MPC tuning for systems with right half plane zeros

WINSTON. GARCA-GABN
Escuela de Ingeniera Elctrica
Universidad de Los Andes
Av. Tulio Febres Cordero, Mrida 5101
VENEZUELA
EDUARDO.F. CAMACHO
Departamento de Ingeniera de Sistemas y Automtica
Universidad de Sevilla
Camino de los Descubrimientos, Sevilla 41092
ESPAA
Abstract: - This paper demonstrates how instability problem will be present for a Model Predictive Control
(MPC) in the case of it is applied to non-minimum phase systems. Instability appears when the control horizon
has the same value that the prediction horizon and the control weight is zero, because MPC achieves its
performance by cancelling the plant zeros including the unstable zeros, which leads to a loss of internal
stability of the feedback system. It is demonstrated that the instability problem can be solved with an adequate
tuning of horizon values without using a big control weight as it is suggested in the literature. The results are
illustrated with a non-linear example.
Key-Words: - Model predictive control, Non-minimum phase systems, stability, tuning controller.
1 Introduction
Generalized Predictive Control was proposed by
Clarke, et al.[1] and has become one of the most
popular MPC methods both in industry and
academia. It has been successfully implemented in
many industrial applications, showing good
performance. The basic idea of MPC is to calculate
a sequence of future control signals in such a way
that it minimizes a multistage cost function defined
over a prediction horizon. The index to be optimized
is the expectation of a quadratic function measuring
the distance between the predictive systems output
and some predictive reference sequence over the
horizon plus a quadratic function measuring control
effort. In order to implement a MPC, a model of the
plant is used to predict the future plant outputs. This
prediction is based on past and current values of the
input and the output of the plant. The process model
plays, in consequence, a decisive role in the
controller performance, and thus it is desirable to
choose a model quite similar to the plant in order to
predict accurately.
A system is said to be a non-minimum phase process
if it has Right Half Plane Zeros (RHPZ) or in the
discrete case if at least one of the zeros of the
transfer function is located outside the unit circle.
These processes are common in industrial
applications and they are characterized by their
inverse response. The control engineers must beware
with this kind of process, because they are an
important source of problems in practical
applications. It is well known that non-minimum
phase systems present difficulty in applying control
strategies, because they have an initial inverse
response to step input in the opposite direction from
the steady state, [2]. The presence of unstable zero
in a process transfer function is thus identified as
being responsible for its difficult dynamic behavior;
it is also the source of a considerable amount of
difficulty in controller design. Another aspect of
controlling a process with unstable zero is the
instability problem, which arises in order to achieve
high performance when the controller contains an
inverse of the process model [3].
The instability problems applying MPC to Single
Input Single Output (SISO) systems with RHPZ
have been reported in literature.[4,5,6,7,8,9] showed
that non-minimum phase systems produce instability
when prediction horizon and control horizon are
equal to one. Also an unstable behaviour in SISO
MPC has been reported by [10,11] when it is applied
to a non minimum phase systems. This can be
solved using a control weight parameter
[1,6,7,12,13].
This paper demonstrated how instability problem
will be present for whatever non-minimum phase
systems, when the control horizon has same value
that the prediction horizon and control weight is
zero, because MPC achieves its performance by
canceling the plant zeros, including the unstable
zeros, which leads to a loss of internal stability of
the feedback system. It is demonstrated that the
instability problem can be solved with an adequate
tuning of horizon values without use a big control
weight as it is suggested in the literature.
This article is organized as follows: Section 2 gives
a brief review of formulation of MPC, Section 3
shows the procedure used to obtain MPC closed
loop relationships. Section 4 shows the application
of this controller for linear and non-linear non-
minimum phase systems. Finally, the conclusions
are presented.
2 Model Predictive Control
Most of the SISO plants, when considering
operation around a particular set-point and after
linearization, can be described by:
( ) ( )
( )
t t t
z C
u z B y z A

1
1
1 1
(1)
Where:
t
y
: Output signal process
t
u
: Input signal process
t

: Zero mean white noise

A, B, C are the following polynomials in the
backward shift operator z
-1

A(z
-1
) = 1 + a
1
z
-1
+ + a
n
z
-na

B(z
-1
) = b
0
+ b
1
z
-1
+ + b
n
z
-nb

C(z
-1
) = 1 + c
1
z
-1
+ + c
n
z
-nc

This model is known as the CARIMA Model
(Controller Auto-Regressive Integrated Moving-
Average). It has been argued that for many industrial
applications in which disturbances are non-
stationary an integrated CARIMA model is more
appropriate [14].
The MPC algorithm consists of applying a control
sequence that minimizes a multistage cost function
(2). This minimization produces u(t), u(t+1) ,,
u(t+Nu), but only u(t) is applied. At time, t+1 a
new minimization problem is solved. This
implementation is called the Receding Horizon
controller.

