Sie sind auf Seite 1von 17

Asian Journal of Control, Vol. 18, No. 5, pp.

117, September 2016


Published online in Wiley Online Library (wileyonlinelibrary.com) DOI: 10.1002/asjc.1252

Brief Paper

A GENERAL ROBUST MPC DESIGN FOR THE STATE-SPACE MODEL:


APPLICATION TO PAPER MACHINE PROCESS
S. Hassan HosseinNia and Michael Lundh
ABSTRACT
Applying model predictive control (MPC) in some cases such as complicated process dynamics and/or rapid sampling leads us to poorly numerically conditioned solutions and heavy computational load. Furthermore, there is always
mismatch in a model that describes a real process. Therefore, in this paper in order to prevail over the mentioned difficulties, we design a robust MPC using the Laguerre orthonormal basis in order to speed up the convergence at the
same time with lower computation adding an extra parameter a in MPC. In addition, the Kalman state estimator is
included in the prediction model and accordingly the MPC design is related to the Kalman estimator parameters as well
as the error of estimations which helps the controller react faster against unmeasured disturbances. Tuning the parameters of the Kalman estimator as well as MPC is another achievement of this paper which guarantees the robustness of
the system against the model mismatch and measurement noise. The sensitivity function at low frequency is minimized
to tune the MPC parameters since the lower the magnitude of the sensitivity function at low frequency the better command tracking and disturbance rejection results. The integral absolute error (IAE) and peak of the sensitivity are used as
constraints in optimization procedure to ensure the stability and robustness of the controlled process. The performance
of the controller is examined via the controlling level of a Tank and paper machine processes.
Key Words: Model predictive control, MPC tuning, optimization, orthonormal basis function, Laguerre network.

I. INTRODUCTION
Model predictive control (MPC), also known as
Receding Horizon Control and Moving Horizon Optimal Control, has been widely adopted in industry for
advanced process control [112]. MPC is a control technology which utilizes a model of the system to predict
the process output over some future horizon and solve a
quadratic optimization problem with the control signal
as decision variables. The ideas of MPC can be traced
back to the 1960s [13], but interest in this field started
to surge only in late 1970s introducing IDCOM [14],
dynamic matrix control (DMC) [15], and generalized
predictive control (GPC) [16,17]. These algorithms were
based on step or impulse response models. Since the
state-space approach provides a unified framework for
discussion of the various predictive control algorithms
and is well suited for stability analysis, interest in MPC

Manuscript received January 9, 2015; revised June 21, 2015; accepted


September 29, 2015.
Hassan HosseinNia (corresponding author, e-mail: s.h.hosseinniakani@
tudelft.nl) is with the Department of Precision and Microsystem Engineering,
Delft University of Technology, Delft, The Netherlands.
Michael Lundh is with ABB AB, Corporate Research, Vsters, Sweden.
The authors would like to thank Dr. Winston Garcia-Gabin from ABB corporate research for his useful discussion.

implementations based on state-space models were created in [18,19].


When the constraints in model predictive control
are activated, the control law effectively becomes a nonlinear control problem. Thus, the stability properties of
linear time-invariant systems do not apply. However, a
remarkable property of model predictive control is that
one can establish stability of the closed-loop system
under certain conditions. A Lyapunov MPC is designed
in [20,21], based on uniting receding horizon control
with explicit Lyapunov-based nonlinear controller design
techniques. In order to guarantee the closed-loop stability, in the design of the controller, constraints based on
Lyapunov functions are incorporated in the controller
formulations. In [22], a contractive MPC proposed in
which the stabilizing state and output feedback MPC
algorithms for constrained continuous-time nonlinear
systems with discrete observations. Most of the MPC
schemes in the literature that are designed for constrained
nonlinear systems involve the solution of nonlinear programming problems (NLPs). Due to the difficulties inherent to solving NLPs and since MPC requires the optimal (feasible) solution to be computed on-line and at
each time step, it is important to simplify the controller
from a computational point of view. Furthermore, many
nonlinear MPC controllers do not provide any stability
or performance guarantees for the closed-loop system.

2015 Chinese Automatic Control Society and Wiley Publishing Asia Pty Ltd

Asian Journal of Control, Vol. 18, No. 5, pp. 117, September 2016

Hence, the contractive MPC propose a stabilizing MPC


controller, which can be implemented as a quadratic programming problem for nonlinear system. In our paper,
the Hildreths quadratic programming is used to solve the
constraint problem (see Appendix I). The results of our
proposed MPC are analyzed using a similar Lyapunov
method as that in [23].
Design of the MPC, essentially, includes the development of an effective plant model that can be used for
predicting the future output of the plant with good accuracy. Therefore, MPCs rely on dynamic models of the
process, most often linear empirical models obtained by
system identification. Although, the major part of industrial MPC is generation of data and identification of
models, system identification and model predictive control have largely developed as two separate disciplines.
In [2427], MPC is designed using ARMAX state-space
models. The main advantage of this approach is that
the states are estimated at the same time as MPC is
designed. But, it is just applicable to the system with the
ARMAX model. In this paper we extend this approach
to be applied to all systems with the state-space model
and use a Kalman filter as the state estimator. Modeling and identification of the classical dynamical systems by means of orthonormal basis functions has been
given a remarkable research attention both from control
theoreticians and practicians involved in modeling and
control [2832]. Recently, a novel mechanism has been
presented for managing the performance and feasibility trade off within MPC based on orthonormal basis
function such as Laguerre and Kautz function [23,33,34].
Specifically, it demonstrates that the orthonormal basis
functions offer simple and effective alternative to achieve
low computation burden at the same time with good feasibility and good performance. Therefore in this paper,
orthonormal basis function is added to the design of
our proposed MPC to improve the performance of the
controller and speed up the process.
Tuning of the MPC has always been a challenge,
especially, in complex industrial processes in which a
large number of parameters should be tuned. One of the
pioneer works studied the tuning of the DMC parameter in order to achieve set point tracking with minimal
overshoot and modest manipulated input move sizes
[35]. In [36], practical on/off-line tuning rules are studied for state-space MPC controllers. One can find more
information about the tuning of the MPC in [3739]. In
[25,26,40], the authors introduced a tuning method minimizing IAE of the system-limiting peak of sensitivity as
a constraint in the optimization procedure. The advantage of this method is guaranteeing the robustness of the
system against the model mismatch and measurement
noises. In this paper, we propose a similar optimization

procedure but minimizing the sensitivity function at low


frequency, which guarantees the good command tracking and disturbance rejection. Optimizing the sensitivity function at low frequency unlike the IAE does not
depend on the time horizon and consequently results in
lower computation. The parameters are optimized limiting: (i) IAE, to maintain the stability; and (ii) peak of
sensitivity, to maintain the robustness against the model
mismatch and measurement noises. To summarize the
objective of this paper in one sentence, it will try to introduce a general robust MPC including a Kalman observer
in the prediction model where the parameters are tuned
by an optimization process.
The rest of the paper is organized as follows. In the
next section, the prediction model will be defined. After
introducing the prediction model, the unconstrained
MPC and Laguerre-based MPC will be proposed in
Sections III and IV, respectively. The proposed MPC
will be extended to the constraint problem in Section V.
Section VI will disscuss the procedure used to adjust
the MPC parameters. To show the performance of the
proposed MPC as well as the tuning procedure, control
of a single input single output (SISO) tank level process and a multiple input multiple output (MIMO) paper
machine process are studied in Section VII. Finally, this
paper is concluded in Section VIII by some remarks and
future works.

