Model Predictive

Control
edited by
Tao ZHENG

SCIYO
Model Predictive Control
Edited by Tao ZHENG

Published by Sciyo
Janeza Trdine 9, 51000 Rijeka, Croatia

Copyright © 2010 Sciyo

All chapters are Open Access articles distributed under the Creative Commons Non Commercial Share
Alike Attribution 3.0 license, which permits to copy, distribute, transmit, and adapt the work in any
medium, so long as the original work is properly cited. After this work has been published by Sciyo,
authors have the right to republish it, in whole or part, in any publication of which they are the author,
and to make other personal use of the work. Any republication, referencing or personal use of the work
must explicitly identify the original source.

Statements and opinions expressed in the chapters are these of the individual contributors and
not necessarily those of the editors or publisher. No responsibility is accepted for the accuracy of
information contained in the published articles. The publisher assumes no responsibility for any
damage or injury to persons or property arising out of the use of any materials, instructions, methods
or ideas contained in the book.

Publishing Process Manager Jelena Marusic

Technical Editor Sonja Mujacic
Cover Designer Martina Sirotic
Image Copyright Richard Griffin, 2010. Used under license from Shutterstock.com

First published September 2010

Printed in India

A free online edition of this book is available at www.sciyo.com

Additional hard copies can be obtained from [email protected]

Model Predictive Control, Edited by Tao ZHENG

  p.  cm.
ISBN 978-953-307-102-2
SCIYO.COM
WHERE KNOWLEDGE IS FREE
free online editions of Sciyo
Books, Journals and Videos can
be found at www.sciyo.com
Contents
Preface  VII

Chapter 1 Robust Model Predictive Control Design  1

Vojtech Veselý and Danica Rosinová

Chapter 2 Robust Adaptive Model Predictive Control of Nonlinear Systems  25

Darryl DeHaan and Martin Guay

Chapter 3 A new kind of nonlinear model predictive control algorithm

enhanced by control lyapunov functions  59
Yuqing He and Jianda Han

Chapter 4 Robust Model Predictive Control Algorithms for Nonlinear

Systems: an Input-to-State Stability Approach  87
D. M. Raimondo, D. Limon, T. Alamo and L. Magni

Chapter 5 Model predictive control of nonlinear processes  109

Author Name

Chapter 6 Approximate Model Predictive Control for Nonlinear

Multivariable Systems  141
JonasWitt and HerbertWerner

Chapter 7 Multi-objective Nonlinear Model Predictive Control:

Lexicographic Method  167
Tao ZHENG, Gang WU, Guang-Hong LIU and Qing LING

Chapter 8 Model Predictive Trajectory Control for High-Speed Rack Feeders  183
Harald Aschemann and Dominik Schindele

Chapter 9 Plasma stabilization system design on the base of model predictive

control  199
Evgeny Veremey and Margarita Sotnikova

Chapter 10 Predictive Control of Tethered Satellite Systems  223

Paul Williams

Chapter 11 MPC in urban traffic management  251

Tamás Tettamanti, István Varga and Tamás Péni
VI

Chapter 12 Off-line model predictive control of dc-dc converter  269

Tadanao Zanma and Nobuhiro Asano

Chapter 13 Nonlinear Predictive Control of Semi-Active Landing Gear  283

Dongsu Wu, Hongbin Gu, Hui Liu
Preface
Since Model Predictive Heuristic Control (MPHC), the earliest algorithm of Model Predictive
Control (MPC), was proposed by French engineer Richalet and his colleagues in 1978, the
explicit background of industrial application has made MPC develop rapidly to satisfy the
increasing request from modern industry. Different from many other control algorithms, the
research history of MPC is originated from application and then expanded to theoretical field,
while ordinary control algorithms often has applications after sufficient theoretical research.

Nowadays, MPC is not just the name of one or some specific computer control algorithms, but
the name of a specific thought in controller design, from which many kinds of computer control
algorithms can be derived for different systems, linear or nonlinear, continuous or discrete,
integrated or distributed. The basic characters of the thought of MPC can be summarized as
the model used for prediction, the online optimization based on prediction and the feedback
compensation for model mismatch, while there is no special demands on the form of model,
the computational tool for online optimization and the form of feedback compensation.

After three decades’ developing, the MPC theory for linear systems is now comparatively
mature, so its applications can be found in almost every domain in modern engineering. While,
MPC with robustness and MPC for nonlinear systems are still problems for scientists and
engineers. Many efforts have been made to solve them, though there are some constructive
results, they will remain as the focuses of MPC research for a period in the future.

In first part of this book, to present the recent theoretical improvements of MPC, Chapter 1
will introduce the Robust Model Predictive Control and Chapter 2 to Chapter 5 will introduce
some typical methods to establish Nonlinear Model Predictive Control, with more complexity,
MPC for multi-variable nonlinear systems will be proposed in Chapter 6 and Chapter 7.

To give the readers an overview of MPC’s applications today, in second part of the book,
Chapter 8 to Chapter 13 will introduce some successful examples, from plasma stabilization
system to satellite system, from linear system to nonlinear system. They can not only help the
readers understand the characters of MPC, but also give them the guidance for how to use
MPC to solve practical problems.

Authors of this book truly want to it to be helpful for researchers and students, who are
concerned about MPC, and further discussions on the contents of this book are warmly
welcome.
VIII

Finally, thanks to SCIYO and its officers for their efforts in the process of edition and
publication, and thanks to all the people who have made contributes to this book.

Editor

Tao ZHENG
University of Science and Technology of China
Robust Model Predictive Control Design 1

0
1

Robust Model Predictive Control Design

Vojtech Veselý and Danica Rosinová
Institute for Control and Industrial Informatics, Faculty of Electrical Engineering and
Information Technology, Slovak University of Technology, Ilkovičova 3, 81219 Bratislava
Slovak Republic

1. Introduction
Model predictive control (MPC) has attracted notable attention in control of dynamic systems
and has gained the important role in control practice. The idea of MPC can be summarized as
follows, (Camacho & Bordons, 2004), (Maciejovski, 2002), (Rossiter, 2003) :
• Predict the future behavior of the process state/output over the finite time horizon.
• Compute the future input signals on line at each step by minimizing a cost function
under inequality constraints on the manipulated (control) and/or controlled variables.
• Apply on the controlled plant only the first of vector control variable and repeat the
previous step with new measured input/state/output variables.
Therefore, the presence of the plant model is a necessary condition for the development of
the predictive control. The success of MPC depends on the degree of precision of the plant
model. In practice, modelling real plants inherently includes uncertainties that have to be
considered in control design, that is control design procedure has to guarantee robustness
properties such as stability and performance of closed-loop system in the whole uncertainty
domain. Two typical description of uncertainty, state space polytope and bounded unstruc-
tured uncertainty are extensively considered in the field of robust model predictive control.
Most of the existing techniques for robust MPC assume measurable state, and apply plant
state feedback or when the state estimator is utilized, output feedback is applied. Thus, the
present state of robustness problem in MPC can be summarized as follows:
Analysis of robustness properties of MPC.
(Zafiriou & Marchal, 1991) have used the contraction properties of MPC to develop necessary-
sufficient conditions for robust stability of MPC with input and output constraints for SISO
systems and impulse response model. (Polak & Yang, 1993) have analyzed robust stability of
MPC using a contraction constraint on the state.
MPC with explicit uncertainty description.
( Zheng & Morari, 1993), have presented robust MPC schemes for SISO FIR plants, given un-
certainty bounds on the impulse response coefficients. Some MPC consider additive type of
uncertainty, (delaPena et al., 2005) or parametric (structured) type uncertainty using CARIMA
model and linear matrix inequality, (Bouzouita et al., 2007). In (Lovas et al., 2007), for open-
loop stable systems having input constraints the unstructured uncertainty is used. The robust
stability can be established by choosing a large value for the control input weighting matrix R
in the cost function. The authors proposed a new less conservative stability test for determin-
ing a sufficiently large control penalty R using bilinear matrix inequality (BMI). In (Casavola
2 Model Predictive Control

et al., 2004), robust constrained predictive control of uncertain norm-bounded linear systems
is studied. The other technique- constrained tightening to design of robust MPC have been
proposed in (Kuwata et al., 2007). The above approaches are based on the idea of increasing
the robustness of the controller by tightening the constraints on the predicted states.
The mixed H2 /H∞ control approach to design of MPC has been proposed by (Orukpe et al.,
2007) .
Robust constrained MPC using linear matrix inequality (LMI) has been proposed by (Kothare et
al., 1996), where the polytopic model or structured feedback uncertainty model has been used.
The main idea of (Kothare et al., 1996) is the use of infinite horizon control laws which guar-
antee robust stability for state feedback. In (Ding et al., 2008) output feedback robust MPC
for systems with both polytopic and bounded uncertainty with input/state constraints is pre-
sented. Off-line, it calculates a sequence of output feedback laws based on the state estimators,
by solving LMI optimization problem. On-line, at each sampling time, it chooses an appro-
priate output feedback law from this sequence. Robust MPC controller design with one step
ahead prediction is proposed in (Veselý & Rosinová , 2009). The survey of optimal and robust
MPC design can be consulted in (Mayne et al., 2000). Some interesting results for nonlinear
MPC are given in (Janík et al., 2008).
In MPC approach generally, control algorithm requires solving constrained optimization prob-
lem on-line (in each sampling period). Therefore on-line computation burden is significant
and limits practical applicability of such algorithms to processes with relatively slow dynam-
ics. In this chapter, a new MPC scheme for an uncertain polytopic system with constrained
control is developed using model structure introduced in (Veselý et al., 2010). The main con-
tribution of the first part of this chapter is that all the time demanding computations of output
feedback gain matrices are realized off-line ( for constrained control and unconstrained control
cases). The actual value of control variable is obtained through simple on-line computation of
scalar parameter and respective convex combination of already computed matrix gains. The
developed control design scheme employs quadratic Lyapunov stability to guarantee the ro-
bustness and performance (guaranteed cost) over the whole uncertainty domain.
The first part of the chapter is organized as follows. A problem formulation and preliminaries
on a predictive output/state model as a polytopic system are given in the next section. In
Section 1.2, the approach of robust output feedback predictive controller design using linear
matrix inequality is presented. In Section 1.3, the input constraints are applied to LMI feasi-
ble solution. Two examples illustrate the effectiveness of the proposed method in the Section
1.4. The second part of this chapter addresses the problem of designing a robust parameter
dependent quadratically stabilizing output/state feedback model predictive control for linear
polytopic systems without constraints using original sequential approach. For the closed-loop
uncertain system the design procedure ensures stability, robustness properties and guaran-
teed cost. Finally, conclusions on the obtained results are given.
Hereafter, the following notational conventions will be adopted: given a symmetric matrix
P = P T ∈ Rn×n , the inequality P > 0( P ≥ 0) denotes matrix positive definiteness (semi-
definiteness). Given two symmetric matrices P, Q, the inequality P > Q indicates that
P − Q > 0. The notation x (t + k) will be used to define, at time t, k-steps ahead prediction
of a system variable x from time t onwards under specified initial state and input scenario. I
denotes the identity matrix of corresponding dimensions.
Robust Model Predictive Control Design 3

1.1 Problem formulation and preliminaries

Let us start with uncertain plant model described by the following linear discrete-time uncer-
tain system with polytopic uncertainty domain

x ( t + 1) = A ( α ) x ( t ) + B ( α ) u ( t ) (1)

y(t) = Cx (t)
where x (t) ∈ Rn , u(t)
∈ Rm , y(t)
∈ Rl
are state, control and output variables of the system,
respectively; A(α), B(α) belong to the convex set

S = { A(α) ∈ Rn×n , B(α) ∈ Rn×m } (2)

N N N
{ A(α) = ∑ Aj αj B(α) = ∑ Bj α j , α j ≥ 0}, j = 1, 2...N, ∑ α j = 1
j =1 j =1 j =1

Matrices Ai , Bi and C are known matrices with constant entries of corresponding dimensions.
Simultaneously with (1) we consider the nominal model of system (1) in the form

x (t + 1) = Ao x (t) + Bo u(t) y(t) = Cx (t) (3)

where Ao , Bo are any constant matrices from the convex bounded domain S (2). The nominal
model (3) will be used for prediction, while (1) is considered as real plant description provid-
ing plant output. Therefore in the robust controller design we assume that for time t output
y(t) is obtained from uncertain model (1), predicted outputs for time t + 1, ...t + N2 will be
obtained from model prediction, where the nominal model (3) is used. The predicted states
and outputs of the system (1) for the instant t + k, k = 1, 2, ...N2 are given by
• k=1
x (t + 2) = Ao x (t + 1) + Bo u(t + 1) = Ao A(α) x (t) + Ao B(α)u(t) + Bo u(t + 1)
• k=2
x (t + 3) = A2o A(α) x (t) + A2o B(α)u(t) + Ao Bo u(t + 1) + Bo u(t + 2)
• for k
k −1
x (t + k + 1) = Ako A(α) x (t) + Ako B(α)u(t) + ∑ Ako−i−1 Bo u(t + 1 + i) (4)
i =0
and corresponding output is
y(t + k) = Cx (t + k) (5)
Consider a set of k = 0, 1, 2, ..., N2 state/output model predictions as follows

z ( t + 1) = A f ( α ) z ( t ) + B f ( α ) v ( t ), y f (t) = C f z(t) (6)

where
z(t) T = [ x (t) T ...x (t + N2 ) T ], v(t) T = [u(t) T ...u(t + Nu ) T ] (7)
y f (t) T = [y(t) T ...y(t + N2 ) T ]
and  
B(α) 0 ... 0
 Ao B(α) Bo ... 0 
B f (α) =   (8)
 ... ... ... 0 
AoN2 B(α) AoN2 −1 Bo ... AoN2 − Nu Bo
4 Model Predictive Control

   
A(α) 0 ... 0 C 0 ... 0
 Ao A(α) 0 ... 0   0 C ... 0 
A f (α) =  ,C =   (9)
 ... ... ... ...  f  ... ... ... ... 
AoN2 A(α) 0 ... 0 0 0 ... C
where N2 , Nu are output and control prediction horizons of model predictive control, respec-
tively. Note that for output/state prediction in (6) one needs to put A(α) = Ao , B(α) = Bo .
Matrices dimensions are A f (α) ∈ Rn( N2 +1)×n( N2 +1) , B f (α) ∈ Rn( N2 +1)×m( Nu +1) and C f ∈
Rl ( N2 +1)×n( N2 +1) .
Consider the cost function associated with the system (6) in the form
∞
J= ∑ J (t) (10)
t =0

where
J (t) = ∑kN=2 0 x (t + k ) T Qk x (t + k) + ∑kN=u 0 u(t + k) T Rk u(t + k) =

= z(t) T Qz(t) + v(t) T Rv(t) (11)

Q = blockdiag{ Qi }i=0,1,...N2 R = blockdiag{ Ri }i=0,1,...Nu
The problem studied in this part of chapter can be summarized as follows. Design the robust
model predictive controller with output feedback and input constraints in the form
v(t) = Fy f (t) = FC f z(t) (12)

where F T = [ F0T ...FN

T ],
u
Fi = [ Fi0 ...FiN2 ], i = 0, 1, 2, ...Nu
are the output feedback gain matrices which for given prediction horizon N2 and control hori-
zon Nu ensure the closed-loop system (13) stability, robustness and guaranteed cost.
z(t + 1) = ( A f (α) + B f (α) FC f )z(t) = Ac (α)z(t) (13)
Definition 1. Consider the system (6). If there exists a control law v(t)∗ and a positive scalar J ∗
such that the closed-loop system (13) is stable and the closed-loop cost function (10) value J
satisfies J ≤ J ∗ then J ∗ is said to be the guaranteed cost and v(t)∗ is said to be the guaranteed
cost control law for the system (6).
To guarantee closed-loop stability of uncertain system overall the whole uncertainty domain,
the concept of quadratic stability is frequently used. That is, one Lyapunov function works
for the whole uncertainty domain. Experience and analysis has shown that quadratic stabil-
ity is rather conservative in many cases, therefore robust stability with parameter dependent
Lyapunov function P(α) has been introduced by (Peaucelle et al., 2000). Using the concept of
Lyapunov stability it is possible to formulate the following definition and lemma.
Definition 2. System (13) is robustly stable in the convex uncertainty domain with parameter-
dependent Lyapunov function P(α) if and only if there exists a matrix P(α) = P(α) T > 0 such
that
Ac (α) T P(α) Ac (α) − P(α) < 0 (14)
Lemma 1. (Rosinová et al., 2003), (Krokavec & Filasová, 2003) Consider the closed-loop system
(13) with control algorithm (12). Control algorithm (12) is the guaranteed cost control law if
and only if there exists a positive definite matrix P(α) and matrix F such that the following
condition holds
Be = z(t) T ( Ac (α) T P(α) Ac (α) − P(α) + Q + C Tf F T RFC f )z(t) ≤ 0 (15)
Robust Model Predictive Control Design 5

where the first term of (15) ∆V (t) = z(t) T ( Ac (α) T P(α) Ac (α) − P(α))z(t) is the first difference
of closed-loop system Lyapunov function V (t) = z(t) T P(α)z(t). Moreover, summarizing (15)
from initial time to to t → ∞ the following inequality is obtained

− V (to ) + J ≤ 0 (16)

Definition 1 and (16) imply

J ∗ ≤ V (to ) (17)
Note, that as a receding horizon strategy is used, only u(t) is sent to the real plant control,
control inputs u(t + k), k = 0, 1, 2, ..., Nu are used for predictive outputs y(t + k) calculation.
According to (de Oliviera et al., 2000) there is no general and systematic way to formally deter-
mine P(α) in (15) as a function of Ac (α). Such a matrix P(α) is called the parameter dependent
Lyapunov matrix (PDLM) and for particular structure of P(α) the inequality (15) defines the
parameter dependent quadratic stability (PDQS). Formal approach to choose P(α) for real
convex polytopic uncertainty (2) can be found in the references. One of the approaches is to
take P(α) = P, in this case if the solution is feasible the quadratic stability is obtained. An-
other possibility P(α) = ∑iN=1 Pi αi , ∑iN=1 αi = 1, Pi = PiT > 0 gives the parameter dependent
quadratic stability (PDQS). To decrease the conservatism of PDQS arising from affine parame-
ter dependent Lyapunov function (PDLF), recently, the use of polynomial PDLF (PPDLF) has
been proposed in different forms. For more details see (Ebihara et al., 2006).

1.2 Robust model predictive controller design. Quadratic stability

Robust MPC controller design which guarantees quadratic stability and guaranteed cost of
closed-loop system is based on (15). Using Schur complement formula inequality (15) can be
rewritten to following bilinear matrix inequality (BMI).
 
− P(α) + Q C Tf F T Ac (α) T
 
 FC f − R −1 0 ≤0 (18)
Ac (α) 0 − P(α) − 1

For the quadratic stability P(α) = P = P T > 0 in (18). Using linearization approach for P−1 ,
de Oliviera et al. (2000), the following inequality can be derived

− P−1 ≤ Yk−1 ( P − Yk )Yk−1 − Yk−1 = lin(− P−1 ) (19)

where Yk , k = 1, 2, ... in iteration process Yk = P. We can recast bilinear matrix inequality (18)
to the linear matrix inequality (LMI) using linearization (19). The following LMI is obtained
for quadratic stability
 
−P + Q C Tf F T A Tfi + C Tf F T B Tfi
 
 FC f − R −1 0  ≤ 0 i = 1, 2, ...N (20)
A f i + B f i FC f 0 lin(− P ) − 1

where
N N
A f (α) = ∑ A f j αj B f (α) = ∑ B f j αj
j =1 j =1

We can conclude that if the LMIs (20) are feasible with respect to  ∗ I > P = P T > 0 and
matrix F then the closed-loop system with control algorithm (12) is quadratically stable with
6 Model Predictive Control

guaranteed cost (17). Note that due to control horizon strategy only the first m rows of ma-
trix F are used for real plant control, the other part of matrix F serves for predicted output
variables calculation. Parameter dependent or Polynomial parameter dependent quadratic
stability approach to design robust MPC may decrease the conservatism of quadratic stability.
In this case for PDQS we can use the approaches given in (Peaucelle et al., 2000), (Grman et
al., 2005) and for (PPDLF) see (Ebihara et al., 2006).

