashion. Let us explore further how other
3.1.4. Sensitivity/Reference Trajectory. Although control schemes are related to IMC.
during the rest of the paper exact model description will 3.2. The (Inferential) Smith-Predictor (SP). A
indeed be assumed, in this section we discuss the possi- discrete analog of the celebrated time delay compensator
bility of desensitizing IMC with respect to modelling in- devised by Smith (1957) is given in Figure 4 (Alevisakis
accuracies. As proposed by Brosilow (1979), one can in- and Seborg, 1973). The main idea is to eliminate the time
clude a filter block F ( z ) in the feedback signal to account delay from the characteristic equation so that higher
for the model and proces_smismatch reflected through the controller gains can be achieved. One may note that in
estimated disturbance d ( z ) (see Figure 3). (In the case the light of the discussion in the previous section the SP
of an exact model such a filter could also compensate for "factors out" the pure delays r, then the rest of the model
certain types of disturbance dynamics.) As a result, the inverse is approximated, including the intrinsic sampling
characteristic equation (14) becomes delay.
G,(z)-' + F(z)(G(z) - G'(z)) = 0 (24) Brosilow (1979) has interpreted the SP from an infer-
ential point of view: the effect of disturbances on the
For a given model-plant mismatch F ( z ) can be selected outputs is estimated and handled directly by the controller.
such that (24)has all roots inside the unit circle. A detailed This reasoning has been adapted and exploited in the
design procedure for F ( z ) will not be attempted at this present paper.
point. Nevertheless, we will demonstrate the main idea 3.3. The Linear Quadratic Optimal Controller
by means of the following example. (LQOC). 3.3.1. Delay-Free Plant. Most modern control
Let us assume that the delay-free part of the model theory is centered around the combined problem of linear
( H ( z ) )has all roots inside the unit circle and is known quadratic optimal control and state reconstruction. In-
precisely, while the process has one unit of delay not ac- voking the separation theorem (htrom, 1970)the following
counted for by the model, that is minimization problem is solved
m
G(z) = Z - ~ H ( Z )
min E( C x(k
m(k) k=O
+ l)TQ1x(k + 1)+ Q2m(k)2] (27)
but
G(2) = 2-'H(z) subject to
Choosing the stable controller G,(z) = H(z)-', (24) becomes
x(k + 1) = ax@)+ fm(k) + w,(k)
+
H(z)(l F(z)(z-2 - 2-1)) = 0 (25)
y ( k ) = cx(k) + w2(k)
For F ( z ) = 1this characteristic equation has two roots on where Q1 is a symmetric positive semidefinite matrix, Q2
the unit circle. However, by introducing the stable ex- 1 0, E is the expected value operator, and w , ( k ) and w2(k)
ponential filter are mutually independent Gaussian processes with zero
means. The tildes above system equation matrices @,,'l
and C have been included to emphasize optimization with
respect to a model of the plant. The control law is known
to have the form
those roots become
m(k) -K%(k) + GCL-'(l)s(k) (28)
z1,22 = r2 %(k + 1) = &%(k) + f m ( k ) + K [ y ( k )- &(k)]
which are stable over the range of a given. where y ( k ) is the noise corrupt_edmeasured output, K is
In addition, one can also make the scheme follow a the proportional gain vector, K is the filter gain vector,
reference trajectory by feeding the setpoint to a reference x ( k ) is the estimate of the state x ( k ) ,and G C L ( 1 ) is the
 Ind. Eng. Chem. Process Des. Dev., Vol. 21, No. 2, 1982 313

UPl*i
I
1 I ,d!zi

U
Figure 7. LQOC for a deterministic first order system with delay.

U
shifted to the same position as in IMC, making the analogy
complete.
Figure 5. The LQOC structure for a delay-free SISO system. One remarkable result is the natural occurrence of the
filter F ( z ) introduced above somewhat heuristically
F(z) = (I + z-~fi(z)~-1z-1E(z)
(which is now a vector due to the state reconstruction). Its
function is to account for uncertainties in the measurement
Y!Zi
(w2)and model. Not surprisingly, the filter was found to
play a similar role in the IMC scheme and thus this
analogy provides a fundamental justification for its in-
clusion. The crucial difference is that the filter here is
designed by assuming certain intensities for the white noise
processes w 1 and w2. Robustness is related in a very
complex manner to these choices as well as the selected
values for Q1and Q2and the chosen state space coordinate
system (Doyle, 1978; Safanov and Athans, 1978). The IMC
structure, on the contrary, allows a filter design to achieve
robustness in a direct manner for a specified model-plant
mismatch.
Figure 6. Almost exact resemblance of the LQOC structure with 3.3.2. Delayed Process. The structural similarity
IMC. among the SP and the optimal control law has been
pointed out in the work of Ogunnaike and Ray (1979).
steady-state closed-loop gain included to achieve zero offset Basically, the scheme "compensates" for T delays and
for setpoint changes (Kwakernaak and Sivan; 1972) approximates the rest of the inverse by feedback as was
shown above. However, as we presently demonstrate using
GcL(z) = C(ZI- & + fK)-'f a simple first-order deterministic system as an example,
the optimal control law includes an additional factor which
In transfer function form, eq 28 become makes it perform sub-optimally as a regulator.
m(z) = -K%(z)+ GcL-'(l)s(z) (29) According to standard optimal control theory (Fuller,
1968; Koppel, 1968) the control law for a system with
+
%(z) = z-lD(z)m(z) z-'fi(z)[y(z) - CW] delays in the input is
where m ( k ) = -Ky(k + T )
where y ( k + 1) = # y ( k ) + y m ( k - T), and K is the same
d ( z ) = (I - z-%)-lf proportional gain which solves the delay-free problem. By
R(2) = (I - z-l&)-lR successively substituting for y ( k ) ,one can express y ( k +
T ) as follows
Let us now break up the state %(z) y(k + T) =
%(z) = & ( z ) + 4(2) + +
@'y(k) @"ym(k - T ) ... + # y m ( k - 2) + y m ( k - 1)
such that or in Z-transforms as
%l(z) = z-'D(z)n(z) z7y(z) = ~ ( z+ )X z)
%&) = z-'E(z)[y(z) - C%(z)] where
X ( z ) = (4r-1z-7 + @ r - 2 ~ - r + 1 + ... + @z+ + z-l)ym(z)
With these definitions the block diagram representation
of the optimal controller is shown in Figure 5. But using the identity
After minor rearrangements one discovers an almost
exact resemblance with the IMC structure in Figure 6. It X ( z ) - z-l@X(z)= z-lym(z) - @ r ~ - r - l y m ( ~ )
features the same approximate inverse controller as the the control law becomes
SP but with the setpoint introduced somewhat differently.
However, for a first-order system the setpoint can be m(z) = - K ( # ' ( ~ ( z-) z-"H(z)n(z)) + z - ' H ( ~ ) m ( z ) ](30)
314 Ind. Eng. Chem. Process Des. Dev., Vol. 21, No. 2, 1982

Figure 7 shows the resulting law in block diagram form.

