OPTIMIZATION-FREE CONSTRAINED

NONLINEAR PREDICTIVE CONTROL MINERAL PROCESSING APPLICATIONS

Jocelyn Bouchard ,1 Andr
e Desbiens ,1
Ren
e del Villar ,1

LOOP (Laboratoire dobservation et doptimisation des

procedes - Process observation and optimization
laboratory)

Department of Electrical and Computer Engineering

Department of Mining, Metallurgical and Materials

Engineering
Universite Laval, Pavillon Adrien-Pouliot
Quebec (Quebec), Canada G1K 7P4
Email: [email protected]

Abstract: A difficulty with constrained nonlinear control is the minimization of

the cost function. With complex system representations such as fundamental
models, the required optimization algorithm may be complex to implement,
setting its parameters may be difficult and the calculation time may be long.
To overcome these problems, an innovative optimization-free predictive control
scheme is proposed. The minimization of the cost function is replaced by a simple
and easy-to-compute simulation. Two mineral processing applications illustrate
c 2005 IFAC
the very good performances of this new algorithm.Copyright
Keywords: Nonlinear control, multivariable control, predictive control,
constraints, optimization.

1. INTRODUCTION
For many nonlinear processes, linear controllers
can perform quite adequately. However, nonlinear
control can be justified when the plant behavior is
highly nonlinear and subject to large and frequent
disturbances or when the operating points span
a wide range of nonlinear dynamics (Qin and
Badgwell, 1997). Furthermore, nonlinear predictive control has to be considered as a solution
if safety and actuator constraints exist, which is
always the case for real processes.

Partially supported by NSERC (Canada) and FQRNT

(Qu
ebec).

To describe the dynamic behavior of a process,

two possible ways are physical modeling and empirical modeling, both having attractive characteristics and drawbacks (Soderstrom and Stoica,
1988). Fundamental models are obtained analytically from basic physical laws while empirical
modeling is an experimental approach where the
parameters of an empirical mathematical relation
between the variables of interest are fitted the
recorded data. The main drawback of physical
models is that some processes are so complex that
it is almost impossible to explain their behavior
using only first principles. On the other hand,
empirical models are much easier to obtain and to
use but their parameters do not have any physical

meaning and a priori information is almost completely neglected. Furthermore, unlike fundamental models, they represent adequately the process
only for conditions (operating points, types of
inputs, etc.) similar to those found in the recorded
data. In fact, if the underlying assumptions of
the fundamental models are respected, they can
mimic behaviors outside the range of calibration
and less data are required for their development.
As a result, fundamental models have been used
for nonlinear model predictive control (MPC).
However, the considered plants are almost always
a single unit operation with a relatively simple
dynamic model (Henson, 1998). Writing fundamental models is a difficult task but commercial
dynamic simulators are now available. However,
using commercial dynamic simulators for nonlinear predictive control (NLPC) does not seem to
have been reported yet in the literature (Henson,
1998). The main reason for not using complex
fundamental nonlinear models for designing predictive controllers is certainly that the complexity
of the on-line solution of the nonlinear programming problem increases with the one of the model,
hence leading to computational and reliability difficulties. Another reason why commercial simulators are not used is probably the unavailability
of the model equations to the control designer
(Henson, 1998).
In recent years, many research works have focused
on the nominal stability problem for nonlinear
model predictive control. Most proposed solutions
consist in insuring nominal stability by imposing
penalties or constraints on the terminal state of
the prediction horizon (Qin and Badgwell, 1997;
Mayne et al., 2000). These solutions are usually
computationally quite demanding. Fortunately,
algorithms to reduce the computational effort are
now appearing in the literature (Fontes, 2001;
Magni et al., 2001), but they still remain relatively
difficult to implement. However, most nonlinear
model predictive controllers do not use terminal
state constraints of any kind (Qin and Badgwell,
1997). They instead allow to set the prediction
horizon long enough to go beyond the steady-state
hence approximating the infinite horizon solution,
which leads to nominal stability (Meadows et al.,
1995). Therefore, the proposed controller is presented in its simplest form, without relying on
terminal state constraints.
The proposed scheme transforms the optimization
of the cost function required to solve MPC problems into a control problem. Thus at each sampling time, instead of solving a complex nonlinear
programming problem (NLP) to obtain the control action, a simple closed-loop simulation of the
process with a pure integrator controller is conducted. The resulting optimization-by-simulation

