CONTROL OF INTEGRATING PROCESSES USING

DYNAMIC MATRIX CONTROL*

Y. P. GUPTA
Department of Chemical Engineering, Dalhousie University, Halifax, Canada
D
ynamic Matrix Control (DMC) has been popular for the control of chemical and
petroleum processes. These processes commonly include integrating process units,
which produce a ramp change in the output for a step change in input. When DMC is
used for control of integrating process units, a steady-state offset occurs for sustained load
changes. This offset is not acceptable for certain applications. A simple modication in the
DMC algorithm that eliminates this offset is presented in this paper. The performance of
the proposed algorithm is presented on SISO and MIMO example problems through
simulations and experimental results.
Keywords: level control; model predictive control; dynamic matrix control
INTRODUCTION
The integrating process units are commonly present in
chemical and petroleum processes that produce a ramp
change in the output for a step change in the input. This
results from the process units material or energy imbalance.
Non-self regulatory level processes are typical examples of
integrating process units. Control of these process units is
important for the control of the overall process. Levels are
usually controlled by adjusting an inlet or an outlet ow and
two distinct control objectives are involved. For some
applications the objective is to control the level tightly and
the variations in the manipulated variable are not important.
This objective is encountered in the control of level in
chemical reactors and otation columns. In other applica-
tions the control objective is to minimize the variations in
the manipulated variable while keeping the level within
specied limits. This objective is referred to as averaging
level control and is encountered in the control of level in
surge tanks and distillation columns. Several efcient
control algorithms (Marlin
1
, Campo and Morari
2
, Cheung
and Luyben
3
, Shunta and Feherari
4
) are available to meet
the above two objectives whenever these units can be
considered as SISO systems. However, real-time process
optimization usually requires the operation of the process at
the intersection of constraints. To accomplish this, a
constrained multivariable control scheme needs to be
implemented. The control of the outputs of integrating
process units needs to be considered along with the control
of other process outputs and all of the constraints need to be
considered simultaneously. Decomposition of the problem,
where control of the integrating process units is considered
separately, does not provide the optimal solution that is
otherwise possible. Model Predictive Control (MPC) is best
suited for control of constrained MIMO processes (Prett and
Garcia
5
). Cutler
6
suggested the use of Dynamic Matrix
Control (DMC) for the control of integrating process units.
However, when the DMC algorithm is used for control of
processes containing integrating process units, steady-state
offsets occur for sustained load changes. For applications
where the performance objective is to ensure that the
controlled variables remain very close to their set points,
these steady-state offsets are not acceptable. Lee et al.
7
and
Lundstrom et al.
8
have presented a state-space based MPC
algorithmthat facilitates the handling of integrating process
units, and addresses some other issues. The state-space
model is expressed in terms of step-response parameters for
systems of stable and/or integrating dynamics. A Riccati
equation needs to be solved in case of integrating process
units.
In this paper, an alternative approach to eliminate the
steady-state offsets that are encountered when dealing with
integrating process units is presented. The proposed
approach does not require the formulation of the MPC
problem in state-space form and is much simpler. Because
of this advantage, it can be implemented directly in the step-
response formulation of the DMC algorithm, which is
currently one of the most popular and widely used MPC
algorithms in the process industry. Many industrialists have
preferred the DMC formulation because it is simple,
intuitive and allows the formation of the prediction vector
in a natural way. The elimination of the steady-state offsets
is demonstrated by presenting the performance of the
proposed algorithm on SISO and MIMO example problems
through simulations and experimental results.
THE STEADY-STATE OFFSET
The steady-state offset that results from the use of the
DMC algorithm for sustained load changes can be
determined from its transfer function. A transfer function
of the DMC algorithm for open-loop stable processes is
given by Gupta
9
. For integrating process units, the predicted
response due to past inputs is a straight line (ramp) and a
different expression applies for the change in predicted
output due to past inputs. This changes the transfer function.
46 5
02 63 8 7 62 /98 /$1 0.00+0.00
q Institution of Chemical Engineers
Trans IChemE, Vol 7 6 , Part A, May 1 99 8
* Based on a paper presented at the 4 6th Canadian Chemical Engineering
Conference, Kingston, Ontario, Canada, October 19 9 6.