As expected, we have obtained the SP for a first-order
system except for the additional constant filter block q ! ~ ~
in the feedback path. The optimal controller is designed
with the objective to minimize a quadratic performance
index when the system returns from a nonzero state to the
origin. I t is known that for systems without time delay
the resulting optimal feedback control law yields also good
regulator properties, for example asymptotically vanishing
offset for a decreasing penalty on the control effort.
Let us investigate both the servo- and the regulator
properties of the optimal controller for time delay systems. ii?

For this purpose a unit pulse (UP(z)) will be introduced

into the process (this is equivalent to a nonzero state at
time 7)and also a constant disturbance d. The response I -
of the closed loop system is
Figure S. General predictive control scheme.

more freedom to impose desirable dynamic characteristics

on the closed-loop system.
In order to derive approximate inverses a general pre-
(31) dictive-type control problem is formulated and solved in
this section. Tuning parameters of physical significance
For the minimum variance control problem (Qz= 0 in to the designer are introduced. As a key result, we prove
(27)) the optimal proportional gain can be shown to be K several stability theorems which allow us to make rea-
= 41y, which puts the root of the characteristic equation sonable decisions on those parameters and hence to obtain
precisely at the origin. The response to the unit pulse a control law which is extremely simple to implement and
alone is from (31) has the desirable characteristics. Finally, we demonstrate
that some recently developed algorithms which have been
y ( z ) = z++l)y(l + 4z-l + ... 47z-7)UP(z) shown to perform well in industrial applications are just
which shows dead-beat convergence to zero after 7 1 + special cases of IMC.
4.1. Predictive Control Problem. 4.1.1. Problem
sampling times. No other controller, including the SP of
Figure 4, can do better in driving the state to the origin. Formulation. In many control applications one is in-
If we look at the asymptotic behavior of the response terested in predicting the output trajectory of a process
(31) on the basis of a model as,for example, in order to prevent
the violation of operating constraints. A suitable predictive
lim y ( t ) = (1 - +'+l)d control strategy would consider the desired trajectory yd(k)
t-m
over a horizon of P sampling times into the future. Then
we notice, however, that even without any limits on the the sequence of control actions m(h),m(h + 1) ... m(k +
controller power we will always have offset, which increases P - 1)is calculated so that the predicted output y ( k + Ilh),
with the time delay. Note that this offset occurs even when 1 = 1, ..., P follows yd(k) as closely as possible. This pre-
r = 0 and also in the continuous-time counterpart of the diction is defined as the model prediction using the inputs
minimum variance problem (Fuller, 1968). +
up to h 1 - 7 - 1 plus the available information on the
These results may be summarized as follows. For sys- outputs up to h. If the plant model, which is used for the
tems with time delay the optimal controller cannot be prediction and the control computation, is incorrect the
regarded as an optimal regulator because it is unable to plant response in this open loop control law could deviate
remove offset. Therefore it is inappropriate to use it in significantly from the desired trajectory. Thus it is
a comparison with standard regulators and to draw con- preferable to implement only the present input m(h) and
clusions about optimal control on this basis (cf. Kesten- to resolve the problem again at h + 1with the measured
baum et al., 1976). The perfect controller (19) has not only output y(E + 1)as the new starting point and so on for 6
equally good servo-behavior as the optimal controller, but + 2, k + 3 etc. Figure 8 illustrates one step in this moving
+
also zero offset after T 1intervals. This provides further horizon problem.
justification for using the IMC philosophy. A t each time k the following general problem is solved
4. Approximate Inverses D

According to the IMC strategy one design problem is to

find a stable approximation to the inverse of H ( z ) , the m(l;+ 11,...,
other to design the filter F to preserve stability for the m (k+M-1 )
specified plant-model mismatch. The latter problem is not Ppm(K + 1 - 1)' (32)
dealt with in any depth in this paper, but a discussion is
included in the conclusion section. The need for an ap- subject to
proximation of H(z)-l can arise for two reasons. (1)H ( z ) Y(k + 716) = Y M ( +~ 7) 4- d ( k + TIL) = h,m(k
- 1) +
has roots outside the unit circle and therefore the exact + ... + h"(k - N) + d ( k + T I L )
hzm(k - 2)
inverse would be unstable. (2) The use of an exact inverse,
even when stable, often requires excessive control action m(R + M - 1) = m(&+ M) = ... = m(h + P - 1)
and can lead to an undesirable oscillatory response of the pt=O; 1>M
continuous process between the sampling times.
Although this problem can be partly alleviated by using where P is the horizon (P Il),yd(h + 7 + 1) is the desired
reference model tracking (cf. section 3.1.4), by defining trajectory, y? are time-varying weights on the output error,
suitable approximations to the inverse the designer has p? are time varying weights on the input, M is the input
 Ind. Eng. Chem. Process Des. Dev., Vol. 21, No. 2, 1982 315

N (34)
ypYpE(F+ P ) = [(Z hi)m(Z+N - l)hp
i=l
O=p,m(F)
O= 02m(K+I)

o= P"(K+ N - 1)

terms, so that the equations can be written concisely in

matrix form as

where
Figure 9. IMC structure with predictive controller and offset com-
pensator.

suppression parameter which specifies the number of in-

tervals into the future during which n ( k ) is allowed to
Y t"1
EP(Kt 1)= € ( E +2) ;

vary; it is kept constant afterwards, y ( k + 71k) is the

+
predicted output, yM(k r ) is the output of the internal
+
model, and d(k 71k) is the predicted disturbance.
The best prediction of the disturbance is to set it equal
to the disturbance at the present time
d ( k + T I E ) = d(k) for all k > R (33a)

[-']
d(k) is the feedback signal in the internal model controller
(Figure 2) which is obtained from the current measurement 1) m(F- N + 1)
and the current output of the internal model
d(@ = Y ( R ) - Ydk) (33b)
TM= ; BM = diag(p, ... p ~ ) ;
If only model prediction is used and disturbances are
omitted from the above problem one obtains the linear r, = diag(y , ... yp);
quadratic feedback control problem. Its inability to handle 0s :
persistent disturbances is a direct consequence of this 0 .