(OBS) does not require a NLP solver and is therefore very easy to implement.
Two examples are given to illustrate the method
efficiency. The first one shows its use in a constrained multivariable case with an application
for the linear control of a grinding circuit. The
second one presents the nonlinear control of the
cooling zone of an induration furnace where a
phenomenological simulator is used as the process
model.
2. OPTIMIZATION-FREE CONSTRAINED
NONLINEAR PREDICTIVE CONTROL
2.1 Notation
The process inputs and outputs at time t = k
are respectively u(k) <n and y(k) <n (all
vectors in the paper are columns). The set points
are r(k) <n . The best plant model MN y , possibly based on phenomenological relationships and
therefore probably highly complex and nonlinear,
is described by
xN y (k + 1) = fy (xN y (k), u(k))

(1)

yN (k) = gy (xN y (k))

(2)

Other states or secondary outputs of the plant are

denoted w(k) <nw where nw n. They can be
predicted using the model MN w
xN w (k + 1) = fw (xN w (k), u(k))

(3)

wN (k) = gw (xN w (k))

(4)

If nw < n, zeros are added to the model to obtain

nw = n.
To represent the plant disturbances, the following
stochastic model MS is used
xS (k + 1) = AS xS (k) + BS (k)

(5)

yS (k) = CS xS (k) + DS (k)

(6)

where (k) <n is a zero mean random vector

and yS (k) <n . The model MS usually contains an integration to represent non-stationary
disturbances hence adding an integral action in
the proposed control scheme.
b : H) will refer
In the following, the notation S(h
to the vector of predictions of the signal s over a
future horizon k + h to k + H
b : H) = [ sbT (k + h/k) sbT (k + h + 1/k)
S(1
T (7)
...
sbT (k + H/k) ]
The vector U (0 : H 1) denotes the present and
future values of the plant inputs
U (0 : H 1) = [ uT (k) uT (k + 1)
T (8)
. . . uT (k + H 1) ]

Deterministic
predictor

^
R(1:H)
+

u(k)
Min J

^
E(1:H)
j

y(k)

Process

yN(k) +

GT

Y^N(1:H)j

MNy
-

r(k)

Y^S(1:H)

Stochastic
predictor

GR

^
e(k+H/k)j
K(z)

U(0:H-1)j
MNy

^
R(1:H)
- Y^S(1:H)

yS(k)

u(k)j

(1:H)
(H)

u(H)
U(0:H-1)

xNy(k)

Fig. 1. Control structure

Fig. 2. OBS - without constraints

2.2 Controller design

Step 6 : Update the state vectors xN y and xN w

with (1) and (3). Obviously, better state estimators could be used. Return to Step 1 at the next
sampling time (k = k + 1).

According to a receding horizon procedure, the

objective of the proposed method consists in minimizing at each sampling time the following cost
function
J = [ rb(k + H/k) ybN (k + H/k)
2
b
yS (k + H/k) ]

(9)

The minimization of the cost function is subject

to a unitary control horizon constraint
u(k + i) = u(k) for i = 1 to H 1

(10)

and to the following constraints on wN and u

cN (1 : H) < Wmax
Wmin < W

(11)

umin < u(k) < umax

(12)

To achieve the above objective, a new predictive control scheme is proposed. The control is
calculated by repeating the following procedure
at every sampling time. The control structure
(Figure 1) is similar to the GlobPC presented
by Desbiens et al. (2000) (with the simplification
that the tracking and regulation controllers are
identical).
Step 1 : Measure the process outputs y(k).
Step 2 : Estimate the process disturbance with
the IMC structure
yS (k) = y(k) yN (k)

(13)

where yN (k) is calculated with (2).

Step 3 : Compute the predictions (details are
given in (Desbiens et al., 2000)). Stochastic predictions Ybs (1 : H) are calculated using the model
MS and a regulation reference model GR . Deb : H) are based on a
terministic predictions R(1
tracking reference model GT . Note that GT and
GR both have a unitary gain and they respectively
modify the tracking and regulation performances.
Step 4 : Calculate the control vector action (U (0 :
H 1)) which minimizes the cost function (9)
with respect to the constraints (10), (11) and (12).
This can be accomplished by using the technique
described in Section 2.3.
Step 5 : Apply u(k) to the plant.