1.3 MPC design for input constraints

In this subsection we propose the off-line calculation of two control gain matrices and using
analogy to SVSC approach, (Adamy & Fleming, 2004), we significantly reduce the computa-
tional effort for MPC suboptimal control with input constraints.
To design model predictive control (Adamy & Fleming, 2004), (Camacho & Bordons, 2004)
with constraints on input, state and output variables at each sampling time, starting from the
current state, an open-loop optimal control problem is solved over the defined finite horizon.
The first element of the optimal control sequence is applied to the plant. At the next time step,
the computation is repeated with new measured variables. Thus, the implementation of the
MPC strategy requires a QP solver for the on-line optimization which still requires significant
on-line computational effort, which limits MPC applicability.
In our approach the actual output feedback control gain matrix is computed as a convex com-
bination of two gain matrices computed a priori (off-line) : one for constrained and one for
unconstrained case such that both gains guarantee performance and robustness properties
of closed-loop system. This convex combination is determined by a scalar parameter which
is updated on-line in each step. Based on this idea, in the following, the algorithm for con-
strained control algorithm is developed.
Consider the system (6) where the control v(t) is constrained to evolve in the following set

Γ = {v ∈ RmNu : |vi (t)| ≤ Ui , i = 1, ...mNu } (21)

The aim of this part of chapter is to design the stabilizing output feedback control law for
system (6) in the form
v(t) = FC f z(t) (22)
which guarantees that for the initial state z0 ∈ Ω( P) = {z(t) : z(t) T Pz(t) ≤ θ } control v(t)
belongs to the set (21) for all t ≥ 0, where θ is a positive real parameter which determines the
size of Ω( P). Furthermore, Ω( P) should be such that all z(t) ∈ Ω( P) provide v(t) satisfying
the relation (21), restricting the values of the control parameters. Moreover, the following
ellipsoidal Lyapunov function level set

Ω( P) = {z(t) ∈ RnN2 : z(t) T Pz(t) ≤ θ } (23)

can be proven to be a robust positively invariant region with respect to motion of the closed-
loop system in the sense of the following definition, (Rohal-Ilkiv, 2004), (Ayd et al., 2008) .
Definition 3. A subset So ∈ R(nN2 ) is said to be positively invariant with respect to motion of
system (6) with control algorithm (22) if for every initial state z(0) inside So the trajectory z(t)
remains in So for all t ≥ 0.
Consider that vector f i denotes the i-th row of matrix F and define

L( F ) = {z(t) ∈ R(nN2 ) : | f i C f z(t)| ≤ Ui , i = 1, 2...mNu }

Robust Model Predictive Control Design 7

The above set can be rewritten as

L( F ) = {z(t) ∈ R(nN2 ) : | Di FC f z(t)| ≤ Ui , i = 1, 2...mNu } (24)

where Di ∈ R1×mNu = {dij}, dij = 1, i = j, dij = 0, i = j. The results are summarized in

the following theorem.
Theorem 1. The inclusion Ω( P) ⊆ L( F ) is for output feedback control equivalent to
 
P C T F T DiT
≥ 0, i = 1, 2, ...mNu (25)
Di FC λi

where
Ui2
λi ∈< 0, >
θ

Proof. To prove that the inclusion Ω( P) ⊆ L( F ) is equivalent to (25) we use S− procedure in

the following way. Rewrite (23) and (24) to the following form

p(z) = z T (t) Pz(t) − θ ≤ 0

gi (z) = z T (t)C Tf F T DiT Di FCz(t) − Ui2 ≤ 0

According to S− procedure the above inclusion is equivalent to the existence of a positive
scalar λi such that
gi ( z ) − λ i p ( z ) ≤ 0
or equivalently
z(t) T (C Tf F T DiT Di FC − λi P)z(t) − Ui2 + λi θ ≤ 0 (26)
After some manipulation (26) can be rewritten in the form
 T T T 
C F Di Di FC − λi P 0
2 ≤0 (27)
0 −Ui + λi θ

i = 1, 2, ...mNu
The above inequality for block diagonal matrix is equivalent to two inequalities. Using Schur
complement formula for the first one the inequality (25) is obtained, which proves the theo-
rem.
In order to check the value of θi for i − th input we solve the optimization problem z(t) T Pz(t) →
max, subject to constraints (24), which yields

Ui2
θi = (28)
Di FCP 1 C T F T DiT
−

In the design procedure it should be verified that when parameter θ decreases the obtained
robust positively invariant regions Ω( P) are nested to region obtained for θ + ,  > 0.
Assume that we calculate two output feedback gain matrices: F1 for unconstrained case and F2
for constrained one. Obviously, closed-loop system with the gain matrix F2 gives the dynamic
behavior slower than the one obtained for F1 . Consider the output feedback gain matrix F in
the form
F = γF1 + (1 − γ) F2 , γ ∈ (0, 1) (29)
8 Model Predictive Control

For gain matrices Fi , i = 1, 2 we obtain two closed-loop system in the form (13), Aci = A f +
B f Fi C f , i = 1, 2. Consider the edge between Ac1 and Ac2 , that is

Ac = αAc1 + (1 − α) Ac2 , α ∈< 0, 1 > (30)

The following lemma gives the stability conditions for matrix Ac (30).
Lemma 2. Consider the stable closed-loop system matrices Aci , i = 1, 2.
• If there exists a positive definite matrix Pq such that
T
Aci Pq Aci − Pq ≤ 0, i = 1, 2 (31)

then matrix Ac (30) is quadratically stable.

• If there exist two positive definite matrices P1 , P2 such that they satisfy the parameter
dependent quadratic stability conditions, see (Peaucelle et al., 2000), (Grman et al., 2005)
the closed-loop system Ac is parameter dependent quadratically stable (PDQS).
Remarks
• If closed-loop matrices Aci , i = 1, 2 satisfy (31) the scalar γ in (29) may be changed
with any rate without violating the closed-loop stability.
• If closed-loop matrices Aci , i = 1, 2 are PDQS, the scalar γ in (29) has to be constant
but may be unknown.
• The proposed control algorithm (29) is similar to Soft Variable-Structure Control (SVSC),
(Adamy & Fleming, 2004), but in our case, when |vi | << Ui the feedback gain matrix
F (29) gives rather quicker dynamic behavior of the closed-loop system (unconstrained
case) than when |vi | approaches to Ui .
Algorithm for calculation of γ for (29) may be as follows:

Ui − |vi |
γ = min (32)
i Ui

If accidentally some |vi | > Ui , γ = 0.

The resulting control design procedure is given by the next steps
• Off-line computation stage, compute output feedback gain matrices:
F1 for unconstrained case as a solution to (20), where LMI (20) is solved for unknown P
and F;
F2 for constrained case as a solution to (20) and (25).
• On-line computation- in each step:
compute the actual value of scalar parameter γ, e.g from (32) (where vi is obtained from
(12) for F = F1 ;
compute the actual feedback gain matrix from (29) and respective constrained control
vector from (12). All on-line computations follow general MPC scheme, i.e. the first
part of computed control vector u(t) is applied on real controlled plant and the other
part of control vector is used for model prediction.
Robust Model Predictive Control Design 9

1.4 EXAMPLES
Two examples are presented to illustrate the qualities of the control design procedure pro-
posed above, namely its ability to cope with robust stability and input constraints without
complex computational load. In each example the results of three simulation experiments are
compared for closed-loop with output feedback control:
case 1 Unconstrained case for output feedback gain matrix F1
case 2 Constrained case for output feedback gain matrix F2
case 3 The new proposed control algorithm (29) for output feedback gain matrix F.
The input constraint case is studied, in each case maximal value of u(t) is checked; stability is
assessed using spectral radius of closed-loop system matrix.
First example serves as a benchmark. The model of double integrator turns to (1) where
 
1 0
Ao =
1 1
 
1  
Bo = ,C = 0 1
0
and uncertainty matrices are  
0.01 0.01
A1u =
0.02 0.03
 
0.001
B1u = ,
0
For the case when number of uncertainty is p = 1, the number of the respective polytope
vertices is N = 2 p = 2, the matrices (2) are calculated as follows
A1 = Ao − A1u , A2 = Ao + A1u , B1 = Bo − B1u , B2 = Bo + B1u
For the parameters:  = 20000, N2 = 6, Nu = 6, Q0 = 0.1I, Q1 = 0.5I, Q2 = ... = Q6 = I, R =
I, the following results are obtained for unconstrained and constrained cases
• Unconstrained case: Closed − loopmaxeig = 0.8495. Maximal value of control variable
is about umax = 0.24.
• Constrained case with Ui = 0.1, θ = 1000, Closed − loopmaxeig = 0.9437. Maximal
value of control variable is about umax = 0.04.
Closed-loop step responses for unconstrained and constrained cases are given in Fig.1 and
Fig.2, respectively. Closed-loop step responses for the case of in this chapter proposed algo-
rithm are given in Fig.3. Maximal value of control variable is about umax = 0.08 < 0.1.
Input constraints conditions were applied only for plant control variable u(t).
Second example has been borrowed from (Camacho & Bordons (2004), p.147). The model cor-
responds to the longitudinal motion of a Boeing 747 airplane. The multivariable process is
controlled using a predictive controller based on the output model of the aircraft. Two of the
usual command outputs that must be controlled are airspeed that is, velocity with respect to
air, and climb rate. Continuous model has been converted to discrete time one with sampling
time of 0.1s, the nominal model turns to (1) where
 
.9996 .0383 .0131 −.0322
 −.0056 .9647 .7446 .0001 
Ao =  .002


−.0097 .9543 0
.0001 −.0005 .0978 1
10 Model Predictive Control

Fig. 1. Dynamic behavior of controlled system for unconstrained case for u(t).

 
.0001 .1002  
 −.0615 .0183 

Bo =  C = 1 0 0 0
−.1133 .0586  0 −1 0 7.74
−.0057 .0029
and model uncertainty matrices are
 
0 0 0 0
 0 0.0005 0.0017 0 
A1u 
= 
0 0 0.0001 0 
0 0 0 0
 
0 0.12
 −0.02 0.1 
B1u =  10−3
 −0.12 0 
0 0
For the case when number of uncertainty is p = 1, the number of vertices is N = 2 p = 2, the
matrices (2) are calculated as in example 1. Note that nominal model Ao is unstable. Consider
N2 = Nu = 1,  = 20000 and weighting matrices Q0 = Q1 = 1I, R0 = R1 = I the following
results are obtained:
• Unconstrained case: maximal closed-loop nominal model eigenvalue is Closed − loopmaxeig =
0.9983. Maximal value of control variables are about u1max = 9.6, u2max = 6.3.
• Constrained case with Ui = 1, θ = 40000 Closed − loopmaxeig = 0.9998 Maximal values
of control variables are about u1max = 0.21, u2max = 0.2.
Closed-loop nominal model step responses of the above two cases for the input u(t) are given
in the Fig.4 and Fig.5, respectively. Closed-loop step responses for in the paper proposed
control algorithm (29) and (32) are in Fig.6. Maximal values of control variables are about
u1max = 0.75 < 1, u2max = 0.6 < 1. Input constraint conditions were applied only for plant
control variable u(t). Both examples show that using tuning parameter θ the demanded input
Robust Model Predictive Control Design 11

Fig. 2. Dynamic behavior of controlled system for constrained case for u(t).

constraints can be reached with high accuracy. The initial guess of θ can be obtained from (28).
It can be seen that the proposed control scheme provides reasonable results : the response in
case 3 (Fig.3 , Fig. 6) are quicker than those in case 2 (Fig.2, Fig.5), while the computation load
has not much increased comparing to case 2.

2. ROBUST MPC DESIGN: SEQUENTIAL APPROACH

2.1 INTRODUCTION
In this part a new MPC algorithm is proposed pursuing the idea of (Veselý & Rosinová ,
2009). The proposed robust MPC control algorithm is designed sequentially. The respec-
tive sequential robust MPC design procedure consists from two steps. In the first step and
one step ahead prediction horizon, the necessary and sufficient robust stability conditions
have been developed for MPC and polytopic system with output feedback, using generalized
parameter dependent Lyapunov matrix P(α). The proposed robust MPC algorithm ensures
parameter dependent quadratic stability (PDQS) and guaranteed cost. In the second step of
design procedure, the nominal plant model is used to design the predicted input variables
u(t + 1), ...u(t + N2 − 1) so that the robust closed-loop stability of MPC and guaranteed cost
are ensured. Thus, input variable u(t) guarantees the performance and robustness proper-
ties of closed-loop system and predicted input variables u(t + 1), ...u(t + N2 − 1) guarantee
the performance and closed-loop stability of uncertain plant model and nominal model pre-
diction. Note that within sequentially design procedure the degree of plant model does not
change when the output prediction horizon changes.
This part of chapter is organized as follows: Section 2.2 presents preliminaries and problem
formulation. In Section 2.3 the main results are given and finally, in Section 2.4 two examples
solved using Yalmip BMI solvers show the effectiveness of the proposed method.
12 Model Predictive Control

Fig. 3. Dynamic behavior of controlled system with the proposed algorithm for u(t) .

2.2 PROBLEM FORMULATION AND PRELIMINARIES

For readers convenience, uncertain plant model and respective preliminaries are briefly re-
called. A time invariant linear discrete-time uncertain polytopic system is

x ( t + 1) = A ( α ) x ( t ) + B ( α ) u ( t ) (33)

y(t) = Cx (t)
where x (t) ∈ Rn , u(t) ∈ Rm , y(t) ∈ Rl are state, control and output variables of the system,
respectively; A(α), B(α) belong to the convex set

S = { A(α) ∈ Rn×n , B(α) ∈ Rn×m } (34)

N N N
{ A(α) = ∑ Aj αj B(α) = ∑ Bj α j , α j ≥ 0}, j = 1, 2...N, ∑ α j = 1
j =1 j =1 j =1
Matrix C is constant known matrix of corresponding dimension. Jointly with the system (33),
the following nominal plant model will be used.

x (t + 1) = Ao x (t) + Bo u(t) (35)

y(t) = Cx (t)
where ( Ao , Bo ) ∈ S are any matrices with constant entries. The problem studied in this part
of chapter can be summarized as follows: in the first step, parameter dependent quadratic
stability conditions for output feedback and one step ahead robust model predictive control
are derived for the polytopic system (33), (34), when control algorithm is given as

u(t) = F11 y(t) + F12 y(t + 1) (36)

and in the second step of design procedure, considering a nominal model (35) and a given
prediction horizon N2 a model predictive control is designed in the form:

u(t + k − 1) = Fkk y(t + k − 1) + Fkk+1 y(t + k) (37)

Robust Model Predictive Control Design 13

Fig. 4. Dynamic behavior of unconstrained controlled system for u(t) .

where Fki ∈ Rm×l , k = 2, 3, ...N2 ; i = k + 1 are output (state) feedback gain matrices to be
determined so that cost function given below is optimal with respect to system variables. We
would like to stress that y(t + k − 1), y(t + 1) are predicted outputs obtained from predictive
model (44).
Substituting control algorithm (36) to (33) we obtain

x (t + 1) = D1 ( j) x (t) (38)

where
D1 ( j) = A j + Bj K1 ( j)
K1 ( j) = ( I − F12 CBj )−1 ( F11 C + F12 CA j ), j = 1, 2, ...N
For the first step of design procedure, the cost function to be minimized is given as
∞
J1 = ∑ J1 (t) (39)
t =0

where
J1 (t) = x (t) T Q1 x (t) + u(t) T R1 u(t)
and Q1 , R1 are positive definite matrices of corresponding dimensions. For the case of k = 2
we obtain
u(t + 1) = F22 CD1 ( j) x (t) + F23 C ( Ao D1 ( j) x (t) + Bo u(t + 1))
or
u ( t + 1 ) = K2 ( j ) x ( t )
and closed-loop system

x (t + 2) = ( Ao D1 ( j) + Bo K2 ( j)) x (t) = D2 ( j) x (t), j = 1, 2, ...N

14 Model Predictive Control

Fig. 5. Dynamic behavior of constrained controlled system for u(t) .

Sequentially, for k = N2 ≥ 2 step prediction, we obtain the following closed-loop system

x (t + k) = ( Ao Dk−1 ( j) + Bo Kk ( j)) x (t) = Dk ( j) x (t) (40)

where
D0 = I, Dk ( j) = Ao Dk−1 ( j) + Bo Kk ( j) k = 2, 3, ..., N2 ; j = 1, 2, ...N
−1
Kk ( j) = ( I − Fkk+1 CBo ) ( Fkk C + Fkk+1 CAo ) Dk−1 ( j)
For the second step of robust MPC design procedure and k prediction horizon the cost function
to be minimized is given as
∞
Jk = ∑ Jk (t) (41)
t =0
where
Jk (t) = x (t) T Qk x (t) + u(t + k − 1) T Rk u(t + k − 1)
and Qk , Rk , k = 2, 3, ...N2 are positive definite matrices of corresponding dimensions. We
proceed with two corollaries following from Definition 2 and Lemma 1.
Corollary 1
The closed-loop system matrix of discrete-time system (1) is robustly stable if and only if
there exists a symmetric positive definite parameter dependent Lyapunov matrix 0 < P(α) =
P(α) T < I such that
− P(α) + D1 (α) T P(α) D1 (α) ≤ 0 (42)
where D1 (α) is the closed-loop polytopic system matrix for system (33). The necessary and
sufficient robust stability condition for closed-loop polytopic system with guaranteed cost is
given by the recent result (Rosinová et al., 2003).
Corollary 2
Consider the system (33) with control algorithm (36). Control algorithm (36) is the guaranteed
cost control law for the closed-loop system if and only if the following condition holds

Be = D1 (α) T P(α) D1 (α) − P(α) + Q1 + ( F11 C + F12 CD1 (α)) T R1 ( F11 C + (43)
Robust Model Predictive Control Design 15

Fig. 6. Dynamic behavior for proposed control algorithm (29) and (32) for u(t) .

+ F12 CD1 (α)) ≤ 0

For the nominal model and k = 1, 2, ...N2 the model prediction can be obtained in the form

z ( t + 1) = A f z ( t ) + B f v ( t ) (44)

y f (t) = C f z(t)
where
z(t) T = [ x (t) T ...x (t + N2 − 1) T ]
v(t) T = [u(t) T ...u(t + N2 − 1) T ]
y f (t) T = [y(t) T ...y(t + N2 − 1) T ]
 
Ao 0 0 ... 0
 Ao D1 0 0 ... 0 
  nN2 ×nN2
Af = 
 Ao D2 0 0 ... 0  ∈R
 ... ... ... ... ... 
Ao D N2 −1 0 0 ... 0
B f = blockdiag{ Bo }nN2 ×mN2
C f = blockdiag{C }lN2 ×nN2
Remarks
• Control algorithm for k = N2 is u(t + N2 − 1) = FN2 N2 y(t + N2 − 1).
• If one wants to use control horizon Nu < N2 (Camacho & Bordons, 2004), the control
algorithm is u(t + k − 1) = 0, Kk = 0, FNu+1 Nu+1 = 0, FNu+1 Nu+2 = 0 for k > Nu .
• Note that model prediction (44) is calculated using nominal model (35), that is D0 =
I, Dk = Ao Dk−1 + Bo Kk , Dk ( j) is used robust controller design procedure.
16 Model Predictive Control

2.3 MAIN RESULTS

2.3.1 Robust MPC controller design. First step
The main results for the first step of design procedure can be summarized in the following
theorem.
Theorem 2.
The system (33) with control algorithm (36) is parameter dependent quadratically stable with
parameter dependent Lyapunov function V (t) = x (t) T P(α) x (t) if and only if there exist ma-
trices N11 , N12 , F11 , F12 such that the following bilinear matrix inequality holds.
 
