Model Predictive Controller Tuning via Eigenvalue Placement

Jorge L. Garriga and Masoud Soroush

AbstractThis paper presents an analytical study of the
effect of model predictive control (MPC) tunable parameters
over a wide range, on the closed-loop performance quantied
in terms of the location(s) of closed-loop eigenvalue(s) of a
large set of widely common, single-input single-output, linear
plants whose constraints are inactive. Symbolic manipulation
capabilities of MATHEMATICA are used to obtain analytical
expressions describing the dependence of closed-loop eigenval-
ues on the tunable parameters. This work is rst to investigate
how MPC tuning-parameters affect the location of the eigen-
values of the closed-loop system of a plant in the discrete-
time setting. It is to provide theoretical basis/justication
for many of the existing qualitative MPC tuning rules and
propose new tuning guidelines for MPC. For example, as the
prediction horizon is increased while other tunable parameters
remain constant, a subset of the closed-loop eigenvalues (poles)
move non-monotonically towards the open-loop eigenvalues
(poles) of the plant. If a prediction horizon much longer than
the reference-trajectory time-constant is used, the value of
reference-trajectory time-constant has little effect on the closed-
loop performance. As the weights on the magnitude or the rate
of change of the manipulated input are increased, the closed-
loop eigenvalues move towards the open-loop eigenvalues. As the
control horizon is increased from one, the dominant eigenvalue
of the closed-loop system initially moves towards the origin and
then away from the origin to a location that does not change
with a further increase in the control horizon.
I. INTRODUCTION
Model predictive control (MPC) has been implemented
widely in the process industries since its introduction in the
mid 1970s. Among factors contributing to the success of
MPC are (i) the ability of MPC to handle constraints on
manipulated, state and controlled variables in a systematic
way during the design and implementation of the controller,
and (ii) its enormous exibility that allows the use of
models, objective functions and constraint functionalities in
a variety of forms [1], [2]. The theory of MPC has received
increasingly attention for the past decade, especially from
researchers outside process systems engineering.
A typical model predictive controller has a great number
of tunable parameters. These parameters include the predic-
tion horizon(s), control horizon(s), model horizon, output
penalties, input magnitude and rate of change penalties,
and reference trajectory time constant(s). For the past three
decades, a qualitative understanding of the effect of these
controller tunable parameters on the stability and perfor-
mance of the closed-loop system has been gained. This
understanding together with closed-loop simulation studies
This work was supported by the National Science Foundation.
J.L. Garriga and M. Soroush are with the Department of Chemical and
Biological Engineering, Drexel University, Philadelphia, PA 19104, USA
Corresponding author: M. Soroush, ms1@drexel.edu
has made possible the wide-spread implementation of MPC.
It is known that the effect of these parameters on the closed-
loop performance is sometimes non-monotonic.
There have been several studies on MPC tuning, includ-
ing [3], [4], [5], [6], [7], [8]. Lee and Yu [3] derived expres-
sions for sensitivity equations for state-space models. Using
an optimal Kalman lter and a long horizon control law, they
studied MPC tuning for robust stability and performance.
Shridhar and Cooper [4] derived an analytical expression
that calculates move suppression coefcients as a function of
plant model parameters, other MPC design parameters, and
partitioned block condition numbers of the system matrix.
Their tuning method is applicable to unconstrained mul-
tivariable plants, including non-square systems. They also
proposed an expression for optimal move suppression coef-
cients using a rst-order-plus-dead-time approximation of
process model, and tuning heuristics for prediction horizon,
model horizon, and control horizon [5]. Al-Ghazzawi et
al. [6] presented an approach to tuning MPC on-line based
on sensitivity equations derived from a step response model
with linear constraints. The sensitivity equations were based
on nding optimal move suppression and controlled variable
damping coefcients from process parameters. Wojsznis et
al. [7] developed a heuristic approach to MPC tuning. They
proposed setting penalties on control moves as a function of
plant dead time as the primary factor, with some correction
from plant gain. Trierweiler and Farinab [8] also presented a
tuning strategy for MPC based on the attainable performance