omission. :1
M, P, P I , and y l are tuning parameters of the algorithm
which have a direct influence on stability and dynamic
response. The theoretical basis for their selection will be
given in section 4.2.
4.1.2. Control Law Computation. For simplicity we
will restrict M IN so that inputs have an effect over a
+
total horizon P = N M - 1. This selection has no in-
fluence on the general results and extension to-other value!
+
is straightforward. Defining the error as c(k 1) = yd(k
+ +
+ 7 1) - d ( k r + LIE), we may observe that the least-
squares solution of eq 34 is equivalent to the solution of
(32) for M = N .
In eq 34 the contribution of past inputs (over which we
have no control) lies to the right of the dashed line. The
input suppression constraints (M < N ) can be included in
the formulation by collecting the corresponding input
316 Ind. Eng. Chem. Process Des, Dev., Vol. 21, No. 2, 1982

and Ij is an identity matrix of dimension j X j . Proof. When M = P I N , y i # 0, pi = 0 the set of eq

The matrix T M accounts for the effect of M on the op- 34 can be solved exactly for m(K),m(k + l), ...,m(k + M
timization. For example, for M = 1 (yi = 1, pi = 0), which - 1). Since the matrix to be inverted is triangular, the first
means that the input applied is kept constant throughout, equation gives the desired control law
the equations become simply

or H ( z ) m ( z )= z'+'yd(z) - ( y ( z ) - Y M ( z ) ) . Thus, G,(z) =

Nc(l)/Dc(z)= H(z)-l and N,(z) = l / h l . Q.E.D. The fol-
lowing obvious corollary is included to emphasize the effect
of NMP characteristics on stability.
Corollary 1. If the zeroes of H ( z ) lie outside the unit
circle, for y,# 0, 0, = 0, M = P I N the control law (36)
is unstable.
where aj are the step response coefficients defined in We next prove several theorems which demonstrate the
section 2. We stress the fact that since the problem is stabilizing effect of the different tuning parameters in the
solved at each time, the actual inputs applied to the pro- presence of NMP characteristics of the plant.
cess are not kept constant after M intervals of time. The Theorem 2. Assume that the system has a discrete
factor M is an artifice of the problem statement included monotonic step response and that y,= 1, 6, = 0, P = N.
to produce a more cautious controller. For M chosen sufficiently small the control law (36) is
The input to be implemented is obtained by selecting stable. (Note: a discrete monotonic step response does
the first element of the least-squares solution of (35) not imply that all the roots of H ( z ) are inside the unit
yielding the control law circle.)
Proof. A process with discrete monotonic step response
m ( k ) = bTuM(E) has h, > 0 (or < 0), i = 1,...,N . We will establish the result
= bT(TMTATRpTI'p2RpATM + by showing stability for M = 1.
If y,= 1,8, = 0, M = 1,P = N , the characteristic eq 39
BM2)-lTMTA%pTrpT(tp(& +- 1) - rp\kV(R - 1)) reduces to
bT = (10 ... 0) (36) N N- 1 2
which has the closed form C(q) = (Ea?) + (Ca,h,+l)q+ ... + (Ca,hN+1-&N-2
+
1=1 1=1 r=l
N-1 alhNqN-' = 0 (40)
+
m(R) C61m(K- I ) =
1=1
P
where ak = Ckl,lh, are the step response coefficients.
Cvl(Yd(k + 7 + I ) - (y(k)- y M ( & ) ) ] (37) Starting with the last (a,hN)one can verify that all the
1=1 terms in each coefficient are included in the next. Con-
The general form of the control law in 2-transforms is sequently, if h, > 0, a, > 0 the monotonic conditions in
therefore lemma 1 of Appendix A are satisfied, and thus C(q) has
all roots outside the unit circle. If h, < 0 then all a, < 0
Dc(z)m(z) = z'+'Nc(z)yd(z) - Nc(l)[y(z) - yM(z)l (38) and the terms in the coefficients of (40) are all positive,
where again implying the result. Q.E.D.
+
D,(z) = 1 612-1 + ... + 6N-1z-N+1 Monotonic step response systems are not uncommon in
chemical processes, which are characterized by high order,
N&) = v1 + v2z + ... + vpzp-1 over-damped responses. Besides, by using a sufficiently
large sampling time all h,'s can be made positive for NMP
Stability of the scheme is determined by the roots of systems which exhibit inverse response. Consequently,
D&). By defining the backward shift operator q = 2-l we
equivalently require the roots of the characteristic equation theorem 2 ensures the existence of a stable approximate
inverse of H ( z ) for a large class of systems.
In case a discrete monotonic response cannot be ob-
tained, the following theorem guarantees stability for any
system when a sufficient penalty on the input is intro-
duced.
Theorem 3. There exists a finite p* > 0 such that for
6, Ip* (i = 1, ..., M) the control law (36) is stable for all
M I1, P I1, and y,2 0.
to lie outside the unit circle for stability.
Proof. We will show the result for 0, = 6, i = 1, ..., M .
Then by selecting PI I the control law will remain stable.
The block diagram representation of the control law is The characteristic eq 39 is
shown in Figure 9. The offset compensator is introduced
in accordance with Property 3. Note the explicit inclusion 1 + 61q + ... + 6N-lqN-l = C(q)
of future setpoint values because of the predictive nature where
of the polynomial N(z). The following theorems show the
effect of M , P, pi, and y ion the stability of D,(z) and thus
provide guidelines for their selection. Once chosen, ma-
trices in eq 36 are completely specified allowing the com-
putation of the control law.
4.2. Stability Theorems. The perfect controller is a
special case of (36) as shown by the following theorem.
Theorem 1. For y i # 0, pi = 0, selecting M = P I N
yields the model inverse control G,(z) = H(z)-'.
 Ind. Eng. Chem. Process Des. Dev., Vol. 21, No. 2, 1982 317