2.3 Optimization-by-simulation
To better understand the OBS approach, the
constraints (11) and (12) will first be neglected.
Knowing the actual states of MN y , an integrator
controller iteratively calculate different manipulated variables trying to bring ybN (k + H/k) equal
to rb(k + H/k) ybS (k + H/k), hence minimizing
the cost function (9). More precisely, every time
the cost function must be minimized (at every
sampling time k), the system depicted in Figure 2
is simulated until it converges. Each discrete step
in the simulation (denoted with the subscript j)
is equivalent to an optimization step for usual
optimization algorithms. The integrator controller
is

K
1
0
.
.
.
0

1 z 1

K2

0
...
0
(14)

1
K(z) =
1z

...
...
...
...

Kn
0
0
...
1 z 1
b : H) eb(H) extracts the predicThe block E(1
b : H)j , since
tions eb(k + H/k)j from its input E(1
only the former appears in the cost function and
therefore must be integrated. The block u(H)
U (0 : H 1) builds the complete vector U (0 : H
1)j from u(k)j by respecting the control horizon
constraint (10). At every simulation step j, the
states of the model MN y are reset to their actual
values xN y (k) before calculating YbN (1 : H)j for
the sequence of inputs U (0 : H 1)j using (2)
and (1); indeed the objective is to find the manipulated variables that bring ybN (k + H/k)j to the
set point from its actual (at time k) state xN y (k).
The value of u(k)j when steady-state is reached is
the result of the cost function minimization, i.e. it
corresponds to u(k) that must be applied to the
plant. For linear systems, it would be easy to show
that the solution is optimal since the integrator
controller insures that the outputs reach the set
points.

y2: passing 325 mesh (%)

w2: Overflow density (%)

^
R(1:H)
-Y^S(1:H)
+

u(k)j umax

wmax
+

+
-

+
wmin

u1: Rod mill feed

rate (t/h)

xNy(k)

(1:H)
(H)

MNy
u(H)
U(0:H-1)

K(z)
umin

u2: Water feed

rate (m3/h)

Rod Mill

Ball Mill
y1: Circulating
load (t/h)

Y^N(1:H)j
^ (1:H)
W
N
j

w1: Pump box level (%)

MNw
xNw(k)

Fig. 3. OBS - with constraints

In presence of constraints, the structure is modified accordingly to the pseudo-cascade approach
described by Lestage et al. (1999), as depicted in
Figure 3, where C is a diagonal proportional controller. A limiter ensures that the constraints (12)
are respected. A second limiter is added for the
cN (k +
inner set points, i.e. the set points for W
c
H/k)j , allowing WN (k + H/k)j to remain within
the constraints Wmin and Wmax . Momentary or
persistent constraint trespassing throughout the
prediction horizon may occur in some cases, such
as:
if predictive controller is tuned in such a
cN (k + H/k)j inner set points
way that W
overshooting cannot be avoided (e.g. when
H is much shorter than the system settling
time)(momentary);
if MN w exhibits an important left-half plane
zero (momentary);
if manipulated variable constraints are hit
(persistent).
Note that the inner feedback loop is not effective unless saturation occurs, thus ensuring that
YbN (k + H/k)j reaches its set point (unless of
course the manipulated variables hit their constraints).
The complete algorithm is the following:
(1) Reset the starting point for the research:
u(k)0 = u(k 1).
(2) Reset the optimization step counter: j = 1.
(3) Generate U (0 : H 1)j with u(H) U (0 :
H 1) from u(k)j1 .
(4) Reset MN y and MN w initial conditions:
xN y (k)j = xN y (k) and xN w (k)j = xN w (k).
(5) Calculate the corresponding predictions of
the models MN y and MN w , YbN (1 : H)j
c (1 : H)j , for the sequence of inputs
and W
U (0 : H 1)j using (1), (2), (3) and (4).
(6) Following the subsequent sequence of operations of the flowsheet, calculate u(k)j .
(7) Convergence test:
(a) If the simulation has reached steadystate, then u(k) = u(k)j . Exit and go
to step 5 of Section 2.2.
(b) Else, set j = j + 1 and go back to step
(3).