of a system and its degree of directionality.
The dynamics of a plant in closed-loop is described mainly
by eigenvalues of the Jacobian of the closed-loop system, if
the plant is linear, and by the eigenvalues and nonlinearity
of the closed-loop system, if the plant is nonlinear. In
MPC, there has been very little or no attempt to study
how a controller places closed-loop eigenvalues, while in
other control methods, closed-loop performance, stability,
and robust stability have been studies in terms of the location
of closed-loop eigenvalues.
This paper presents an analytical study of the effect of the
MPC tunable parameters over a wide range, on the closed-
loop performance quantied in terms of the location(s) of
closed-loop eigenvalue(s) of a large set of widely common,
single-input single-output, linear plants whose constraints are
inactive. The tunable parameters include control horizon,
prediction horizon, controlled variable weights, time constant
of reference trajectory, and weights on the magnitude and
rate of change of manipulated variable. Symbolic manipu-
lation capabilities of MATHEMATICA are used to obtain
analytical expressions describing the dependence of closed-
2008 American Control Conference
Westin Seattle Hotel, Seattle, Washington, USA
June 11-13, 2008
WeA13.1
978-1-4244-2079-7/08/$25.00 2008 AACC. 429
loop eigenvalues on the tunable parameters. This work is
rst to investigate how MPC tuning-parameters affect the
location of the eigenvalues of the closed-loop system of a
plant in the discrete-time setting. It is to provide theoretical
basis/justication for many of the existing qualitative MPC
tuning rules and propose new tuning guidelines for MPC.
The organization of this paper is as follows. The scope of
the study and some mathematical preliminaries are given in
Section II. A general moving-horizon optimization problem
is formulated in Section III and is applied to linear plants in
Section IV. Symbolic calculations are described in Section
V. Several widely common plants are considered in Section
VI. Concluding remarks are given in Section VII.
II. SCOPE AND PRELIMINARIES
Consider a single-input single-output (SISO) nonlinear
plant in the form
x(k + 1) = [ x(k), u(k)], x(0) = 0
y(k) = h[ x(k)] +d
_
(1)
where x X
n
denotes the vector of the plant state
variables, u U is the plant manipulated variable
(input), y h(X) is the plant controlled variable
(output), and d D is an unmeasurable constant
disturbance. Here X and U are open connected sets, and D
is a connected set. It is assumed that the discrete-time plant
description is a result of using a sampling period in order of
one-tenth of the dominant time constant of the plant.
We also make the following assumptions: (a) (x
ss
, u
ss
) =
(0, 0) is the nominal equilibrium pair, (b) the nominal
equilibrium pair (0, 0) X U, (c) (x, u) and h(x) are
smooth vector functions on XU and X, respectively, and
(d) for a plant of the form of (1), a discrete-time model of
the following form is available:
x(k + 1) = [x(k), u(k)], x(0) = 0
y(k) = h[x(k)]
_
(2)
where x X
n
denotes the vector of model state
variables and y h(X) is the model controlled variable
(output). Relative order (degree) of the controlled output y
with respect to the manipulated input u is denoted by r,
where the relative order r is the smallest integer for which
y(k +r) depends explicitly on u(k).
There have been several studies on the connections be-
tween MPC and existing analytical control methods, leading
to a better understanding of the effect of MPC tuning param-
eters on the closed-loop performance [9], [10]. A few special
cases of unconstrained SISO MPC are listed in Table I. M
and P denote control and prediction horizons, respectively.
As the prediction horizon is increased and the control horizon
is decreased, a more robust but less aggressive controller is
obtained. It has been shown [9], [10] that in special cases
shortest-prediction-horizon MPC leads to:
Input-output linearizing control laws that inherently in-
clude optimal windup and directionality compensators,
Model state feedback control (MSFC) and modied in-
ternal model control (IMC) laws that inherently include
an optimal directionality compensator,
TABLE I
UNCONSTRAINED SISO MODEL PREDICTIVE CONTROLLERS WITH NO
PENALTY ON MANIPULATED INPUT MAGNITUDE OR RATE OF
CHANGE [9].
P M Reference Model Resulting
Trajectory Controller
r 1 No Deadbeat
r 1 Yes I-O Linearization
r 1 Yes Linear Modied IMC
r 1 Yes Linear MSFC
r = 1 1 Yes Linear PI, PID
>> r >> 1 Yes/No Long Horizon MPC
1 Yes/No Steady State
Proportionalintegral (PI) and proportionalintegral
derivative (PID) controllers that inherently include op-
timal windup and directionality compensators.
III. MODEL PREDICTIVE CONTROL LAW
Consider a moving-horizon minimization problem of the
form
min
U(k)
_
P