Defining e = 1/(N - 1) and di = [Sil, Biz, ..., 8 i ~ - ~ ] ~ theorem 1we know that for Pi = 0, yi # 0, M = P I N the
control law (32) is equal to H(z)-l. By changing the tuning
parameters we obtain different approximations of H(z)-'.
In case A the first goal is to obtain a stable approximation;
in the second step we modify it to obtain a more desirable
response. In case B stability is not an issue and only the
quality of the response is affected.
In view of this it is advisable to start any tuning pro-
cedure by checking the roots of the equation H ( z ) = 0
employing a polynomial root finder. Actually, since all we
1 1 need is the modulus of the largest (smallest) root of D,(z)
-ATA - -(ATA)2 - ...)I (C(q)) to establish stability, an algorithm to find the
P2 P4 spectral radius of its associated companion matrix as, for
where we have expanded the inverse in a Newmann series. example, by successive substitutions, is sufficient (Fad-
Thus deeva, 1959). Due to encountered convergence difficulties
this approach was not used here.
, (1) Sampling Time (2'). Case A Frequent sampling
of a continuous process with RHP zeroes produces an H ( z )
with roots outside the unit circle. It is straightforward to
show, however, that if the sampling time is sufficiently
increased those roots are made stable.
Assuming p2 is chosen sufficiently large such that P2 > Another option is to increase T just enough to ensure
IIATAII the above series will converge. Then there exists hi > 0 (i = 1,..., N) and to obtain a stable approximation
a b2 > IIATA1l such that lSil C e for any A,* and conse- by decreasing M according to theorem 2. Case B Contrary
quently to conventional feedback control the stability of IMC is
N- 1 not affected by T. Larger 7"s generally lead to less extreme
l6il C (N - l ) =~1 excursions of the manipulated variable, but they have a
i=l
detrimental affect on the ability of the system to handle
It follows then from lemma 2 of Appendix A that the frequent disturbance changes.
polynomial (2) Input Suppression Parameter ( M ) . Case A:
According to theorem 2 any system with hi> 0 (i = 1, ...,
1 + 61q + ... + 6N-lqN-1 = C(q) N) can be stabilized by choosing M sufficiently small. Case
has all N - 1 roots outside the unit circle. Q.E.D. B: Perfect control (M = P = N) requires severe variations
We finally give a result which demonstrates the stabi- in the input variable and usually leads to strong oscillations
lizing effects of the horizon limit P > N + M - 1. of the output between the sampling times. Reducing M
Theorem 4. Assume yi = 1, Pi = 0. Then for a suffi- reduces the extreme excursion of the manipulated variables
ciently small A4 and a sufficiently large P > N + M - 1the and thus leads to a more desirable response of the plant.
controller (36) is stable. (3) Input Penalty Parameter (&). Case A According
Proof. The result is established by showing stability to theorem 3 we can always obtain a stable approximation
for M = 1 and P > N sufficiently large. For any P > N, of H(z)-l by making Pi sufficiently large. Case B: In-
and M = 1 the characteristic eq 39 becomes creasing Piwill decrease the actions taken by the manip-
ulated variables and make the system more sluggish.
N P N-1
Furthermore, Pi # 0 will in general lead to offset since in
C(q) = (C(riaJ2+ ( C Y~%N') + ( C aiy?hi+l)q + that case Nc(l)/Dc(l)# l/iY(l). This is corrected by the
i=l N+ 1 ill
2 offset compensators shown in Figure 9.
... + +
(Caiyi2hN+i-2)qN-2 a1y12h"-1 =0 An alternate procedure is to formulate the problem by
i=l
penalizing control action increments thus minimizing the
Hence, since y? = 1, we can find a P > N large enough such objective
that P
Cyr2(y&
1=1
+ + I ) -y(K + 7 + ilk))' +
7

p?(m(k + 1 - 1) - m(k + 1 - 2))' (41)

2

then it is straightforward to show that Nc(l)/Dc(l)= 1/

H(1)for any P1. However, excessive weighting would yield
and consequently by lemma 2 (Appendix A) all roots of a sluggish controller since it puts the largest root of D,(z)
C(q) lie outside the unit circle. Q.E.D. on the unit circle. It is important to note that by penal-
One may look at the effect of a P > N as equivalent to izing control action increments only it is generally not
introducing a penalty on the inputs as may be seen from possible to obtain a stable approximation of H(z)-l for
the problem formulation in eq 35. Thus the results of NMP systems. An example is presented in Appendix B.
theorem 4 can be interpreted in the light of theorem 3. (4) Output Penalty Parameter (yi). We first note
Moreover, it is evident from the proof that the restriction that the use of yi is only meaningful when these parameters
yi = 1 can be eased. For stability it is only required that are time varying. Otherwise one should choose yi = 1. The
the sum CpN+lyi2 be sufficiently large. motivation for introducing yi arises from the work of Reid
4.3. Implementation of IMC. 4.3.1. Tuning Proce- et al. (1979). As we remarked repeatedly, G,(z) = H(z)-'
dures. We should distinguish two cases: Case A. The leads to large excursions of the input and strong oscillations
plant model H(z) is nonminimum phase, some zeroes of of the output between the sampling intervals. The reason
H ( z ) are outside the unit circle, and H(z)-l is unstable. for this undesirable behavior is that the controller drives
Case B. H(z) is minimum phase and H(z1-l is stable. From only the output to zero without taking care of the other
318 Ind. Eng. Chem. Process Des. Dev., Vol. 21, No. 2, 1982

system states (e.g., derivatives of the output). It is well

known that any nth order system can be brought precisely
to rest in no more than n discrete steps. This suggests to
postulate the following predictive control law: ”Let the
inputs be variable over the next n steps and zero after-
wards. Determine the inputs such that the outputs at steps
n, n + 1, ..., 2n - 1are zero”. This implies the following
choice of tuning parameters for a system of order n: P =
2n - 1; M = n + 1; p1 = p2 = ... = p, = 0; pn+l = large; y1
= y2 = ... = Y ~ =- 0;~ yn = yn+l = ... = y2n-1= 1. These
values resulted in superb performance in our test simu- Figure 10. MAC with reference trajectory.
lations. If just the impulse response of a system is iden-
tified the order n is unknown. Nevertheless, we have not until the past few years, after several successful ap-
reduced the tuning procedure to the choice of this one plications, that special attention has been given to its
parameter n from which all other parameters can be de- theoretical properties. The original concept is expounded
termined. by Richalet et al. (1978) and the theory studied by Mehra
(5) Optimization Horizon (P). Case A: Any system et al. (1980), with special attention to nonminimum phase
can be stabilized by selecting M sufficiently small and P systems given by Mehra and Rouhani (1980). Applications
sufficiently large. Case B: For minimum phase systems range from steam generator control (Lecrique et al., 1978),
the length of the optimization horizon has virtually no electric power plants (Mehra and Eterno, 19801, to flight
effect on the performance as soon as P exceeds about twice control (Mehra et al., 1978). A review of the developments
the system order. is presented in Mehra et al. (1981).
4.3.2. On-Line Computations. The control law is of The control law is obtained from the optimization
the form problem
m(k) = min [ y d ( k
m(k)
+ ’i + 1)- y ( k + T + (42)
N-l P
-E &m(k - 1) Cvl(Yd(k 7 I) - ( y ( k ) - Y&))) subject to
1=1 1=1
(37) N

Once the tuning parameters are selected the coefficients

YM(k + ’i + 1) = 1=1
CRlm(k + 1 - I )
ai, vi could be calculated off-line from (36). Alternatively, which is a one-step ahead predictive controller. A reference
if on-line tuning is desired (36) can be solved on-line, re-
quiring the inversion of a matrix of dimension M each time trajectory is defined as
the tuning parameters are modified. If only some of the Yd(k + ? + 1) = (1 - CY)s(k + + I ) + a y ( k + ?lk)
tuning parameters are changed the inversion of a matrix
of lower order can be sufficient after proper partitioning. where 0 ICY < 1specifies an exponential approach from
We can also express the control law (37) in velocity form the predicted output 7 time intervals ahead to the setpoint
to provide bumpless transfer from manual to computer +
s ( k T + 1). The prediction is just
control N -
N-1 Y(I + ’ilk) = Chjm(l- j ) + [ ~ ( k-) Y&)I
Am(k) + i = l 6iAm(k - i)= j=l

