A PRACTICAL PREDICTIVE CONTROL ALGORITHM

FOR A LARGE CLASS OF PROCESSES

Vasile Crtoaje, Sanda Frncu, Gabriel Rdulescu
Department of Automation & Computers
Oil &Gas University of Ploiesti, Romania
E-mail:[email protected]
Abstract. In the last few years, the model based process control (MBPC) methodology was mainly
situated at the upper control level (optimization level, product level or economical level), which requires
more advanced control techniques. Many laboratories and industrial applications have demonstrated that
MBPC has a well performance/cost ratio and it is not difficult to be implemented by means of digital
equipment. Moreover, even non-specialists in control field easily understand the basic principles of MBPC.
As compared to the classical PID strategy, the MBPC strategy uses an explicit process model to predict the
controlled process output. However, due to the computational complexity and the requirement to own a
suitable process model, the PID type algorithms continue to remain the most widely used in lower level
control loops.
In this paper, we will try to prove that it is possible to obtain well control performances using
some standard unconstrained predictive algorithms (like those PID), characterized by fewer calculus and
robustness relating to the process model accuracy. In our control strategy, we propose for process model a
second order continuous-time transfer function with dead time and having the time constants allocation
coefficient in the range 0.6 - 1.0. At first, the user must specify three process parameters: the proportional
constant, the main time constant, and the dead time. During the operation, the user may modify, like in PID
strategy, three tuning parameters: steady state gain adjustment parameter, time constant adjustment
parameter, and dead time adjustment parameter. We show how each of these parameters modifies the
system dynamic performances. The increasing of the steady state gain adjustment parameter leads usually
to a stable system. In order to ensure the stability and good performances of the system, the prediction
horizon should completely overlap the transient time and the number of the sampling periods on the
prediction horizon should be limited at maximum 20.
Key words: MBPC, predictive control, robust control, standard control algorithm, real time
execution, sampling time.
1. Introduction
Model Based Predictive Control represents
a receding horizon control strategy based
on the explicit on-line use of a suitable
process model to predict the effect of
potential control action upon the plant
future state over a long-range time horizon
[1]. At each sampling instant, the updated
plant information is used to solve an open-
loop optimal control problem, but only the
first element of the optimal control vector
is actually applied to the real process. All
other elements of the optimal control
vector can be either not calculated or
forgotten because at the next sampling
instant all calculus-sequences are
performed again based on the new
output measurement (receding horizon
principle). The control vector or only its
first element are calculated in order to
minimize a specified cost function [8],
depending on the future postulated values
of the control variable and the predicted
control errors.
The availability of a suitable process
model and the necessity to optimize the
control objective in real time (at each
sampling instant) are the most stringent
requirements for the use of MBPC policy.
Due to the mathematical convenience, in
this paper we use the optimization
quadratic criterion, which presents the
advantages of an intuitive and reliable
analytical solution. Instead of an accurate
but less robust optimal algorithm, we will
use an algorithm based on analytical
solution [2], which allows the quick
computation of suboptimal control actions.
2. The proposed MBPC algorithm
The proposed algorithm consists of two
parts: one off-line computation part and
other on-line computation part.
Off-line computation
In the proposed control strategy, we consider
the process dynamic model as a second order
continuous-time transfer function with dead
time

) 1 )( 1 (
e
) (
2 1
+ +
=

s s
s H
T T
K
s t
. (1)
This off-line computation part consists in
finding the following elements: the process
model coefficients, the sampling period,
the prediction horizon, and the unit step
response of the process. The initial data
are the following: K process steady
state gain (proportionality constant),
T
0
process dominant time constant,
t process dead time, c - time constants
allocation coefficient, T sampling
period. The time constant T
0
can be
approximated by the process time
constants sum or by the ratio (T
tr
t)/3,
where T
tr
is the process transient time (for
step response and 5 percent limits).
The sampling time is chosen so that the
integer value of the ratio (t+3T
0
)/T should
be in the range 15 20. This integer
represents the prediction horizon and it is
denoted by N.
The time constants of the process model
are given by the relations:
T
1
=(10.5c)X
t
T
0
, T
2
=0.5cX
t
T
0
. (2)
In these relations, X
t
is the adjustment
coefficient of T
0
(default value equal to 1).
We define the integer dead time m as being
equal to the integer value of the ratio
] [