=r
w
y

[y
d
(k +) y(k +)]
2
+
M1

=0
w
u
+1
[u(k +)]
2
+
M1

=0
w
u
+1
[u(k +) u(k + 1)]
2
_
(3)
subject to
x(k + 1) = [x(k), u(k)]
y(k) = h[x(k)]
u(k +) = u(k + 1), = M, , P r
where
P r,
U(k) = [u(k) u(k +M 1)]
T
,
w
yr
, , w
yP
, w
u1
, , w
uM
, w
u1
, , w
uM
0,
y
d
is a reference trajectory, given by
y
d
(k +j) =(1 +)
r
y
sp
(k +j)
r1

=0
_
r
r
_

r
y
d
(k +j r +), j = r, , P
y
d
(k +) = y(k +), = 0, , r 1
is a tunable scalar parameter that sets the speed of
the reference trajectory and is chosen such that 1 <
< 0.
y
sp
Y is the output set-point. Y is the set of
all y
sp
for which for every d D there exist a u
ss

U and an x
ss
X satisfying x
ss
= (x
ss
, u
ss
) and
y
sp
= h(x
ss
) +d.

_
a
b
_

a!
b!(a b)!
430
y is the predicted future value of the controlled output,
given by
y(k +)

= y(k) +h

[x(k)] h[x(k)], = 0, , r 1
y(k +r)

= y(k) +h
r
[x(k), u(k)] h[x(k)]
y(k +)

= y(k) +h

[x(k), u(k), , u(k + r)]

h[x(k)], = r + 1, , P
(4)

h
0
[x(k)]

= h[x(k)]
h

[x(k)]

= h
1
[(x(k), u(k))],
= 1, , r 1
h
r
[x(k), u(k)]

= h
r1
[(x(k), u(k))]
h
r+1
[x(k), u(k), u(k + 1)]

= h
r
[(x(k), u(k)), u(k + 1)]
.
.
.
h
P
[x(k), u(k), , u(k+Pr)]

=
h
P1
[(x(k), u(k)), u(k + 1), , u(k +P r)]
When measurements of the state variables are available, in
the preceding prediction equations we set x = x. Otherwise,
the values of the state variables have to be estimated, for
example, via on-line simulation of the plant model (use of an
open-loop state estimator) or by using an extended Kalman
lter. An estimate of the unmeasured output disturbance,