P and we note that y ( k ) - yM(k) is the estimated disturbance

CviIAYd(k
i=l
+ 7 + i) - ACY(k) - Y d k ) ) ) at time k (eq 33).
Solution of (42) yields the control law
where Ax&) = x ( k ) - x ( k - 1). Observe that past input
changes need to be stored and the future desired trajectory
m(2) = A(z)-’[(1 - ff)Zf+lS(Z) + +
CY(z-’A(z)m(z) ( y ( 2 ) -
y.af(z))J - (Yk) - Y&))l (43)
specified for implementation.
For a reduction of the computational burden it is further which is IMC’s model inverse controller with an expo-
possible not to move the horizon after each step but to nential trajectory Yd(k +
? + 1) (see theorem 1). The
implement the first k control moves before redoing the resemblance with the general IMC structure is more ob-
calculations. This, however, makes the performance more vious if we rearrange (43) in the form
sensitive to modelling errors.
It was also suggested to “block”the inputs for the com-
putations, i.e., to allow them to vary only each kth step.
With proper weighting excellent results were achieved
(Reid et al., 1980). with block diagram shown in Figure 10. Observe that this
4.4. Special Cases. Two recently developed algorithms is precisely the structure in Figure 3 with
which have been reported to perform well in chemical 1-CY
process applications are discussed here. In fact, we dem- R(2) = F(2) =-
onstrate both schemes to be special cases of IMC for 1 - az-1
specific choices of tuning parameters. Their excellent G,(z) = A(z)-l
performance is explained in the light of the exposed
properties of IMC. Thus MAC’S reference trajectory is equivalent to using an
4.4.1. Model Algorithmic Control (MAC). MAC was exponential filter in the feedback path of IMC. By ad-
developed in France in the late 60’s inside the chemical justing the time constant of the reference trajectory the
process industry. Mainly a heuristic algorithm (then called operator can affect the robustness of the controller in a
Model Predictive Heuristic Control), it was interpreted as direct manner. As would be expected, a slow reference
a dual of the least-squares identification method and it was trajectory makes the scheme more robust.
 Ind. Eng. Chem. Process Des. Dev., Vol. 21, No. 2, 1982 319

As we already know, the perfect controller obtained in

MAC is unstable for NMP systems and thus modifications
have to be introduced. From eq 24 the characteristic
equation for this scheme is
(1- az-')A(z) + (1- a)[z-'-'H(z)- z-"'A(z)] = 0 (45)
and consequently when 'i # 7 , a(z) # H ( z ) , a could be
manipulated in the attempt to put the roots of (45) inside
the unit circle. However, for exact model description (45)
becomes ( z - a ) H ( z )= 0 and thus a affects the dynamic
characteristics of the system but has no influence on the -
"unstable" zeroes of H(z). t-7
0
Mehra and Rouhani (1980) propose two methods for
handling NMP systems. Their pole placement method
factors H ( z ) as
-2 , 1 1 1 I , 1 , I , 1 , 1 r7-1 I , , , I , , ,~.
H ( z ) = H,(z)H,(z)
where H,(z) and H J z ) collect the stable and unstable roots TIME
of H(z), respectively. By choosing suitable polynomial Figure 11. Effect of M N for the MP system ($5.1): ( O ) ,M = 10,
factor P&) for the internal model, H(z) = H,(z)Pdz),one M = 1, p = 0.41 (P= N = 10,
p = 0.77; (A),hf = 2, p = 0.54; (O),
can make the roots of H,(z)((l- a)H,(z) + (z - 1 ) P M ( Z ) ) 0, = 0, yi = 1, CY = 0).
= 0 stable. This method has two disadvantages. One is
that there are no guidelines for choosing PMto obtain a control optimizer with excellent results.
good dynamic response. The second one is more serious: The scheme is precisely the IMC predictive controller
the internal model used for output prediction is with P = N and M < N . Due to the model description
changed-it barely resembles the plant model-and thus employed, input changes are penalized as in eq 41.
the main attractive feature of IMC, i.e., subtracting the Our new results provide convenient tuning guidelines
effect of the inputs from the measurements, is destroyed. for DMC and also point out many generalizations for
The other method they suggest also changes the internal performance improvement.
model. Th_einternal model H ( z ) (and the corresponding 5. Examples
controller H(z)-') are selected as "close" as possible to the In order to demonstrate the features of IMC several
plant H(z) but such that the roots of (45) lie inside the unit simulation examples were run on a digital computer. In
circle. Mathematically, the problem can be stated as all cases the systems start at steady state (y, = m, = 0) and
1 m N a disturbance d(1) of magnitude 0.5 is introduced. The
sampling time (7')is chosen as 10 units with a delay of 7
= 2 (0 = 20). We include the magnitude of the largest root
where m(l - j ) is the system input obtained from the MAC of D J z ) ( p ) in all results for comparison.
control law (44). Problem (46) is solved off-line using 5.1. Minimum Phase System. A second-order process
optimal control theory requiring the solution of a Riccati with transfer function
equation. Although "optimal", solving a Riccati equation
of order N (number of terms used in the impulse response
G(s) = e-20S/(100s2 12s + + 1)
= 40-80) presents quite a computational burden. This is is controlled using IMC. We would like to compare the
specially inconvenient because the Riccati equation has responses obtained using the model inverse controller,
to be solved each time the on-line tuning parameter a is G,(z) = H(z)-', vs. the proposed inverse approximations.
modified. The 10th-order truncation model ( N = 10)
In terms of IMC, the approximation to the inverse we
suggest either by reducing M o r increasing PI provides no
y ( k + 1) = 0.322m(k - 2) 0.461m(k - 3) + +
0.255m(k- 4) + 0.056m(k - 5) - 0.034m(k - 6) -
additional complications and conserves the inferential
0.043m(k - 7) - 0.023m(k - 8) - 0.004m(k - 9) -
structure. In addition, the approach is also optimal for
a given optimization horizon. 0.003m(k - 10) - 0.002m(k - 11)
Before concluding this discussion we should mention a was used to obtain sufficient accuracy.
recent paper by h r o m (1980) where robustness of MAC Effect of M . Using yl = 1, Pl = 0, and P = 10, the
is demonstrated with a simplified first-order truncation model inverse controller is given by M = 10. As shown in
model Figure 11, it gives perfect setpoint compensation after 7
A(%)= z-'bm(z) + 1= 3 discrete time intervals but at the expense of severe
input variations. Reducing M to values of 2 and 1 gives
Stability is proven for continuous monotonic response better input dynamics and less oscillations in the output
systems although by proper selection of the sampling time although the setpoint is not satisfied exactly at the sam-
it should be possible to handle inverse response systems pling periods.
as well (see discussion on sampling time T above). Effect of Penalty on Inputs m . For M = P = 10 a
4.4.2. Dynamic Matrix Control (DMC). Using a step constant penalty parameter of P1 = 5 was tried. Without
response model, Cutler and Ramaker (1979) of Shell Oil proper compensation, a significant offset is observed (see
Co. developed an algorithm which they applied successfully Figure 12). Introducing the offset compensator (Figure
to the control of chemical processes such as catalytic- 9) a smooth response is obtained with good input dynam-
cracking units. The scheme is a predictive-type controller ics.
capable of following a time varying trajectory and exhib- Effect of Penalty on Am. Using the same weight as
iting zero offset to persistent disturbances. It was used in Figure 12 (PI = 5), a penalty on input changes Am(k)
by Prett and Gillette (1979) also of Shell in a constraint yields a sluggish controller (Figure 13). A better response
320 Ind. Eng. Chem. Process Des. Dev., Vol. 21, No. 2, 1982