II. PREDICTION MODEL


In this section, the prediction model is obtained in
the state-space.
Considering the state-space model as
{
xm (k + 1) = Axm (k) + Bu(k) + Bd d(k) + w(k),
y(k) = Cxm (k) + v(k)
(1)
and defining u(k) = u(k) u(k 1), d(k) = d(k)
d(k 1), w(k) = w(k) w(k 1), v(k) = v(k) v(k 1)
and xm (k) = xm (k) xm (k 1), equation (1) can be
rewritten as:
{
xm (k + 1) = Axm (k)+ Bu(k) + Bd d(k) + w(k),
y(k + 1) = Cxm (k + 1) + y(k) + v(k)
(2)
[
]T
Defining new state variable x(k) = xm (k) y(k) , an
augmented model can be obtained as follows:
{
x(k + 1) = x(k) + u(k) + d d(k) + w w(k)
y(k) = (k) + v(k)
(3)

2015 Chinese Automatic Control Society and Wiley Publishing Asia Pty Ltd

S. H. HosseinNia and M. Lundh: A General Robust MPC Design for the State-Space Model

]
[ ]
[
]
[ ]
A 0
B
Bd
I
where, =
,=
, d =
, w =
CA I
CB
CBd
C
[ ]
and = 0 I .
Using (3) as the prediction model rejects any disturbances with constant value since there is an integrator
controller.The Block diagram of the closed loop control system is shown in Fig. 1. The estimation and the
prediction model may for the state space model (3) be
represented as
(k + j + 1|k) = (k + j|k 1) + u(k + j|k)

+d d(k + j|k) + ()e(k + j|k)

y (k + j) = (k + j|k), j = 0, , Np 1

(4)
{
K(),
j=0
where () =
and (k) =
0, j = 1, , Np 1
[
]
x m (k) y (k) is states and output estimation, e(k) =
y(k) y (k|k 1) is the observer error, Np is the prediction horizon and K() is the Kalman filter gain matrix
obtained in which the observer poles of the system are
reside inside the circle of radios 0 < 1. Therefore, the
observer gain matrix be obtained as
(
K() =

( ) ( )T )1
( )T (

,
P
R2 +
P

(5)

where,
( )
P=

( )(
( ) ( )T )1 ( )

PP
R2 + P
P

( )T

+ R1 ,

III. UNCONSTRAINED MPC FOR


STATE-SPACE FORM MODEL
In this section, the MPC will be designed for
the state-space prediction model (4). The designed
MPC is variable of the states (k), the error e(k) and
disturbance d(k).
3.1 Design framework
The future control trajectory, and future output are
respectively denoted
by:
[
]
u(k) u(k + 1) u(k + Nc 1) , and
U
=
[
]
Y = y(k + 1|k) y(k + Np |k) where, Nc is the
control horizon.
Based on the state-space model (, , ), the predicted output calculated sequentially using the set of
future control parameters:
y (k + 1) = (k + 1|k) = (k) + d d(k)
+u(k) + K()e(k)
y (k + 2) = (k + 2|k) = 2 (k) + d d(k)
+u(k) + u(k + 1) + K()e(k)

y (k + Np ) = (k + Np |k) = Np (k)+Np 1 d d(k)


+Np 1 u(k) + + Np Nc u(k + Nc 1)
+Np 1 K()e(k)
(7)
We collect (7) together in a compact matrix form as:
Y = F(k) + Ee(k) + d d(k) + U

(6)

Fig. 1. Block diagram of the closed loop system.

in which, R1 = E{w(k)w(k)T } and R2 = E{v(k)v(k)T }


are covariance of the w(k) and v(k), respectively. It should
be noted that there is no cross correlation between w(k)
and v(k).

(8)


K()
2
K()
where, F =
, E =
, d =


Np
Np 1

K()
d
d

, and,

Np 1


0
0

0
= 2
. Let

Np 1 Np 2 Np 3 Np Nc

2015 Chinese Automatic Control Society and Wiley Publishing Asia Pty Ltd

Asian Journal of Control, Vol. 18, No. 5, pp. 117, September 2016

control principle, we only take the first element of U at


time k as the incremental control, thus

the objective of MPC be:


J = (Rs Y )T Qy (Rs Y ) + U T RU

(9)

Np

[
]
where in the case of SISO system Rs = 1 1 1 r(k) =
s r(k) and r(k) is set-point signal at time k. In the case
R
of MIMO system (ni input and no output with n states)
the matrix T has dimension ni Nc ni Nc and T F has
s equals the last no columns
dimension ni Nc n, and T R
of T F. The weight matrices R and Qy are a block matrices with ni and no blocks and has their dimensions equal
to the dimension of T and T , respectively. The
]T
[
set-point signal is r(k) = r1 (k) r2 (k) rno (k) as the
no set-point signals to the multi-output system.
To find the optimal U that will minimize J, by
using (8), J is expressed as:
J=
(
(Rs F(k) Ee(k) d d(k))T Qy Rs F(k)
)
Ee(k) d d(k)
2U T T Qy (Rs F(k) Ee(k)
(
)
d d(k)) + U T T Qy + R U
(10)
From the first derivative of the cost function J:
J
= 2T Qy (Rs F(k) Ee(k) d d(k))
U
)
(
+ 2 T Qy + R U,
(11)
the necessary condition of the minimum J is obtained as
J
= 0, from which we find the optimal solution for the
U
control signal as:
(

Qy + R
T

)1

U =
Qy (Rs F(k) Ee(k) d d(k)),
T

(12)
)1 T
(
Q y Rs ,
It is obvious from (12), the T Qy + R
( T
)1 T
( T
)1 T
Qy + R
Qy E and, Qy + R
Qy d
correspond to the set-point change, error of tracking, and
(
)1 T
disturbance, respectively, while, T Qy + R
Qy F
corresponds to the state feedback control within the
framework of predictive control. All depend on the
system parameters, hence are constant matrices for a
time-invariant system. Because of the receding horizon

Nc

[
]
u(k) = I O O

( T
)1 T
s F(k) Ee(k) d d(k)),
Qy + R
Qy (R
= Kr r(k) Ke e(k) Kff d(k) Kmpc (k)
(13)
where I and O are Identity and zeros matrices with
dimension ni ni , Kr , Ke and Kff are the first element of
( T
)1 T
(
)1 T
Qy + R
Q y Rs , T Q y + R
Qy E and,
)1 T
( T
Qy d , respectively, while, Kmpc is the
Qy + R
( T
)1 T
first row of Qy + R
Qy F. It must be emphasized that many available expressions for linear-quadratic
controllers misses the term Ke and Kff that arise due to
observer gain matrix K() and disturbance gain matrix
d in (4), respectively.