d, is then calculated from

d = y h(x). The use of this
disturbance estimate in the model predictive control method
leads to a model predictive controller with integral action.
For the sake of simplicity, the state estimator dynamics are
not accounted for in the results presented herein.
Solving the moving-horizon minimization problem of (3)
leads to a feedforward/state feedback. It is assumed that
for every x X, every y
sp
Y , and every d D, the
optimization problem of (3) is feasible; that is, there is a
u U that minimizes the performance index in (3) globally.
IV. APPLICATION TO LINEAR PLANTS
Consider the class of time-invariant, linear plants with a
model of the form
x(k + 1) = Ax(k) +bu(k), x(0) = 0
y(k) = cx(k)
_
(5)
where A, b and c are nn, n1 and 1n constant matrices
respectively. This class of plants is a special case of (1).
For this class of plants, the predicted future values of the
controlled output are given by:
y(k + 1) = y(k) + [cA c]x(k) +cbu(k)
y(k + 2) = y(k) + [cA
2
c]x(k) +cAbu(k) +cbu(k + 1)
.
.
.
y(k +P) = y(k) + [cA
P
c]x(k) +cA
P1
bu(k) +
+cA
r1
bu(k +P r) + +cbu(k +P 1)
(6)
and the current and future values of the reference trajectory
by
y
d
(k) = y(k)
y
d
(k + 1) = y(k) + [cA c]x(k)
.
.
.
y
d
(k +r 1) = y(k) + [cA
r1
c]x(k)
y
d
(k +r) = (1 +)
r
y
sp
(k +r)
r1

=0
_
r
r
_

r
y
d
(k +)
.
.
.
y
d
(k +P) = (1 +)
r
y
sp
(k +P)
r1

=0
_
r
r
_

r
y
d
(k +P r +)
V. SYMBOLIC CALCULATIONS
After entering values for A, b, and c matrices (unless the
plant is rst order), the following symbolic calculations are
performed:
(a) After setting values of P and M, calculate the perfor-
mance index J(x(k), y
sp
(k), u(k), , u(k +M1)).
(b) Calculate [u(k), , u(k + M 1)]
T
=
Q(x(k), y
sp
(k)) that globally minimizes
J(x(k), y
sp
(k), u(k), , u(k + M 1). This
involves taking partial derivatives of J with respect
to u(k), , u(k + M 1), setting the resulting
partial derivatives to zero, and then solving for
u(k), , u(k +M 1).
(c) Calculate the Jacobian of the closed-loop system; i.e.,
J
cl
= A+bk
1
, where k
1
is the rst row of the M n
matrix Q(x, y
sp
)/x.
(d) Calculate the eigenvalues of J
cl
.
This approach can be extended to calculate eigenvalues of the
closed-loop system of an input-constrained plant. For such a
plant, steps (b) and (c) are different; the optimization in step
(b) is constrained, and when an input constraint is active,
k
1
= 0; J
cl
= A, which is the Jacobian of the plant in open-
loop (J
ol
). A sufcient condition for asymptotic stability of
an input-constrained plant is that all eigenvalues of (a) J
cl
when the same plant is unconstrained and (b) J
ol
= A lie
inside the unit circle. In the case that the order of the plant is
higher than one and the values of the matrices A, b, and c are
not set, MATHEMATICA calculates symbolic expressions
for the closed-loop Jacobian that may be hundreds of pages
long.
Sample analytical expressions for the closed-loop Jacobian
(eigenvalue), J
cl
, of a general rst-order plant of the form
(5) for several low values of P and M are given in Table
II. They show how M, P, w
y1
, , w
yn
, w
u1
, , w
uM
,
w
u1
, , w
uM
, and affect the location of the closed-
loop eigenvalue. Closed-loop Jacobian expressions for higher
values of P and M are not presented here because of their
larger sizes and the 6-page limit on the length of ACC
431
TABLE II
ANALYTICAL EXPRESSIONS FOR THE CLOSED-LOOP EIGENVALUE OF A
GENERAL FIRST-ORDER LINEAR PLANT.

P=1,
M=1

P=2,
M=1

P=3,
M=1

P=4,
M=1

P=5,
M=1

proceedings papers. For example, the closed-loop Jacobian
(eigenvalue) for P = 5 and M = 3 is more than four
pages long. As can be seen in the sample expressions, for
every combination of P and M values, sum of the 2
M
terms A(
1j

Mj
), j = 1, 2, where
i1
= w
ui
and

i2
= w
ui
, i = 1, M, appear in the numerator of J
cl
,
and sum of the 2
M
terms (
1j