G11 G12
Be = T ≤0 (45)
G12 G22
where
T
G22 = N12 Ac (α) + Ac (α) T N12 − P(α) + Q1 + C T F11
T
R1 F11 C
T
G12 = Ac (α) T N11 + N12
T
Mc (α) + C T F11
T
R1 F12 C
T
G11 = N22 Mc (α) + Mc (α) T N22 + C T F12
T
R1 F12 C + P(α)
Mc (α) = B(α) F12 C − I
Ac (α) = A(α) + B(α) F11 C
Note that (45) is affine with respect to α. Substituting (34) and P(α) = ∑iN=1 αi Pi to (45) the
following BMI is obtained for the polytopic system
 
G11i G12i
Bie = T ≤ 0 i = 1, 2, ...N (46)
G12i G22i
where
T T
G11i = N22 Mci + Mci N22 + C T F12
T
R1 F12 C + Pi
T T T
G12i = Aci N22 + N12 Mci + C T F11
T
R1 F12 C
T T
G22i = N12 Aci + Aci N12 − Pi + Q1 + C T F11
T
R1 F11 C
Mci = Bi F12 C − I Aci = Ai + Bi F11 C
Proof. For the proof of this theorem see the proof of Theorem 3 .
If the solution of (46) is feasible with respect to symmetric matrices Pi = PiT > 0, i = 1, 2...N,
and matrices N11 , N12 , within the convex set defined by (34), the gain matrices F11 , F12 ensure
the guaranteed cost and parameter dependent quadratic stability (PDQS) of closed-loop poly-
topic system for one step ahead predictive control.
Note that:
• For concrete matrix P(α) = ∑iN=1 αi Pi BMI robust stability conditions "if and only if" in
(45) reduces in (46) to BMI conditions " if".
• If in (46) Pi = Pj = P, i = j = 1, 2...N, the feasible solution of (46) with respect to
matrices N11 , N12 , and symmetric positive definite matrix P gives the gain matrices
F11 , F12 guaranteeing quadratic stability and guaranteed cost for one step ahead pre-
dictive control for the closed-loop polytopic system within the convex set defined by
(34). Quadratic stability gives more conservative results than PDQS. Conservatism of
real results depend on the concrete examples.
Assume that the BMI solution of (46) is feasible, then for nominal plant one can calculate
matrices D1 and K1 using (38). For the second step of MPC design procedure, the obtained
nominal model will be used.
Robust Model Predictive Control Design 17

2.3.2 Model predictive controller design. Second step

The aim of the second step of predictive control design procedure is to design gain matrices
Fkk , Fkk+1 , k = 2, 3, ...N2 such that the closed-loop system with nominal model is stable with
guaranteed cost. In order to design model predictive controller with output feedback in the
second step of design procedure we proceed with the following corollary and theorem.
Corollary 3
The closed-loop system (40) is stable with guaranteed cost iff the following inequality holds

Bek (t) = ∆Vk (t) + x (t) T Qk x (t) + u(t + k − 1) T Rk u(t + k − 1) ≤ 0 (47)

where ∆Vk (t) = Vk (t + k) − Vk (t) and Vk (t) = x (t) T Pk x (t), Pk = PkT > 0, k = 2, 3, ...N2 .
Theorem 3
The closed-loop system (40) is robustly stable with guaranteed cost iff for k = 2, 3, ...N2 there
exist matrices
Fkk , Fkk+1 , Nk1 ∈ Rn×n , Nk2 ∈ Rn×n
and positive definite matrix Pk = PkT ∈ Rn×n such that the following bilinear matrix inequality
holds  
Gk11 Gk12
Be2 = T ≤0 (48)
Gk12 Gk22
where
T T
Gk11 = Nk1 Mck + Mck Nk1 + C T Fkk
T
+1 Rk Fkk+1 C + Pk

T
Gk12 = Dk−1 ( j) T C T Fkk
T
Rk Fkk+1 C + Dk−1 ( j) T Ack
T T
Nk1 + Nk2 Mck

Gk22 = Qk − Pk + Dk−1 ( j) T C T Fkk

T
Rk Fkk CDk−1 ( j)
T
+ Nk2 Ack Dk−1 ( j) + Dk−1 ( j) T Ack
T
Nk2
and
Mck = B0 Fkk+1 C − I; Ack = A0 + B0 Fkk C
Dk ( j) = A0 Dk−1 ( j) + B0 Kk ( j)
Kk ( j) = ( I − Fkk+1 CB0 )−1 ( Fkk C + Fkk+1 CA0 ) Dk−1 ( j), j = 1, 2, ...N
Proof. Sufficiency.
The closed-loop system (40) can be rewritten as follows

x (t + k) = −( Mck )−1 Ack Dk−1 ( j) x (t) = Aclk x (t) (49)

Since the matrix (j is omitted)

UkT = [− DkT−1 Ack

T
( Mck )−1 I]

has full row rank, multiplying (48) from left and right side the inequality equivalent to (47) is
obtained. Multiplying the results from left by x (t) T and right by x (t), taking into account the
closed-loop matrix (49), the inequality (47) is obtained, which proves the sufficiency.
Necessity.
Suppose that for k-step ahead model predictive control there exists such matrix 0 < Pk =
18 Model Predictive Control

PkT < Iρ that (48) holds. Necessarily, there exists a scalar β > 0 such that for the first difference
of Lyapunov function in (47) holds
T T
Aclk Pk Aclk − Pk ≤ − β( Aclk Aclk ) (50)

The inequality (50) can be rewritten as

T
Aclk ( Pk + βI ) Aclk − Pk ≤ 0

Using Schur complement formula we obtain

 T ( P + βI )

− Pk − Aclk k ≤0 (51)
( Pk + βI ) Aclk −( Pk + βI )

taking
Nk1 = −( Mck )−1 ( Pk + βI/2)
T −1 T −1
Nk2 = − DkT−1 Ack
T
( Mck ) Mck β/2
one obtains
T
− Aclk ( Pk + βI ) = DkT−1 Ack
T T
Nk1 + Nk2 Mck
T
− Pk = − Pk + Nk2 Ack Dk−1 + DkT−1 (52)
T −1 T −1
Ack Nk2 + β( DkT−1 Ack
T
( Mck ) Mck Ack Dk−1 )
−( Pk + βI ) = 2Mck Nk1 + Pk
Substituting (52) to (51) for β → 0 the inequality (48) is obtained for the case of Qk = 0, Rk = 0.
If one substitutes to the second part of (47) for u(t + k − 1) from (37), rewrites the obtained
result to matrix form and takes sum of it with the above matrix, inequality (48) is obtained,
which proves the necessity. It completes the proof.
If there exists a feasible solution of (48) with respect to matrices Fkk , Fkk+1 , Nk1 ∈ Rn×n , Nk2 ∈
Rn×n , k = 2, 3, ...N2 and positive definite matrix Pk = PkT ∈ Rn×n , then the designed MPC
ensures quadratic stability of the closed-loop system and guaranteed cost.

Remarks
• Due to the proposed design philosophy, predictive control algorithm u(t + k), k ≥ 1 is
the function of corresponding performance term (39) and previous closed-loop system
matrix.
• In the proposed design approach constraints on system variables are easy to be imple-
mented by LMI using a notion of invariant set (Ayd et al., 2008), (Rohal-Ilkiv, 2004) (see
Section 1.3).
• The proposed MPC with sequential design is a special case of classical MPC. Sequential
MPC may not provide "better" dynamic behavior than classical one but it is another
approach to the design of MPC.
• Note that in the proposed MPC sequential design procedure, the size of system does
not change when N2 increases.
• If there exists feasible solution for both steps in the convex set (34), the proposed con-
trol algorithm (37) guarantees the PDQS and robustness properties of closed-loop MPC
system with guaranteed cost.
Robust Model Predictive Control Design 19

The sequential robust MPC design procedure can be summarized in the following steps:
• Design of robust MPC controller with control algorithm (36) by solving (46).
• Calculate matrices K1 , D1 and K1 ( j), D1 ( j), j = 1, 2, ...N given in (38) for nominal and
uncertain model of system.
• For a given k = 2, 3, ...N2 and control algorithm (37), sequentially calculate Fkk , Fkk+1 by
solving (48) with Kk , Dk given in (40).
• Calculate matrices A f , B f , C f (44) for model prediction.

2.4 EXAMPLES
Example 1. First example is the same as in section 1.5, it serves as a benchmark. The model of
double integrator turns to (35) where
 
1 0
Ao =
1 1
 
1  
Bo = ,C = 0 1
0
and uncertainty matrices are  
0.01 0.01
A1u =
0.02 0.03
 
0.001
B1u = ,
0
For the case when number of uncertainties p = 1, the number of vertices is N = 2 p = 2, the
matrices (34) are calculated as

A1 = An − A1u , A2 = An + A1u

B1 = Bn − B1u , B2 = Bn + B1u
For the parameters:  = 20000, prediction and control horizons N2 = 4, Nu = 4, performance
matrices R1 = ...R4 = 1, Q1 = .1I, Q2 = .5I, Q3 = I, Q4 = 5I, the following results are
obtained using the sequential design approach proposed in this part :
• For prediction k = 1, the robust control algorithm is given as

u(t) = F11 y(t) + F12 y(t + 1)

From (46), one obtains the gain matrices F11 = 0.9189; F12 = −1.4149. The eigenvalues
of closed-loop first vertex system model are as follows

Eig(Closed − loop) = {0.2977 ± 0.0644i }

• For k = 2, control algorithm is

u(t + 1) = F22 y(t + 1) + F23 y(t + 2)

In the second step of design procedure control gain matrices obtained solving (48) are
F22 = 0.4145; F23 = −0.323. The eigenvalues of closed-loop first vertex system model
are
Eig(Closed − loop) = {0.1822 ± 0.1263i }
20 Model Predictive Control

• For k=3, control algorithm is

u(t + 2) = F33 y(t + 2) + F34 y(t + 3)

In the second step of design procedure the obtained control gain matrices are F33 =
0.2563; F34 = −0.13023. The eigenvalues of closed-loop first vertex system model are

Eig(Closed − loop) = {0.1482 ± 0.051i }

• For prediction k = N2 = 4, control algorithm is

u(t + 3) = F44 y(t + 3) + F45 y(t + 4)

In the second step the obtained control gain matrices are F44 = 0.5797; F45 = 0.0. The
eigenvalues of closed-loop first vertex model system are

Eig(Closed − loop) = {0.1002 ± 0.145i }

Example 2. Nominal model for the second example is

 
0.6 0.0097 0.0143 0 0
 0.012 0.9754 0.0049 0 0 
 
Ao =  −0.0047 0.01 0.46 0 0 

 0.0488 0.0002 0.0004 1 0 
−0.0001 0.0003 0.0488 0 1
   
0.0425 0.0053 1 0 0 0 0
 0.0052 0.01   0 0 1 0 0 
Bo = 
 0.0024
C =  
0.0001   0 0 0 1 0 
0 0.0012 0 0 0 0 1
The linear affine type model of uncertain system (34) is in the form

Ai = An + θ1 A1u ; Bi = Bn + θ1 B1u

Ci = C, i = 1, 2
where A1u , B1u are uncertainty matrices with constant entries, θ1 is an uncertain real parame-
ter θ1 ∈< θ1 , θ1 > . When lower and upper bounds of uncertain parameter θ1 are substituted
to the affine type model, the polytopic system (33) is obtained. Let θ1 ∈< −1, 1 > and
 
0.025 0 0 0 0
 0 0.021 0 0 0 
 
A1u =  0 0 0.0002 0 0  
 0.001 0 0 0 0 
0 0 0.0001 0 0
 
0.0001 0
 0 0.001 
 
B1u = 
 0 0.0021 

 0 0 
0 0
Robust Model Predictive Control Design 21

In this example two vertices (N = 2) are calculated. The design problem is: Design two PS(PI)
model predictive robust decentralized controllers for plant input u(t) and prediction horizon
N2 = 5 using sequential design approach. The cost function is given by the following matrices

Q1 = Q2 = Q3 = I, R1 = R2 = R3 = I,

Q4 = Q5 = 0.5I, R4 = R5 = I
In the first step, calculation for the uncertain system (33) yields the robust control algorithm

u(t) = F11 y(t) + F12 y(t + 1)

where matrix F11 with decentralized output feedback structure containing two PS controllers,
is designed. From (46), the gain matrices F11 , F12 are obtained
 
−18.7306 0 −42.4369 0
F11 =
0 8.8456 0 48.287

where decentralized proportional and integral gains for the first controller are

K1p = 18.7306, K1i = 42.4369

and for the second one

K2p = −8.8456, K2i = −48.287
Note that in F11 sign - shows the negative feedback. Because predicted output y(t + 1) is
obtained from prediction model (44), for output feedback gain matrix F12 there is no need to
use decentralized control structure
 
−22.0944 20.2891 −10.1899 18.2789
F12 =
−29.3567 8.5697 −28.7374 −40.0299

In the second step of design procedure, using (48) for nominal model, the matrices (37) Fkk , Fkk+1 , k =
2, 3, 4, 5 are calculated. The eigenvalues of closed-loop first vertex system model for N2 =
Nu = 5 are

Eig(Closed − loop) = {−0.0009; −0.0087; 0.9789; 0.8815; 0.8925}

Feasible solutions of bilinear matrix inequality have been obtained by YALMIP with PENBMI
solver.

3. CONCLUSION
The first part of chapter addresses the problem of designing the output/state feedback robust
model predictive controller with input constraints for output and control prediction horizons
N2 and Nu . The main contribution of the presented results is twofold: The obtained robust
control algorithm guarantees the closed-loop system quadratic stability and guaranteed cost
under input constraints in the whole uncertainty domain. The required on-line computa-
tion load is significantly less than in MPC literature (according to the best knowledge of au-
thors), which opens possibility to use this control design scheme not only for plants with slow
dynamics but also for faster ones. At each sample time the calculation of proposed control
algorithm reduces to a solution of simple equation. Finally, two examples illustrate the effec-
tiveness of the proposed method. The second part of chapter studies the problem of design
22 Model Predictive Control

a new MPC with special control algorithm. The proposed robust MPC control algorithm is
designed sequentially, the degree of plant model does not change when the output predic-
tion horizon changes. The proposed sequential robust MPC design procedure consists of two
steps: In the first step for one step ahead prediction horizon the necessary and sufficient ro-
bust stability conditions have been developed for MPC and the polytopic system with output
feedback, using generalized parameter dependent Lyapunov matrix P(α). The proposed ro-
bust MPC ensures parameter dependent quadratic stability (PDQS) and guaranteed cost. In
the second step of design procedure the uncertain plant and nominal model with sequential
design approach is used to design the predicted input variables u(t + 1), ...u(t + N2 − 1) so
that to ensure the robust closed-loop stability of MPC with guaranteed cost. Main advantages
of the proposed sequential method are that the design plant model degree is independent on
prediction horizon N2 ; robust controller design procedure ensures PDQS and guaranteed cost
and the obtained results are easy to be implemented in real plant. In the proposed design
approach, constraints on system variables are easy to be implemented by LMI (BMI) using a
notion of invariant set. Feasible solution of BMI has been obtained by Yalmip with PENBMI
solver.

4. ACKNOWLEDGMENT
The work has been supported by Grant N 1/0544/09 of the Slovak Scientific Grant Agency.

5. References
Adamy, J. & Flemming, A. ( 2004) Soft variable-structure controls: a survey, Automatica, 40,
1821-1844.
Ayd, H., Mesquine, F. & Aitrami, M. (2008) Robust control for uncertain linear systems with
state and control constraints. In:Proc. of the 17th World Congress IFAC, Seoul, Korea,
2008, 1153-1158.
Bouzouita, B., Bouani, F. & Ksouri, M. (2007) Efficient Implementation of Multivariable MPC
with Parametric Uncertainties, In:Proc. ECC 2007, Kos, Greece, TuB12.4, CD-ROM.
Camacho, E.F & Bordons, C. (2004)Model predictive control, Springer-Verlag London Limited.
Casavola, A., Famularo, D. & Franze, G. (2004) Robust constrained predictive control of un-
certain norm-bounded linear systems. Automatica, 40, 1865-1776.
Clarke, D.W. & Mohtadi, C. (1989) Properties of generalized predictive control. Automatica,
25(6), 859-875.
Clarke, D.W. & Scattolini, R.(1991) Constrained Receding-horizon Predictive Control.
Proceedings IEE 138,(4), 347-354.
Dermicioglu, H. & Clarke, D.W. (1993) Generalized predictive control with end-point weight-
ing. IEE Proc , 140, Part D(4): 275-282, 1993.
Ding, B., Xi, Y., Cychowski, M.T. & O’Mahony, T. (2008) A synthesis approach for output
robust constrained model predictive control, Automatica,44, 258-264.
Ebihara, Y., Peaucelle, D., Arzelier, D. & Hagivara, T. (2006) Robust H2 Performance Analysis
of Uncertain LTI Systems via Polynomially Parameter Dependent Lyapunov func-
tions. In:Preprint of the 5th IFAC Symposium on Robust Control Design, ROCOND
06, Toulouse, France, July 5-7, 2006 CD-ROM.
Grman, L., Rosinová, D., Veselý, V. & Kozáková, A. (2005) Robust stability conditions for
polytopic systems. Int. Journal of Systems Science, Vol36, N15, 961-973.
Robust Model Predictive Control Design 23

Janík, M., Miklovicová, E. & Mrosko, M. (2008) Predictive control of nonlinear systems. ICIC
Express Letters, Vol. 2, N3, 239-244.
Kothare, M.V., Balakrishnan, V, & Morari, M. (1996) Robust Constrained Model Predictive
Control using Linear Matrix Inequalities, Automatica , Vol 32, N10, 1361-1379.
Krokavec, D. & Filasová, A. (2003) Quadratically stabilized discrete-time robust LQ control.
In:Control System Design , Proc. of the 2nd IFAC Conf., Bratislava, 375-380.
Kuwata, Y., Richards, A. & How, J. (2007) Robust Receding Horizon using Generalized Con-
straint Tightening, In:Proc. ACC, New Yourk, CD-ROM.
Lovas, Ch. , Seron, M.M. & Goodwin, G.C. (2007) Robust Model Predictive Control of
Input-Constrained Stable Systems with Unstructured Uncertainty, In:Proc. ECC, Kos,
Greece, CD-ROM.
Maciejovski, J.M. (2002) Predictive Control with Constraints. Prentice Hall
Mayne, D.Q., Rawlings, J.B.,Rao, C.V. & Scokaert, P.O.M. (2000) Contrained model predictive
control: stability and optimality. Automatica 36: 789-814.
de Oliveira, M.C., Camino, J.F. & Skelton, R.E. (2000) A convexifying algorithm for the de-
sign of structured linear controllers. In:Proc.39th IEEE Conference on Decision and
Control, Sydney , 2781-2786.
Orukpe, P.E., Jaimoukha, I.M. & El-Zobaidi, H.M.H. (2007) Model Predictive Control Based
on MixedH2 /H∞ Control Approach, In:Proc. ACC, New York July, 2007, CD-ROM.
Peaucelle, D., Arzelier, D., Bachelier, O. & Bernussou, J. (2000) A new robust D-stability con-
dition for real convex polytopic uncertainty, Systems and Control Letters, 40, 21-30.
delaPena, D.M., Alamo, T., Ramirez, T. & Camacho E. (2005) Min-max Model Predictive Con-
trol as a Quadratic Program, In: Proc. of 16th IFAC World Congress, Praga, CD-ROM.
Polak, E. & Yang, T.H. (1993) Moving horizon control of linear systems with input saturation
and plant uncertainty, Int. J. Control , 53, 613-638.
Rawlings, J. & Muske, K. (1993) The stability of constrained Receding Horizon Control. IEEE
Trans. on Automatic Control 38, 1512-1516.
Rohal-Ilkiv, B. (2004) A note on calculation of polytopic invariant and feasible sets for linear
continuous -time systems. Annual Rewiew in Control , 28, 59-64.
Rosinová, D., Veselý, V. & Kuc̆era, V. (2003) A necessary and sufficient condition for static
output feedback stabilizability of linear discrete-time systems, Kybernetika, Vol39,
N4, 447-459.
Rossiter, J.A. (2003) Model Based Predictive Control: A Practical Approach,Control Series.
Veselý, V., Rosinová, D. & Foltin, M. (2010) Robust model predictive control design with input
constraints. ISA Transactions,49, 114-120.
Veselý, V. & Rosinová, D. (2009) Robust output model predictive control design : BMI ap-
proach, IJICIC Int. Vol 5, 4, 1115-1123.
Yanou, A., Inoue,A., Deng, M. & Masuda,S. (2008) An Extension of two Degree-of-freedom of
Generalized Predictive Control for M-input M-output Systems Based on State Space
Approach IJICIC, Vol4, N12, 3307-3318.
Zafiriou, E. & Marchal, A. (1991) Stability of SISO quadratic dynamic matrix control with hard
output constraints, AIChE J. 37, 1550-1560.
Wang, Z., Chen, Z. Sun, Q. & Yuan, Z.(2006) GPC design technique based on MQFT for MIMO
uncertain system, Int. J. of Innovitive Computing, Information and Control, Vol2, N3,
519-526.
Zheng, Z.Q. & Morari, M. (1993) Robust Stability of Constrained Model Predictive Control, In
Proc. ACC, San Francisco, CA, 379-383.
24 Model Predictive Control
Robust Adaptive Model Predictive Control of Nonlinear Systems 25

0
2

Robust Adaptive Model Predictive

Control of Nonlinear Systems
Darryl DeHaan and Martin Guay
Dept. Chemical Engineering, Queen’s University
Canada

1. Introduction
When faced with making a decision, it is only natural that one would aim to select the course
of action which results in the “best" possible outcome. However, the ability to arrive at a de-
cision necessarily depends upon two things: a well-defined notion of what qualities make an
outcome desirable, and a previous decision1 defining to what extent it is necessary to charac-
terize the quality of individual candidates before making a selection (i.e., a notion of when a
decision is “good enough"). Whereas the first property is required for the problem to be well
defined, the later is necessary for it to be tractable.
The process of searching for the “best" outcome has been mathematically formalized in the
framework of optimization. The typical approach is to define a scalar-valued cost function,
that accepts a decision candidate as its argument, and returns a quantified measure of its
quality. The decision-making process then reduces to selecting a candidate with the lowest
(or highest) such measure.

1.1 The Emergence of Optimal Control

The field of “control" addresses the question of how to manipulate an input u in order to drive
the state x of a dynamical system

ẋ = f ( x, u) (1)

to some desired target. Ultimately this task can be viewed as decision-making, so it is not sur-
prising that it lends itself towards an optimization-based characterization. Assuming that one
can provide the necessary metric for assessing quality of the trajectories generated by (1), there
exists a rich body of “optimal control" theory to guide this process of decision-making. Much
of this theory came about in the 1950’s and 60’s, with Pontryagin’s introduction of the Mini-
mum (a.k.a. Maximum) Principle Pontryagin (1961), and Bellman’s development of Dynamic
Programming Bellman (1952; 1957). (This development also coincided with landmark results
for linear systems, pioneered by Kalman Kalman (1960; 1963), that are closely related). How-
ever, the roots of both approaches actually extend back to the mid-1600’s, with the inception
of the calculus of variations.