Therefore, the applicable transfer function of the DMC
algorithm is rst derived for a SISO case.
The control move for the DMC algorithm is given by
Cutler and Ramaker
1 0
:
D
u(k)
=
Ce
c
(k) (1)
where the C matrix is calculated from the dynamic matrix
(A) of the process. The size of the A matrix is N
p
N
m
.
To suppress the control moves, the diagonal elements of the
A
T
A matrix are increased by multiplying them by a move
suppression factor f. Let us say that the matrix C in equation
(1) includes the effect of the move suppression used. The rst
control move that is implemented can be expressed as:
D
u(k)
=
Np
i
=
1
c
i
e
c
(k
+
i) (2)
where c
i
represents the i
t h
element in the rst row of C
matrix and e
c
(k + i) represents the error at the i
t h
step
that needs to be compensated. The error, e
c
(k + i), is given
by,
e
c
(k + i) = Desired value of Adjusted predicted value
output at point of output at point
(k + i) (k + i) due to past control
moves
= y
SP
y
p
(k
+
i) (3 )
where,
y
p
(k
+
i)
=
y(k)
+
D
y(k
+
i) i
=
1
,
2
, ...,
N
p
(4)
y(k) represents the measured value of output at the current
control instant and
D
y(k+i ) represents the expected change
in output at i steps ahead due to past control moves.
D
y(k+i )
is the residual effect of past control moves that has not
already shown up in the measured value of output. To
account for unmeasured disturbances and model mismatch,
the predicted response in the DMC algorithm is adjusted by
an amount equal to the difference between the actual output
and the model prediction. Note that the expression for the
predicted response given above includes this adjustment.
Since the predicted response due to past inputs is a straight
line, its slope depends only upon the input implemented at
the previous control instant. The other past inputs have no
effect. Therefore,
D
y(k+i ) is give by,
D
y(k
+
i)
=
ia
1
u(k
-
1) i
=
1
,
2
, ...,
N
p
(5)
where u(k
-
1) represents the input at the previous control
instant as a deviation from the initial steady state, and a
1
is
the rst step response coefcient. If the input has not
changed over the last N control intervals,
D
y(k + i) equals
zero in the case of open-loop stable processes. Note that this
is not necessarily true for integrating units. By combining
equations (2 ) through (5 ),
D
u(k)
=
Np
i
=
1
c
i
[y
SP -
y(k)
-
ia
1
u(k
-
1)
]
(6)
=
Np
i
=
1
c
i
[e(k)
-
ia
1
u(k
-
1)
]
(7)
=
e(k)
Np
i
=
1
c
i -
a
1
u(k
-
1)
Np
i
=
1
ic
i
(8)
=
b
1
e(k)
-
a
1
b
2
u(k
-
1) (9)
where
b
1 =
Np
i
=
1
c
i
(1 0)
b
2 =
Np
i
=
1
ic
i
(1 1)
The control move can also be expressed as,
D
u(k)
;
u(k)
-
u(k
-
1) (1 2)
Substituting for
D
u(k) in equation (9 ) and by taking the z-
transform of both sides, gives the transfer function of the
DMC controller,
D(z)
;
u(z)/e(z)
=
b
1
/[1
+
(a
1
b
2 -
1)z
-
1
]
(1 3)
Now the steady-state offset can be calculated that results
in the case of sustained disturbances. Consider a liquid-level
control system where the inlet ow (F
i
) is the load variable
and the outlet ow (F
0
) is the manipulated variable. The
open-loop transfer functions for level (h) are given by
h(s)/F
0
(s)
;
G
p
(s)
= -
1/(As) (1 4)
h(s)/F
1
(s)
;
G
d
(s)
=
1/(As) (1 5)
where A is the cross-sectional area of the tank.