IV. GENERALIZATION OF MPC USING


LAGUERRE NETWORK
In essence, the core technique in the design of
discrete-time MPC is based on optimizing the future control trajectory, that is the difference of the control signal,
u(k). By assuming a finite control horizon Nc , the difference of the control signal u(k) for k = 0, 1, 2, , Nc 1
is captured by the control vector U while the rest of
the u(k) for k = Nc , Nc + 1, ..., Np is zero. In some
cases where the neglected trajectory u(k) are not zero,
however, they are small in magnitude. In addition, as a
consequence, in the case of rapid sampling, complicated
process dynamics and/or high demands on closed-loop
performance, satisfactory approximation of the control
signal u may require a very large number of parameters, leading to poorly numerically conditioned solutions
and heavy computational load when implemented
on-line. Instead, a more appropriate approach is to use
orthonormal basis networks in the design of model predictive control. The idea is now to generalize the design
procedure by introducing a set of discrete Laguerre networks into the design. This generalization will help us in

Fig. 2. Discrete Laguerre network.

2015 Chinese Automatic Control Society and Wiley Publishing Asia Pty Ltd

S. H. HosseinNia and M. Lundh: A General Robust MPC Design for the State-Space Model

reformulating the predictive control problem and simplifying the solutions, in addition to tuning the predictive
control system. Furthermore, a long control horizon can
be realized without using a large number of parameters.
4.1 Laguerre filter
There are several orthonormal basis filters that can
be used for development of linear models. The selection
of the appropriate type of filter depends on the dynamic
behaviour of the system to be modelled. In this paper,
we focus on the Laguerre filters. The Laguerre filters are
first-order lag filters with one real pole. They are, therefore, more appropriate for well damped processes. The
Laguerre filters are given by
(
) (1 az)i1
i (z) =
1 a2
, |a| < 1
(14)
(z a)i
where a is the chosen pole which is related to the time constant, , and the sampling interval Ts of the system [32].
Letting li (k) denote the inverse z-transform of i (z, a) this
set of discrete-time Laguerre filters can be expressed in a
vector form as
[
]
L(k) = l1 (k), l2 (k), , lN (k) ,
(15)
where, N is the length of Laguerre network shown in
Fig. 2. Now, let us rewrite (14) in state space as
L(j + 1) = Al L(j),

(16)

0 0 0
a

a 0 0
a

a 0 with dimenwhere, matrix Al =


.
.

0
N2

a a a
sion of (N N) is a function of parameters a and =
(1 a2 ), and the initial condition is given by L(0) =

[
]
1, a, a2 , a3 , , (1)N1 aN1 .
4.2 Design framework
At time k , the control trajectory u(k), u(k +
1), u(k + 2), , u(k + j), is regarded as the impulse
response of a stable dynamic system. Thus, a set of
Laguerre functions, l1 (j), l2 (j), , lN (j) are used to capture this dynamic response with a set of Laguerre coefficients that will be determined from the design process.
More precisely, at an arbitrary future sample instant j,
u(k + j) =

i=1

ci li (j),

(17)

with k being the initial time of the moving horizon


window and j being the future sampling instant; and
ci , i = 1, 2, , N, are the coefficients, and they are functions of the initial time of the moving horizon window,
k. Within this design framework, the control horizon Nc
from the previous approach has vanished. Instead, the
number of terms N is used to describe the complexity of
the trajectory in conjunction with the parameter a. For
instance, a larger value of a can be selected to achieve a
long control horizon with a smaller number of parameters N required in the optimization procedure. We note
that when a = 0, N = Nc , and we recover the approach
of the previous section.
Theorem 1. Consider state-space system (1) and assume
Np
Np
(j)Q(j)T + R, 1 = j=1
(j)Qj , 2 =
= j=1
Np

N
p
(j)Qj1 K(), 3 =
(j)Qj1 d and the
j=1
j=1
vector form of Laguerre approximation of input signal as
u(k + j) = L(j)T ,

(18)

where the parameter


vector ]comprises N Laguerre coef[
ficients = c1 c2 cN . The incremental control
(
)
(
)
u(k) = L(0)T 1 1 (k) L(0)T 1 2 e(k)
(
)
L(0)T 1 3 d(k) is the optimal solution to minimize
(9).
Proof. When using Laguerre functions, the state-space
model (3) with u(.) as the input signal with the initial state variable information (k), the prediction of
the future state variable, (k + j|k) at sampling instant
j, becomes
(k + j|k) =
j (k) + j1 K()e(k) + j1 d d(k) + (j)T ,
(19)
where, we replaced the function u(k + i) by L(i) and
simplified it by
j1 ji1
(j)T =

L(i)T . Similarly, upon obtaining


i=0
the prediction of state variables, the prediction for the
plant output at future sample j is
y (k + j|k) =
j (k) + j1 K()e(k) + j1 d d(k) + (j)T .
(20)
With this formulation, both predictions of state variable
and output variable are expressed in terms of the coefficient vector of the Laguerre network, instead of U as

2015 Chinese Automatic Control Society and Wiley Publishing Asia Pty Ltd

Asian Journal of Control, Vol. 18, No. 5, pp. 117, September 2016

in previous section. Thus, the coefficient vector will be


optimized and computed in the design.
Using (18) , the cost function (9) can be rewritten as:
Np
j=1

J=

Np
Np
where, = j=1
(j)Q(j)T + R, 1 = j=1
(j)Qj ,
Np
Np
(j)Qj1 K(), and 3 = j=1
(j)Qj1 d .
2 = j=1
Upon obtaining the optimal parameter vector , the
receding horizon control law is realized as

(r(k) y (k + j|k)) Qy (r(k) y (k + j|k)) + T R


T

u(k) = L(0)T

(21)
where Qy is a positive definite matrix with proper dimension regarding to the number of output. The cost function (21) is based on the minimization of the error
between the set-point signal and the output signal. The
reasons for this choice include simplicity, practicality of
the cost and its relevance to applications, and its similarity to the classical predictive control systems. Here, the
cost function is re-formulated with a link to discrete-time
linear quadratic regulators (DLQR), where the objective is to find the coefficient vector to minimize the
cost function:
J=