Mj
), j = 1, 2, in
the denominator. This explains clearly the previously-known
qualitative and intuitive understanding that:
As w
u1
, , w
uM
, J
cl
A = J
ol
.
As w
u1
, , w
wu
M
, J
cl
A = J
ol
.
VI. ILLUSTRATIVE EXAMPLES
The application of the approach described above to several
specic simple plants is presented in this section.
A. Examples
Example 1. A minimum-phase, asymptotically stable,
rst-order, linear plant:
A = [0.905] , b = [1.903] , c = [1]
which has the transfer function
G(z) =
1.903
z 0.905
Example 2. A non-minimum phase, asymptotically stable,
second-order, linear plant:
A =
_
0 1
0.125 0.75
_
, b =
_
0
1
_
, c =
_
2 1

;
G(z) =
2 z
(z 0.5)(z 0.25)
Example 3. A non-minimum phase, asymptotically stable,
second-order, linear plant:
A =
_
0 1
0 0.5
_
, b =
_
0
1
_
, c =
_
2 1

;
G(z) =
2 +z
z(z 0.5)
Example 4. A minimum phase, unstable, second-order,
linear plant:
A =
_
0 1
1.6 2.8
_
, b =
_
0
1
_
, c =
_
0.5 1

;
G(z) =
z 0.5
(z 2)(z 0.8)
Example 5. A non-minimum phase, unstable, second-
order, linear plant:
A =
_
0 1
1.6 2.8
_
, b =
_
0
1
_
, c =
_
1.5 1

;
G(z) =
z 1.5
(z 2)(z 0.8)
Example 6. A minimum phase, asymptotically stable,
fourth-order, linear plant:
A =
_