Fig. 4. Grinding circuit

In practice, the convergence test can be selected
as follows

ku(k)j u(k)j1 k2 < n

(15)
where > 0. The parameter is therefore the
desired precision for the elements of the solution,
equivalent to the termination tolerance for usual
optimization algorithms.

3. MINERAL PROCESSING APPLICATIONS

3.1 Grinding circuit - Constrained MPC
The purpose of this example is to illustrate how
the proposed technique may be used to control a
constrained multivariable system. The considered
process, a post-classification grinding circuit involving a rod mill and a ball mill, is depicted in
Figure 4.
The process outputs are:
y1 : the circulating load (material fed back to
the ball mill in t/h), denoted CL;
y2 : the size distribution of the overflow (i.e.
the quality of the product), described by the
% passing 325 mesh (47 m), denoted -325.
Even if it is generally not considered, the circulating load should always be controlled or constrained because mills efficiency decreases drastically outside the optimal range of throughput. In
the present case, the circulating load is estimated
with a material mass balance around the cyclone
using:

F xF = O xO + U xU

(16)

where is a pulp flow rate and x, a percent solid.

The subscripts F , O, U are respectively used for
the feed, overflow and underflow. F , O , xF and
xO are measured, therefore the circulating load
(U xU ) may be estimated from (16). Note that
the use of a steady-state mass balance is justified
because the cyclone dynamics are much faster
than the grinding mill dynamics.

 the rod mill feed rate (RMF) in t/h (u1 );

the water feed rate (WF) in m3 /h (u2 ).
The process is described by the following models
MN y and MN w (Lestage et al., 1999)

u2: WF [m3/h]

u1: RMF [t/h]

150
145
140

y2: 325 [%]

850

800
Outputs
Set points

750
80

w2: O/FD [%]

Only two process variables can be manipulated to

control the circuit:

155

y1: CL [t/h]

the pump box level (PBL): 25 % w1 75

% to avoid pumping problems and pump
box overflows;
the overflow density (O/FD): 47 % w2
53 % required for upstream processes and
to avoid underflow roping.

w1: PBL [%]

In order to insure a proper circuit operation, two

other process variables are included as constraints
in the control strategy:

60
40
20
0

Outputs
Constraints

time [s]

100

80

60
50

45

40
55

50

45
0

x 10

time [s]

3
4

x 10

Fig. 5. Constrained CL & size distribution control

because the process has an important left-half
plane zero.

yN (s)=

13,8
(1+5700s)(1+400s)

4,2(1700s)
(1+5000s)(1+5s)

0,2(1900s)e600s
(1+5200s)(1+750s)

0,012(1+39500s)
(1+4400s)(1+50s)

u(s) (17)

3.2 Induration furnace - Unconstrained NLPC

u(s) (18)

The second example shows how an optimizationfree predictive controller may be efficient when
based on a phenomenological simulator.

The same models are used for the plant simulation (no process-model mismatch). The sampling
period Ts is 200 s and the prediction horizon is
H = 12. The stochastic model is given by


1 0, 8z 1
I (k)
(19)
yS (k) =
1 z 1

The plant to be controlled is a pellet cooling phenomenological simulator described by Pomerleau

et al. (2003). It simulates the cooling zone of
an induration furnace used for the agglomeration
of iron ore oxide pellets. The two manipulated
variables are the shutter positions for two fans
forcing the air circulation through the moving bed
of pellets (one above and one below). The gas
temperature and pressure above the pellet bed are
the controlled variables. The simulator is based on
energy balance equations for the pellets and the
gas. The pressure drop in the bed is calculated
with the Erguns model. The fans characteristics,
the pressure drops in the shutters and the pressure loss in the outlet resistance are explained by
nonlinear empirical relationships.

5,749
(1+5500s)(1+210s)

1,962
(1+4700s)

0,0255(15600s)
(1+5300s)(1+750s)

0,14(1+4050s)
(1+3200s)(1+60s)

wN (s)=

with the following operating points:

u1op = 142 t/h;

u2op = 85 m3 /h;
y1op = 800 t/h;
y2op = 48 %;
w1op = 60 %;
w2op = 50 %.