Np

((k + j|k))T Q ((k + j|k)) + T R

(22)

j=1

[
]
(k + j|k) = x m (k + j|k) y (k + j|k) r(k)

u(k) = Kmpc (k) Ke e(k) Kff d(k)

(27)

which is simplified in notation as:


Kmpc = L(0)T 1 1 ,
Ke = L(0)T 1 2 ,
and,

(N
)
p
T + R +
(j)Q(j)
J = T
j=1
(
[ j
])
T Np
j1
j1
2
(j)Q

(k)
+
K()e(k)+

d(k)
d
j=1
)T
Np ( j
+ j=1
(k) + j1 K()e(k) + j1 d d(k) Q
( j
)
(k) + j1 K()e(k) + j1 d d(k) .
(24)
(N

(j)Q(j)T + R
j=1
p

)1
exists, when

= 0, the optimal solution of parameter vector is:


(
)
= 1 1 (k) + 2 e(k) + 3 d(k)

Since the prediction of future states is based on the current information on (k), the set-point information is
contained in (k).
More specifically,
[
]
(k + j|k) = x m (k + j|k) y (k + j|k) r(k) and e(k) =
y(k) y (k|k 1).

(23)

then the cost function for minimization of output errors


(21) is identical to the cost function (22) with Q =
T Qy . It should be emphasized that this inclusion of
the set-point signal in the cost function does not alter the
model used for prediction.
Substituting the (19) in (22) we obtain:

The control u(k) can be written in the form of linear state feedback control ( is a function of state variable
(k) and e(k)). Namely,

Ke = L(0)T 1 3 ,

In order to include the set-point signal in the cost function (22), we need to re-define the state variable (k+j|k).
With the inclusion of a set-point signal r(k), within the
optimization window, if the state variable (k + j|k) is
chosen to be

Assuming that

(26)

(25)

V. LAGUERRE BASED MPC WITH


CONSTRAINT
With parameterization of the control signal trajectory using Laguerre functions, we have the flexibility to choose the locations of the future constraints.
This could potentially reduce the number of constraints
within the prediction horizon, and hence the online
computational load for large-scale systems. In addition,
because of the existing exponential decay factor in the
Laguerre functions, the difference of the control signal
is ensured to converge to zero after the transient period.
Thus, it is sufficient in the majority of cases that the
constraints are imposed in the transient period of the
response. This in turn will reduce the number of constraints [23]. This section will show how to incorporate
constraints in the design and implementation of model
predictive control systems. The approach is based on
application of Hildreths quadratic programming method
(see Appendix I).

2015 Chinese Automatic Control Society and Wiley Publishing Asia Pty Ltd

S. H. HosseinNia and M. Lundh: A General Robust MPC Design for the State-Space Model

5.1 Constraints on the difference of control signals

5.3 Constraints on the outputs of the system

In the MPC design with constraints on the difference of the input variable the aim is to minimize the
cost function:

We can also specify the operating range for the plant


output. For instance, supposing that the future output
y(k + j) has an upper limit vector ymax and a lower limit
ymin , then the output constraints are specified as:

)
(
J = T + 2 T 1 (k) + 2 e(k) + 3 d(k) ,
(28)
such that,
umin u(k + j) umax , j = 0, 1, 2, .

(29)

Since using Laguerre functions in the design, the incremental control signal can be represented as:
u(k + j) = (j), j = 0, 1, 2, ,

(30)

0
0
L1 (j)T
0 L2 (j)T
0
where, (j) =
, j denotes the

0 Lm (j)T
0
set of future time instants at which we wish to impose the
limits on u and m is the number of inputs. Therefore,
the input constraints can be rewritten as:
]
]
[
(j)
umax
, j = 0, 1, 2, .

(j)
umin

(31)

In the case of SISO system (j) = LT (j) as in (18).


5.2 Constraints on the control signals
Noting that the increment of the control signal is
j1
u(j) = i=0 u(i). Similar to constraints on the difference
of the input variable the constraint of input variable can
be represented as:
umin (j) + u(k 1) umax , j = 1, 2, .

(32)

j1 L1 (i)T

0
j1
i=0

T
0
L (i)
0
.
i=0 2
where (j) =

j1

T
L
(i)
0
0

i=0 m
Therefore, the inequality constraints on control signal
can be rewritten in the following form:
[

]
[
]
(j)
umax u(k 1)

, j = 1, 2, . (33)
(j)
umin + u(k 1)

where j is the future sample and u(k 1) is the previous


control signal.

ymin
j (k) + j1 K()e(k) + j1 d d(k) + (j)T
ymax , j = 1, 2,
(34)
In order to avoid the conflict in the constraint of
control or incremental control variables, output constraints are often implemented as soft constraints in
the way that a slack variable sv > 0 is added to the
constraints, forming
(ymin sv ) j (k) j1 K()e(k) j1 d d(k)
(j)T
j
(ymax + sv ) (k) j1 K()e(k) j1 d d(k),
j = 1, 2,
(35)
Finally, the inequality constraints on output can be
rewritten as
[
]
(j)T

(j)T
[
]
j
j1
(ymax + sv ) (k) K()e(k)j1 d d(k)
.
(ymin + sv )+j (k)+j1 K()e(k)+j1 d d(k)
j = 1, 2,
(36)
5.4 Stability analysis
Recall that the model predictive control is established using the principle of receding horizon control.
That is, at current sample time k the future of the control
trajectory u(k + m), m = 0, 1, 2, ..., Np is optimized by
minimizing the cost function (9) subject to constraints,
and upon obtaining the optimal control sequence, only
the first sample u(k) is implemented. At the next sample time k + 1, the same optimization and implementation procedures are repeated. The core to establishing
closed-loop stability is based on an equality constraint on
the terminal state, which is x(k + Np |k) = 0.
Theorem 2. Assuming an additional constraint is placed
on the final state of the receding horizon optimization
problem: x(k + Np |k) = 0, where x(k + Np |k) is the terminal state resulting from the control sequence u(k + m) =
L(m)T , m = 0, 1, 2, ..., Np ; and for each sampling instant,

2015 Chinese Automatic Control Society and Wiley Publishing Asia Pty Ltd

Asian Journal of Control, Vol. 18, No. 5, pp. 117, September 2016

k, there exists a solution such that the cost function J is


minimized subject to the inequality constraints and terminal equality constraint x(k + Np |k) = 0. Subject to the
assumptions, the closed-loop model predictive control
system is asymptotically stable.
Proof. See chapter 3.7 of [23].