1 The recursiveness of this definition is of course ill-posed until one accepts that at some level, every
decision is ultimately predicated upon underlying assumptions, accepted entirely in faith.
26 Model Predictive Control

The tools of optimal control theory provide useful benchmarks for characterizing the notion
of “best" decision-making, as it applies to control. However applied directly, the tractability of
this decision-making is problematic. For example, Dynamic Programming involves the con-
struction of a n−dimensional surface that satisfies a challenging nonlinear partial differential
equation, which is inherently plagued by the so-called curse of dimensionality. This method-
ology, although elegant, remains generally intractable for problems beyond modest size. In
contrast, the Minimum Principle has been relatively successful for use in off-line trajectory
planning, when the initial condition of (1) is known. Although it was suggested as early as
1967 in Lee & Markus (1967) that a stabilizing feedback u = k( x ) could be constructed by
continuously re-solving the calculations online, a tractable means of doing this was not im-
mediately forthcoming.

1.2 Model Predictive Control as Receding-Horizon Optimization

Early development (Richalet et al. (1976),Richalet et al. (1978),Cutler & Ramaker (1980)) of the
control approach known today as Model Predictive Control (MPC) originated in the process
control community, and was driven much more by industrial application than by theoret-
ical understanding. Modern theoretical understanding of MPC, much of which developed
throughout the 1990’s, has clarified its very natural ties to existing optimal control theory. Key
steps towards this development included such results as Chen & Allgöwer (1998a;b); De Nico-
lao et al. (1996); Jadbabaie et al. (2001); Keerthi & Gilbert (1988); Mayne & Michalska (1990);
Michalska & Mayne (1993); Primbs et al. (2000), with an excellent unifying survey in Mayne
et al. (2000).
At its core, MPC is simply a framework for implementing existing tools of optimal control.
Taking the current value x (t) as the initial condition for (1), the Minimum Principle is used
as the primary basis for identifying the “best" candidate trajectory by predicting the future
behaviour of the system using model (1). However, the actual quality measure of interest in
the decision-making is generally the total future accumulation (i.e., over an infinite future) of
a given instantaneous metric, a quantity rarely computable in a satisfactorily short time. As
such, MPC only generates predictions for (1) over a finite time-horizon, and approximates the
remaining infinite tail of the cost accumulation using a penalty surface derived from either a
local solution of the Dynamic Programming surface, or an appropriate approximation of that
surface. As such, the key benefit of MPC over other optimal control methods is simply that its
finite horizon allows for a convenient trade-off between the online computational burden of
solving the Minimum Principle, and the offline burden of generating the penalty surface.
In contrast to other approaches for constructive nonlinear controller design, optimal control
frameworks facilitate the inclusion of constraints, by imposing feasibility of the candidates
as a condition in the decision-making process. While these approaches can be numerically
burdensome, optimal control (and by extension, MPC) provides the only real framework for
addressing the control of systems in the presence of constraints - in particular those involving
the state x. In practice, the predictive aspect of MPC is unparalleled in its ability to account
for the risk of future constraint violation during the current control decision.

1.3 Current Limitations in Model Predictive Control

While the underlying theoretical basis for model predictive control is approaching a state
of relative maturity, application of this approach to date has been predominantly limited to
“slow" industrial processes that allow adequate time to complete the controller calculations.
There is great incentive to extend this approach to applications in many other sectors, moti-
Robust Adaptive Model Predictive Control of Nonlinear Systems 27

vated in large part by its constraint-handling abilities. Future applications of significant inter-
est include many in the aerospace or automotive sectors, in particular constraint-dominated
problems such as obstacle avoidance. At present, the significant computational burden of
MPC remains the most critical limitation towards its application in these areas.
The second key weakness of the model predictive approach remains its susceptibility to un-
certainties in the model (1). While a fairly well-developed body of theory has been devel-
oped within the framework of robust-MPC, reaching an acceptable balance between computa-
tional complexity and conservativeness of the control remains a serious problem. In the more
general control literature, adaptive control has evolved as an alternative to a robust-control
paradigm. However, the incorporation of adaptive techniques into the MPC framework has
remained a relatively open problem.

2. Notational and Mathematical Preliminaries

Throughout the remainder of this dissertation, the following is assumed by default (where
s ∈ Rs and S represent arbitrary vectors and sets, respectively):
• all vector norms are Euclidean, defining balls B(s, δ)  {s | s − s  ≤ δ}, δ ≥ 0.
• norms of matrices S ∈ R m×s are assumed induced as S  maxs=1 S s.
• the notation s[ a,b] denotes the entire continuous-time trajectory s(τ ), τ ∈ [ a, b], and
likewise ṡ[ a,b] the trajectory of its forward derivative ṡ(τ ).
• For any set S ⊆ Rs , define
i) its closure cl{S }, interior S̊, and boundary ∂S = cl{S } \ S̊
ii) its orthogonal distance norm sS  infs ∈S s − s 
iii) a closed δ-neighbourhood B(S, δ)  {s ∈ Rs | sS ≤ δ}
←
−
iv) an interior approximation B (S, δ)  {s ∈ S | infs ∈∂S s − s  ≥ δ}
v) a (finite, closed, open) cover of S as any (finite) collection {Si } of (open,
closed) sets Si ⊆ Rs such that S ⊆ ∪i Si .
vi) the maximal closed subcover cov {S } as the infinite collection {Si } contain-
ing all possible closed subsets S i ⊆ S; i.e., cov {S } is a maximal “set of sub-
sets".
Furthermore, for any arbitrary function α : S → R we assume the following definitions:
• α(·) is C m+ if it is at least m-times differentiable, with all derivatives of order m yielding
locally Lipschitz functions.
• A function α : S → (−∞, ∞] is lower semi-continuous (LS -continuous) at s if it satisfies
lim inf α(s ) ≥ α(s) (2)
(see Clarke et al. (1998)): s →s

• a continuous function α : R ≥0 → R ≥0 belongs to class K if α(0) = 0 and α(·) is strictly

increasing on R >0 . It belongs to class K∞ if it is furthermore radially unbounded.
• a continuous function β : R ≥0 × R ≥0 → R ≥0 belongs to class KL if i) for every fixed
value of τ, it satisfies β(·, τ ) ∈ K, and ii) for each fixed value of s, then β(s, ·) is strictly
decreasing and satisfies limτ →∞ β(s, τ ) = 0.
• the scalar operator satba (·) denotes saturation of its arguments onto the interval [ a, b],
a < b. For vector- or matrix-valued arguments, the saturation is presumed by default
to be evaluated element-wise.
28 Model Predictive Control

3. Brief Review of Optimal Control

The underlying assumption of optimal control is that at any time, the pointwise cost of x
and u being away from their desired targets is quantified by a known, physically-meaningful
function L( x, u). Loosely, the goal is to then reach some target in a manner that accumulates
the least cost. It is not generally necessary for the “target" to be explicitly described, since
its knowledge is built into the function L( x, u) (i.e it is assumed that convergence of x to
any invariant subset of { x | ∃u s.t. L( x, u) = 0} is as acceptable). The following result,
while superficially simple in appearance, is in fact the key foundation underlying the optimal
control results of this section, and by extension all of model predictive control as well. Proof
can be found in many references, such as Sage & White (1977).
Definition 3.1 (Principle of Optimality:). If u∗[t ,t ] is an optimal trajectory for the interval t ∈
1 2
[t1 , t2 ], with corresponding solution x[∗t ,t ] to (1), then for any τ ∈ (t1 , t2 ) the sub-arc u∗[τ, t ] is
1 2 2
necessarily optimal for the interval t ∈ [τ, t2 ] if (1) starts from x ∗ (τ ).

4. Variational Approach: Euler, Lagrange

and Pontryagin
Pontryagin’s Minimum principle (also known as the Maximum principle, Pontryagin (1961))
represented a landmark extension of classical ideas of variational calculus to the problem of
control. Technically, the Minimum Principle is an application of the classical Euler-Lagrange
and Weierstrass conditions2 Hestenes (1966), which provide first-order necessary conditions to
characterize extremal time-trajectories of a cost functional.3 . The Minimum Principle there-
fore characterizes minimizing trajectories ( x[0,T ] , u[0,T ] ) corresponding to a constrained finite-
horizon problem of the form
 T
VT ( x0 , u[0,T ] ) = L( x, u) dτ + W ( x ( T )) (3a)
0
s.t. ∀τ ∈[0, T ] :
ẋ = f ( x, u), x (0) = x0 (3b)
g( x (τ )) ≤ 0, h( x (τ ), u(τ )) ≤ 0, w( x ( T )) ≤ 0 (3c)
where the vectorfield f (·, ·) and constraint functions g(·), h(·, ·), and w(·) are assumed suffi-
ciently differentiable.
Assume that g( x0 ) < 0, and, for a given ( x0 , u[0,T ] ), let the interval [0, T ) be partitioned into
p
(maximal) subintervals as τ ∈ ∪i=1 [ti , ti+1 ), t0 = 0, t p+1 = T, where the interior ti represent
intersections g < 0 ⇔ g = 0 (i.e., the {ti } represent changes in the active set of g). Assuming
that g( x ) has constant relative degree r over some appropriate neighbourhood, define the fol-
lowing vector of (Lie) derivatives: N ( x )  [ g( x ), g(1) ( x ), . . . g(r−1) ( x )] T , which characterizes
additional tangency constraints N ( x (ti )) = 0 at the corners {ti }. Rewriting (3) in multiplier
form
 T
VT = H( x, u) − λ T ẋ dτ + W ( x ( T )) + µw w( x ( T )) + ∑ µ TN (ti ) N ( x (ti )) (4a)
0 i

H  L( x, u) + λ T f ( x, u) + µh h( x, u) + µ g g(r) ( x, u) (4b)
2 phrased as a fixed initial point, free endpoint problem
3 ∂p
i.e., generalizing the NLP necessary condition ∂x = 0 for the extrema of a function p( x ).
Robust Adaptive Model Predictive Control of Nonlinear Systems 29

overa Taking the first variation of the right-hand sides of (4a,b) with respect to perturbations
in x[0,T ] and u[0,T ] yields the following set of conditions (adapted from statements in Bert-
sekas (1995); Bryson & Ho (1969); Hestenes (1966)) which necessarily must hold for VT to be
minimized:

Proposition 4.1 (Minimum Principle). Suppose that the pair (u∗[0,T ] , x[∗0,T ] ) is a minimizing solu-
tion of (3). Then for all τ ∈ [0, T ], there exists multipliers λ(τ ) ≥ 0, µh (τ ) ≥ 0, µ g (τ ) ≥ 0, and
constants µw ≥ 0, µiN ≥ 0, i ∈ I , such that
i) Over each interval τ ∈ [ti , ti+1 ], the multipliers µh (τ ), µ g (τ ) are piecewise continuous, µ N (τ )
is constant, λ(τ ) is continuous, and with (u∗[t , t ] , x[∗t , t ] ) satisfies
i i +1 i i +1

∗ ∗ ∗ ∗
ẋ = f ( x , u ), x (0) = x0 (5a)
T T ∗ ∗
λ̇ = ∇ x H a.e., with λ ( T ) = ∇ x W ( x ( T )) + µw ∇ x w( x ( T )) (5b)

where the solution λ[0,T ] is discontinuous at τ ∈ {ti }, i ∈ {1, 3, 5...p}, satisfying

λ T (ti− ) = λ T (ti+ ) + µ TN (ti+ )∇ x N ( x (ti )) (5c)

ii) H( x ∗ , u∗ , λ, µh , µ g ) is constant over intervals τ ∈ [ti , ti+1 ], and for all τ ∈ [0, T ] it satisfies
 
(where U ( x )  {u | h( x, u) ≤ 0 and g(r) ( x, u) ≤ 0 if g( x ) = 0 } ):

H( x ∗ , u∗ , λ, µh , µ g ) ≤ min H( x ∗ , u, λ, µh , µ g ) (5d)
u ∈U ( x )

∇u H( x ∗ (τ ), u∗ (τ ), λ(τ ), µh (τ ), µ g (τ )) = 0 (5e)

iii) For all τ ∈ [0, T ], the following constraint conditions hold

g( x ∗ ) ≤ 0 h( x ∗ , u∗ ) ≤ 0 w( x ∗ ( T )) ≤ 0 (5f)
(r ) ∗ ∗ ∗ ∗ ∗
µ g (τ ) g (x , u ) = 0 µh (τ )h( x , u ) = 0 µw w( x ( T )) = 0 (5g)
 
µ TN (τ ) N ( x ∗ ) = 0 and N ( x ∗ ) = 0, ∀τ ∈ [ti , ti+1 ], i ∈ {1, 3, 5...p} (5h)

The multiplier λ(t) is called the co-state, and it requires solving a two-point boundary-value
problem for (5a) and (5b). One of the most challenging aspects to locating (and confirming)
a minimizing solution to (5) lies in dealing with (5c) and (5h), since the number and times of
constraint intersections are not known a-priori.

5. Dynamic Programming: Hamilton, Jacobi,

and Bellman
The Minimum Principle is fundamentally based upon establishing the optimality of a partic-
ular input trajectory u[0,T ] . While the applicability to offline, open-loop trajectory planning
is clear, the inherent assumption that x0 is known can be limiting if one’s goal is to develop
a feedback policy u = k( x ). Development of such a policy requires the consideration of all
possible initial conditions, which results in an optimal cost surface J ∗ : R n → R, with an asso-
ciated control policy k : R n → R m . A constructive approach for calculating such a surface,
referred to as Dynamic Programming, was developed by Bellman Bellman (1957). Just as the
30 Model Predictive Control

Minimum Principle was extended out of the classical trajectory-based Euler-Lagrange equa-
tions, Dynamic Programming is an extension of classical Hamilton-Jacobi field theory from
the calculus of variations.
For simplicity, our discussion here will be restricted to the unconstrained problem:
 ∞
V ∗ ( x0 ) = min L( x, u) dτ (6a)
u[0,∞) 0

s.t. ẋ = f ( x, u), x (0) = x0 (6b)

with locally Lipschitz dynamics f (·, ·). From the Principle of Optimality, it can be seen that
(6) lends itself to the following recursive definition:
 t+∆t 
∗ ∗
V ( x (t)) = min L( x (τ ), u(τ ))dτ + V ( x (t + ∆t)) (7)
u[t, t+∆t] t

Assuming that V ∗ is differentiable, replacing V ∗ ( x (t + ∆t) with a first-order Taylor-series and

the integrand with a Riemannian sum, the limit ∆t → 0 yields
 
∂V ∗
0 = min L( x, u) + f ( x, u) (8)
u ∂x
Equation (8) is one particular form of what is known as the Hamilton-Jacobi-Bellman (HJB)
equation. In some cases (such as L( x, u) quadratic in u, and f ( x, u) affine in u), (8) can
be simplified to a more standard-looking PDE by evaluating the indicated minimization in
closed-form4 . Assuming that a (differentiable) surface V ∗ : R n → R is found (generally
by off-line numerical solution) which satisfies (8), a stabilizing feedback u = k DP ( x ) can be
constructed from the information contained in the surface V ∗ by simply defining5 k DP ( x ) 
∗
{u | ∂V
∂x f ( x, u ) = − L ( x, u )}.
Unfortunately, incorporation of either input or state constraints generally violates the as-
sumed smoothness of V ∗ ( x ). While this could be handled by interpreting (8) in the context
of viscosity solutions (see Clarke et al. (1998) for definition), for the purposes of application to
model predictive control it is more typical to simply restrict the domain of V ∗ : Ω → R such
that Ω ⊂ R n is feasible with respect to the constraints.

6. Inverse-Optimal Control Lyapunov Functions

While knowledge of a surface V ∗ ( x ) satisfying (8) is clearly ideal, in practice analytical so-
lutions are only available for extremely restrictive classes of systems, and almost never for
systems involving state or input constraints. Similarly, numerical solution of (8) suffers the
so-called “curse of dimensionality" (as named by Bellman) which limits its applicability to
systems of restrictively small size.
An alternative design framework, originating in Sontag (1983), is based on the following:
Definition 6.1. A control Lyapunov function (CLF) for (1) is any C1 , proper, positive definite
function V : R n → R ≥0 such that, for all x = 0:
∂V
inf f ( x, u) < 0 (9)
u ∂x
4 In fact, for linear dynamics and quadratic cost, (8) reduces down to the linear Ricatti equation.
5 k DP (·) is interpreted to incorporate a deterministic selection in the event of multiple solutions. The
existence of such a u is implied by the assumed solvability of (8)
Robust Adaptive Model Predictive Control of Nonlinear Systems 31

Design techniques for deriving a feedback u = k( x ) from knowledge of V (·) include the well-
known “Sontag’s Controller" of Sontag (1989), which led to the development of “Pointwise
Min-Norm" control of the form Freeman & Kokotović (1996a;b); Sepulchre et al. (1997):

∂V
min γ(u) s.t. f ( x, u) < −σ( x ) (10)
u ∂x
where γ, σ are positive definite, and γ is radially unbounded. As discussed in Freeman
& Kokotović (1996b); Sepulchre et al. (1997), relation (9) implies that there exists a function
L( x, u), derived from γ and σ, for which V (·) satisfies (8). Furthermore, if V ( x ) ≡ V ∗ ( x ), then
appropriate selection of γ, σ (in particular that of Sontag’s controller Sontag (1989)) results in
the feedback u = k cl f ( x ) generated by (9) satisfying k cl f (·) ≡ k DP (·). Hence this technique is
commonly referred to as “inverse-optimal" control design, and can be viewed as a method for
approximating the optimal control problem (6) by replacing V ∗ ( x ) directly.

7. Review of Nonlinear MPC based on Nominal Models

The ultimate objective of a model predictive controller is to provide a closed-loop feedback u =
κmpc ( x ) that regulates (1) to its target set (assumed here x = 0) in a fashion that is optimal
with respect to the infinite-time problem (6), while enforcing pointwise constraints of the form
( x, u) ∈ X × U in a constructive manner. However, rather than defining the map κmpc : X →
U by solving a PDE of the form (8) (i.e thereby pre-computing knowledge of κmpc ( x ) for every
x ∈ X), the model predictive control philosophy is to solve for, at time t, the control move
u = κmpc ( x (t)) for the particular value x (t) ∈ X. This makes the online calculations inherently
trajectory-based, and therefore closely tied to the results in Section 4 (with the caveat that the
initial conditions are continuously referenced relative to current (t, x )). Since it is not practical
to pose (online) trajectory-based calculations over an infinite prediction horizon τ ∈ [t, ∞), a
truncated prediction τ ∈ [t, t + T ] is used instead. The truncated tail of the integral in (6) is
replaced by a (designer-specified) terminal penalty W : Xf → R ≥0 , defined over any local
neighbourhood Xf ⊂ X of the target x = 0. This results in a feedback of the form:

u = κmpc ( x (t))  u∗[t, t+T ] (t) (11a)

where u∗[t, t+T ] denotes the solution to the x (t)-dependent problem:

  t+T

p p p p
u∗[t, t+T ]  arg min
p
VT ( x (t), u[t, t+T ] )  L( x , u ) dτ + W ( x (t+T )) (11b)
u[t, t+T ] t

d p
s.t. ∀τ ∈ [t, t+T ] : dτ x = f ( x p , u p ), x p (t) = x (t) (11c)
p p
( x (τ ), u (τ )) ∈ X × U (11d)
x p ( t + T ) ∈ Xf (11e)

Clearly, if one could define W ( x ) ≡ V ∗ ( x ) globally, then the feedback in (11) must satisfy
κmpc (·) ≡ k DP (·). While W ( x ) ≡ V ∗ ( x ) is generally unachievable, this motivates the selection
of W ( x ) as a CLF such that W ( x ) is an inverse-optimal approximation of V ∗ ( x ). A more
precise characterization of the selection of W ( x ) is the focus of the next section.
32 Model Predictive Control

8. General Sufficient Conditions for Stability

A very general proof of the closed-loop stability of (11), which unifies a variety of earlier, more
restrictive, results is presented6 in the survey Mayne et al. (2000). This proof is based upon
the following set of sufficient conditions for closed-loop stability:

Criterion 8.1. The function W : Xf → R ≥0 and set Xf are such that a local feedback k f : Xf → U
exists to satisfy the following conditions:
C1) 0 ∈ Xf ⊆ X, Xf closed (i.e., state constraints satisfied in Xf )
C2) k f ( x ) ∈ U, ∀ x ∈ Xf (i.e., control constraints satisfied in Xf )
C3) Xf is positively invariant for ẋ = f ( x, k f ( x )).
∂W
C4) L( x, k f ( x )) + ∂x f ( x, k f ( x )) ≤ 0, ∀ x ∈ Xf .
Only existence, not knowledge, of k f ( x ) is assumed. Thus by comparison with (9), it can be
seen that C4 essentially requires that W ( x ) be a CLF over the (local) domain Xf , in a manner
consistent with the constraints.
In hindsight, it is nearly obvious that closed-loop stability can be reduced entirely to con-
ditions placed upon only the terminal choices W (·) and Xf . Viewing VT ( x (t), u∗[t,t+T ] ) as a

Lyapunov function candidate, it is clear from (3) that VT contains “energy" in both the L dτ
and terminal W terms. Energy dissipates from the front of the integral at a rate L( x, u) as time
t flows, and by the principle of optimality one could implement (11) on a shrinking horizon
(i.e., t + T constant), which would imply V̇ = − L( x, u). In addition to this, C4 guarantees that
the energy transfer from W to the integral (as the point t + T recedes) will be non-increasing,
and could even dissipate additional energy as well.