U
T

-115 I , , , , , I I T I , , , I , , , , I , ,,,,
I 50 IOU I15 150 0 3 50 :5 IOU 325 I58

TIMF

-9 I
N
0 P

- 2 - , , , , , , I , I , , , , 1 1 1 , , I , , , , I

8 ZS se 7s 108 1 5 150 i is
TIME TIME
Figure 12. Effect of penalty on m(k) for system ($5.1): ( O ) ,6 = 0, Figure 14. Reference trajectory for system ($5.1): (O),
a = 0; (A),
p = 0.77; (A),6 = 5, p = 0.52 no offset compensation; (01, 6 = 5, a = 0.45; (0),a = 0.70 ( P = M = N = 10, 61= 0, 71 = 1,I/ = 0.77).
p = 0.52 with offset compensation (P = M = N = 10, yI = 1, a = 0).

Tl'4F
B O

1
H
P - 8 5--. ~ - ..--.
u
1

-18 I ,
- , , I t , , I , , , I , , , I , , , 1 , 1 1 1

a 25 58 7s lee 125 150

0 n sb 7s 108 125 158
TIRE
TIME
Figure 15. Dead beat controller with reference trajectory for system
Figure 13. Effect of penalty on Am(k) for system ($5.1): (01, 6= ($5.1): (0)a = 0; (A),(Y = 0.45; (n),a = 0.70 (n = 2 in $4.3.1, no.
0, p = 0.77; (A),6 = 5, p = 0.83; (O),6 = 1.5, p = 0.58 ( P = ?d= 4).
N = 10, = 1, = 0).
1 a0

is obtained by using a P1 = 1.5. Note the difference in input

dynamics as compared to Figure 12.
Effect of Reference Trajectory. Figure 14 shows the
response when CY is increased. The exponential discrete
approach to the setpoint causes decreased input ampli-
0 0 SO
tudes; however, a much smoother response is obtained by U
reducing M or by using input penalties. This is the result T
P
of a not being able to move the roots of H(z). U
19 25
Effect of Penalty on Outputs. According to our
discussion above we choose the following tuning parame-
ters for this second-order system. P = M = 3, PI = P2 = e 88
0, P3 = very large, y1 = 0, y2 = y3 = 1. The excellent
performance of this control scheme for different reference
trajectories is shown in Figure 15. Note that DMC with -8 25
M = 2 and P = N (Figure 11) is almost as good as dead d s'0 I ho
beat control. This is to be expected due to the similar way TIME
the input moves are restricted in both schemes.
5.2. Nonminimum Phase System. 5.2.1. Discrete Figure 16. Unit step response for NMP system in $5.2.1: (O),
indicate discrete samples for T = 10.
Monotonic Response System. Stabilizability with M .
The following system with a RHP zero the response to a unit step change. In order to guarantee
- 5s e-20s stability for a sufficiently small M a sampling time of 10
G(s) =
(10s 1)2 + units was chosen producing a discrete monotonic response.
The largest root of H ( z ) (using N = 10) has a magnitude
was controlled using the IMC approach. Figure 16 shows of 3.97 for this sampling time and hence the inverse con-
 Ind. Eng. Chem. Process Des. Dev., Vol. 21, No. 2, 1982 321

T
P
y 0 0 ----
0 TlMF ,
tiI 0 0 1
I
H
P -1 b P
U
-I .-
U T
T

TIME
TIRE Figure 19. Stable controllers for NMP system of $5.2.2 by intro-
Figure 17. Stable controllers for NMP system for Figure 16 by ducing input penalties and increasing the horizon: (O),penalty on
reducing M. ( O ) ,M = 9, p = 0.51; ( A ) ,M = 2, p = 0.42; ( O ) ,M = m(k) with @ = 1 and offset correction (P= M = N = 10, yi = 1, (Y
1, p = 0.36 (P= N = 10, pi = 0, y f = 1, (Y = 0). = 0, p = 0.51); ( A ) ,penalty on Am(k) with 0 = 0.5 (P= M = N =
10, = I, (Y = 0, p = 0.46); (0), P = 11 ( M = N = 10, = 0, 71 =
1, CY = 0, p = 0.62).

Table I. Normalized Impulse Response of

Furnace Outlet Temperature to Fuel Flow
(Cutler and Ramaker, 1979)
time
1 2 3 4 5 6
d io 4e de se ioe 150 lie normalized 0.014 0.072 0.128 0.200 0.186 0.136
temperature
time
7 8 9 10 11 12
normalized 0.100 0.068 0.045 0.037 0.014 0.0
temperature