Assuming the dynamics of the measuring sensor and nal
control element to be negligible, for a zero-order hold, the
pulse transfer function, HG
p
(z), and the discrete-time
closed-loop response of the system to load changes, h(z),
are as follows (Stephanopoulos
1 1
):
HG
p
(z)
= -
(1/A)Tz
-
1
/(1
-
z
-
1
) (1 6)
h(z)
=
Z[G
d
(s)F
i
(s)
]
/[1
+
HG
p
(z)D(z)
]
(1 7)
For a unit step change in load,
Z[G
d
(s)F
i
(s)
]
=
Z[(1/A)t
]
=
(1/A)Tz
-
1
/(1
-
z
-
1
)
2
(1 8)
By substituting equations (1 3 ), (1 6 ) and (1 8 ) into equation
(1 7 ), then,
h(z)
=
(1/A)Tz
-
1
[1
+
(a
1
b
2 -
1)z
-
1
]
(1
-
z
-
1
)[(1
-
z
-
1
)(1
+
(a
1
b
2 -
1)z
-
1
)
-
(1/A)Tb
1
z
-
1
]
(1 9)
Using nal-value theorem, the steady-state offset for a unit
step change in load is given by,
offset
=
a
1
b
2
/b
1
(2 0)
where b
1
and b
2
are dened by equations (1 0) and (11 ). This
expression for the steady-state offset was checked against
closed-loop simulation results obtained on the system in
Example Problem 3 . Several runs were conducted using
different values of the move suppression factor. In each
case, the observed offset was the same as given by equation
(2 0).
PROPOSED CONTROL ALGORITHM
The adjusted predicted response of outputs due to past
control moves is needed for calculating the errors to be
compensated at any control instant. For a SISO case, this is
shown in equation (3). In the proposed algorithm, the
4 6 6 GUPTA
Trans IChemE, Vol 76, Part A, May 1998
adjusted predicted response of outputs that exhibit a ramp
for a step change in input is calculated in a different manner.
The rest of the procedure for control calculation remains the
same. Advantage is taken of the fact that the predicted
response due to past inputs is a straight line passing through
the output at current control instant. The slope of the
predicted response is determined from the slope of the
output trajectory between the current and the previous
control instants. Note that this slope includes the effect of
unmeasured disturbances and any model mismatch that may
be present. The adjusted predicted response y
p
(k + i) at i
steps ahead is thus calculated by using the following
equation.
y
p
(k
+
i)
=
y(k)
+
i[y(k)
-
y(k
-
1)
]
i
=
1
,
2
, ...,
N
p
(2 1)
Note that the above calculation is affected by the noise on
the output signal. The effect of noise and its reduction is
considered in Example Problem 1 . The adjusted predicted
response of outputs that are open-loop stable is determined
in the usual way. Let us consider a three-inputthree-output
case where the rst output exhibits a ramp for a step change
in input and the other two outputs are open-loop stable. In
this case, the adjusted predicted response of the rst output
due to past control moves will be calculated using equation
(2 1 ). For the last two outputs, the adjusted
predicted response will be calculated from the following
equation,
y
p
(k
+
i)
=
[y(k)
+
D
y(k
+
i)
]
i
=
1
,
2
, ...,
N
p
(2 2)
where,
D
y(k
+
i
=
D
A
i
+
1
D
u(k
-
1)
+
D
A
i
+
2
D
u(k
-
2)
+ ...
+
D
A
i
+
N
-
1
D
u(k
-
N
+
1) (2 3)
and
D
_
A
i
+
j ;
a
2
,
1
,
i
+
j -
a
2
,
1
,
j
a
2
,
2
,
i
+
k -
a
2
,
2
,
j
a
2
,
3
,
i
+
j -
a
2
,
3
,
j
a
3
,
1
,
i
+
j -
a
3
,
1
,
j
a
3
,
2
,
i
+
j -
a
3
,
2
,
j
a
3
,
3
,
i
+
j -
a
3
,
3
,
j
(2 4)
In equation (24 ), a
2 , 1 , j
refers to the change in the 2nd output
due to a unit step change in 1st input occurring j steps ahead.
a
3 , 2 , j
refers to the change in the 3rd output due to a unit step
change in 2nd input occurring j steps ahead, and so on. Note
that whenever j > N, a
2 , 1 , j
equals a
2 , 1 ,N
, a
3 ,2 , j
equals a
3 ,2 , N
and so on, because the open-loop response is assumed to
settle in N control intervals.