X
d
t
, where X
d
is the adjustment
coefficient of t (default value equal to 1).
When the user modifies t. he must observe
if the integer m is also modified.
Introducing the coefficients
1
/
1
e
T T
c

= ,
2
/
2
e
T T
c

= , a
1
=(c
1
+c
2
) ,
a
2
=c
1
c
2
, b = X
k
K(1+a
1
+a
2
) , (3)
the process discrete model has the form
y
t
+ a
1
y
t1
+ a
2
y
t2
= bu
tm
.

(4)
In (3), the parameter X
k
is the adjustment
coefficient of K and, in order to ensure the
system stability, it must be greater than 1
(default value equal to 1.5).
The unit step response g of the process can
be computed by the recursive equations:

> =
= =
> + =
= =
+
+ +
0 ,
1 , 1 , 0
0 ,
,
2 1 1 2
0 1 1 0
i h g
m i g
i h a h a b h
h a b h b h
i m i
i
i i i
(5)
The last off-line computation is the
expression
2 2
1
2
...
N m m
g g g E + + + =
+
. (6)
When the operator changes one of these
three adjustment parameters, again the off-
line computation set must be done.
On-line computation
One linear form of the predictive control
law is the following

=
+
+ =
N
m i
i t i
p
t t
e g u u
E
X
1

(7)
where t is the current discrete time, X
p
- the
control weighting coefficient, E - the
expression (6), g - the process unit step
response, and e - the predicted error
e
t+i
=w
t+i
y
t+i .
(8)
In (8), w denotes the setpoint and y denotes
the predicted free response of the process
(fig. 1). The free response y(t) represents
the effect of the past control [8], because
the input remains constant for the future
time, i.e. U
t
=U
t1
, U
t+1
= U
t1
,
The prediction can be achieved by means
of the process model (4), using the
previous measured values Y
t
, Y
t1
and Y
t2
,
as well as the previous values of the input
U
t1
, U
t2
, , U
tm
.
The free response can be calculated by the
recursive equations:
y
t2
= Y
t2
, y
t1
= Y
t1
,

y
t
= Y
t
y
t+k
=

y
t+k1
a
1
(y
t+k1
y
t+k2
)



(9)
a
2
(y
t+k2
y
t+k3
)

+

bz
t+k
, k >1
where

>
<
=
+ +
+
m k
m k u u
z
m k t m k t
k t
, 0
,
1
(10)

Figure 1. The predicted free response y
t+i
=f(Y
t2
,Y
t1
,Y
t
; U
t1
,
t2
, , U
tm
)
The weighting coefficient X
p
must be less
than 1 (its default value is 0.6) and it is
used to filter the control variable u.
The control law (7) can be obtained by
minimizing the quadratic objective
function
2
) (
i t
N
m i
i t
y w F
+
=
+
=

, (11)
where y represents the predicted response
of the system to the input u having a single
postulated value U over all prediction
horizon N (fig. 2). If there are two
postulated values for u, one
1
U for the first
interval [t, t+1] and the other one
2
U for
the interval [t+1, t+N], then the predictive
control law has the following form:

=
+
+ =
N
m i
i t i i
p
t t
e ag g u u
E
X
) (
1 1
1
(12)
where
E
1
=EaF
1
,

=
=
N
m i
i
g E
2
, (13)

=
+
=
1
1 1
N
m i
i i
g g F ,
2
N
g
a
E
F

= . (14)
Furthermore, considering that u has the
value
'
1
U for the interval [t, t+2] and the
value
'
2
U for the remaining interval
[t+2, t+N], then the predictive control law
becomes

=
+
+ =
N
m i
i t i i
p
t t
e ag g u u
E
X
) (
2
2
1
(15)
where
E
2
=EaF
2
,

=
=
N
m i
i
g E
2
, (16)

=
+
=
2
2 2
N
m i
i i
g g F ,
2 2
1
2
N N
g g
a
E
F

. (17)