VI. ROBUST TUNING OF THE MPC


PARAMETERS
In this section, the MPC parameters will be tuned
minimizing the sensitivity function at low frequencies
as reference signals and disturbances occur. Let us first
define the sensitivity function.
6.1 Sensitivity function

]
Kx Kr , Kx
[
]T

y (k) , then,
Rni n , Kr Rni no and recall (k) = x(k)
rewrite the controller (27) as:
Let us now define Kmpc

u(k) = Kmpc (k) Ke e(k) Kff d(k) + Kr r(k)


= Kmpc (k) Ke (y(k) y (k)) Kff d(k) + Kr r(k)
= K
( mpc (k) K)e (y(k) (k)) Kff d(k) + Kr r(k)
= Kmpc + Ke (k) Ke y(k) Kff d(k) + Kr r(k).
(37)
Recalling (4), equation (37) can be represented in
state-space form:
{
(k + 1) = Ac (k) + Bcy y(k) + Bcr r(k) + Bff d(k),
u(k) = Cc (k) + Dcy y(k) + Dcr r(k) + Dff d(k),
(38)
(

where, Ac = K() + Kmpc + Ke , Bcy = K()


D =
Ke , Bcr = Kr , Bff = BKff , Cc = Kmpc + Ke C,
cy
Ke , Dcr = Kr , Dff = Kff .
The output sensitivity function, S(z), is defined as:

function should be small i.e. S(j) 0 (in MIMO system


singular value decomposition of S(j) should be small).
In addition, for good disturbance rejection, the disturbance inputs should have a negligible effect upon the output. Thus, similar to reference tracking we require large
open loop value and small sensitivity function. Therefore,
since the command following and disturbance rejection
occur at low frequency, we will minimize the sensitivity
function at low frequencies to achieve the best command following and disturbance rejection. Fig. 3 shows
the crossover region and the region for good command
following and disturbance rejection.
Relative stability measures, such as gain and phase
margins, are in general inversely related to the peak value
of |S(j)|, so large peak values are to be avoided in order
to achieve good transient performance. In addition, the
controller parameters are typically matched to the process characteristics and since the process may change, it
is important that the controller parameters are chosen in
such a way that the closed loop system is not sensitive
to variations in process dynamics. Therefore, in order to
maintain the stability against the model mismatch, the
peak value of the sensitivity function (in MIMO systems
peak value in singular value decomposition of sensitivity matrix) should be limited to a certain value while the
sensitivity function being minimized in low frequency.
Prediction horizon Np must be selected greater than
the control horizon, i.e., Np > N to avoid a deadbeat effect. With move suppression present, increasing
Np from this lower limit has a significant effect on
closed loop performance, ranging from damped to underdamped to damped when Np is close to N and both are
small. On the other hand, a larger prediction horizon
improves nominal stability of the closed loop. For practical applications this translates to using a reasonably large
but finite prediction horizon, Np , equal to the open loop

S(z) = (I + G(z)C(z))1
(39)
(
)
where, C(z) = (I z1 )1 Cc (zI Ac )1 Bcy + Dcy and,
G(z) denote the z-transfer function of system and controller, respectively.
6.2 Optimization
For good command following, we require the output to track the input reference signals. To accomplish
this, we desire that y(z) r(z) 0. For this to be true, the
loop G(j)C(j) should be large and therefore, sensitivity

Fig. 3. Typical sensitivity function.

2015 Chinese Automatic Control Society and Wiley Publishing Asia Pty Ltd

S. H. HosseinNia and M. Lundh: A General Robust MPC Design for the State-Space Model

settling time in samples of the slowest subprocess [35].


Length of Laguerre network N can be chosen arbitrary
value less than Np (the smaller value for N, the faster
computation). The observer parameter is set as R1 = I
where is a small positive value. The remaining tuning
parameters i.e. R, Qy , , R2 and a are obtained solving
the constraint optimization problem:
Jt (, R2 , a, R, Qy ) =
)
nu (
(S(j))|

(S
ui (j))|<s ,
<s +
i=1

In the case of having uncertain process these value can be


multiplied by the number of nominal plus uncertain processes. The low frequency s can be set by the operator
in regard to each process.
In addition, in order to avoid searching the scaled
value for R and Qy i.e. R and Qy , > 0 and therefore,
to speed up the optimization we set the first element of R
matrix to 1.

min

,R2 ,a,R,Qy

VII. SOME EXAMPLES


In this section, we study the performance of the
designed MPC as well as the tuning procedure through
two SISO and MIMO examples.

subject to,

< Ms
S(z) = sup (S(j))

Example 1. Level control of a tank.

ui (j)) < Ms , i = 1, , nu
Sui (z) = sup (S
R

(
)
nu | j
no Nf |
|
j
r (k) y i (k)|| + j=1 |ri (k) y ui (k)| < Mi
i=1
k=0 | i
|
|
0<1
0 a Iv
R2 > 0
Qy > 0
R > 0,
(40)

where, Iv is a vector of ones with the length of input, (.)

denotes the singular value decomposition, nu is number


of uncertain systems, Ms is peak of the sensitivity func

tion, S(z) , Sui (z) , i = 1, , nu are the H norm

of the sensitivity function of the nominal system S(z) and


uncertain systems Sui (z) and Nf is final sample. The estimated outputs of the nominal system and uncertain systems are denoted by y (k) and y ui (k). As it can be seen in
the constraint optimization sum of IAE in nominal and
no Nf |
uncertain closed loop systems i=1
r (k) y i (k)||
k=0 | i
no Nf |
|
and i=1 k=0 |ri (k) y ui (k)| are limited by Mi in order
|
|
to manage the time response of the system. One way to
choose the Mi is to guess the desired response of the system and obtain the IAE of the desired response plus a
tolerance value due to indefinite transient response. For
instance, if we consider a SISO system in which the unit
step response of the system is desired to be settled in
15 samples with delay of 5 samples and sample time is
Ts = 2s, then one guess for Mi is
Mi Ts

Settling time-Delay
+ Delay + Indefinite transient
2
(
)
2 155
+
5
+
2
= 24.
2

Let us consider the tank model transfer function


as [41]:
G(s) =

1.12e45s
,
87s + 1

and apply the designed MPC.


The optimization process has been carried out using
the global optimisation toolbox in Matlab. As discussed
before, the prediction horizon is set to the settling time of

+L

system i.e. Np = 2T where, Ts = 5 is sampling time, L


s
is the process delay and is time constant of the FOPDT
systems. Therefore, in this example Np = 34 would be a
reasonably large prediction horizon for the MPC design.
Therefore, applying following constraints:
6 u(k) 6,
2 u(k) 2,
0 sv y(k) 5.5 + sv , sv = 0.1,

)
response ,

2015 Chinese Automatic Control Society and Wiley Publishing Asia Pty Ltd

Fig. 4. Sensitivity function.