( ) [ ]
[ ]

+ +

+
u
N
j
j t
N
N j
j t j t
u
r y t u J
1
2
1
2
2
1
,

(2)
Subject to ut+j=0, j=Nu, ..., N2
Where:
N
1
: Minimum prediction horizon
N
2
: Maximum prediction horizon
N
u
: Control horizon
:Control weight
r : Reference trajectory
In order to solve the problem posed by the
minimization of (2), the y(t+1) has been computed.
The j-step ahead output for j = N
1
...N
2
based on the
information known at time t and the future values of
the control increments. The following Diophantine
equation is considered,
( ) ( ) ( ) ( )
1 1 1 1
+ z F z z A z E z C
j
j
j
(3)
The polynomials E
j
and F
j
are uniquely defined with
degrees j-1 and na respectively.
Combining the plant model (1), and Diophantine
equation (3), the follow prediction output equation
can be obtained,
1

+ +
+
j t
j
t
j
j t
u
C
B E
y
C
F
y
(4)
In this expression
j t
y
+

is a function of a known
signal values at time t and of future control inputs,
which have not been computed yet. To distinguish
past and future control values a second Diophantine
equation (5) has been used.
( ) ( ) ( ) ( )
( )
1
1 1 1 1

+
z z
z C z G z B z E
j
j
j j
(5)
The following expression of the prediction was
obtained.
f
t j
f
t j j t j j t
y F u u G y + +
+ + 1 1

(6)
Where
f
t
u
and
f
t
y
are
( )
( )
t
f
t
t
f
t
y z C y
u z C u
1 1
1 1

(7)
(8)
Finally, (6) can be rewritten as:
( )
t j t j t j j t
y u z G y
+ +

+
+
1
1
(9)
Where
t j t
y
+

is the free response prediction of

j t
y
+

assuming that future control increments after time t-
1 will be zero,
( ) ( )
f
t j
f
t j t j t
y z F u z y
1
1
1

+
+
(10)
Substituting Ej(q-1) of (3) into (5), this yields
( )
( )
1
1 1

+ +
C A BF z
C z G A B
j
j
j
j
j
(11)
Define the vector f, composed of the free response
predictions,
[ ]
T
t N t t t t t
y y y f

2 1
2
, , ,
+ + +

(12)
the vector of future control increments,
[ ]
T
N t t t
u
u u u u

1 1
, , ,
~
+ +

(13)
and the vector of the predicted plant outputs,
[ ]
T
N t t t
y y y y

2 1
2
, , ,
+ + +

(14)
From the prediction (10) the predicted input-output
relationship of the plant can be written as the vector
equation,
f u G y +
~

(15)
Where the matrix G is composed of the step
response parameters g
i
of the plant model.

1
1
1
1
1
1
1
1
]
1

u
u u
N N N N
N N
g g g
g g g
g g
g
G
2 2 2
2 1
0 2 1
0 1
0
0
0 0

(16)
The quadratic minimization of (2) becomes a direct
problem of linear algebra, assuming there are no
constraints on the control signal, which leads to:
) ( ) (
~ 1
f r G I G G u
T T
+

(17)
3 Closed loop relationships
Closed loop relationships can be obtained for the
MPC in order to show how the tuning parameters
N1, N2, might affect the stability of the controlled
plant. MPC is a receding horizon controller so
therefore only components
i
of the first row of the
matrix equation (17) were considered. They can be
rewritten as,

2
2
2
2 2
1
1
1
1
N t
N
i
i N
i
t
N
i
i i t i
N
i
i
r z C
y F u z C
+

,
_

,
_

,
_

(18)
With definitions for the polynomials R, S and T1,
(18) is given by,
2
1 N t t t
r CT Sy u R
+
+
(19)
Substituting (19) into CARIMA model (1), the close
loop relationship was obtained,
t
c
N t
c
t
CP
R
r
CP
BT
y +
+
2
1
(20)
Where, the characteristic polynomial can be defined
as:
( )
(
( )
c
N
i
i
i i
N
i
i i i
CP
z G A B
A C
z BF A
C A BSz R A