_
0.9 0 0 0
1 0 0 0
0 1 0 0
0 0 1 0
_

_
, b =
_

_
1.9
0
0
0
_

_
, c =
_
0 0 0 1

;
G(z) =
1.9
z
3
(z 0.9)
B. Results
Results of the application of the approach to the plant
examples are presented in Figure 1 and Tables III-VIII.
C. Discussions
Figure shows how the closed-loop eigenvalue(s) of Exam-
ples 1-5 and the dominant closed-loop eigenvalue of Example
6 change(s) with P. When w
u1
= w
u1
= 0 and M = 1, as
P increases, the closed-loop eigenvalue(s) move towards the
open-loop eigenvalue(s), irrespective of the value of that
sets the speed of the reference trajectory. Furthermore, if a
prediction horizon much longer than the reference-trajectory
time-constant is used, the value of the reference-trajectory
time constant has little effect on the closed-loop performance
(locations of the closed-loop eigenvalues). When P = 1,
r closed-loop eigenvalues are placed at z = , and the
remaining (n r) eigenvalues at the plant zeros. For this
432
0
0.2
0.4
0.6
0.8
1
0 20 40 60 80 100
E
i
g
e
n
v
a
l
u
e
Prediction Horizon (P)
J=0 J=-0.3
J=-0.6 J=-0.9 (Ex.1)
0.0
0.2
0.4
0.6
0.8
1.0
0 20 40 60 80 100
E
i
g
e
n
v
a
l
u
e
Prediction Horizon (P)
J=0.0 J=-0.3
J=-0.6 J=-0.9
0.0
0.5
1.0
1.5
2.0
0 20 40 60 80 100
E
i
g
e
n
v
a
l
u
e
Prediction Horizon (P)
J=0.0 J=-0.3
J=-0.6 J=-0.9
J 0 0 J 0 3
1.0
(Ex.2)
-0.5
0.0
0.5
1.0
0 20 40 60 80 100
E
i
g
e
n
v
a
l
u
e
Prediction Horizon (P)
J=0.0 J=-0.3
J=-0.6 J=-0.9
-2.0
-1.0
0.0
0 20 40 60 80 100
E
i
g
e
n
v
a
l
u
e
Prediction Horizon (P)
J=0.0 J=-0.3
J=-0.6 J=-0.9
(Ex.3)
0.0
0.2
0.4
0.6
0.8
1.0
0 20 40 60 80 100
E
i
g
e
n
v
a
l
u
e
Prediction Horizon (P)
J=0.0 J=-0.3
J=-0.6 J=-0.9
0.0
0.2
0.4
0.6
0.8
1.0
0 20 40 60 80 100
E
i
g
e
n
v
a
l
u
e
Prediction Horizon (P)
J=0.0 J=-0.3
J=-0.6 J=-0.9
1.0 1.6
J 0 0 J 0 3
(Ex.4)
0.0
0.2
0.4
0.6
0.8
0 20 40 60 80 100
E
i
g
e
n
v
a
l
u
e
Prediction Horizon (P)
J=0.0 J=-0.3
J=-0.6 J=-0.9
0.8
1.0
1.2
1.4
0 20 40 60 80 100
E
i
g
e
n
v
a
l
u
e
Prediction Horizon (P)
J=0.0 J=-0.3
J=-0.6 J=-0.9
(Ex.5)
-3
2
7
12
17
22
0 20 40 60 80 100
E
i
g
e
n
v
a
l
u
e
Prediction Horizon (P)
J=0.0 J=-0.3
J=-0.6 J=-0.9
(Ex.6)
Fig. 1. Closed-loop eigenvalue(s) of Examples 1-5 and the dominant
closed-loop eigenvalue of Example 6 for several values of P and [M = 1,
wu
1
= 0, w
u
1
= 0, wyr
= = wy
P
= 1].
TABLE III
RESULTS FOR EXAMPLE 1.
(A) Closed-loop eigenvalue and the performance index J (with ysp =
0 and x = 5) for several values of M, = 0.3, and P =
7, wu
1
= 0, w
u
1
= 0, wyr
= = wy
P
= 1.
M 1 2 3 4 5
Eigenvalues 0.75 0.20 0.30 0.30 0.30
J 15.7928 0.59747 0.04267 0.00274 0.00015
M 6 7
Eigenvalues 0.30 0.30
J 0.00005 0.00000
(B) Closed-loop eigenvalue for several values of wu
1
and w
u
1
, and =
0.0, P = 7, M = 3, wyr
= = wy
P
= 1.
wu
1
Eigenvalue w
u
1
Eigenvalue
0 0.00000 0 0.00000
1 0.17015 1 0.20051
2 0.26461 2 0.26825
3 0.32912 3 0.310569
4 0.37778 4 0.342283
TABLE IV
RESULTS FOR EXAMPLE 2.
(A) Closed-loop eigenvalues for several values of M, = 0.3, and
P = 7, wu
1
= 0, w
u
1
= 0, wyr
= = wy
P
= 1.
M 1 3 7
Eigenvalues 0.2553, 0.4509 0.2359, 0.2998 0.2373, 0.3000
(B) Closed-loop eigenvalues for several values of wu
1
and w
u
1
, and
= 0.0, P = 7, M = 3, wyr
= = wy
P
= 1.
wu
1
(w
u
1
=0) Eigenvalues w
u
1
(wu
1
=0) Eigenvalues
0 0.0000, 0.5013 0 0.0000, 0.5013
1 0.0495, 0.5010 1 0.0800, 0.5008