Figure 2. The predicted response y to the input having one or two future values
3. The simulation results
The previous predictive algorithms have
been implemented in a real time
multitasking application. We have
considered that the process has the model
) 1 50 )( 1 60 )( 1 90 (
)e .2(1 1
) (
90
4
+ + +

s s s
s
s H
s
T
.
We have chosen the starting data:
K

=1.2; T
0
=200 s; t =90 s; c=0.6; T=40 s
and
X
k
=1.5; X
t
=1.0; X
d
=1.0; X
p
=0.6 .
From these data, we obtain
] [

X
d
m
t
= =2 , 17
3
0
=

+
=
t
T
T
N .
In figures 310, the simulation results for
the minimum phase process (T
4
=0) and the
control law (7) are shown.
Because the system model does not
describe exactly the real system dynamics,
the controlled system can be unstable for
small values of X
k
. Increasing X
k
, the
system becomes usually stable, but slower
(fig. 3 and fig. 4). The introduction of the
reference trajectory has not an important
impact on the loop dynamics (fig. 5). The
system may become unstable for both large
and small values of the X
d
and X
t
respectively (figures 6, 7, 8). The change
of the time constants allocation coefficient
c does not reveal major effects upon the
system response form (fig. 9). However, it
can be observed that the best response is
obtained for maximum value of c. In fig.
10 is presented the effect of the control
weighting constant X
p
upon the system
response. For a small X
p
, the response is
too slow; it becomes faster as X
p
increases
to 1. In consequence, for many processes
X
p
can be fixed in advance at a value in the
range 0.5 - 0.8.
In the figure 11 is shown that for T
4
=60 s
(non-minimum phase process), the
predictive control system may be stabilized
using a large value of X
k
(fig. 11) .
4. Conclusions
In this paper, we have proved that it
is possible to obtain well control
performances using some standard
unconstrained predictive algorithms (like
those PID), characterized by fewer
calculus and robustness relating to the
process model accuracy.
We have shown that a second order
transfer function with dead time and
having the time constants allocation
coefficient c in the range 0.6 1.0 can be
successfully used to describe the process
dynamics.
The user must firstly specify three process
parameters: K
.
T
0
, t .
In order to get the right performances, the
user may modify during the operation time
only three tuning parameters: X
k
. X
t
, X
d
(which are initialized at the values 1.5, 1.0,
1.0 respectively). Usually, the control
weighting constant X
p
(initialized at a
value in the range 0.5- 0.8) is not necessary
to be adjusted during the operation time.
For both large and small values of the
tuning parameters X
d
and X
t
respectively,
the system may become unstable. But the
system may be usually stabilized by
increasing the steady state gain adjustment
parameter X
k
.
In order to ensure good performances it is
necessary that the prediction horizon
should completely overlap the transient
time and the number of the sampling
periods on the prediction horizon should
be chosen in the range 15 - 20.
References
[1] De Keyser R., A Model Based
Predictive Control Tutorial, In
Proceedings of the 12
th
International
Conference CSCS, Vol.1, pp. 1-12,
Bucharest, 1999.
[2] Boucher P., Dumur D., La commande
predictive, Editions Technip, Paris,
1996.
[3] Marlin T., Process Control Designing
Processes and Control Systems for
Dynamic Performance, McGraw-Hill,
Inc., pp. 604-630, 1995.
[4] Lazar C., Arabolu C., Model Predictive
Control with Amplitude Constraints,
In Proceedings of the 9
th
SIMSIS 96,
pp. 77-82, Galati, 1999.
[5] Richalet J., Industrial Applications of
Model Based Predictive Control,
Automatica, Vol. 29, No. 5, pp. 1251-
1274, 1993.
[6] Tsang T., Clarke D, Generalised
Predictive Control with Input
Constraint, IEE Proceedings, Vol.
135, No. 6, pp. 451-460, 1988.
[7] Chwn H., Allgower F., Nonlinear
Model Predictive Control Scchemes
with Garanteed Stability, NATO
Advanced Study Institute Proceedings,
pp. 465-495, Antalya, Turkey, 1997.
[8] Camacho E, Bordons C., Model
Predictive Control, Advanced
Textbooks in Control and Signal
Processing, Springer-Verlag , London
Limited, 1999.