No reference trajectory filters (GT and GR ) are

used and the OBS is set as follows
"
#
1
0, 1638
0
K(z) =
0
6, 7248
1 z 1
"
#
(20)
0, 0729
0
C =
0
2, 0407

= 0, 5E 4

Figure 5 shows an example of the grinding circuit

behavior under control. As expected, the circulating load and the size distribution respect their set
point when constrained variables are kept within
their upper and lower limits. Note that a slight
constraint trespassing occurs for w1 because of
the relatively short prediction horizon and for w2 ,

The nonlinear model MN y is identical to the plant.

The sampling period is 10 seconds. The horizon
is H = 25 and the stochastic model is (19).
Again, no reference models are used. Figure 6
compares the optimization-free controller (Optimfree) and a nonlinear predictive controller (Optimal) which directly minimizes (9) using lsqnonlin
from Matlab. Both approaches lead to similar
results but the optimization-free controller allows
a significant computation time reduction (about
four times faster) for similar termination tolerances.
4. CONCLUSION
A new constrained nonlinear predictive controller
is presented and simulation examples illustrate

75
74
73
72
71
0

50

100

76
74
72
70
68
0

50

100

1348

y2: Temperature [K]

y1: Pressure [Pa]

100

u2: Shutter #2 positon [%]

u1: Shutter #1 positon [%]

76

50

1347

1346

50

Optimal
Optimfree
Set point

100
150
0

50

Optimal
Optimfree
Set point

1345

100

1344
0

time [s]

50

100

time [s]

Fig. 6. Nonlinear cooling zone control

its very good performances in terms of efficiency,
optimality and computation time.
The novelty of the proposed scheme lies in the
computation of the control action at each sampling time period. The optimization of the cost
function is replaced by a simulation of the process
in closed-loop with an integral controller, using a
pseudo-cascade scheme.
The main advantages of using the proposed
optimization-free predictive controller are:
A nonlinear programming solver is not required.
A phenomenological simulator may easily be
used as the process model.
It is easy to tune: the only parameters to
select are the gains of the OBS controllers
and .
It converges rapidly.
With the optimization-free controller, the optimization problem is seen as a decentralized control
problem. This means that as well as the selection
of the pairing, the choice of the optimizer gains
(C and K) are issues that cannot be avoided
by further investigations. How the OBS structure
deals with constraints incompatibility and how
the solutions are affected in term of optimality
in such a case remain also to be studied. In other
words, neither the convergence conditions nor the
optimality of the solutions reached with the OBS
technique have been studied yet. A complete analysis is required to evaluate its properties and limitations. Nevertheless, the OBS approach could
become a useful and practical solution for constrained nonlinear multivariable predictive control
design.

REFERENCES
Plamondon.
A. Desbiens, D. Hodouin, and E.
Global predictive control: A unified control

structure for decoupling setpoint tracking, feedforward compensation and disturbance rejection dynamics. IEE Proceedings on Control
Theory and Applications, 147:465475, 2000.
F.A.C.C. Fontes. A general framework to design stabilizing nonlinear model predictive controllers. Systems & Control Letters, 42:127143,
2001.
M.A. Henson. Nonlinear model predictive control:
Current status and future directions. Computers & Chemical Engineering, 23:187202, 1998.
R. Lestage, A. Pomerleau, and A. Desbiens. Improved constrained cascade control for parallel
processes. Control Engineering Practice, 7:969
974, 1999.
L. Magni, G. De Nicolao, and L. Magnani. A
stabilizing model-based predictive control algorithm for nonlinear systems. Automatica, 37:
13511362, 2001.
D.Q. Mayne, J.B. Rawlings, and C.V. Rao. Constrained model predictive control: Stability and
optimality. Automatica, 36:789814, 2000.
E.S. Meadows, M.A. Henson, J.W. Eaton, and
J.B. Rawlings. Receding horizon control and
discontinuous state feedback stabilization. International Journal of Control, 62:12171229,
1995.
Poulin. PerforD. Pomerleau, D. Hodouin, and E.
mance analysis of a dynamic phenomenological
controller for a pellet cooling process. Journal
of Process Control, 13:137151, 2003.
S.J. Qin and T.A. Badgwell. An overview of nonlinear model predictive control applications. In
J.C. Kantor, C.E. Garcia, and B. Carnahan, editors, Fifth International Conference on Chemical Process Control, volume 93, pages 232256,
Englewood Cliffs, 1997. AIChE and CACHE.
T. Soderstrom and P. Stoica. System identification, chapter 1. Prentice Hall, New Jersey, 1988.