Now, let us derive the transfer function for the proposed
algorithm for a SISO case. By substituting equation (21 )
into equation (3 ), the error e
c
(k + i) is given by,
e
c
(k
+
i)
=
y
SP -
y(k)
-
i[y(k)
-
y(k
-
1)
]
i
=
1
,
2
, ...,
N
p
(2 5)
=
e(k)
-
i[y(k)
-
y(k
-
1)
]
i
=
1
,
2
, ...,
N
p
(2 6)
By adding and subtracting y
S P
, the above equation can be
written as,
e
c
(k
+
i)
=
e(k)
-
i[y
SP -
y(k
-
1)
-
(y
SP -
y(k))
]
(2 7)
=
e(k)
-
i[e(k
-
1)
-
e(k)
]
(2 8)
By substituting equation (2 5 ) into equation (2 ),
D
u(k)
=
Np
i
=
1
c
i
[e(k)
-
i(e(k
-
1)
-
e(k))
]
(2 9)
Combining equations (1 0), (1 1 ), (1 2 ) and (29 ) yields,
u(k)
-
u(k
-
1)
=
b
1
e(k)
-
b
2
[e(k
-
1)
-
e(k)
]
(3 0)
By taking z - transform of both sides, the transfer function of
the proposed algorithm is:
D(z)
;
u(z)/e(z)
=
(b
1 +
b
2
)
-
b
2
z
-
1
(1
-
z
-
1
(3 1)
Therefore, it can be seen that in SISO cases, the proposed
modication reduces the DMC controller to a PI controller
and thus eliminates the offset. For MIMO situations, the
proposed modication is made within the MPC framework,
as shown above for the three-inputthree-output case.
Therefore, all of the advantages that the MPC framework
offers in dealing with a constrained multivariable optimiza-
tion problem, e.g., considering all constraints collectively,
the ability to use the desired weights on different outputs in
the objective function, the handling of process interactions,
are preserved.
CONTROL PERFORMANCE
The performance of the proposed control algorithm is
presented on three example problems. The load changes
were considered to be `unmeasured, that is, their effect was
not fed forward in the control calculations. In all of the
gures, the changes considered occur at time equal to zero.
The negative time region shows the initial steady-state
values of the controlled and manipulated variables. The
results obtained are compared with the results from the
regular DMC algorithm. The values of the tuning para-
meters (move suppression, control horizon, prediction
horizon, weights on outputs) were the same in the two
algorithms. These values and other details are given in the
Appendix.
Example Problem 1
This example considers an isothermal CSTR where a
rst-order chemical reaction A
!
B takes place. The total
mass balance and the component mass balance are described
by the following differential equations.
dh/dt
=
(F
i -
F)/(Area) (3 2)
dC
A
/dt
=
F
i
(C
Af -
C
A
)/V
-
kC
A
(3 3)
A two-input-two-output system is considered where the
level in the reactor and concentration of component A are
controlled by manipulating outlet ow and inlet feed
composition. The simulation results obtained for step
changes in set points are shown in Figure 1 . Since the
responses of the two algorithms are essentially the same, the
response curves lie on top of each other and therefore cannot
be distinguished in the gure. This comparison shows that
the proposed modication does not change the performance
of the DMC algorithm for set point changes. The simulation
results obtained for a step increase of 0.2 litre/second in feed
ow to the reactor are shown in Figure 2 . In this case, the
DMC algorithm results in a steady-state offset not only in
4 6 7 CONTROL OF INTEGRATING PROCESSES USING DYNAMIC MATRIX CONTROL
Trans IChemE, Vol 76, Part A, May 1998
level but in concentration also. The proposed algorithm
brings both of the outputs back to their set points.
The effect of noise on the output signal can now be
considered. It is assumed that the noise is normally distributed
and that approximately 99.75% of the measurements are
within 61 % of the true value of the output. To simulate the
effect of the noise, normally distributed random numbers
with a mean of zero and standard deviation (r) of 0.3 3 %
were generated and added to the outputs. These outputs
were then used in the control calculations. The results
obtained for a step increase of 0.2 litre/second in feed ow
are shown in Figure 3 . As expected, the proposed algorithm
was affected more than the DMC algorithm by the noise.
The effect of the noise can be reduced by using a analog
lter and/or by reading the outputs a number of times and
using the average value. The second approach is very easy
to implement through the data acquisition software. At any
control instant, the output channels can be read a number of
times in a very short time and an average value for each of
the outputs can be calculated. The averaging reduces the
standard deviation in the output. To study the effect of this
approach, the outputs were read in the following order,
y
1 ,
y
2 ,
y
1 ,
y
2
, and so on. Each of the two outputs was read 5 0
times and the averaged values for the outputs were used in
the control algorithms. The resulting responses for a step
increase of 0.2 litre/second in feed ow are shown in
Figure 4 and are fairly smooth. This simulation shows that
the adverse effect of the noise can be reduced.