2 0.0821, 0.5009 2 0.1151, 0.5011
3 0.1054, 0.5007 3 0.1362, 0.5016
4 0.1230, 0.5006 4 0.1506, 0.5019
reason, prediction horizons greater than one should be used
for non-minimum-phase plants (Examples 2, 3 and 5; Figure
1) to ensure closed-loop stability. When a plant (such as
Example V) is unstable and non-minimum-phase, prediction
and control horizons much longer that those that are adequate
to stabilize unstable or non-minimum-phase plants, should be
used to ensure that the closed-loop system is asymptotically
stable (e.g., P = 11 and M = 10 for Example V). As
the control horizon is increased from one, as shown in
Table III(A), the dominant eigenvalue of the closed-loop
system initially moves towards the origin and then away
from the origin to a location that does not change with a
further increase in the control horizon. Note that the value
of the performance index, J, evaluated at the minimizing
u, as expected, decreases monotonically as M increases.
In other words, J monotonically decreases with M, while
the eigenvalue rst decreases, then increases, and nally
does not change. Tables III(B)-VIII(B) show how closed-
loop eigenvalue(s) vary as w
u1
and w
u1
are increased while
= 0.0, P = 7, M = 3, and w
yr
= = w
yP
= 1.
VII. CONCLUDING REMARKS
This paper presented an analytical study of the effect of the
MPC tunable parameters over a wide range, on the closed-
loop performance quantied in terms of the location(s) of
closed-loop eigenvalue(s) of a large set of widely common
433
TABLE V
RESULTS FOR EXAMPLE 3.
(A) Closed-loop eigenvalues for several values of M, = 0.3, and
P = 7, wu
1
= 0, w
u
1
= 0, wyr
= = wy
P
= 1.
M 1 3 7
Eigenvalues -0.0217, 0.4477 0.3012, -0.6495 0.3000, -0.6953
(B) Closed-loop eigenvalues for several values of wu
1
and w
u
1
, and
= 0.0, P = 7, M = 3, wyr
= = wy
P
= 1.
wu
1
(w
u
1
=0) Eigenvalues w
u
(wu
1
=0) Eigenvalues
0 0.0000, 0.5011 0 0.0000, 0.5011
1 0.0000, 0.5008 1 0.0000, -0.0888
2 0.0000, 0.5006 2 0.0000, 0.0276
3 0.0000, 0.5005 3 0.0000, 0.0919
4 0.0000, 0.5004 4 0.0000, 0.1353
TABLE VI
RESULTS FOR EXAMPLE 4.
(A) Closed-loop eigenvalues for several values of M, = 0.3, and
P = 7, wu
1
= 0, w
u
1
= 0, wyr
= = wy
P
= 1.
M 1 3 6
Eigenvalues 0.975 0.075i 0.384 0.050 0.300, 0.500
(B) Closed-loop eigenvalues for several values of wu
1
and w
u
1
, and
= 0.0, P = 7, M = 3, wyr
= = wy
P
= 1.
wu
1
(w
u
1
=0) Eigenvalues w
u
1
(wu
1
=0) Eigenvalues
0 0.0000, 0.4809 0 0.0000, 0.4908
1 0.4976, 0.6130 1 0.667 0.214i
2 0.5660, 0.6468 2 0.737 0.211i
3 0.6151, 0.6584 3 0.782 0.199i
4 0.659 0.011i 4 0.813 0.184i
plants whose constraints are inactive. Symbolic manipula-
tion capabilities of MATHEMATICA were used to obtain
analytical expressions describing the dependence of closed-
loop eigenvalues on the tunable parameters. This work is
rst to investigate how MPC tuning-parameters affect the
location of the eigenvalues of the closed-loop system of
a plant in the discrete-time setting. It provides theoretical
basis/justication for many of the existing qualitative MPC
tuning rules and propose new tuning guidelines for MPC.
For example, as the prediction horizon is increased while
other tunable parameters remain constant, a subset of the
closed-loop eigenvalues (poles) move non-monotonically to-
wards the open-loop eigenvalues (poles) of the plant. If a