(41)

Asian Journal of Control, Vol. 18, No. 5, pp. 117, September 2016

and setting the robustness bound as Ms = 6dB, the limit


for IAE as Mi = 450, final sample as Nf = 300, and
N = 4 as the length of Laguerre network then, solution of
(40) yields the following tuning parameters: = 0.7442,
R2 = 7.6159, a = 0.7352, R = 1.7324, Qy = 7.1314.
The minimum value for cost function i.e. the sensitivity of the system in low frequency (s = 109 ) is

obtained 145.12. Fig. 4 shows that peak of sensitivity


function which is less than the Ms = 6dB.
The simulation results are shown in Fig. 5 where
they are compared with the response of the mismatch
models. In the mismatch case, the time constant in G(s)
are 75% and 125% of the nominal values for the simulation model. In addition, a measurement noise with

Fig. 5. Time response of the system.

Fig. 6. Wet-end overview.


2015 Chinese Automatic Control Society and Wiley Publishing Asia Pty Ltd

S. H. HosseinNia and M. Lundh: A General Robust MPC Design for the State-Space Model

power of 0.1 is considered and a disturbance with amplitude of 5 is applied in time 750s. The tuning is robust,
since the deviations between the nominal and mismatch
case is marginal. Furthermore, it can be concluded, that
the MPC with the selected tuning rejects the disturbance nicely and have good tracking properties. Applying
the controller the poles of the closed loop system not
considering the observer are placed in 0.7486 0.1951i
which verifies the low overshoot in the corresponding
step response of the system shown in Fig. 5.
Example 2. Paper machine control.
System description. A paper machine is a complex process
that requires advanced control to reach the best possible

performance. Different types of control mechanisms are


therefore used. The most important is the machine direction (MD) control, for control of the paper weight, the
ash content in the paper, and the moisture in the paper.
The leftmost part of the paper machine is the head-box,
from which the thin stock is spread out on the wire.
The thin stock concentration is about 0.5%. The rest is
basically water. After the press section, the paper goes
into the drying section where it passes several cylinders
that are heated with steam. After drying, the paper is
rolled onto the reel, which is the rightmost part of the
paper machine. Before the reel there is a measurement
frame with a traversing sensor, that travelles forth and
back over the paper. The measurements are the paper

Fig. 7. Open loop step response of the system.


2015 Chinese Automatic Control Society and Wiley Publishing Asia Pty Ltd

Asian Journal of Control, Vol. 18, No. 5, pp. 117, September 2016

weight, the ash content in the paper, and the moisture in


the paper.
Fig. 6 shows an overview of the wet-end of the paper
machine. In particular it shows how the pulp is prepared
before it enters the head box. Various types of pulp are
mixed in the blend chest, where different chemicals are
added. The pulp flow is controlled with the stock valve,
which is the main control signal for the paper weight.
Before entering the head-box, clay (filler) and various
chemicals are added. The clay addition is used for controlling the ash content in the paper. The thin stock
flow from the head-box onto the wire consist of a huge
amount of water. Most of the water passes through the
wire. The water is collected in the white water silo, and is
then used for dilution of the thick stock flow. However,
this short circulation does not only contain water. Both
fiber and filler are present in the white water recirculation.
The retention, i.e. the amount of material that stays on
the wire, is controlled by the addition of retention chemicals. From another point of view, it controls how much
material that is recirculated in the wet end.
Fig. 6 also gives an overview of the available signals for control. Variables in red are the signals that can
be manipulated, variables in blue are measurements that
may or may not have a set-point, and variables in green
are disturbances that are measured and can be used for
feed-forward compensation. The variables encircled with
dashed lines are used in this study.

Fig. 7 depicts the step response of the system identified in state-space in form (1) with 44 states. The process
variables are:
1. Ash weight, in g/m2 , of the paper at the reel denoted
by y1 .
2. Fiber weight, in g/m2 , of the paper at reel denoted
by y2 .
3. Moisture, in %, in the paper at reel denoted by y3 .
4. Moisture, in %, in the paper between the main bank
dryers and the after dryers denoted by y4 .
The manipulated variables are:
1.
2.
3.
4.

Flow of chalk, used as filter denoted by u1 .


Flow of thick stock denoted by u2 .
Steam pressure for the after dryers denoted by u3 .
Steam pressure for the main bank dryers denoted
by u4 .
The disturbances variables for feed-forward are:

1. Machine chest consistency in g/l.


2. Machine speed in m/min.
Sample time of the identified system is Ts = 6s.
Tuning. The parameters tuned setting Ms = 6, Mi =
1000, s = 109 . It is obvious from Fig. 7 that the
slower subprocess (y1 vs u2 ) settles at about 150s, then
choosing Np = 1506 = 25 will be a fairy large prediction horizon.
The optimization has been carried out using the
gs function (global search) in the global optimization
Toolbox of Matlab and the best results obtained for
both uncertain and nominal systems. The uncertainties applied in the B matrix which are 70% and 130%
of the nominal matrix. Thus, the tuned parameters are
obtained as:

a
R
Qy
R2

Fig. 8. Closed loop poles of the controlled system.

=
=
=
=
=

0.6907,
([
])
diag 0.91751 0.42211 0.71974 0.1001 ,
([
])
diag 1 457616.33 14714.89 123233.14 ,
([
])
diag 1288.07 1153.81 4291.60 4107.106 ,
([
])
diag 1.7335 0.33091 0.81975 0.66713 .
(42)

Fig. 8 shows the closed loop poles of the


controller
system i.e.] eign(Acl ) where, Acl
=
[
K()
0
is the closed loop matrix of the
Kmpc Kmpc
system with the Kaman observer. As can be observed,
the closed loop poles are settled inside the unity cycle
which verifies the stability of the designed controller.

2015 Chinese Automatic Control Society and Wiley Publishing Asia Pty Ltd

S. H. HosseinNia and M. Lundh: A General Robust MPC Design for the State-Space Model

The initial value for process variables are 20g/m2 ,


60g/m2 , 5% and 2% and initial value for the machine
chest consistency and machine speed are 5g/l and
750m/min, respectively. A measurement noise with power
of 0.1 is considered for the outputs of the processes.
In order to see the performance of the designed controller, the setpoints are changed from their initial
values to 22g/m2 , 70g/m2 , 6% and 3%. In addition,
the measured disturbances i.e. the machine chest consistency and machine speed are changed from their
initial value to 5.5g/l and 760m/min, at 1170s and
1560s, respectively.
Simulation results are shown in Fig. 9. As can
be seen, the MPC with the selected parameters rejects
the disturbance nicely and has good tracking properties.

Fig. 10 compares the uncertain and nominal system


which verifies the stability and robustness of the designed
controller. the sum of IAE of the nominal and uncertain system is 886.26 which is less than Mi . The singular value decomposition of sensitivity also shown in
Fig. 11 wherein the peak of the singular value is less than
Ms which ensures the robustness of the system against
model mismatch.