Table 11. Effect of Tuning Procedures o n Magnitude o f

Largest Control-Law Root ( p ) for System of Table I
reduction increase in
TIM of M (P = penalty o n penalty o n horizon P
N = 11, m ( h ) (P= A m ( h ) (P= (M = N =
Figure 18. Stable controllers for nonmonotonic NMP system of Pi=O) M=N=11) M=N=11) Il,Pl=O)
$5.2.2 by reducing M (0),M = 9, p = 0.62; ( A ) ,M = 2, p = 0.52;
(O),M 1, p = 0.40 (P = N 10, = 0, 71= 1, (Y = 0). M P PI P PI P P P
11 3.4 0 3.4 0 3.4 11 3.4
troller is unstable. Figure 17 shows the stable responses io 1.3 10-3 1.2 10-3 1.2 12 1.3
obtained when M is reduced. We should mention that 9 0.95 lo-' 0.80 10" 0.77 13 0.82
using a sampling time of 5 which yields an hl < 0, a stable 8 0.82 lo-' 0.73 10.' 0.74 17 0.82
controller is obtained for M = 1. Also, for T = 20 the 7 0.82 1 0.59 1 0.70 21 0.82
6 0.82 10 0.33 3.5 0.63
magnitude of the largest root of H ( z ) is brought down to 5 0.83 10 0.94
0.62 and thus the exact model inverse controller could be 4 0.81
used. 3 0.72
5.2.2. Nonmonotonic Response System. In order to 2 0.68
demonstrate the effectiveness of the tuning procedures the 1 0.51
process
The responses when other methods of stabilization are
(1 - used are shown in Figure 19. As expected, a penalty of
G(s) = Pr = 1on m(k) produces a stable controller. Moreover, we
1oos2 + 12s +1 were also able to stabilize this system by penalizing Am&)
was controlled using IMC, which exhibits inverse response and increasing P.
and a nonmonotonic profile. For T = 10, p = 6.0 and hence 5.3. Dynamic Matrix Control. In their original article,
the exact inverse controller cannot be used. We also em- Cutler and Ramaker (1979) give a furnace outlet tem-
ployed a 10th-order truncation model here. perature response to a step change in fuel flow which ex-
Although stability by reducing M for P = N , PI = 0 is hibits a discrete monotonic response (see Table I).
ensured only for systems with hi> 0, in Figure 18 we show However, using the response coefficientsgiven, an unstable
that M = 9, 2, and 1provide stable controllers, with the inverse controller would result since for M = P = N = 11,
same dynamic characteristics as before, for this nonmo- p = 3.4. In Table I1 we show how p varies with the dif-
notonic response system. ferent tuning parameters.
322 Ind. Eng. Chem. Process Des. Dev., Vol. 21, No. 2, 1982

Observe that the decrease in magnitude of the largest to identify a low order parametric model and then use the
root is nonmonotonous with M and it is not until M = 8 impulse response from this model in IMC. Furthermore,
that a value significantly less than 1is obtained. Moreover, it would be interesting to employ the IMC structure in an
by penalizing Am&) as done in DMC, p decreases to a adaptive control law.
minimum and then approaches 1. Also, observe that an We have not discussed how to design the filter (or
increase in P is able to stabilize this NMP system. equivalently how to choose the reference trajectory) to
compensate for a specific model-plant mismatch. An ex-
6. Conclusions cellent summary of the available results is provided by
IMC consists of three parts: (1)internal model to pre- Mehra et al. (1979). It was shown that by making the filter
dict the effect of the manipulated variables on the output; time constant sufficiently large robustness can virtually
(2) filter to achieve a desired degree of robustness; (3) always be achieved. The use of alternate more complex
control algorithm to compute future values of the manip- filters should be explored in order to improve the control
ulated variable such that the process follows a desired quality.
trajectory closely. The IMC structure carries over to the multivariable case
This structure gives IMC a highly intuitive appeal: (1) in a straightforward fashion. The factorization of the
A model based prediction of the output is attractive in the transfer matrix into an invertible and a noninvertible part,
presence of output constraints. Any constraint violations the construction of approximate inverses, and the filter
can be anticipated and corrective action can be taken. (2) design are less trivial. Some light has been shed on these
Introducing a filter is equivalent to requiring the output questions in a recent report by Zames (1979).
to follow a reference trajectory. This interpretation gives
the filter constants physical meaning and therefore allows Acknowledgment
on-line tuning by the operating personnel. (3) The ob- Support of this research through NSF Grant ENG-
jective of the control law is transparent; input constraints 7906353 is gratefully acknowledged. Many of the ideas
can easily be incorporated without affecting the stability. appearing in this paper were stimulated through discus-
It was shown that for MP systems the adjustment of only sions with C. Brosilow. K. Astrom, P. Belanger, J. J.
one tuning parameter (the “system order” n) is sufficient Haydel, R. K. Mehra, and R. E. Otto provided many in-
to obtain a very good response. teresting comments during the Engineering Foundation
From a theoretical point of view IMC has the following Conference “Chemical Process Control 11”. Finally, we
attractive features: (1)The IMC structure allows affecting found the criticism by W. H. Ray and B. Ogunnaike here
the quality of the response and the robustness virtually at Wisconsin very useful.
independently through the design of the control block and
the filter, respectively. These objectives are hopelessly Appendix A
intertwined in most other design procedures. (2) The The following lemmas were used in the proofs of stability
stability of the closed loop system is not an issue, only the theorems 2,3, and 4 of 4.2. Their proofs are found in the
robustness. (3) The use of an impulse response model is book by Jury (1964, p 116).
advantageous because a structural model identification is Lemma 1. Monotonic Conditions. The real polyno-
not required and the nonminimal representation adds mial
robustness to the scheme. Furthermore, this model rep-
resentation is very suitable for a theoretical robustness P(x) = a# + afl-lxn-l+ ... + q x + a. (Al)
analysis and thus for the tuning of the filter.
We showed that many conventional control schemes has roots outside the unit circle if a. > a1 > ... > a , > 0.
have a structure identical with IMC. The design philos- Lemma 2. Dominant Coefficient. Consider the real
ophy, the computation of the inputs, and therefore the polynomial (Al). If the kth coefficient is such that
robustness are completely different, however.
We have not dealt in any depth with certain other as-
Iakl ’lad + 1%-11 + a** + b 1 1 + b’OI
pects of IMC because no or only preliminary results are then P ( r ) has k roots inside the unit circle, ( ( n- k) out-
available to date. Let us discuss them here and point out side).
some areas for future research along the way.
The control law formulation used in this work does not
allow one to account for output or state constraints directly
but only through penalty terms in the objective. Inspection
Appendix B
It can easily be verified that in the limit as P1 m , the
criterion (41) yields a control law with a root on the unit
-
of the output predicted by the model allows the operating circle. We show by means of an example that in general
personnel to anticipate violations and to change the pen- an NMP system cannot be stabilized by increasing P t
alty terms on-line to prevent them. An alternate more sufficiently.
attractive way is to add inequality constraints to the op- Let us assume an impulse response model with N = 2
timization problem (32) and to solve the resulting quad- ~ ( +k 1) = hlm(k) + hzm(k - 1) (h1,h2 # 0 )
ratic program on-line each time a new value for the input
has to be found. A gradient projection method has been and apply criterion (41) for yI = 1,P1 = P to compute the
developed for this purpose (Mehra et al., 1981), but no control law. For P = M = N = 2, minimization of (41)
results have been reported. If the absolute values of the yields a control system with the characteristic root
deviations of the output from the reference are minimized
instead of the squares, we obtain a linear program. This
might be more attractive for on-line computations.
The currently suggested identification methods for the
impulse response coefficients (e.g., Richalet et al., 1978) As a check, zl(0)= -hz/hl, the open-loop zero, and z l ( m )
are unsatisfactory in the presence of low signal/noise ratios. = 1, as expected. If Ih,/h,J 2 1the process is NMP and
It is probably altogether undesirable to identify an impulse thus the process inverse controller (0 = 0) is unstable. We
response model because the lack of parsimony is likely to would like to see how that unstable root reaches the lim-
give rise to large errors in the estimates. It might be better iting value of unity as p2 is increased.
 Ind. Eng. Chem. Process Des. Dev., Vol. 21, No. 2, 1982 323