Example Problem 2
This example considers a nonisothermal CSTR where an
exothermic rst-order chemical reaction A
!
B takes
place. The total mass balance, the component mass balance
and the energy balance are described by the following
differential equations.
dh/dt
=
(F
i -
F)/(Area) (3 4)
dC
A
/dt
=
F
i
(C
Af -
C
A
)/V
-
kC
A
(3 5)
4 6 8 GUPTA
Trans IChemE, Vol 76, Part A, May 1998
Figure 1. Comparison of control performance for step changes in set points.
Figure 2. Comparison of control performance for a step increase in feed ow.
Figure 3. Effect of noise on control performance.
dT/dt
=
F
i
(T
i -
T)/V
+
JkC
A -
Q/(qc
p
V) (3 6)
where J
=
(
-
D
H)/(qc
p
) (3 7)
k
=
k
0
e
-
E/RT
(3 8)
A three-input-three-output system is considered where
the level, concentration of component A and reactor
temperature are controlled by manipulating outlet ow,
inlet feed composition and the amount of heat removed by
the coolant. Again, normally distributed random numbers
with a mean of zero and standard deviation of 0.33 % were
generated and added to the outputs as noise. Each of the
three outputs was read 1 00 times and the averaged values
were used in the control algorithms. The simulation results
obtained for a step increase of 0.1 litre/second in feed ow
to the reactor are shown in Figure 5. The DMC algorithm
again results in a steady-state offset not only in the level but
also in the temperature. The offsets resulted even when
noise was not added. The proposed algorithm brings all of
the outputs back to their set points. The simulation results
were also checked for step changes in set points. Again, the
two algorithms performed in essentially the same way.
Example Problem 3
This example problem considers the control of water
level in an actual column, 1 0 cm in diameter and 2 .5 m in
height. The outlet ow from the bottom of the column was
manipulated by changing the speed of a pump. The column
was interfaced with an IBM type personal computer where
the control algorithms were executed. To reduce the effect
of noise, the output was read 5 0 times at each control
calculation and the averaged value of the output was used in
the control algorithms. The experimental results obtained
for a step decrease of about 2 litres/minute in feed to the
column are shown in Figure 6 . With the DMC algorithm,
the level continuously drops for this sustained disturbance
and results in a steady-state offset of 1 1.5 cm. With the
proposed algorithm, the level is controlled and brought back
to its set point. Step changes in set point were also
considered. Both of the algorithms took the level to the new
set points and their transient responses were close to each
other.
OTHER COMMENTS
(a) To avoid the steady-state offsets that result from the
use of the DMC algorithm, the control of the outputs of
integrating process units can be considered separately from
the control of other outputs. However, this decomposition
would restrict the determination of the true optimum of the
optimization problem that needs to be solved at every
control instant. The proposed algorithm allows the con-
sideration of all inputs, outputs and constraints in one
optimization problem.
4 6 9 CONTROL OF INTEGRATING PROCESSES USING DYNAMIC MATRIX CONTROL
Trans IChemE, Vol 76, Part A, May 1998
Figure 4. Reduction in the effect of noise on control performance.
Figure 5. Comparison of control performance for a step increase in feed ow.
(b) The constraints are not considered in this study
because the proposed modication only affects the predic-
tion part of the algorithm. The optimization part remains the
same.
(c) The comparison made in the paper shows that when
the regular DMC algorithm is used in MIMO situations, the
offset occurs not only in the liquid-level but also in other
outputs. The comparison also shows that the proposed
modication does not alter the performance of the DMC for
set point changes.
CONCLUSIONS
The proposed modication is very simple and can be
implemented directly in the step-response formulation used
in the DMC algorithm. For processes containing integrating
process units, the modication eliminates the steady-state
offsets that are otherwise encountered for sustained load
changes.