VIII. CONCLUSIONS
In this paper, a general Laguerre-based MPC is
designed for state-space systems and the MPC parameters are tuned through an optimization procedure. The
main advantage of the proposed controller can be summarized as follows:

Fig. 9. Response of the system.


2015 Chinese Automatic Control Society and Wiley Publishing Asia Pty Ltd

Asian Journal of Control, Vol. 18, No. 5, pp. 117, September 2016

Fig. 10. Comparing the response of the uncertain and nominal systems.

Fig. 11. Singular value decomposition of sensitivity.

The MPC is designed based on the state-space model


since it easily handles shifts, time-varying parameters of some static model.

The Kalman state estimator is included in the prediction model to help the controller react faster
against disturbances.
The Laguerre network used in MPC design speeds
up the computation.
The MPC parameters is automatically tuned
using an optimization procedure which makes the
designed controller easy to use.

In this paper we tried to combine all mentioned items


and design an MPC to fulfil the industrial requirements
which are lower computation, applicability to wide range
processes and easy to use. Applying the designed MPC
on paper machine process shows the efficiency of the
proposed method on complicated processes.
Studying/reforming the constraints in tuning
parameters can be considered as a future work to
improve the performance robustness of the controller
and speed up the optimization process.

2015 Chinese Automatic Control Society and Wiley Publishing Asia Pty Ltd

S. H. HosseinNia and M. Lundh: A General Robust MPC Design for the State-Space Model

REFERENCES
1. Qin, S. J. and T. A. Badgwell, An overview of industrial model predictive control technology, Vol. 93,
AICHE Symposium Series, American Institute of
Chemical Engineers, New York, pp. 232256 (1997).
2. Camacho, E. F. and C. A. Bordons, Model Predictive Control in the Process Industry, Springer-Verlag,
New York (1997).
3. De Don, J. A., G. C. Goodwin, and M. M. Seron,
Anti-windup and model predictive control: Reflections and connections, Eur. J. Control, Vol. 6, No. 5,
pp. 467477 (2000).
4. Findeisen, R., L. Imsland, F. Allgower, and B. A.
Foss, State and output feedback nonlinear model
predictive control: An overview, Eur. J.Control,
Vol. 9, No. 2, pp. 190206 (2003).
5. Borrelli, F., A. Bemporad, M. Fodor, and D. Hrovat,
An MPC/hybrid system approach to traction control, IEEE Trans. Control Syst. Technol., Vol. 14,
No. 3, pp. 541552 (2006).
6. Peng, H., K. Nakano, and H. Shioya, Nonlinear predictive control using neural nets-based local
linearization ARX modelstability and industrial
application, IEEE Trans. Control Syst. Technol.,
Vol. 15, No. 1, pp. 130143 (2007).
7. Venkat, A. N., I. A. Hiskens, J. B. Rawlings, and S. J.
Wright, Distributed MPC strategies with application to power system automatic generation control,
IEEE Trans. Control Syst. Technol., Vol. 16, No. 6,
pp. 11921206 (2008).
8. Liu, X., P. Guan, and C. W. Chan, Nonlinear multivariable power plant coordinate control by constrained predictive scheme, IEEE Trans. Control
Syst. Technol., Vol. 18, No. 5, pp. 11161125 (2010).
9. Zhang, L. and J. Wang, A novel multi-step model
predictive control for multi-input systems, Asian J.
Control, Vol. 17, No. 2, pp. 707715 (2015).
10. Widd, A., H. H. Liao, J. C. Gerdes, P. Tunestl,
and R. Johansson, Hybrid model predictive control of exhaust recompression hcci, Asian J. Control,
Vol. 16, No. 2, pp. 370381 (2014).
11. Alexis, K., G. Nikolakopoulos, and A. Tzes, On
trajectory tracking model predictive control of an
unmanned quadrotor helicopter subject to aerodynamic disturbances, Asian J. Control, Vol. 16, No. 1,
pp. 209224 (2014).
12. Lin, C. K., T. H. Liu, L. C. Fu, and Hsiao C. F.
et al. Model-free predictive current control for
interior permanent-magnet synchronous motor
drives based on current difference detection technique, IEEE Trans. Ind. Electron., Vol. 61, No. 2,
pp. 667681 (2014).

13. Garcia, C. E., D. M. Prett, and M. Morari, Model


predictive control: theory and practicea survey,
Automatica, Vol. 25, No. 3, pp. 335348 (1989).
14. Richalet, J., A. Rault, J. Testud, and J. Papon,
Model predictive heuristic control: Applications to
industrial processes, Automatica, Vol. 14, No. 5,
pp. 413428 (1978).
15. Cutler, C. R. and B. L. Ramaker, Dynamic matrix
control-a computer control algorithm, Vol. 1, Proc.
Joint Autom. Control Conf., American Automatic
Control Council Piscataway, NJ, pp. Wp5B (1980).
16. Clarke, D. W., C. Mohtadi, and P. Tuffs, Generalized predictive controlpart I. the basic algorithm,
Automatica, Vol. 23, No. 2, pp. 137148 (1987).
17. Clarke, D., C. Mohtadi, and P. Tuffs, Generalized
predictive controlpart II extensions and interpretations, Automatica, Vol. 23, No. 2, pp. 149160
(1987).
18. Muske, K. R. and J. B. Rawlings, Model predictive control with linear models, AICHE J., Vol. 39,
No. 2, pp. 262287 (1993).
19. Mayne, D. Q., J. B. Rawlings, C. V. Rao, and P. O.
Scokaert, Constrained model predictive control:
Stability and optimality, Automatica, Vol. 36, No. 6,
pp. 789814 (2000).
D., D. Munoz,

20. la Pena,
and P. D. Christofides,
Lyapunov-based model predictive control of nonlinear systems subject to data losses, IEEE Trans.
Autom. Control, Vol. 53, No. 9, pp. 20762089 (2008).

21. Christofides, P. D., J. Liu, and D. M. de la Pena,


Lyapunov-based model predictive control, Networked and Distributed Predictive Control, Springer,
pp. 1345 (2011).
22. De Oliveira Kothare, S. L. and M. Morari, Contractive model predictive control for constrained nonlinear systems, IEEE Trans. Autom. Control, Vol. 45,
No. 6, pp. 10531071 (2000).
23. Wang, L., Model Predictive Control System Design
and Implementation using Matlab Grimble, M. J.
and M. A. Johnson (Eds.), Springer-Verlag, London
(2009).
24. Jrgensen, J. B., J. K. Huusom, and J. B. Rawlings, Finite horizon MPC for systems in innovation
form, 50th IEEE Conf. Decis. Control & Eur. Control
Conf. (CDC-ECC), IEEE, Hilton Orlando Bonnet
Creek Orlando, FL, USA, pp. 18961903 (2011).
25. Huusom, J. K., N. K. Poulsen, S. B. Jrgensen,
and J. B. Jrgensen, Noise modelling and MPC
tuning for systems with infrequent step disturbances, Proc. 18th IFAC World Cong., Milan, Italy,
pp. 1122611232 (2011).
26. Huusom, J. K., N. K. Poulsen, S. B. Jrgensen, and
J. B. Jrgensen, Tuning SISO offset-free model pre-

2015 Chinese Automatic Control Society and Wiley Publishing Asia Pty Ltd

Asian Journal of Control, Vol. 18, No. 5, pp. 117, September 2016

27.