9. Robustness Considerations
As can be seen in Proposition 4.1, the presence of inequality constraints on the state variables
poses a challenge for numerical solution of the optimal control problem in (11). While locating
the times {ti } at which the active set changes can itself be a burdensome task, a significantly
more challenging task is trying to guarantee that the tangency condition N ( x (ti+1 )) = 0 is
met, which involves determining if x lies on (or crosses over) the critical surface beyond which
this condition fails.
As highlighted in Grimm et al. (2004), this critical surface poses more than just a computa-
tional concern. Since both the cost function and the feedback κmpc ( x ) are potentially discon-
tinuous on this surface, there exists the potential for arbitrarily small disturbances (or other
plant-model mismatch) to compromise closed-loop stability. This situation arises when the
optimal solution u∗[t,t+T ] in (11) switches between disconnected minimizers, potentially result-
ing in invariant limit cycles (for example, as a very low-cost minimizer alternates between
being judged feasible/infeasible.)
A modification suggested in Grimm et al. (2004) to restore nominal robustness, similar to the
idea in Marruedo et al. (2002), is to replace the constraint x (τ ) ∈ X of (11d) with one of the
form x (τ ) ∈ X o (τ − t), where the function X o : [0, T ] → X satisfies X o (0) = X, and the strict
containment X o (t2 ) ⊂ X o (t1 ), t1 < t2 . The gradual relaxation of the constraint limit as future
predictions move closer to current time provides a safety margin that helps to avoid constraint
violation due to small disturbances.
6 in the context of both continuous- and discrete-time frameworks
Robust Adaptive Model Predictive Control of Nonlinear Systems 33

The issue of robustness to measurement error is addressed in Tuna et al. (2005). On one hand,
nominal robustness to measurement noise of an MPC feedback was already established in
Grimm et al. (2003) for discrete-time systems, and in Findeisen et al. (2003) for sampled-data
implementations. However, Tuna et al. (2005) demonstrates that as the sampling frequency
becomes arbitrarily fast, the margin of this robustness may approach zero. This stems from
the fact that the feedback κmpc ( x ) of (11) is inherently discontinuous in x if the indicated
minimization is performed globally on a nonconvex surface, which by Coron & Rosier (1994);
Hermes (1967) enables a fast measurement dither to generate flow in any direction contained
in the convex hull of the discontinuous closed-loop vectorfield. In other words, additional
attractors or unstable/infeasible modes can be introduced into the closed-loop behaviour by
arbitrarily small measurement noise.
Although Tuna et al. (2005) deals specifically with situations of obstacle avoidance or stabi-
lization to a target set containing disconnected points, other examples of problematic noncon-
vexities are depicted in Figure 1. In each of the scenarios depicted in Figure 1, measurement
dithering could conceivably induce flow along the dashed trajectories, thereby resulting in
either constraint violation or convergence to an undesired equilibrium.
Two different techniques were suggested in Tuna et al. (2005) for restoring robustness to the
measurement error, both of which involve adding a hysteresis-type behaviour in the optimiza-
tion to prevent arbitrary switching of the solution between separate minimizers (i.e., making
the optimization behaviour more decisive).

Fig. 1. Examples of nonconvexities susceptible to measurement error

10. Robust MPC

10.1 Review of Nonlinear MPC for Uncertain Systems
While a vast majority of the robust-MPC literature has been developed within the framework
of discrete-time systems7 , for consistency with the rest of this thesis most of the discussion
will be based in terms of their continuous-time analogues. The uncertain system model is

7 Presumably for numerical tractability, as well as providing a more intuitive link to game theory.
34 Model Predictive Control

therefore described by the general form

ẋ = f ( x, u, d) (12)

where d(t) represents any arbitrary L∞ -bounded disturbance signal, which takes point-wise8
values d ∈ D . Equivalently, (12) can be represented as the differential inclusion model ẋ ∈
F ( x, u)  f ( x, u, D).
In the next two sections, we will discuss approaches for accounting explicitly for the distur-
bance in the online MPC calculations. We note that significant effort has also been directed
towards various means of increasing the inherent robustness of the controller without requir-
ing explicit online calculations. This includes the suggestion in Magni & Sepulchre (1997)
(with a similar discrete-time idea in De Nicolao et al. (1996)) to use a modified stage cost
L( x, u)  L( x, u) + ∇ x VT∗ ( x ), f ( x, u) to increase the robustness of a nominal-model imple-
mentation, or the suggestion in Kouvaritakis et al. (2000) to use an prestabilizer, optimized
offline, of the form u = Kx + v to reduced online computational burden. Ultimately, these ap-
proaches can be considered encompassed by the banner of nominal-model implementation.

10.1.1 Explicit robust MPC using Open-loop Models

As seen in the previous chapters, essentially all MPC approaches depend critically upon the
Principle of Optimality (Def 3.1) to establish a proof of stability. This argument depends inher-
p
ently upon the assumption that the predicted trajectory x[t, t+T ] is an invariant set under open-
p
loop implementation of the corresponding u[t, t+T ] ; i.e., that the prediction model is “perfect".
Since this is no longer the case in the presence of plant-model mismatch, it becomes necessary
p p
to associate with u[t, t+T ] a cone of trajectories { x[t, t+T ] }D emanating from x (t), as generated by
(12).
Not surprisingly, establishing stability requires a strengthening of the conditions imposed on
the selection of the terminal cost W and domain Xf . As such, W and Xf are assumed to satisfy
Criterion (8.1), but with the revised conditions:
C3a) Xf is strongly positively invariant for ẋ ∈ f ( x, k f ( x ), D).
∂W
C4a) L( x, k f ( x )) + ∂x f ( x, k f ( x ), d) ≤ 0, ∀( x, d) ∈ Xf × D .
While the original C4 had the interpretation of requiring W to be a CLF for the nominal sys-
tem, so the revised C4a can be interpreted to imply that W should be a robust-CLF like those
developed in Freeman & Kokotović (1996b).
Given such an appropriately defined pair (W, Xf ), the model predictive controller explicitly
p
considers all trajectories { x[t, t+T ] }D by posing the modified problem

u = κmpc ( x (t))  u∗[t, t+T ] (t) (13a)

where the trajectory u∗[t, t+T ] denotes the solution to

 
p
u∗[t, t+T ]  arg min
p
max VT ( x (t), u[t, t+T ] , d[t, t+T ] ) (13b)
u
[t, t+T ]
d[t, t+T ] ∈D
T ∈[0,Tmax ]

8 The abuse of notation d[t1 , t2 ] ∈ D is likewise interpreted pointwise

Robust Adaptive Model Predictive Control of Nonlinear Systems 35

p
The function VT ( x (t), u[t, t+T ] , d[t, t+T ] ) appearing in (13) is as defined in (11), but with (11c) re-
placed by (12). Variations of this type of design are given in Chen et al. (1997); Lee & Yu (1997);
Mayne (1995); Michalska & Mayne (1993); Ramirez et al. (2002), differing predominantly in the
manner by which they select W (·) and Xf .
If one interprets the word “optimal" in Definition 3.1 in terms of the worst-case trajectory in
p
the optimal cone { x[t, t+T ] }∗D , then at time τ ∈ [t, t + T ] there are only two possibilities:
p
• the actual x[t,τ ] matches the subarc from a worst-case element of { x[t, t+T ] }∗D , in which
case the Principle of Optimality holds as stated.
p
• the actual x[t,τ ] matches the subarc from an element in { x[t, t+T ] }∗D which was not the
worst case, so implementing the remaining u∗[τ, t+T ] will achieve overall less cost than
the worst-case estimate at time t.
One will note however, that the bound guaranteed by the principle of optimality applies only
to the remaining subarc [τ, t+ T ], and says nothing about the ability to extend the horizon. For
the nominal-model results of Chapter 7, the ability to extend the horizon followed from C4)
p
of Criterion (8.1). In the present case, C4a) guarantees that for each terminal value { x[t, t+T ] (t +
T )}∗D there exists a value of u rendering W decreasing, but not necessarily a single such value
p
satisfying C4a) for every { x[t, t+T ] (t + T )}∗D . Hence, receding of the horizon can only occur at
the discretion of the optimizer. In the worst case, T could contract (i.e., t + T remains fixed)
p
until eventually T = 0, at which point { x[t, t+T ] (t + T )}∗D ≡ x (t), and therefore by C4a) an
∗
appropriate extension of the “trajectory" u[t,t] exists.
Although it is not an explicit min-max type result, the approach in Marruedo et al. (2002)
makes use of global Lipschitz constants to determine a bound on the the worst-case distance
between a solution of the uncertain model (12), and that of the underlying nominal model es-
timate. This Lipschitz-based uncertainty cone expands at the fastest-possible rate, necessarily
p
containing the actual uncertainty cone { x[t, t+T ] }D . Although ultimately just a nominal-model
approach, it is relevant to note that it can be viewed as replacing the “max" in (13) with a
simple worst-case upper bound.
Finally, we note that many similar results Cannon & Kouvaritakis (2005); Kothare et al. (1996)
in the linear robust-MPC literature are relevant, since nonlinear dynamics can often be ap-
proximated using uncertain linear models. In particular, linear systems with polytopic de-
scriptions of uncertainty are one of the few classes that can be realistically solved numerically,
since the calculations reduce to simply evaluating each node of the polytope.

10.1.2 Explicit robust MPC using Feedback Models

Given that robust control design is closely tied to game theory, one can envision (13) as rep-
resenting a player’s decision-making process throughout the evolution of a strategic game.
However, it is unlikely that a player even moderately-skilled at such a game would restrict
themselves to preparing only a single sequence of moves to be executed in the future. Instead,
a skilled player is more likely to prepare a strategy for future game-play, consisting of several
“backup plans" contingent upon future responses of their adversary.
To be as least-conservative as possible, an ideal (in a worst-case sense) decision-making pro-
cess would more properly resemble

u = κmpc ( x (t))  u∗t (14a)

36 Model Predictive Control

where u∗t ∈ R m is the constant value satisfying

 
p
u∗t  arg min max min VT ( x (t), u[t, t+T ] , d[t, t+T ] ) (14b)
ut d[t, t+T ] ∈D u p ∈U (ut )
[t, t+T ]

p
with the definition U (ut )  {u[t, t+T ] | u p (t) = ut }. Clearly, the “least conservative" prop-
erty follows from the fact that a separate response is optimized for every possible sequence
the adversary could play. This is analogous to the philosophy in Scokaert & Mayne (1998),
for system x + = Ax + Bu + d, in which polytopic D allows the max to be reduced to select-
ing the worst index from a finitely-indexed collection of responses; this equivalently replaces
the innermost minimization with an augmented search in the outermost loop over all input
responses in the collection.
While (14) is useful as a definition, a more useful (equivalent) representation involves mini-
mizing over feedback policies k : [t, t + T ] × X → U rather than trajectories:

u = κmpc ( x (t))  k∗ (t, x (t)) (15a)

 
k∗ (·, ·)  arg min max VT ( x (t), k (·, ·), d[t, t+T ] ) (15b)
k(·,·) d[t, t+T ] ∈D
 t+T
VT ( x (t), k (·, ·), d[t, t+T ] )  L( x p , k (τ, x p (τ ))) dτ + W ( x p (t+T )) (15c)
t
d p
s.t. ∀τ ∈ [t, t+T ] : dτ x = f ( x p , k (τ, x p (τ )), d), x p (t) = x (t) (15d)
p p
( x (τ ), k(τ, x (τ ))) ∈ X × U (15e)
x p ( t + T ) ∈ Xf (15f)

There is a recursive-like elegance to (15), in that κmpc ( x ) is essentially defined as a search over
future candidates of itself. Whereas (14) explicitly involves optimization-based future feedbacks,
the search in (15) can actually be (suboptimally) restricted to any arbitrary sub-class of feed-
backs k : [t, t+ T ] × X → U. For example, this type of approach first appeared in Kothare et al.
(1996); Lee & Yu (1997); Mayne (1995), where the cost functional was minimized by restricting
the search to the class of linear feedback u = Kx (or u = K (t) x).
p
The error cone { x[t, t+T ] }∗D associated with (15) is typically much less conservative than that of
(13). This is due to the fact that (15d) accounts for future disturbance attenuation resulting
from k (τ, x p (τ )), an effect ignored in the open-loop predictions of (13). In the case of (14) and
(15) it is no longer necessary to include T as an optimization variable, since by condition C4a
one can now envision extending the horizon by appending an increment k ( T + δt, ·) = k f (·).
This notion of feedback MPC has been applied in Magni et al. (2003; 2001) to solve H∞ dis-
turbance attenuation problems. This approach avoids the need to solve a difficult Hamilton-
Jacobi-Isaacs equation, by combining a specially-selected stage cost L( x, u) with a local HJI
approximation W ( x ) (designed generally by solving an H∞ problem for the linearized sys-
tem). An alternative perspective of the implementation of (15) is developed in Langson et al.
(2004), with particular focus on obstacle-avoidance in Raković & Mayne (2005). In this work,
p
a set-invariance philosophy is used to propagate the uncertainty cone { x[t, t+T ] }D for (15d) in
the form of a control-invariant tube. This enables the use of efficient methods for constructing
control invariant sets based on approximations such as polytopes or ellipsoids.
Robust Adaptive Model Predictive Control of Nonlinear Systems 37

11. Adaptive Approaches to MPC

The sectionr will be focused on the more typical role of adaptation as a means of coping with
uncertainties in the system model. A standard implementation of model predictive control
using a nominal model of the system dynamics can, with slight modification, exhibit nominal
robustness to disturbances and modelling error. However in practical situations, the sys-
tem model is only approximately known, so a guarantee of robustness which covers only
“sufficiently small" errors may be unacceptable. In order to achieve a more solid robustness
guarantee, it becomes necessary to account (either explicitly, or implicitly) for all possible
trajectories which could be realized by the uncertain system, in order to guarantee feasible
stability. The obvious numerical complexity of this task has resulted in an array of different
control approaches, which lie at various locations on the spectrum between simple, conser-
vative approximations versus complex, high-performance calculations. Ultimately, selecting
an appropriate approach involves assessing, for the particular system in question, what is an
acceptable balance between computational requirements and closed-loop performance.
Despite the fact that the ability to adjust to changing process conditions was one of the ear-
liest industrial motivators for developing predictive control techniques, the progress in this
area has been negligible. The small amount of progress that has been made is restricted to
systems which do not involve constraints on the state, and which are affine in the unknown
parameters. We will briefly describe two such results.

11.1 Certainty-equivalence Implementation

The result in Mayne & Michalska (1993) implements a certainty equivalence nominal-model9
MPC feedback of the form u(t) = κmpc ( x (t), θ̂ (t)), to stabilize the uncertain system

ẋ = f ( x, u, θ )  f 0 ( x, u) + g( x, u)θ (16)

subject to an input constraint u ∈ U. The vector θ ∈ R p represents a set of unknown con-

stant parameters, with θ̂ ∈ R p denoting an identifier. Certainty equivalence implies that the
nominal prediction model (11c) is of the same form as (16), but with θ̂ used in place of θ.
At any time t ≥ 0, the identifier θ̂ (t) is defined to be a (min-norm) solution of
 t    t
g( x (s), u(s)) T ẋ (s)− f 0 ( x (s), u(s)) ds = g( x (s), u(s)) T g( x (s), u(s))ds θ̂ (17)
0 0

which is solved over the window of all past history, under the assumption that ẋ is mea-
sured (or computable). If necessary, an additional search is performed along the nullspace
t
of 0 g( x, u) T g( x, u)ds in order to guarantee θ̂ (t) yields a controllable certainty-equivalence
model (since (17) is controllable by assumption).
The final result simply shows that there must exist a time 0 < t a < ∞ such that the regressor
t T
0 g ( x, u ) g ( x, u ) ds achieves full rank, and thus θ̂ ( t ) ≡ θ for all t ≥ t a . However, it is only by
assumption that the state x (t) does not escape the stabilizable region during the identification
phase t ∈ [0, t a ]; moreover, there is no mechanism to decrease t a in any way, such as by
injecting excitation.

9 Since this result arose early in the development of nonlinear MPC, it happens to be based upon a
terminal-constrained controller (i.e., Xf ≡ {0}); however, this is not critical to the adaptation.
38 Model Predictive Control

11.1.1 Stability-Enforced Approach

One of the early stability results for nominal-model MPC in (Primbs (1999); Primbs et al.
(2000)) involved the use of a global CLF V ( x ) instead of a terminal penalty. Stability was
enforced by constraining the optimization such that V ( x ) is decreasing, and performance
achieved by requiring the predicted cost to be less than that accumulated by simulation of
pointwise min-norm control.
This idea was extended in Adetola & Guay (2004) to stabilize unconstrained systems of the
form

ẋ = f ( x, u, θ )  f 0 ( x ) + gθ ( x )θ + gu ( x )u (18)

Using ideas from robust stabilization, it is assumed that a global ISS-CLF10 is known for the
nominal system. Constraining V ( x ) to decrease ensures convergence to a neighbourhood of
the origin, which gradually contracts as the identification proceeds. Of course, the restrictive-
ness of this approach lies in the assumption that V ( x ) is known.

12. An Adaptive Approach to Robust MPC

Both the theoretical and practical merits of model-based predictive control strategies for non-
linear systems are well established, as reviewed in Chapter 7. To date, the vast majority of
implementations involve an “accurate model" assumption, in which the control action is com-
puted on the basis of predictions generated by an approximate nominal process model, and
implemented (un-altered) on the actual process. In other words, the effects of plant-model
mismatch are completely ignored in the control calculation, and closed-loop stability hinges
upon the critical assumption that the nominal model is a “sufficiently close" approximation of
the actual plant. Clearly, this approach is only acceptable for processes whose dynamics can
be modelled a-priori to within a high degree of precision.
For systems whose true dynamics can only be approximated to within a large margin of un-
certainty, it becomes necessary to directly account for the plant-model mismatch. To date, the
most general and rigourous means for doing this involves explicitly accounting for the error
in the online calculation, using the robust-MPC approaches discussed in Section 10.1. While
the theoretical foundations and guarantees of stability for these tools are well established,
it remains problematic in most cases to find an appropriate approach yielding a satisfactory
balance between computational complexity, and conservatism of the error calculations. For
example, the framework of min-max feedback-MPC Magni et al. (2003); Scokaert & Mayne
(1998) provides the least-conservative control by accounting for the effects of future feedback
actions, but is in most cases computationally intractable. In contrast, computationally simple
approaches such as the openloop method of Marruedo et al. (2002) yield such conservatively-
large error estimates, that a feasible solution to the optimal control problem often fails to exist.
For systems involving primarily static uncertainties, expressible in the form of unknown (con-
stant) model parameters θ ∈ Θ ⊂ R p , it would be more desirable to approach the problem in
the framework of adaptive control than that of robust control. Ideally, an adaptive mechanism
enables the controller to improve its performance over time by employing a process model
which asymptotically approaches that of the true system. Within the context of predictive
control, however, the transient effects of parametric estimation error have proven problematic

10 i.e., a CLF guaranteeing robust stabilization to a neighbourhood of the origin, where the size of the
neighbourhood scales with the L∞ bound of the disturbance signal
Robust Adaptive Model Predictive Control of Nonlinear Systems 39

towards developing anything beyond the limited results discussed in Section 11. In short, the
development of a general “robust adaptive-MPC" remains at present an open problem.
In the following, we make no attempt to construct such a “robust adaptive" controller; in-
stead we propose an approach more properly referred to as “adaptive robust" control. The
approach differs from typical adaptive control techniques, in that the adaptation mechanism
does not directly involve a parameter identifier θ̂ ∈ R p . Instead, a set-valued description of
the parametric uncertainty, Θ, is adapted online by an identification mechanism. By gradually
eliminating values from Θ that are identified as being inconsistent with the observed trajecto-
ries, Θ gradually contracts upon θ in a nested fashion. By virtue of this nested evolution of Θ,
it is clear that an adaptive feedback structure of the form in Figure 2 would retain the stability
properties of any underlying robust control design.

Identifier

Robust Controller for Plant

Fig. 2. Adaptive robust feedback structure

The idea of arranging an identifier and robust controller in the configuration of Figure 2 is
itself not entirely new. For example the robust control design of Corless & Leitmann (1981),
appropriate for nonlinear systems affine in u whose disturbances are bounded and satisfy the
so-called “matching condition", has been used by various authors Brogliato & Neto (1995);
Corless & Leitmann (1981); Tang (1996) in conjunction with different identifier designs for
estimating the disturbance bound resulting from parametric uncertainty. A similar concept
for linear systems is given in Kim & Han (2004).
However, to the best of our knowledge this idea has not been well explored in the situation
where the underlying robust controller is designed by robust-MPC methods. The advantage
of such an approach is that one could then potentially imbed an internal model of the identi-
fication mechanism into the predictive controller, as shown in Figure 3. In doing so the effects
of future identification are accounted for within the optimal control problem, the benefits of
which are discussed in the next section.