prediction horizon much longer than the reference-trajectory
time-constant is used, the value of reference-trajectory time-
constant has little effect on the closed-loop performance. As
the weights on the magnitude or the rate of change of the
manipulated input are increased, the closed-loop eigenvalues
move towards the open-loop eigenvalues.
REFERENCES
[1] Allgower, F., Zheng, A. Nonlinear Model Predictive Control,
Progress in Systems and Control Theory series, Vol. 26, Birkhauser-
Verlag, Birkhauser Verlag, Basel, 2000
[2] Qin, J., Badgwell, T.A. A survey of industrial model predictive
control technology. Control Engineering Practice, 11, 733764, 2003
[3] Lee, J.H., Yu, Z. H. Tuning of model predictive controllers for robust
performance. Computers & Chemical Engineering, 18(1), 1537,
1994
[4] Shridar, R., Cooper, D.J. A tuning strategy for unconstrained mul-
tivariable model predictive control. Ind. Eng. Chem. Res., 37(10),
40034016, 1998
[5] Shridhar, R., Cooper, D.J. A novel tuning strategy for multivariable
model predictive control. ISA Transactions, 36(4), 273280, 1998
TABLE VII
RESULTS FOR EXAMPLE 5.
(A) Closed-loop eigenvalues for several values of M, = 0.3, and
P = 7, wu
1
= 0, w
u
1
= 0, wyr
= = wy
P
= 1.
M 1 3 6
Eigenvalues 0.8850, 1.2045 0.3177, 1.2452 0.3000, 1.4844
(B) Closed-loop eigenvalues for several values of wu
1
and w
u
1
, and
= 0.0, P = 7, M = 3, wyr
= = wy
P
= 1.
wu
1
(w
u
1
=0) Eigenvalues w
u
1
(wu
1
=0) Eigenvalues
0 -0.0002, 0.9239 0 -0.0002, 0.9238
1 0.3307, 0.7094 1 0.7025 0.1088i
2 0.3991, 0.7262 2 0.7604 0.1314i
3 0.4288, 0.7387 3 0.7936 0.1277i
4 0.4452, 0.7478 4 0.8167 0.1190i
TABLE VIII
RESULTS FOR EXAMPLE 6.
(A) Closed-loop eigenvalues for several values of M, = 0.3, and
P = 7, wu
1
= 0, w
u
1
= 0, wyr
= = wy
P
= 1.
M 1 3 7
Eigenvalues -0.017 0.121i 0.302 0.002i 0.302 0.002i
0.077, 0.772 0.298 0.002i 0.298 0.002i
(B) Closed-loop eigenvalues for several values of wu
1
and w
u
1
, and
= 0.0, P = 7, M = 3, wyr
= = wy
P
= 1.
wu
1
(w
u
1
=0) Eigenvalues w
u
1
(wu
1
=0) Eigenvalues
0 0.0000, 0.0000 0 0.0000, 0.0000
0.0000, 0.0000 0.0000, 0.0000
1 0.0000, 0.0000 1 0.0000, 0.0000
0.0000, 0.1709 0.0000, 0.2980
2 -0.000 0.000i 2 0.0000, 0.0000
0.0000, 0.2669 0.0000, 0.4021
3 -0.000 0.000i 3 0.0000, 0.0000
0.0000, 0.3324 0.0000, 0.4614
4 0.0000, 0.0000 4 -0.000 0.000i
0.0000, 0.3816 0.0000, 0.5009
[6] Al-Ghazzawi, A., Ali, E., Nouh, A., Zariou, E. On-line tuning
strategy for model predictive controllers. J. of Process Control, 11,
265284, 2001
[7] Wojsznis, W., Gudaz, J., Blevins, T., Mehta, A. Practical approach
to tuning MPC. ISA transactions (ISA trans.), 42(1), 149162, 2003
[8] Trierweiler, J.O., Farinab, L.A. RPN tuning strategy for model
predictive control. J. of Process Control, 13(7), 591598, 2003
[9] Soroush, M., Muske, K. Analytical model predictive control. in
Nonlinear Model Predictive Control, F. Allgower and A. Zheng
(eds.), Progress in Systems and Control Theory Series, Vol. 26,
Birkhauser-Verlag, Birkhauser Verlag, Basel, 163179, 2000
[10] Soroush, M., Soroush, H. M. Long-prediction-horizon nonlinear
model predictive control. in Nonlinear Predictive Control: Theory
and Practice, B. Kouvaritakis and M. Cannon (eds.), IEE Control
Engineering Book Series, 61, 190201, 2001
434