APPENDIX: PARAMETERS
Example 1:
Area =0.05 m
2
k = 0.2 5 s
1
N= 100 N
p
=1 00 N
m
= 2
f = 1 .00001 T =6 s
w
1
=1 w
2
= 1
Example 2:
Area =0.05 m
2
D
H= 50 000 000 q = 1 000 000 g/m
3
cal/kmole
c
p
=1 cal/(g 8K) E/R =1 000 8K k
0
=4 .8 72 9 9 75 8 4 s
1
N= 30 N
p
=3 0 N
m
= 2
f = 1 .005 T =3 s
w
1
=1 w
2
=2 .5 w
3
=1 /4 00
Heat removal is plotted as Q/(q c
p
)
Example 3:
Area =0.0081 m
2
N= 25 N
p
=2 5 N
m
= 2
f = 1 .3 T =2 s
NOMENCLATURE
A area of cross-section
a
1
change in the output, one step ahead, due to a unit step change
in the input
b
1
, b
2
dened by equations (1 0) and (1 1)
c
i
i
th
element in the rst row of C matrix
e(k) difference in the set point and measured value of output
e
c
(k+i) error in the output at i steps ahead that needs to be compensated
f move suppression factor
h level
k rate constant
(k) denotes current control instant
N number of control intervals in which the open-loop response
settles
N
p
number of points on each output trajectory at which the error is
minimized
N
m
number of control moves into the future including the current
one
T control interval
u(k) manipulated variable
w
i
weight on the i
th
controlled variable
y(k) measured value of the controlled (output) variable at current
control instant
y
p
(k + i) adjusted predicted value of output at point (k+i ) due to past
inputs
y
SP
set point of the controlled variable
Dy(k+i) change in output at i steps ahead from current value due to
past control moves
boldface indicates a vector or matrix
REFERENCES
1. Marlin, T. E., 19 9 5, Process Control, (McGraw-Hill, New York),
p. 5 81 59 9.
2. Campo, P. and Morari, M.,1 98 9 , Model predictive optimal averaging
level control, AIChE J, 3 5: 5 79 59 1.
3. Cheung, T. F. and Luyben, W., 1 97 9 , Liquid-level control in single
tanks and cascades of tanks with proportional-only and proportional-
integral feedback controllers, IEC Fund, 1 8: 1 5 2 1.
4. Shunta, J. and Feherari, W., 1 97 6 , Non-linear control of liquid level,
Instr Techn, 4 3 48 .
5. Prett, D. M. and Garcia, C. E., 1 9 88 , Fundamental Process Control,
(Butterworths-Heinemann, Boston, MA), p. 3.
6. Cutler, C. R., 1 98 2 , Dynamic matrix control of imbalanced systems,
ISA Trans, 2 1 : 1 6 .
7. Lee, J. H., Morari, M., Garcia, C. E., 1 99 4 , State-space interpretation of
model predictive control, Automatica, 3 0: 707 7 17 .
8. Lundstrom, P., Lee, J. H., Morari, M. and Skogestad, S., 19 9 5,
Limitations of dynamic matrix control, Comput Chem Eng, 1 9: 4 09
4 21 .
9. Gupta, Y. P., 1 99 3 , Characteristic equations and robust stability of a
simplied predictive control algorithm, Can J Chem Eng, 71 : 6 17
6 24 .
1 0. Cutler, C. R. and Ramaker, B. L., 1 98 0, Dynamic matrix controlA
computer control algorithm, Proc JACC, Paper WP5 -B.
1 1 . Stephanopoulos, G., 19 8 4, Chemical Process Control, (Prentice-Hall,
Englewood, NJ), p. 6 17 62 3.
ACKNOWLEDGEMENTS
The nancial support provided by the Natural Sciences and Engineering
Research Council of Canada is gratefully acknowledged The help of Ms M.
Bhagi in obtaining the experimental results is appreciated.
ADDRESS
Correspondence concerning this paper should be addressed to Professor
Y. P. Gupta, Department of Chemical Engineering, DalTech, Dalhousie
University, Halifax, Nova Scotia, Canada B3 J 2 X4 (E-mail: yash.gupta@
dal.ca).
The manuscript was received 3 November 1997 and accepted for
publication after revision 17 April 1998.
4 7 0 GUPTA
Trans IChemE, Vol 76, Part A, May 1998
Figure 6. Comparison of control performance for a step decrease in feed ow.