observe its sign as p2 -

Hence, we are interested in calculating dzl(P2)/dp2and
m. We can easily perform the
differentiation and obtain
Horowltz, I . M.; Shaked, U. IEEE Trans. Autom. Control 1975, AC-20, 84.
Jury, E. I.“Theory and Application of the 2-transform Method”; R. E. Ktieger,
Ed.; Huntlngton: New York, 1964; p 116.
Kestenbaum, A.; Shinnar, R.; Thau, F. E. Ind. Eng. Chem. Process Des.
Dev. 1978, 15, 2.
d21(02) -
-- Koppei, L. B. “Introduction to Control Theory wlth Applications to Process
Control”; Prentice Hall: Englewood Cliffs, NJ, 1968; Chapter 8.
dP2 Kwakernaak, H.; Sivan, R. ”Linear Optlmal Control Systems”; Wiley-Intersci-
+ +
{(2h1 h 2 ) P 2h13p2+ h13(h12+ hlh2 h Z 2 ) ) + ence: New York, 1972; Chapter 6.
Lecrlque, M.; Rault, A.; Tessier, M.; Testud, J. L. “Multivariable Regulation of
(” h2)
+
+
(p4 (3h12+ 2hlh2 h2’)p2 hl4l2 + + a Thermal Power Plant Steam Generator”; Int. Fed. Autom. Control World
Congress, Helsinkl, 1978.
Lee, W.; Weekman, V. W., Jr. AIChE J. 1978, 22, 27.
It is straightforward to verify that for any MP system Mehra, R. K.; Eterno, J. S. “Model Algorithmic Control for Electric Power
(Ih2/hll< 1)this derivative is positive for any finite p2 and Plants”; IEEE Conf. on Dec. and Control, Albuquerque, NM, 1980.
Mehra, R. K.; Kessel, W. C.; Rault, A.; Richaiet, J.; Papon, J. “Alternatives for
hence, the controller cannot be made unstable by in- Linear Multivariable Control”; Saln, et ai. Ed.; NEC Inc.: Chicago, 1978; p
creasing /3 (as expected). Also, NMP systems with mon- 317.
Mehra, R. K.; Rouhani, R. “Theoretical Considerations on Model Algorithmic
otonic step responses can be stabilized using p. However, Control for Non-minimum Phase Systems”; Joint Automatic Control Con-
one can easily concoct examples where an unstable root ference Proceedings, San Francisco, CA, 1980.
approaches unity from above, e.g. for hl = -0.4, h2 = 0.7, Mehra, R. K.; Rouhani, R.; Eterno, J.; Rlchaiet, J.; Rault, A. ”Model Algorlthm
IC Control: Revlew and Recent Developments”; Eng. Foundation Conf. on
q(0) = and Chemlcal Process Control 11, Sea Island, GA, 1981.
Mehra, R. K.; Rouhanl, R.; Praly, L. ”New Theoretical Developments in Multi-
dz1(P2)
-- - variable Predictive Algorithmic Control”; Joint Automatic Control Confer-
ence Proceedings, San Francisco, CA, 1980.
dS2 Mehra, R. K.; Rouhani, R.; Rault, A.; Reid, J. G. “Model Algorithmic Control:
((-0.1)p4 + (-0.128)p2+ (-0.064)(0.37)] Theoretical Results on Robustness”; Joint Automatic Control Conference
Proceedings, Denver, CO, 1979.
(0.3) <O
(p4+ 0.41p2 + 0.0256)
Ogunnaike, 8. A.; Ray, W. H. AIChE J. 1979, 2 5 , 1043.
Palmor, 2 . J.; Shinnar, R. Ind. Eng. Chem. Process Des. Dev. 1979, 18, 8 .
Popov, V. M. Aufom. Remote Control 1983, 2 4 , 1.
for all finite values of p2. Prett, D. M.; Glllette, R. D. “Optlmlzatlon and Constrained Multivariable Control
of a Catalytic Cracking Unit”; AIChE 86th National Meetlng. April, 1979;
Literature Cited also In Jolnt Automatic Control Conference Proceedings, San Francisco,
Alevlsakis, G.: Seborg, D. E. Int. J. Confro/ 1073, 3 , 541. CA, 1980.
Arkun, Y.; Stephanopoulos, G. A I C M J. 1080, 26. 975. Reid, J. G.; Chaffin, D. E.; Silverthron, J. T. “Output Predlctive Algorithmlc
Astr6m, K. J. ”Introduction to Stochastic Control Theory”; Academic Press: Control: Predictlon Tracking with Application to Terrain Following”, Joint
New York, 1970 p 275. Automatic Control Conference Proceedings, San Francisco, CA, 1980.
Astr6m, K. J. Automatics 1080, 16, 313. Reid, J. G.; Mehra, R. K.; Kirkwood, E. “Robustness Properties of Output
Bohl, A. ti.; McAvoy, T. J. Ind. Eng. Chem. Process Des. Dev. 1078, 15, Predictive Dead-Beat Control: SISO Case”, Proc. IEEE Conf. on Dec.
24. and Control, Fort Lauderdale, FL, 1979, p 307.
Brosllow. C. 8. “The Structure arid Deslgn of Smith Predictors from the View- Richaiet, J. A.; Rault, A.; Testud, J. D.; Papon, J. Aotomatlca 1978, 14, 413.
point of Inferential Control”; Jolnt Automatic Control Conference Proceed- Safonov, M. G.; Athans, M. IEEE Trans. Autom. Control 1978, AC-23, 717.
Ings, Denver, CO, 1979. Smlth, 0 . J. M. Chem. Eng. Prog. 1957, 5 3 , 217.
Cutler, C. R.; Ramaker. B. L. “Dynamic Matrix Control-A Computer Control Zames, 0. “Feedback and Optimal Sensitivity: Model Reference Transforma-
Algorithm”; AIChE 86th National Meetlng, April, 1979; also in Joint Auto- tions, Multlpllcative Seminorms, and Approximate Inverses”; Report, Elec-
matic Control Conference Proceedings, San Francisco, CA, 1980. trical Engineering Department, McGIli University, Montreal, 1979.
Doyle. J. C. IEEE Trans. Autom. Control 1978, AC-23, 756.
Faddeeva, V. N. “Computational Methods of Linear Algebra”; Dover Pubiica- Received for review March 30, 1981
tions, Inc.: New York, 1959; p 202.
Foss, A. S. AIChE J . 1973, 19, 209. Revised manuscript received September 30, 1981
Fuller, A. T. Int. J. Control 1968, 8 , 145. Accepted November 30,1981