28.

29.

30.

31.

32.

33.

34.

35.

36.

37.

38.

dictive control based on ARX models, J. Process


Control, Vol. 22, No. 10, pp. 19972007 (2012).
HosseinNia, S. H. and M. Lundh, Robust MPC
design using orthonormal basis function for the processes with armax model, 19th IEEE Int. Conf.
Emerg. Technol. Factory Autom., Spain, pp. 18
(2014).
Latawiec, K., C. Marciak, W. Hunek, and R.
Stanisawski, A new analytical design methodology
for adaptive control of nonlinear block-oriented systems, Vol. 11, Proc. 7th World Multi-Conf. Systemics
Cybern. Informatics, Orlando, Florida, pp. 215220
(2003).
Gmez, J. C. and E. Baeyens, Identification of
block-oriented nonlinear systems using orthonormal
bases, J. Process Control, Vol. 14, No. 6, pp. 685697
(2004).
Heuberger, P. S., P. M. Van den Hof, and B.
Wahlberg, Modelling and Identification with Rational Orthogonal Basis Functions, Springer, London
(2005).
Stanisawski, R. and K. J. Latawiec, Orthonormal basis and radial basis functions in modeling
and identification of nonlinear block-oriented systems, Robotics and Automation, Vienna, Austria,
pp. 97888 (2007).
Lemma, D. T., M. Ramasamy, and M. Shuhaimi,
System identification using orthonormal basis filters, J. Appl. Sci., Vol. 10, No. 21, pp. 25162522
(2010).
Rossiter, J. A. and L. Wang, Exploiting laguerre
functions to improve the feasibility/performance
compromise in MPC, Proc. 47th IEEE Conf. Decis.
Control, Cancun, Mexico, pp. 47374742 (2008).
Khan, B., J. A. Rossiter, and G.-5. Valencia-Palomo,
Exploiting kautz functions to improve feasibility in
MPC, Proc. 18th IFAC World Cong., Milan, Italy,
pp. 67776782 (2011).
Shridhar, R. and D. J. Cooper, A tuning strategy for unconstrained multivariable model predictive
control, Ind. Eng. Chem. Res., Vol. 37, No. 10,
pp. 40034016 (1998).
Lee, J. and Z. Yu, Tuning of model predictive controllers for robust performance, Comp. Chem. Eng.,
Vol. 18, No. 1, pp. 1537 (1994).
Rowe, C. and J. Maciejowski, Tuning MPC using
H loop shaping, Vol. 2, Amer. Control Conf.,
IEEE, Chicago, Illinois, USA, pp. 13321336 (2000).
Wojsznis, W., J. Gudaz, T. Blevins, and A. Mehta,
Practical approach to tuning MPC, ISA Trans.,
Vol. 42, No. 1, pp. 149162 (2003).

39. Shah, G. and S. Engell, Tuning MPC for desired


closed-loop performance for mimo systems, Amer.
Control Conf., IEEE, pp. 44044409 (2011).
40. Olesen, D. H., Tuning methods for model predictive
controllers, Ph.D. Thesis (2012).
41. Camacho, E. F. and C. B. Alba, Model Predictive
Control, Springer-Verlag, London (2013).
42. Luenberger, D. G., Optimization by Vector Space
Methods, John Wiley & Sons, United States of America (1969).
43. Wismer, D. A. and R. Chattergy, Introduction to Nonlinear Optimization: A Problem Solving Approach,
North-Holland New York, Amsterdam (1978).

IX. APPENDIX I
9.1 Hildreths quadratic programming
A simple algorithm, called Hildreths quadratic programming procedure ([42,43]), was proposed for solving
this dual problem. In this algorithm, the direction vec[tors were selected ]to be equal to the basis vectors ei =
0 0 1 0 0 . Then, the vector can be varied one
component at a time. At a given step in the process, having obtained a vector 0, we fix our attention on
a single component i . The objective function may be
regarded as a quadratic function in this single component. We adjust i to minimize the objective function.
If that requires i < 0, we set i = 0. In either case,
the objective function is decreased. Then, we consider
the next component i+1 . If we consider one complete
cycle through the components to be one iteration taking the vector m to m+1 , the method can be expressed
explicitly as:
= max(0, m+1
),
m+1
i
i
and

[
m+1
i

1
=
hii

ki +

i1

i=1

(43)

hij m+1
j

]
nhij m
j

(44)

j=i+1

where the scalar hij is the ijth element in the matrix


H = ME 1 M T , and ki is the ith element in the vector
K = b + ME 1 F. Also note that in (44) there are two sets
of values in the computation: one involves m and one
involves the updated m+1 . Because the converged vector contains either zero or positive values of the Lagrange
multipliers, we have:
x = E 1 (F + M T )

(45)

There are a few comments to be made. First, Hildreths quadratic programming algorithm is based on an

2015 Chinese Automatic Control Society and Wiley Publishing Asia Pty Ltd

S. H. HosseinNia and M. Lundh: A General Robust MPC Design for the State-Space Model

element-by-element search, therefore, it does not require


any matrix inversion. As a result, if the active constraints
are linearly independent and their number is less than
or equal to the number of decision variables, then the
dual variables will converge. However, if one or both of
these requirements are violated, then the dual variables
will not converge to a set of fixed values. The iteration will terminate when the iterative counter reaches
its maximum value. Because there is no matrix inversion, the computation will continue without interruption.
This is one of the key strengths of using this approach
in real-time applications, because the algorithms ability to automatically recover from an ill-conditioned
constrained problem is paramount for the safety of
plant operation.
When the conditions are satisfied, the onedimensional search technique in Hildreths quadratic

programming procedure has been shown to converge


to the set of , where contains zeros for inactive
constraints and the positive components corresponding to the active constraints. The positive component collected as a vector is called act with its value
defined by
(
) (
)
T 1
act = Mact E 1 Mact
bact + Mact E 1 F , (46)
where Mact and bact are the constraint data matrix and
vector with the deletion of the row elements that corresponding to the zero elements in . The proof of the
convergence relies on the existence of a set of bounded
act . This is virtually determined by the existence of the
T 1
(Mact E 1 Mact
) [23].

2015 Chinese Automatic Control Society and Wiley Publishing Asia Pty Ltd

Das könnte Ihnen auch gefallen