,
_

+
+
+ +

2
2
1
1
1
1
1

(21)
When the control weight is zero, (17) is given by,

) ( ) (
~ 1
f r G G G u
T T


(22)
If the prediction horizon and the control horizon
have the same value N
u
=N
2
=N, the first row of
u
~
is
given by,


xN
g
1
0
0 ... 0 0
1
1
]
1

(23)
Therefore, the characteristic polynomial becomes,
( )
0
0
0
1
g
B
C g A B
g
A C P
C

,
_

+
(24)
The characteristic polynomial has the numerator
polynomial of the process model B. Consequently,
close loop poles are the zeros of the plant model.
3.1 Solution using tuning parameters
The unstable behavior produced by RHPZ when
N
2
=N
u
with =0 can be solved with an adequate
difference between prediction horizon and control
horizon. This can be confirmed as follow,
Increasing the prediction horizon (N
2

) and
using an unitary control horizon (N
u
= 1) the control
signal during the prediction horizon are,
u t u
N t u t u t u t u
+
+ + +
) 1 (
) 1 ( ) 2 ( ) 1 ( ) (
2

(25)
the control increment u is calculated to achieved
the final values of process output y(t+N
2
) must be
the reference r(t+N
2
) , hence using a big prediction
horizon,
) ( ) ( ) (
1
k t w k t u z G Lim
z
+ +

(26)
thus,
) ( ) (
1
k t r G k t u + +

(27)
Then, the closed loop response in the prediction
horizon is obtained,
) ( ) 1 ( ) ( ) (
1
k t r G z G k t y + +

(28)
Thus, the output dynamics is given by the open loop
poles.
The above results can be summarized as follow,
when MPC is applied to process with RHP zeros,
the tuning parameters can be adjusted to avoid
N
u
=N
2
with =0 because this produce the
cancellation of RHP zeros with RHP poles. Stable
behavior can be obtained using a prediction horizon
bigger than control horizon. Using a big prediction
horizon and unitary control horizon the closed loop
poles are given by the open loop poles.
4. Simulation examples
Two examples have been used to illustrate the
results, a linear system to show the influence of the
tuning parameters in the pole placement and a non-
linear system to compare the tuning methods.
4.1 linear non-minimum phase system
The following linear non-minimum phase system
was considered to obtain the pole placement of the
controller in function of its tuning parameters.
1
1
1
9 . 0 1
5 . 1 1
) (

z
z
z G
(29)
Using the closed loop polynomial characteristic (21)
the Biggest Pole Magnitude (BPM) is obtained for
each tuning parameters set. Figure 1, shows the
BPM when the Prediction horizon (N
2
) was changed
from N
2
=1 to N
2
=100 with N
u
=1 and =0. When
N
2
=1 the controller has its pole in z = 1.5, it has the
same value of RHPZ of the plant model, this
produces an unstable behavior. Subsequently, the
prediction horizon was increased. When the N
2
value
is close to N
u
, the controller has an unstable pole, if
N
2
is bigger than N
u
the pole will be inside the unit
circle (the magnitude is smaller than one). Note that
the closed loop pole tends towards the open loop
pole when N
2
is increased.
Fig.1 BPM vs. Prediction horizon
The control weight parameter has the function of
including a penalty on the control signal of the
multistage cost function (2). Fig. 2, shows the BPM
obtained for the control weight change from =0 to
=10, using the same prediction horizon and control
horizon.
Fig.3. BPM vs Control weight (). N
2
=N
u
=5.
Despite of the horizons have the same value, when
the control weight parameter was increased the BPM
of the controller was changed tending to inside the
unit circle. This is the solution proposed in the
literature to allow a stable behavior [1,6,7,12,13].
4.2 Non linear reactor
In order to compare the tuning methods a non-linear,
non minimum phase process has been used. The
isothermal Van de Vussen reaction systems involve
series and parallel reactions. The equations that
govern the systems are:
( )
V
F
C C C k C k
dt
dC
a in a a a
a
+
2
3 1
(26)
V
F
C C k C k
dt
dC
b b a
b

2 1
(27)
The desired output is the concentration of B, C
b
[mol/l], C
a
and C
a in
are the concentrations of A
[mol/l] in the reactor and in the feed respectively,
the manipulate input,
F is the dilution rate [l/min], V is the volume [l], and
the rate constants are given by k
1
=5/6 [min
-1
], k
2
=5/3
[min
-1
],k
3
=1/6 [mol/(litermin)] [15]
After linearizing model (26,27) about the operating
point, the physical model gives the following
transfer function (30). The discretization has been
made with a sampling rate (T
s
=0.2 min).
2 1
1
3951 . 0 2573 . 1 1
1745 . 0 0939 . 0
) (

+
+

z z
z
z G (30)
The transfer function has a zero in 1.8584.