13. A Minimally-Conservative Perspective

13.1 Problem Description
The problem of interest is to achieve robust regulation, by means of state-feedback, of the
system state to some compact target set Σox ∈ R n . Optimality of the resulting trajectories are
measured with respect to the accumulation of some instantaneous penalty (i.e., stage cost)
L( x, u) ≥ 0, which may or may not have physical significance. Furthermore, the state and
input trajectories are required to obey pointwise constraints ( x, u) ∈ X × U ⊆ R n × R m .
40 Model Predictive Control

Identifier

Robust-MPC Plant
Identifier

Fig. 3. Adaptive robust MPC structure

It is assumed that the system dynamics are not fully known, with uncertainty stemming from
both unmodelled static nonlinearities as well as additional exogenous inputs. As such, the
dynamics are assumed to be of the general form

ẋ = f ( x, u, θ, d(t)) (19)

where f is a locally Lipschitz vector function of state x ∈ R n , control input u ∈ R m , dis-

turbance input d ∈ R d , and constant parameters θ ∈ R p . The entries of θ may represent
physically meaningful model parameters (whose values are not exactly known a-priori), or
alternatively they could be parameters associated with any (finite) set of universal basis func-
tions used to approximate unknown nonlinearities. The disturbance d(t) represents the com-
bined effects of actual exogenous inputs, neglected system states, or static nonlinearities lying
outside the span of θ (such as the truncation error resulting from using a finite basis).

Assumption 13.1. θ ∈ Θo , where Θo is a known compact subset of R p .

Assumption 13.2. d(·) ∈ D∞ , where D∞ is the set of all right-continuous L∞ -bounded functions
d : R → D ; i.e., composed of continuous subarcs d[ a,b) , and satisfying d(τ ) ∈ D , ∀τ ∈ R, with
D ⊂ R d a compact vectorspace.
Unlike much of the robust or adaptive MPC literature, we do not necessarily assume exact
knowledge of the system equilibrium manifold, or its stabilizing equilibrium control map.
Instead, we make the following (weaker) set of assumptions:

Assumption 13.3. Letting Σou ⊆ U be a chosen   assume that L : X × U → R ≥0 is

compact set,
continuous, L(Σox , Σou ) ≡ 0, and L( x, u) ≥ γ L ( x, u)Σox×Σou , γ L ∈ K∞ . As well, assume that
 
L( x, u) c2
min ≥ ∀ x ∈ X \ B(Σox , c1 ) (20)
(u,θ,d)∈U ×Θ ×Do  f ( x, u, θ, d )  x Σox

Definition 13.4. For each Θ ⊆ Θo , let Σ x (Θ) ⊆ Σox denote the maximal (strongly) control-invariant
subset for the differential inclusion ẋ ∈ f ( x, u, Θ, D), using only controls u ∈ Σou .

Assumption 13.5. There exists a constant NΣ < ∞, and a finite cover of Θo (not necessarily unique),
denoted {Θ}Σ , such that
Robust Adaptive Model Predictive Control of Nonlinear Systems 41

i. the collection {Θ̊}Σ is an open cover for the interior Θ̊o .

ii. Θ ∈ {Θ}Σ implies Σ x (Θ) = ∅.
iii. {Θ}Σ contains at most NΣ elements.

The most important requirement of Assumption 13.3 is that, since the exact location (in R n ×
R m ) of the equilibrium11 manifold is not known a-priori, L( x, u) must be identically zero on
the entire region of equilibrium candidates Σox × Σou . One example of how to construct such
a function would be to define L( x, u) = ρ( x, u) L( x, u), where L( x, u) is an arbitrary penalty
satisfying ( x, u) ∈ Σox × Σou =⇒ L( x, u) > 0, and ρ( x, u) is a smoothed indicator function of
the form




0 ( x, u) ∈ Σox × Σou
( x,u)Σox×Σou
ρ( x, u) = δρ 0 < ( x, u)Σox×Σou < δρ (21)


1 ( x, u) o o ≥ δΣ x×Σu ρ

The restriction that L( x, u) is strictly positive definite with respect to Σox × Σou is made for con-
venience, and could be relaxed to positive semi-definite using an approach similar to that of
Grimm et al. (2005) as long as L( x, u) satisfies an appropriate detectability assumption (i.e.,
as long as it is guaranteed that all trajectories remaining in { x | ∃u s.t. L( x, u) = 0} must
asymptotically approach Σox × Σou ).
The first implication of Assumption 13.5 is that for any θ ∈ Θo , the target Σox contains a
stabilizable “equilibrium" Σ(θ ) such that the regulation problem is well-posed. Secondly, the
openness of the covering in Assumption 13.5 implies a type of “local-ISS" property of these
equilibria with respect to perturbations in small neighbourhoods Θ of θ. This property ensures
that the target is stabilizable given “sufficiently close" identification of the unknown θ, such
that the adaptive controller design is tractable.

13.2 Adaptive Robust Controller Design Framework

13.2.1 Adaptation of Parametric Uncertainty Sets
Unlike standard approaches to adaptive control, this work does not involve explicitly gener-
ating a parameter estimator θ̂ for the unknown θ. Instead, the parametric uncertainty set Θo is
adapted to gradually eliminate sets which do not contain θ. To this end, we define the infimal
uncertainty set

Z (Θ, x[ a,b] , u[ a,b] )  { θ ∈ Θ | ẋ (τ ) ∈ f ( x (τ ), u(τ ), θ, D), ∀τ ∈ [ a, b]} (22)

By definition, Z represents the best-case performance that could be achieved by any iden-
tifier, given a set of data generated by (19), and a prior uncertainty bound Θ. Since exact
online calculation of (22) is generally impractical, we assume that the set Z is approximated
online using an arbitrary estimator Ψ. This estimator must be chosen to satisfy the following
conditions.

Criterion 13.6. Ψ(·, ·, ·) is designed such that for a ≤ b ≤ c, and for any Θ ⊆ Θo ,
C13.6.1 Z ⊆ Ψ
C13.6.2 Ψ(Θ, ·, ·) ⊆ Θ, and closed.
11 we use the word “equilibrium" loosely in the sense of control-invariant subsets of the target Σox , which
need not be actual equilibrium points in the traditional sense
42 Model Predictive Control

C13.6.3 Ψ(Θ1 , x[ a,b] , u[ a,b] ) ⊆ Ψ(Θ2 , x[ a,b] , u[ a,b] ), for Θ1 ⊆ Θ2 ⊆ Θo

C13.6.4 Ψ(Θ, x[ a,b] , u[ a,b] ) ⊇ Ψ(Θ, x[ a,c] , u[ a,c] )
C13.6.5 Ψ(Θ, x[ a,c] , u[ a,c] ) ≡ Ψ(Ψ(Θ, x[ a,b] , u[ a,b] ), x[b,c] , u[b,c] )

The set Ψ represents an approximation of Z in two ways. First, both Θo and Ψ can be restricted
a-priori to any class of finitely-parameterized sets, such as linear polytopes, quadratic balls, etc.
Second, contrary to the actual definition of (22), Ψ can be computed by removing values from
Θo as they are determined to violate the differential inclusion model. As such, the search for
infeasible values can be terminated at any time without violating C13.6.
The closed loop dynamics of (19) then take the form

ẋ = f ( x, κmpc ( x, Θ(t)), θ, d(t)), x ( t0 ) = x0 (23a)

o
Θ ( t ) = Ψ ( Θ , x [ t0 , t ] , u [ t0 , t ] ) (23b)

where κmpc ( x, Θ) represents the MPC feedback policy, detailed in Section 13.2.2. In practice,
the (set-valued) controller state Θ could be generated using an update law Θ̇ designed to
gradually contract the set (satisfying C13.6). However, the given statement of (23b) is more
general, as it allows for Θ(t) to evolve discontinuously in time, as may happen for example
when the sign of a parameter can suddenly be conclusively determined.

13.2.2 Feedback-MPC framework

In the context of min-max robust MPC, it is well known that feedback-MPC, because of its abil-
ity to account for the effects of future feedback decisions on disturbance attenuation, provides
significantly less conservative performance than standard open-loop MPC implementations.
In the following, the same principle is extended to incorporate the effects of future parameter
adaptation.
In typical feedback-MPC fashion, the receding horizon control law in (23) is defined by mini-
mizing over feedback policies κ : R ≥0 × R n × cov {Θo } → R m as

u = κmpc ( x, Θ)  κ ∗ (0, x, Θ) (24a)

κ  arg min J ( x, Θ, κ )
∗
(24b)
κ (·,·,·)

where J ( x, Θ, κ ) is the (worst-case) cost associated with the optimal control problem:
 T
p
J ( x, Θ, κ )  max L( x p , u p )dτ + W ( x f , Θ̂ f ) (25a)
θ ∈Θ 0
d(·)∈D∞
s.t. ∀τ ∈ [0, T ]
d p
dτ x = f ( x p , u p , θ, d), x p (0) = x (25b)
p p
Θ̂(τ ) = Ψp (Θ(t), x[0,τ ] , u[0,τ ] ) (25c)
p
x (τ ) ∈ X (25d)
u p (τ )  κ (τ, x p (τ ), Θ̂(τ )) ∈ U (25e)
p p
xf  x ( T ) ∈ Xf (Θ̂ f ) (25f)
p p
Θ̂ f  Ψf (Θ(t), x[0,T ] , u[0,T ] ) (25g)
Robust Adaptive Model Predictive Control of Nonlinear Systems 43

Throughout the remainder, we denote the optimal cost J ∗ ( x, Θ)  J ( x, Θ, κ ∗ ), and further-

more we drop the explicit constraints (25d)-(25f) by assuming the definitions of L and W have
been extended as follows:

L( x, u) < ∞ ( x, u) ∈ X × U
L( x, u) = (26a)
+∞ otherwise

W ( x, Θ) < ∞ x ∈ Xf (Θ)
W ( x, Θ) = (26b)
+∞ otherwise

The parameter identifiers Ψp and Ψf in (25) represent internal model approximations of the
actual identifier Ψ, and must satisfy both C13.6 as well as the following criterion:

Criterion 13.7. For identical arguments, Z ⊆ Ψ ⊆ Ψf ⊆ Ψp .

Remark 13.8. We distinguish between different identifiers to emphasize that, depending on the fre-
quency at which calculations are called, differing levels of accuracy can be applied to the identification
calculations. The ordering in Criterion 13.7 is required for stability, and implies that identifiers existing
within faster timescales provide more conservative approximations of the uncertainty set.

There are two important characteristics which distinguish (25) from a standard (non-adaptive)
feedback-MPC approach. First, the future evolution of Θ̂ in (25c) is fed back into both (25b)
and (25e). The benefits of this feedback are analogous to those of adding state-feedback into
the MPC calculation; the resulting cone of possible trajectories x p (·) is narrowed by account-
ing for the effects of future adaptation on disturbance attenuation, resulting in less conserva-
tive worst-case predictions.
The second distinction is that both W and Xf are parameterized as functions of Θ̂ f , which
reduces the conservatism of the terminal cost. Since the terminal penalty W has the inter-
pretation of the “worst-case cost-to-go", it stands to reason that W should decrease with de-
creased parametric uncertainty. In addition, the domain Xf would be expected to enlarge
with decreased parametric uncertainty, which in some situations could mean that a stabilizing
CLF-pair (W ( x, Θ), Xf (Θ)) can be constructed even when no such CLF exists for the original
uncertainty Θo . This effect is discussed in greater depth in Section 14.1.1.

13.2.3 Generalized Terminal Conditions

To guide the selection of W ( xf , Θ̂ f ) and Xf (Θ̂ f ) in (25), it is important to outline (sufficient)
conditions under which (23)-(25) can guarantee stabilization to the target Σox . The statement
given here is extended from the set of such conditions for robust MPC from Mayne et al. (2000)
that was outlined in Sections 8 and 10.1.1.
For reasons that are explained later in Section 14.1.1, it is useful to present these conditions in
a more general context in which W (·, Θ) is allowed to be LS -continuous with respect to x, as
may occur if W is generated by a switching mechanism. This adds little additional complexity
to the analysis, since (25) is already discontinuous due to constraints.

Criterion 13.9. The set-valued terminal constraint function Xf : cov {Θo } → cov {X } and terminal
penalty function W : R n × cov {Θo } → [0, +∞] are such that for each Θ ∈ cov {Θo }, there exists
k f (·, Θ) : Xf → U satisfying
C13.9.1 Xf (Θ) = ∅ implies that Σox ∩ Xf (Θ) = ∅, and Xf (Θ) ⊆ X is closed
C13.9.2 W (·, Θ) is LS -continuous with respect to x ∈ R n
44 Model Predictive Control

C13.9.3 k f ( x, Θ) ∈ U, for all x ∈ Xf (Θ).

 
C13.9.4 Xf (Θ) and Σ x (Θ) ⊆ Σox ∩ Xf (Θ) are both strongly positively invariant with
respect to the differential inclusion ẋ ∈ f ( x, k f ( x, Θ), Θ, D).
C13.9.5 ∀ x ∈ Xf (Θ), and denoting F  f ( x, k f ( x, Θ), Θ, D),
 
 
W ( x +δv, Θ)−W ( x,Θ)
maxL( x, k f ( x, Θ))+ lim inf δ
≤ 0
f ∈F v→ f
δ ↓0

Although condition C13.9.5 is expressed in a slightly non-standard form, it embodies the stan-
dard interpretation that W must be decreasing by at least an amount − L( x, k f ( x, Θ)) along
all vectorfields in the closed-loop differential inclusion F ; i.e., W ( x, Θ) is a robust-CLF (in
an appropriate non-smooth sense) on the domain Xf (Θ). Lyapunov stability involving LS -
continuous functions is thoroughly studied in Clarke et al. (1998), and provides a meaningful
sense in which W can be considered a “robust-CLF" despite its discontinuous nature.
It is important to note that for the purposes of Criterion 13.9, W ( x, Θ) and Xf (Θ) are param-
eterized by the set Θ, but the criterion imposes no restrictions on their functional dependence
with respect to the Θ argument. This Θ-dependence is required to satisfy the following crite-
ria:

Criterion 13.10. For any Θ1 , Θ2 ∈ cov {Θo } such that Θ1 ⊆ Θ2 ,

C13.10.1 Xf (Θ2 ) ⊆ Xf (Θ1 )
C13.10.2 W ( x, Θ1 ) ≤ W ( x, Θ2 ), ∀ x ∈ Xf (Θ2 )

Designing W and Xf as functions of Θ satisfying Criteria 13.9 and 13.10 may appear pro-
hibitively complex; however, the task is greatly simplified by noting that neither criterion im-
poses any notion of continuity of W or Xf with respect to Θ. A constructive design approach
exploiting this fact is presented in Section 14.1.1.

13.2.4 Closed-loop Stability

Theorem 13.11 (Main result). Given system (19), target Σox , and penalty L satisfying Assumptions
13.1, 13.2, 13.3, 13.5, assume the functions Ψ, Ψp , Ψf , W and Xf are designed to satisfy Criteria
13.6, 13.7, 13.9, and 13.10. Furthermore, let X0  X0 (Θo ) ⊆ X denote the set of initial states, with
uncertainty Θ(t0 ) = Θo , for which (25) has a solution. Then under (23), Σox is feasibly asymptotically
stabilized from any x0 ∈ X0 .

Remark 13.12. As indicated by Assumption 13.5, the existence of an invariant target set Σox (Θo ),
robust to the full parametric uncertainty Θo , is not required for Theorem 13.11 to hold. The identifier
Θ̂ f must be contained in a sufficiently small neighbourhood of (the worst-case) θ such that nontrivial
Xf (Θ̂ f ) and W (·, Θ̂ f ) exist, for (25) to be solvable. While this imposes a minimum performance
requirement on Ψ f , it enlarges the domain X0 for which the problem is solvable.
Robust Adaptive Model Predictive Control of Nonlinear Systems 45

14. Computation and Performance Issues

14.1 Excitation of the closed-loop trajectories
Contrary to much of the adaptive control literature, including adaptive-MPC approaches such
as Mayne & Michalska (1993), the result of Theorem 13.11 does not depend on any auxiliary
excitation signal, nor does it require any assumptions regarding the persistency or quality of
excitation in the closed-loop behaviour.
Instead, any benefits to the identification which result from injecting excitation into the input
signal are predicted by (25c) and (25g), and thereby are automatically accounted for in the
posed optimization. In the particular case where Ψp ≡ Ψf ≡ Ψ, then the controller generated
by (25) will automatically inject the exact type and amount of excitation which optimizes the
cost J ∗ ( x, Θ); i.e., the closed-loop behaviour (23) could be considered “optimally-exciting".
Unlike most a-priori excitation signal design methods, excess actuation is not wasted in trying
to identify parameters which have little impact on the closed-loop performance (as measured
by J ∗ ).
As Ψp and Ψf deviate from Ψ, the convergence result of Theorem 13.11 remains valid. How-
ever, the non-smoothness of J ∗ ( x, Θ) (with respect to both x and Θ) makes it difficult to quan-
tify the impact of these deviations on the closed-loop behaviour. Qualitatively, small changes
to Ψp or Ψf yielding increasingly conservative identification would generally result in the
optimal control solution injecting additional excitation to compensate for the de-sensitized
identifier. However, if the changes to Ψp or Ψf are sufficiently large such that the injection of
additional excitation is insufficient to prevent a discontinuous increase in J ∗ , then it is possi-
ble that the optimal solution may suddenly involve less excitation than previously, to instead
reduce actuation energy. Clearly this behaviour is the result of nonconvexities in the optimal
control problem (24), which is inherently a nonconvex problem even in the absence of the
adaptive mechanisms proposed here.

14.1.1 A Practical Design Approach for W and Xf

Proposition 14.1. Let {(W i , Xfi )} denote a finitely-indexed collection of terminal function candidates,
with indices i ∈ I , where each pair (W i , Xfi ) satisfies Criteria 13.9 and 13.10. Then

W ( x, Θ)  min {W i ( x, Θ)}, Xf ( Θ )  {Xfi (Θ)} (27)
i ∈I
i ∈I

satisfy Criteria 13.9 and 13.10.

Using Proposition 14.1, it is clear that one approach to constructing W (·, ·) and Xf (·) is to use
a collection of pairs of the form
 
  W i ( x ), Xfi Θ ⊆ Θi
i i
W ( x, Θ), Xf (Θ) =
(+∞, ∅) otherwise

This collection can be obtained as follows:

1. Generate a finite collection {Θi } of sets covering Θo
• The elements of the collection can, and should, be overlapping, nested, and ranging
in size.
• Categorize {Θi } in a hierarchical (i.e., “tree") structure such that
46 Model Predictive Control

i. level 1 (i.e., the top) consists of Θo . (Assuming Θo ∈ {Θi } is w.l.o.g., since

W (·, Θo ) ≡ +∞ and Xf (Θo ) = ∅ satisfy Criteria 13.9 and 13.10)
ii. every set in the l’th vertical level is nested inside one or more “parents" on level
l−1
iii. at every level, the “horizontal peers" constitute a cover12 of Θo .
2. For every set Θ j ∈ {Θi }, calculate a robust CLF W j (·) ≡ W j (·, Θ j ), and approximate its
j j
domain of attraction Xf ≡ Xf (Θ j ).

• Generally, W j (·, Θ j ) is selected first, after which Xf (Θ j ) is approximated as either a

maximal level set of W j (·, Θ j ), or as some other controlled-invariant set.
• Since the elements of {Θi } need not be unique, one could actually define multiple
(W i , Xfi ) pairs associated with the same Θ j .
• While not an easy task, this is a very standard robust-control calculation. As such,
there is a wealth of tools in the robust control and viability literatures (see, for exam-
ple Aubin (1991)) to tackle this problem.
3. Calculate W (·, Θ) and Xf (Θ) online:
 
i. Given Θ, identify all sets that are active: I ∗ = I ∗ (Θ)  j | Θ ⊆ Θ j . Using the
hierarchy, test only immediate children of active parents.
j
ii. Given x, search over the active indices to identify I ∗f = I ∗f ( x, I ∗ )  { j ∈ I ∗ | x ∈ Xf }.
Define W ( x, Θ)  min j∈I ∗f W j ( x ) by testing indices in I ∗f , setting W ( x, Θ) = +∞ if
I ∗f = ∅.

j
Remark 14.2. Although the above steps assume that Θ j is selected before Xf , an alternative approach
would be to design the candidates W j (·) on the basis of a collection of parameter values θ̂ j . Briefly,
this could be constructed as follows:
1. Generate a grid of values {θ i } distributed across Θo .
2. Design W j (·) based on a certainty-equivalence model for θ̂ j (for example, by linearization).
j
Specify Xf (likely as a level set of W j ), and then approximate the maximal neighbourhood Θ j of
θ̂ j such that Criterion 13.9 holds.
j
3. For the same (θ j , W j ) pair, multiple (W j , Xf ) candidates can be defined corresponding to differ-
ent Θ j .

14.2 Robustness Issues

One could argue that if the disturbance model D in (19) encompasses all possible sources
of model uncertainty, then the issue of robustness is completely addressed by the min-max
formulation of (25). In practice this is not realistic, since it is generally desirable to explicitly
consider significant disturbances only, or to exclude D entirely if Θ represents the dominant
uncertainty. The lack of nominal robustness to model error in constrained nonlinear MPC is
a well documented problem, as discussed in Grimm et al. (2004). In particular, Grimm et al.
12 specifically, the interiors of all peers must together constitute an open cover
Robust Adaptive Model Predictive Control of Nonlinear Systems 47

(2003); Marruedo et al. (2002) establish nominal robustness (for “accurate-model", discrete-
time MPC) in part by implementing the constraint x ∈ X as a succession of strictly nested
sets. We present here a modification to this approach, that is relevant to the current adaptive
framework.
In addition to ensuring robustness of the controller itself, using methods similar to those men-
tioned above, it is equally important to ensure that the adaptive mechanism Ψ, including its
internal models Ψf and Ψp , exhibits at least some level of nominal robustness to unmodelled
disturbances. By Criterion 13.6.4, the online estimation must evolve in a nested fashion and
therefore the true θ must never be inadvertently excluded from the estimated uncertainty set.
Therefore, just as Z in (22) defined a best-case bound around which the identifiers in the pre-
vious sections could be designed, we present here a modification of (22) which quantifies the
type of conservatism required in the identification calculations for the identifiers to possess
nominal robustness.
For any γ, δ ≥ 0, and with τa  τ − a, we define the following modification of (22):

Z δ,γ (Θ, x[ a,b] , u[ a,b] )  { θ ∈ Θ | B( ẋ, δ + γτa ) ∩ f ( B( x, γτa ), u, θ, D) = ∅, ∀τ } . (28)

Equation (28) provides a conservative outer-approximation of (22) such that Z ⊆ Z δ,γ . The
definition in (28) accounts for two different types of conservatism that can be introduced into
the identification calculations. First, the parameter δ > 0 represents a minimum tolerance
for the distance between actual derivative information from trajectory x[ a,b] and the model
(19) when determining if a parameter value can be excluded from the uncertainty set. For
situations where the trajectory x[ a,b] is itself a prediction, as is the case for the internal models
Ψf and Ψp , the parameter γ > 0 represents increasingly relaxed tolerances applied along the
length of the trajectory. Throughout the following we denote Z δ ≡ Z δ,0 , with analogous
notations for Ψ, Ψf , and Ψp .
The following technical property of definition (28) is useful towards establishing the desired
robustness claim:

Claim 14.3. For any a < b < c, γ ≥ 0, and δ ≥ δ ≥ 0, let x[a,c] be an arbitrary, continuous perturbation
of x[ a,b] satisfying

 γ(τ − a) τ ∈ [ a, b]
i.  x (τ ) − x (τ ) ≤
γ(b − a) τ ∈ [b, c]

 δ − δ + γ(τ − a) τ ∈ [ a, b]
ii.  ẋ (τ ) − ẋ (τ ) ≤
γ(b − a) τ ∈ [b, c]
Then, Z δ,γ satisfies
  
Z δ,γ Z δ (Θ, x[a,b] , u[ a,b] ), x[b,c] , u[b,c] ⊆ Z δ,γ (Θ, x[ a,c] , u[ a,c] ). (29)

Based on (28), we are now able to detail sufficient conditions under which the stability claim of
Theorem 13.11 holds in the presence of small, unmodelled disturbances. For convenience, the
following proposition is restricted to the situation where the only discontinuities in W ( x, Θ)
and Xf (Θ) are those generated by a switching mechanism (as per Prop. 14.1) between a set of
48 Model Predictive Control

candidates {W i ( x, Θ), Xfi (Θ)} that are individually continuous on x ∈ Xfi (Θ) (i.e., a strength-
ening of C13.9.2). With additional complexity, the proposition can be extended to general
LS -continuous penalties W ( x, Θ).
Proposition 14.4. Assume that the following modifications are made to the design in Section 13.2:
i. W ( x, Θ) and Xf (Θ) are constructed as per Prop. 14.1, but with C13.9.2 strengthened to require
the individual W i ( x, Θ) to be continuous w.r.t x ∈ Xfi (Θ).
ii. For some design parameter δx > 0, (26) and (27) are redefined as:
 ←−
L( x, u) ( x, u) ∈ B (X, δx Tτ ) × U
L̃(τ, x, u) =
+∞ otherwise
 ←−
W i ( x ) x ∈ B (Xfi (Θ), δx )
W̃ i ( x, Θ) =
+∞ otherwise

iii. The individual sets Xfi are specified such that there exists δ f > 0, for which C13.9.4 holds for every
← −
inner approximation B (Xfi (Θ), δx ), δx ∈ [0, δx ], where positive invariance is with respect to all
flows generated by the differential inclusion ẋ ∈ B( f ( x, kif ( x, Θ), Θ, D), δ f )
iv. Using design parameters δ > δ > 0 and γ > 0, the identifiers are modified as follows:

• Ψ in (23b) is replaced by Ψδ ≡ Ψδ, 0

 

δ,γ δ,γ
• Ψp and Ψf in (25) are replaced by Ψp and Ψf , respectively

where the new identifiers are assumed to satisfy C13.6, C13.7, and a relation of the form (29).
Then for any compact subset X̄0 ⊆ X0 (Θo ), there exists c∗ = c∗ (γ, δx , δ f , δ, δ , X̄0 ) > 0 such that,
for all x0 ∈ X̄0 and for all disturbances d2  ≤ c ≤ c∗ , the target Σox and the actual dynamics

ẋ = f ( x, κmpc ( x, Θ(t)), θ, d(t)) + d2 (t), x ( t0 ) = x0 (30a)

δ o
Θ(t) = Ψ (Θ , x[t0 ,t] , u[t0 ,t] ) (30b)

are input-to-state stable (ISS); i.e., there exists αd ∈ K such that x (t) asymptotically converges to
B(Σox , αd (c)).

14.3 Example Problem

To demonstrate the versatility of our approach, we consider the following nonlinear system:
 
ẋ1 = − x1 + 2 sin( x1 + πθ1 ) + 1.5θ2 − x1 + x2  x1 + d1 (t)
ẋ2 = 10 θ4a θ4b x1 (u + θ3 ) + d2 (t)

The uncertainty D is given by |d1 | , |d2 | ≤ 0.1, and Θo by θ1 , θ2 , θ3 ∈ [−1, 1], and θ4a ∈
{−1, +1}, θ4b ∈ [0.5, 1]. The control objective is to achieve regulation of x1 to the set x1 ∈
[−0.2, 0.2], subject to the constraints X  {| x1 | ≤ M1 and | x2 | ≤ M2 }, U  {|u| ≤ Mu }, with
M1 , M2 ∈ (0, +∞] and Mu ∈ (1, +∞] any given constants. The dynamics exhibit several
challenging properties: i) state constraints, ii) nonlinear parameterization of θ1 and θ2 , iii) po-
tential open-loop instability with finite escape, iv) uncontrollable linearization, v) unknown
Robust Adaptive Model Predictive Control of Nonlinear Systems 49

sign of control gain, and vi) exogenous disturbances. This system is not stabilizable by any
non-adaptive approach (MPC or otherwise), and furthermore fits very few, if any, existing
frameworks for adaptive control.
One key property of the dynamics (which is arguably necessary for the regulation objective
to be well-posed) is that for any known θ ∈ Θ the target is stabilizable and nominally robust.
This follows by observing that the surface

s  2 sin( x1 + πθ1 ) + 1.5θ2 − x1 + x2 = 0

defines a sliding mode for the system, with a robustness margin |s| ≤ 0.5 for | x1 | ≥ 0.2. This
motivates the design choices:

Xf (Θ)  { x ∈ X |−M2 ≤ Γ( x1 , Θ) ≤ x2 ≤ Γ( x1 , Θ) ≤ M2 }
avg avg
Γ  x1 − 1.5θ 2 − 2 sin( x1 + πθ1 ) − 2π (θ 1 − θ1 ) + 0.5
avg avg
Γ  x1 − 1.5θ 2 − 2 sin( x1 + πθ1 ) − 2π (θ 1 − θ1 ) − 0.5

where θ i , θ i denote upper and lower bounds corresponding to Θ ⊆ Θo , and θ avg  θ + θ

2 . The
set Xf (Θ) satisfies C13.10 and is nonempty for any Θ such that θ 2 − θ 2 + π (θ 1 − θ 1 ) ≤ 0.5,
that defines minimum thresholds for the performance of Ψf and the amount of excitation in
solutions to (25).
∀θ ∈Θo
It can be shown that |s| ≤ 0.5 =⇒ | x1 − x2 | ≤ 4, and that Xf (Θ) is control-invariant us-
ing u ∈ [−1, 1], as long as the sign θ4a is known. This motivates the definitions Σou 
[−1, 1], Σ1 = [−0.2, 0.2], Σ12 = [−4, 4], and Σox  { x | ( x1 , x1 − x2 ) ∈ Σ1 × Σ12 }, plus
the modification of Xf (Θ) above to contain the explicit requirement Θ4a = {−1, +1} =⇒
Xf (Θ) = ∅. Then on x ∈ Xf (Θ), the cost functions W ( x, Θ)  12  x1 2Σ1 and L( x, u) 
 
1 2 2 2
2  x1 Σ1 +  x1 − x2 Σ12 +  u Σo satisfy all the claims of C13.9, since W ≡ L ≡ 0 on x ∈
u
Xf ∩ Σox , and on x ∈ Xf \ Σox one has:
   
Ẇ ≤  x1 Σ1 − 12  x1  + 0.1 ≤ − 12  x1 2Σ1 ≤ − L( x, u).

15. Conclusions
In this chapter we have demonstrated the methodology for adaptive MPC, in which the ad-
verse effects of parameter identification error are explicitly minimized using a robust MPC
approach. As a result, it is possible to address both state and input constraints within the
adaptive framework. Another key advantage of this approach is that the effects of future pa-
rameter estimation can be incorporated into the optimization problem, raising the potential
to significantly reduce the conservativeness of the solutions, especially with respect to design
of the terminal penalty. While the results presented here are conceptual, in that they are gen-
erally intractable to compute due to the underlying min-max feedback-MPC framework, this
chapter provides insight into the maximum performance that could be attained by incorpo-
rating adaptation into a robust-MPC framework.
50 Model Predictive Control

16. Proofs for Section 13

16.1 Proof of Theorem 13.11
This proof will follow the so-called “direct method" of establishing stability by directly prov-
ing strict decrease of J ∗ ( x (t), Θ(t)), for all x ∈ Σox . Stability analysis involving LS -continuous
Lyapunov functions (for example, (Clarke et al., 1998, Thm4.5.5)) typically involves the prox-
imal subgradient ∂ p J ∗ (a generalization of ∇ J), which is a somewhat ambiguous quantity in
the context here given (23b). Instead, this proof exploits an alternative framework involv-
ing subderivates (generalized Dini-derivatives), which is equivalent by (Clarke et al., 1998,
Prop4.5.3). Together, the following two conditions can be shown sufficient to ensure decrease
of J ∗ , where F  f ( x, κmpc ( x, Θ(t)), Θ(t), D)
−
→ J ∗ ( x+δv,Θ(t+δ))− J ∗ ( x,Θ(t))
(i.) max D J ∗ ( x,Θ)  max lim inf δ <0
f ∈F f ∈F v→ f
δ ↓0

←
− J ∗ ( x−δv,Θ(t−δ))− J ∗ ( x,Θ(t))
(ii.) min D J ∗ ( x,Θ)  min lim sup δ >0
f ∈F f ∈F v→ f
δ ↓0

i.e., J ∗ is decreasing on both open future and past neighborhoods of t, for all t ∈ R, where
−
→ ∗ ← −
D J , D J ∗ ∈ [−∞, +∞].
To prove condition (i.), let x p , L p , W p , Θ̂ p correspond to any worst-case minimizing solution of
p
J ∗ ( x (t), Θ(t)), defined on τ ∈ [0, T ]. Additional notations which will be used: Tδ  T +δ, Θ̂ T 
p
Θ̂ f ( T ), Θ̂ Tδ  Θ̂ f ( Tδ ); i.e., both sets represent solutions of the terminal identifier Ψ f , evaluated
p p p p
along x[0,T ] and x[0,T ] , respectively. Likewise, for an arbitrary argument S ∈ {Θ̂ T , Θ̂ Tδ }, we
δ
p p
define WT (S )  W ( x p ( T ), S ) and WTδ (S )  W ( x p ( Tδ ), S ).
p
With the above notations, it can be seen that if the minimizing solution x[0,T ] were extended
p p
to τ ∈ [0, Tδ ] by implementing the feedback u p (τ ) = k f ( x p (τ ), θ̂ T ) on τ ∈ [ T, Tδ ] (i.e., with θ̂ T
fixed), then Criterion C13.9.5 guarantees the inequality
 
p p p p p p p
lim 1δ δL( x T , k f ( x T , Θ̂ T )) + WTδ (Θ̂ T ) − WT (Θ̂ T ) ≤ 0.
δ ↓0

Using this fact, the relationship (i.) follows from:

−→  T 
p p
max D J ∗ ( x, Θ) = max lim inf 1δ J ∗ ( x + δv, Θ(t + δ)) − 0 L p dτ − WT (Θ̂ T )
f ∈F f ∈F v→ f
δ ↓0
 δ T p p
≤ max lim inf 1δ J ∗ ( x + δv, Θ(t + δ)) − 0 L p dτ − δ L p dτ − WT (Θ̂ T)
f ∈F v→ f
δ ↓0
 
p p p p p p p
− δL( x T , k f ( x T , Θ̂ T )) + WTδ (Θ̂ T ) − WT (Θ̂ T )
 T T 
p p
≤ max lim inf 1δ J ∗ ( x + δv, Θ(t + δ))− δ L p dτ − T δ L p dτ − WTδ (Θ̂ T )− δL p |δ
f ∈F v→ f
δ ↓0
 T 
p p
≤ max lim 1δ J ∗ ( x p (δ), Θ̂ p (δ))− δ δL p dτ − WTδ (Θ̂ Tδ ) − δL p |δ
f ∈F δ↓0

≤ − L( x, κmpc ( x, Θ))

The final inequalities are achieved by recognizing:

Robust Adaptive Model Predictive Control of Nonlinear Systems 51


• the L p dτ + W p term is a (potentially) suboptimal cost on the interval [δ, Tδ ], starting
from the point ( x p (δ), Θ̂ p (δ)).
p p
• The relation Θ̂ Tδ ⊆ Θ̂ T holds by Criterion C13.6.4, which implies by Criterion C13.10.2
p p p p
that WTδ (Θ̂ Tδ ) ≤ WTδ (Θ̂ T )
• by C13.7, Θ(t + δ)  Ψ(Θ(t), x[0,δ] , u[0,δ] ) ⊆ Ψ p (Θ(t), x[0,δ] , u[0,δ] ), along any locus con-
necting x and x + δv.
• the lim infv applies over all sequences {vk } → f , of which the particular sequence
x p (δk )− x
{v(δk ) = δ } is a member.
• there exists an arbitrary perturbation of the sequence {v(δk )} satisfying
Ψ p (Θ(t), x[0,δ] ) = Θ̂ p (δ). The lim infv includes the limiting cost J ∗ ( x p (δ), Θ̂ p (δ)) of any
such perturbation of {v(δk )}.
• The cost J ∗ ( x p (δ), Θ̂ p (δ)) is optimal on [δ, Tδ ], and passes through the same point ( x p (δ), Θ̂ p (δ))
as the trajectory defining the L p and W p expressions. Thus, the bracketed expression is
non-positive.
For the purposes of condition (ii.), let x v denote a solution to the prediction model (25b) for
initial condition x v (−δ) = x − δv. Condition (ii.) then follows from:
←−  
T−δ
min D J ∗ ( x, Θ) = min lim sup 1δ −δ Lv dτ + WTv−δ (Θ̂vT−δ ) − J ∗ ( x, Θ)
f ∈F f ∈F v→ f
δ ↓0
  T−δ
≥ min lim sup 1δ δLv |−δ + 0 Lv dτ + WTv−δ (Θ̂vT−δ ) − J ∗ ( x, Θ)
f ∈F v→ f
δ ↓0
 
+ δL( x vT−δ , k f ( x vT−δ , Θ̂vT−δ )) + WTv (Θ̂vT−δ ) − WTv−δ (Θ̂vT−δ )
 T 
≥ min lim sup 1δ δLv |−δ + 0 Lv dτ + WTv (Θ̂vT−δ ) − J ∗ ( x, Θ)
f ∈F v→ f
δ ↓0
 T 
p p
≥ min lim 1δ δL p |−δ + 0 L p dτ + WT (Θ̂ T ) − J ∗ ( x, Θ)
f ∈F δ↓0

≥ L( x, κmpc ( x, Θ))
←−
The above derivation made use of the fact that the reverse subderivate D W satisfies
  
W ( x−δv, Θ)−W ( x,Θ)
min lim sup − L( x − δv, k f ( x − δv, Θ))+ δ ≥0
f ∈F v→ f
δ ↓0

which follows from a combination of C13.9.5 and the LS -continuity of W.

←− −
→
Using the above inequalities for D J ∗ ( x, Θ) and D J ∗ ( x, Θ) together with Assumption 13.3, it
follows that J ∗ (t) is strictly decreasing on x ∈ Σox and non-increasing on x ∈ Σox . It follows
that limt→∞ ( x, Θ) must converge to an invariant subset of Σ0x × cov {Θo }. Assumption 13.1
guarantees that such an invariant subset exists, since it implies ∃δ∗ > 0 such Σ x ( B(θ ∗ , δ∗ )) =
∅, with θ ∗ the actual unknown parameter in (19). Continued solvability of (25) as ( x (t), Θ(t))
evolve follows by: 1) x (τ ) ∈ X0 (Θ(τ )) ⇒ J ∗ (τ ) = +∞, and 2) if x (t) ∈ X0 (Θ(t)) and
x (t ) ∈ X0 (Θ(t )), then (t − t) ↓ 0 contradicts either condition (i.) at time t, or (ii.) at time t .
52 Model Predictive Control

16.2 Proof of Proposition 14.1

The fact that C13.10 holds is a direct property of the union and min operations for the closed
sets Xfi , and the fact that the Θ-dependence of individual (W i , Xfi ) satisfies C13.10. For the
purposes of C13.9, the Θ argument is a constant, and is omitted from notation. Properties
C13.9.1 and C13.9.2 follow directly by (27), the closure of Xfi , and (2). Define

I f ( x )& = {i ∈ I | x ∈ Xfi and W ( x ) = W i ( x )}

Denoting F i  f ( x, kif ( x ), Θ, D), the following inequality holds for every i ∈ I f ( x ):

W ( x +δv)−W ( x ) W i ( x +δv)−W ( x )
max lim inf δ ≤ max lim inf δ ≤ − L( x, kif ( x ))
f i ∈F i v→ f i f i ∈F i v→ f i
δ ↓0 δ ↓0

i( x)
It then follows that u = k f ( x )  k f ( x ) satisfies C13.9.5 for any arbitrary selection rule
i ( x ) ∈ I f ( x ) (from which C13.9.3 is obvious). Condition C13.9.4 follows from continuity of
the x (·) flows, and observing that by (26), C13.9.5 would be violated at any point of departure
from Xf .

16.3 Proof of Claim 14.3

By contradiction, let θ ∗ be a value contained in the left-hand side of (29), but not in the right-
hand side. Then by (28), there exists τ ∈ [ a, c] (i.e., τa ≡ (τ − a) ∈ [0, c − a]) such that

f ( B( x, γτa ), u, θ ∗ , D) ∩ B( ẋ, δ + γτa ) = ∅ (31)

Using the bounds indicated in the claim, the following inclusions hold when τ ∈ [ a, b]:

f ( x  , u, θ ∗ , D) ⊆ f ( B( x, γτa ), u, θ ∗ , D) (32a)
B( ẋ  , δ ) ⊆ B( ẋ, δ + γτa ) (32b)

Combining (32) and (31) yields


f ( x  , u, θ ∗ , D) ∩ B( ẋ  , δ ) = ∅ =⇒ θ ∗ ∈ Z δ (Θ, x[a,τ ] , u[ a,τ ] ) (33)

which violates the initial assumption that θ ∗ is in the LHS of (29). Meanwhile, for τ ∈ [b, c]
the inclusions

f ( B( x  , γτb ), u, θ ∗ , D) ⊆ f ( B( x, γτa ), u, θ ∗ , D) (34a)

B( ẋ  , δ + γτb ) ⊆ B( ẋ, δ + γτa ) (34b)

yield the same contradictory conclusion:

f ( B( x  , γτb ), u, θ ∗ , D) ∩ B( ẋ  , δ + γτb ) = ∅ (35a)

  
=⇒ θ ∗ ∈ Z δ,γ Z δ (Θ, x[a,b] , u[ a,b] ), x[b,τ ] , u[b,τ ] (35b)

It therefore follows that the containment indicated in (29) necessarily holds.