Fig. 1 shows a step changes in the reference of
controlled variable when MPC is tuning with N
2
=8,
N
u
=8, =0. The reactor has an unstable response,
this behavior verified the cancellation of the
unstable zero with an unstable pole that was
demonstrated in Section 3.
Fig. 3. MPC tuning with N
2
=8, N
u
=8, =0
To contrast the controller performance with two
different sets of tuning parameters a set-point
change was produced. Fig. 4 shows the proposed
tuning method against the tuning method proposed
in the literature. The proposed method avoids the
instability using a prediction horizon bigger than
control horizon. The Fig. 4 illustrate that the raised
time is a 28 % of the raised time required by the
method proposed in the literature, that avoids the
instability increasing the control weight. Thus, as
can be observed the MPC with a high value of
control weight produces a soft control signal as well
as a soft output process. The proposed method
allows adjusting a small control weight as
consequence the raised time is reduced considerably.
Fig. 4. Proposed tuning method (- -) vs method that
increase the control weight (-)
4 Conclusion
This paper has presented how the presence of RHP
zeros in the process model produces an unstable
dynamics in MPC. The instability is produced by the
cancellation RHP zeros with RHP poles. A tuning
method is proposed to solve it. The method is based
in use a prediction horizon bigger than control
horizon. It is showed that an appropriate selection of
the difference between the prediction horizon and
the control horizon produce a stable performance. In
spite of the method proposed in the literature
produces a stable behavior, a better raised time is
obtained employing the proposed method. The bad
performance is obtained because the control weight
parameter produces soft moving of manipulate
variable as consequence a soft behavior in the
controlled variable.
In summary, small values of control weight
parameter with a convenient difference between
prediction horizon and control horizon, it seems to
work well for MPC applied to non-minimum phase
systems.
References:
[1] Clarke, D., C. Mohtadi and P. Tuffs,
Generalized predictive Control-Part I. Basic
Algorithm, Automatica, 23, 1987,137-148.
[2] Ogunaike, B., W. Harmon Process Dynamics.
modeling, and Control. Oxford University
Press,1994.
[3] Bardley, R., and M. Morari Design of
Resilient processing Plants. Chemical
Engineering Science, 40,1985, 59-74.
[4] R. K. Mehra and R. Rouhani. Theoretical
considerations on model algorithmic control for
non minimum phase systems. Joint Automatic
Control Conference,1980.
[5] Clarke, D., C. Mohtadi and P. Tuffs,
Generalized predictive Control-Part II Extensions
and Interpretations, Automatica, 23,1987,149-
160.
[6] M.J. Grimble. Generalized Predictive
Optimal Control: an introduction to the
advantages and limitations. International Journal
of Systems Science, 23, 1992,85-98.
[7] W.S. Levine (Ed). The Control Handbook
CRC Press,1996.
[8] J. M. Martn and J. Rodellar, Adaptive
predictive control: from the concepts to plant
optimisation,Prentice Hall,1996
[9] J.M. Maciejowski, Predictive Control with
Constraints, Prentice Hall,2001
[10] D. Seborg and E. Thomas and D.
Mellichamp, Process Dynamics and Control.
John Wiley and Sons,1989
[11] T. E. Marlin. Process Control: Designing
Processes and Control Systems for Dynamic
Performance, McGraw-Hill, 1995.
[12] Bitmead, R., M. Gevers, and V. Wertz,
Adaptative Optimal Control. Prentice Hall, 1990.
[13] R. Soeterboek. Predictive Control: A
Unified Approach, Prentice Hall,1991.
[14] Camacho, E.F., and C. Bordons. Model
Predictive Control, Springer 1999.
[15] B. Wayne Bequette, Process Dynamics:
Modeling, Analysis, and Simulation, Prentice
Hall,1998.