Robust Adaptive Model Predictive Control of Nonlinear Systems 53

16.4 Proof of Proposition 14.4

It can be shown that Assumption 13.3, together with the compactness of Σ x , is sufficient for an
∗ interpreted in a min − max sense). In other words,
analogue of Claim ?? to hold (i.e., with J∞
∗
the cost J ( x, Θ) satisfies

αl ( x Σox , Θ) ≤ J ∗ ( x, Θ) ≤ αh ( x Σox , Θ)

for some functions αl , αh which are class-K∞ w.r.t. x, and whose parameterization in Θ satis-
↑
fies αi ( x, Θ1 ) ≤ αi ( x, Θ2 ), Θ1 ⊆ Θ2 . We then define the compact set X̄0  { x | minΘ∈cov{Θo } J ∗ ( x, Θ) <
0
maxx0 ∈X̄0 αh ( x0 Σox , Θ )}.
By a simple extension of (Khalil, 2002, Thm4.19), the ISS property follows if it can be shown
that there exists αc ∈ K such that J ∗ ( x, Θ) satisfies
 −
→
↑ o max f ∈Fc D J ∗ ( x, Θ) < 0
x ∈ X̄0 \ B(Σ x , αc (c)) ⇒ ←
− (36)
min f ∈Fc D J ∗ ( x, Θ) > 0

where Fc  B( f ( x, κmpc ( x, Θ(t)), Θ(t), D), c). To see this, it is clear that J decreases until
x (t) enters B(Σox , αc (c)). While this set is not necessarily invariant, it is contained within an
invariant, compact levelset Ω(c, Θ)  { x | J ∗ ( x, Θ) ≤ αh (αc (c), Θ)}. By C13.6.4, the evolution
of Θ(t) in (30b) must approach some constant interior bound Θ∞ , and thus limt→∞ x (t) ∈
Ω(c, Θ∞ ). Defining αd (c)  maxx∈Ω(c,Θ∞ )  x Σox completes the Proposition, if c∗ is sufficiently
↑
small such that B(Σox , αd (c∗ )) ⊆ X̄0 .
Next, we only prove decrease in the forward direction, since the reverse direction follows
analogously, as it did in the proof of Theorem 13.11. Using similar procedure and notation as
p
the Thm 13.11 proof, x[0,T ] denotes any worst-case prediction at (t, x, Θ), extended to [ T, Tδ ]
via k f , that is assumed to satisfy the specifications of Proposition 14.4. Following the proof of
Theorem 13.11,
−→  T 
p p
max D J ∗ ( x, Θ) ≤ max lim inf 1δ J ∗ ( x + δv, Θ(t + δ))− δ δL p dτ − WTδ (Θ̂ T ) − L p |δ
f ∈Fc∗ f ∈F v→ f
δ ↓0
 T 
≤ max lim inf 1δ J ∗ ( x + δv, Θ(t + δ))− δ δLv dτ − WTvδ (Θ̂vTδ ) − L p |δ
f ∈F v→ f
δ ↓0
  Tδ p 
1 Tδ v v v p p
+ δ δ L dτ + WTδ ( Θ̂ Tδ )− δ L dτ − WTδ ( Θ̂ T ) (37)

where Lv , W v denote costs associated with a trajectory x[v0,T ] satisfying the following:
δ

• initial conditions x v (0) = x, Θv (0) = Θ.

p
• generated by the same worst-case θ̂ and d(·) as x[0,T ]
δ

• dynamics of form (30) on τ ∈ [0, δ], and of form (25b),(25c) on τ ∈ [δ, Tδ ], with the
trajectory passing through x v (δ) = x + δv, Θvp (δ) = Θ(t + δ).
• the minκ in (25) is constrained such that κ v (τ, x v , Θv ) = κ p (τ, x p , Θ p ); i.e., uv[0,T ] ≡
δ
p
u[0,T ] ≡ u[0,Tδ ] .
δ
54 Model Predictive Control

↑
Let Kf denote a Lipschitz constant of (19) with respect to x, over the compact domain X̄0 ×
Θo ×D . Then, using the comparison lemma (Khalil, 2002, Lem3.4) one can derive the bounds

 x v − x p  ≤ Kcf (eKf τ − 1)
τ ∈ [0, δ] : (38a)
 ẋ v − ẋ p  ≤ c eKf τ

 x v − x p  ≤ Kcf (eKf δ − 1) eKf (τ −δ)
τ ∈ [δ, Tδ ] : (38b)
 ẋ v − ẋ p  ≤ c (eKf δ − 1) eKf (τ −δ)

As δ ↓ 0, the above inequalities satisfy the conditions of Claim 14.3 as long as c∗ < min{γ, (δ −
δ ), γeKf T , Kγf eKf T }, thus yielding

p p
Θ̂vf = Ψ f (Ψδ (Θ, x[v0,δ] , u[0,δ] ), x[vδ,Tδ ] , u[δ,Tδ ] ) ⊆ Ψ f (Θ, x[0,T ] , u[0,Tδ ] ) = Θ̂ f
δ,γ  δ,γ
δ

p
as well as the analogue Θ̂vp (τ ) ⊆ Θ̂ p (τ ), ∀τ ∈ [0, Tδ ].
p
Since x[0,T ] is a feasible solution of the original problem from (t, x, Θ) with τ ∈ [0, T ], it follows
for the new problem posed at time t + δ that x v is feasible with respect to the appropriate inner
p p
approximations of X and Xfi (Θ̂ T ) ⊆ Xf (Θ̂vTδ ) (where i∗ denotes an active terminal set for x f )
∗

if

v p δ δTx τ ∈ [δ, T ]
x − x  ≤
δ δ f τ ∈ [ T, Tδ ]

which holds by (38) as long as c∗ < min{δ f , δTx } e−Kf T . Using arguments from the proof
Theorem 13.11, the first term in (37) can be eliminated, leaving:
−
→  T 
T p p
max D J ∗ ( x, Θ) ≤ max lim inf 1δ δ δLv dτ + WTvδ (Θ̂vTδ )− δ δL p dτ − WTδ (Θ̂ T ) − L p |δ
f ∈Fc f ∈F v→ f
δ ↓0
 
Tδ
≤ max lim inf 1δ δ K L  x v − x p dτ + KW  x v ( T ) − x p ( T )− L p |δ
f ∈F v→ f
δ ↓0
 
c ( e K f δ − 1)  
≤ lim KW + TK L eK f T − L p |δ
δ ↓0 Kf δ
≤ − L( x, k MPC ( x, Θ)) + c(KW + TK L )eK f T
<0 ∀ x ∈ X̄0↑ \ B(Σox , αc (c))

with αc ∈ K given by
 
1 Kf T
αc (c)  γ−
L
c ( K W + TK L ) e

↑
where KW is a Lipschitz constant of W i ( x, Θ) over the compact domain X̄0 ∩ Xfi (Θ), maximal
∗ ∗

over all Θ ∈ cov {Θo }. Likewise, K L is a Lipschitz constant of L( x, u) with respect to x,

maximal over u ∈ U.
This proves the forward case in (36), with the reverse case following similarly. As argued
previously, this is sufficient to yield the ISS property of (30) with respect to d2  ≤ c ≤ c∗ ,
which completes the proof.
Robust Adaptive Model Predictive Control of Nonlinear Systems 55

17. References
Adetola, V. & Guay, M. (2004). Adaptive receding horizon control of nonlinear systems, Proc.
IFAC Symposium on Nonlinear Control Systems, Stuttgart, Germany, pp. 1055–1060.
Aubin, J. (1991). Viability Theory, Systems & Control: Foundations & Applications, Birkhäuser,
Boston.
Bellman, R. (1952). The theory of dynamic programming, Proc. National Academy of Science,,
number 38, USA.
Bellman, R. (1957). Dynamic Programming, Princeton Press.
Bertsekas, D. (1995). Dynamic Programming and Optimal Control, Vol. I, Athena Scientific, Bel-
mont, MA.
Brogliato, B. & Neto, A. T. (1995). Practical stabilization of a class of nonlinear systems with
partially known uncertainties, Automatica 31(1): 145 – 150.
Bryson, A. & Ho, Y. (1969). Applied Optimal Control, Ginn and Co., Waltham, MA.
Cannon, M. & Kouvaritakis, B. (2005). Optimizing prediction dynamics for robust MPC,
50(11): 1892–1897.
Chen, H. & Allgöwer, F. (1998a). A computationally attractive nonlinear predictive control
scheme with guaranteed stability for stable systems, Journal of Process Control 8(5-
6): 475–485.
Chen, H. & Allgöwer, F. (1998b). A quasi-infinite horizon nonlinear model predictive control
scheme with guaranteed stability, Automatica 34(10): 1205–1217.
Chen, H., Scherer, C. & Allgöwer (1997). A game theoretic approach to nonlinear robust
receding horizon control of constrained systems, Proc. American Control Conference.
Clarke, F., Ledyaev, Y., Stern, R. & Wolenski, P. (1998). Nonsmooth Analysis and Control Theory,
Grad. Texts in Math. 178, Springer-Verlag, New York.
Corless, M. J. & Leitmann, G. (1981). Continuous state feedback guaranteeing uniform ulti-
mate boundedness for uncertain dynamic systems., IEEE Trans. Automat. Contr. AC-
26(5): 1139 – 1144.
Coron, J. & Rosier, L. (1994). A relation between continuous time-varying and discontinuous
feedback stabilization, Journal of Mathematical Systems, Estimation, and Control 4(1): 67–
84.
Cutler, C. & Ramaker, B. (1980). Dynamic matrix control - a computer control algorithm,
Proceedings Joint Automatic Control Conference, San Francisco, CA.
De Nicolao, G., Magni, L. & Scattolini, R. (1996). On the robustness of receding horizon control
with terminal constraints, IEEE Trans. Automat. Contr. 41: 454–453.
Findeisen, R., Imsland, L., Allgöwer, F. & Foss, B. (2003). Towards a sampled-data theory
for nonlinear model predictive control, in C. Kang, M. Xiao & W. Borges (eds), New
Trends in Nonlinear Dynamics and Control, and their Applications, Vol. 295, Springer-
Verlag, New York, pp. 295–313.
Freeman, R. & Kokotović, P. (1996a). Inverse optimality in robust stabilization, SIAM Journal
of Control and Optimization 34: 1365–1391.
Freeman, R. & Kokotović, P. (1996b). Robust Nonlinear Control Design, Birkh auser.
Grimm, G., Messina, M., Tuna, S. & Teel, A. (2003). Nominally robust model predictive control
with state constraints, Proc. IEEE Conf. on Decision and Control, pp. 1413–1418.
Grimm, G., Messina, M., Tuna, S. & Teel, A. (2004). Examples when model predictive control
is non-robust, Automatica 40(10): 1729–1738.
56 Model Predictive Control

Grimm, G., Messina, M., Tuna, S. & Teel, A. (2005). Model predictive control: for want of a
local control lyapunov function, all is not lost, IEEE Trans. Automat. Contr. 50(5): 617–
628.
Hermes, H. (1967). Discontinuous vector fields and feedback control, in J. Hale & J. LaSalle
(eds), Differential Equations and Dynamical Systems, Academic Press, New York,
pp. 155–166.
Hestenes, M. (1966). Calculus of Variations and Optimal Control, John Wiley & Sons, New York.
Jadbabaie, A., Yu, J. & Hauser, J. (2001). Unconstrained receding-horizon control of nonlinear
systems, IEEE Trans. Automat. Contr. 46(5): 776 – 783.
Kalman, R. (1960). Contributions to the theory of optimal control, Bol. Soc. Mat. Mexicana
5: 102–119.
Kalman, R. (1963). Mathematical description of linear dynamical systems, SIAM J. Control
1: 152–192.
Keerthi, S. S. & Gilbert, E. G. (1988). Optimal, infinite horizon feedback laws for a general class
of constrained discrete time systems: Stability and moving-horizon approximations,
Journal of Optimization Theory and Applications 57: 265–293.
Khalil, H. (2002). Nonlinear Systems, 3rd edn, Prentice Hall, Englewood Cliffs, N.J.
Kim, J.-K. & Han, M.-C. (2004). Adaptive robust optimal predictive control of robot manipu-
lators, IECON Proceedings (Industrial Electronics Conference) 3: 2819 – 2824.
Kothare, M., Balakrishnan, V. & Morari, M. (1996). Robust constrained model predictive con-
trol using linear matrix inequalities, Automatica 32(10): 1361–1379.
Kouvaritakis, B., Rossiter, J. & Schuurmans, J. (2000). Efficient robust predictive control, IEEE
Trans. Automat. Contr. 45(8): 1545 – 1549.
Langson, W., Chryssochoos, I., Raković, S. & Mayne, D. (2004). Robust model predictive
control using tubes, Automatica 40(1): 125 – 133.
Lee, E. & Markus, L. (1967). Foundations of Optimal Control Theory, Wiley.
Lee, J. & Yu, Z. (1997). Worst-case formulations of model predictive control for systems with
bounded parameters, Automatica 33(5): 763–781.
Magni, L., De Nicolao, G., Scattolini, R. & Allgöwer, F. (2003). Robust model predictive con-
trol for nonlinear discrete-time systems, International Journal of Robust and Nonlinear
Control 13(3-4): 229–246.
Magni, L., Nijmeijer, H. & van der Schaft, A. (2001). Receding-horizon approach to the non-
linear h∞ control problem, Automatica 37(3): 429 – 435.
Magni, L. & Sepulchre, R. (1997). Stability margins of nonlinear receding-horizon control via
inverse optimality, Systems and Control Letters 32: 241–245.
Marruedo, D., Alamo, T. & Camacho, E. (2002). Input-to-state stable MPC for constrained
discrete-time nonlinear systems with bounded additive uncertainties, Proc. IEEE
Conf. on Decision and Control, pp. 4619–4624.
Mayne, D. (1995). Optimization in model based control, Proc. IFAC symposium on dynamics
and control, chemical reactors and batch processes (DYCORD), Oxford: Elsevier Science.,
pp. 229–242. plenary address.
Mayne, D. Q. & Michalska, H. (1990). Receding horizon control of non-linear systems, IEEE
Trans. Automat. Contr. 35(5): 814–824.
Mayne, D. Q. & Michalska, H. (1993). Adaptive receding horizon control for constrained
nonlinear systems, Proc. IEEE Conf. on Decision and Control, pp. 1286–1291.
Mayne, D. Q., Rawlings, J. B., Rao, C. V. & Scokaert, P. O. M. (2000). Constrained model
predictive control: Stability and optimality, Automatica 36: 789–814.
Robust Adaptive Model Predictive Control of Nonlinear Systems 57

Michalska, H. & Mayne, D. (1993). Robust receding horizon control of constrained nonlinear
systems, IEEE Trans. Automat. Contr. 38(11): 1623 – 1633.
Pontryagin, L. (1961). Optimal regulation processes, Amer. Math. Society Trans., Series 2 18: 321–
339.
Primbs, J. (1999). Nonlinear Optimal Control: A Receding Horizon Approach, PhD thesis, Califor-
nia Institute of Technology, Pasadena, California.
Primbs, J., Nevistic, V. & Doyle, J. (2000). A receding horizon generalization of pointwise
min-norm controllers, IEEE Trans. Automat. Contr. 45(5): 898–909.
Raković, S. & Mayne, D. (2005). Robust time optimal obstacle avoidance problem for con-
strained discrete time systems, Proc. IEEE Conf. on Decision and Control.
Ramirez, D., Alamo, T. & Camacho, E. (2002). Efficient implementation of constrained min-
max model predictive control with bounded uncertainties, Proc. IEEE Conf. on Deci-
sion and Control, pp. 3168–3173.
Richalet, J., Rault, A., Testud, J. & Papon, J. (1976). Algorithmic control of industrial processes,
Proc. IFAC symposium on identification and system parameter estimation, pp. 1119–1167.
Richalet, J., Rault, A., Testud, J. & Papon, J. (1978). Model predictive heuristic control: Appli-
cations to industrial processes, Automatica 14: 413–428.
Sage, A. P. & White, C. C. (1977). Optimum Systems Control, 2nd edn, Prentice-Hall.
Scokaert, P. & Mayne, D. (1998). Min-max feedback model predictive control for constrained
linear systems, IEEE Trans. Automat. Contr. 43(8): 1136–1142.
Sepulchre, R., Jankovic, J. & Kokotovic, P. (1997). Constructive Nonlinear Control, Springer, New
York.
Sontag, E. (1989). A “universal" construction of artstein’s theorem on nonlinear stabilization,
Systems and Control Letters 13: 117–123.
Sontag, E. D. (1983). Lyapunov-like characterization of asymptotic controllability., SIAM Jour-
nal on Control and Optimization 21(3): 462 – 471.
Tang, Y. (1996). Simple robust adaptive control for a class of non-linear systems: an adaptive
signal synthesis approach, International Journal of Adaptive Control and Signal Process-
ing 10(4-5): 481 – 488.
Tuna, S., Sanfelice, R., Messina, M. & Teel, A. (2005). Hybrid MPC: Open-minded but not
easily swayed, International Workshop on Assessment and Future Directions of Nonlinear
Model Predictive Control, Freudenstadt-Lauterbad, Germany, pp. 169–180.
58 Model Predictive Control
A new kind of nonlinear model predictive control
algorithm enhanced by control lyapunov functions 59

x3

A new kind of nonlinear model

predictive control algorithm
enhanced by control lyapunov functions
Yuqing He and Jianda Han
State Key Laboratory of Robotics, Shenyang Institute of Automation,
Chinese Academy of Sciences
P.R.China

1. Introduction
With the abilities of handling constraints and performance of optimization, model based
predictive control (MPC), especially linear MPC, has been extensively researched in theory
and applied in practice since it was firstly proposed in 1970s (Qin & Badgwell, 2003).
However, when used in systems with heavy nonlinearities, nonlinear MPC (NMPC) results
often in problems of high computational cost or closed loop instability due to their
complicated structure. This is the reason why the gaps between NMPC theory and its
applications in reality are larger and larger, and why researches on NMPC theory absorbs
numerous scholars (Chen & Shaw, 1982; Henson, 1998 ; Mayne, et al., 2000 ; Rawlings, 2000).
When the closed loop stability of NMPC is concerned, some extra strategies is necessary,
such as increasing the length of the predictive horizon, superinducing state constraints, or
introducing Control Lyapunov Functions (CLF).
That infinite predictive/control horizon (in this chapter, predictive horizon is assumed
equal to control horizon) can guarantee the closed loop stability is natural with the
assumption of feasibility because it implicates zero terminal state, which is a sufficient
stability condition in many NMPC algorithm (Chen and Shaw, 1982). In spite of the
inapplicability of infinite predictive horizon in real plants, a useful proposition originated
from it makes great senses during the development of NMPC theory, i.e., a long enough
predictive horizon can guarantee the closed loop stability for most systems (Costa & do Val,
2003; Primbs & Nevistic, 2000). Many existing NMPC algorithm is on the basis of this result,
such as Chen & Allgower (1998), Magni et al. (2001). Although long predictive horizon
scheme is convenient to be realized, the difficulty to obtain the corresponding threshold
value makes this scheme improper in many plants, especially in systems with complicated
structure. For these cases, another strategy, superinducing state constraints or terminal
constraints, is a good substitue. A typical predictive control algorithm using this strategy is
the so called dual mode predictive control(Scokaert et al., 1999 ; Wesselowske and Fierro,
2003 ; Zou et al., 2006), which is originated from the predictive control with zero terminal
state constrains and can increase its the stability region greatly. CLF is a new introduced
60 Model Predictive Control

concept to design nonlinear controller. It is firstly used in NMPC by Primbs et al. in 1999 to
obtain two typical predictive control algorithm with guaranteed stability.
Unfortunately, each approach above will result in huge computational burden
simultaneously since they bring either more constraints or more optimizing variables. It is
well known that the high computational burden of NMPC mainly comes from the online
optimization algorithm, and it can be alleviated by decreasing the number of optimized
variables. But this often deteriorates the closed loop stability due to the changed structure of
optimal control problem at each time step.
In a word, the most important problem during designing NMPC algorithm is that the
stability and computational burden are deteriorated by each other. Another problem,
seldom referred to but top important, is that the stability can only be guarangteed under the
condition of perfect optimization algorithm that is impossible in reality. Thus, how to design
a robustly stable and fast NMPC algorithm has been one of the most difficult problems that
many researchers are pursued.
In this chapter, we attempt to design a new stable NMPC which can partially solve the
problems referred to above. CLF, as a new introduced concept to design nonlinear controller
by directly using the idea of Lyapunov stability analysis, is used in this chapter to ensure the
stability. Firstly, a generalized pointwise min-norm (GPMN) controller (a stable controller
design method) based on the concept of CLF is designed. Secondly, a new stable NMPC
algorithm, called GPMN enhanced NMPC (GPMN-ENMPC), is given through
parameterized GPMN controller. The new algorithm has the following two advantages, 1) it
can not only ensure the closed loop stability but also decrease the computational cost
flexibly at the price of sacrificing the optimality in a certain extent; 2) a new tool of guide
function is introduced by which some extra control strategy can be considered implicitly.
Subsequently, the GPMN-ENMPC algorithm is generalized to obtain a robust NMPC
algorithm with respect to the feedback linearizable system. Finally, extensive simulations
are conducted and the results show the feasibility and validity of the proposed algorithm.

2. Concept of CLF
The nonlinear system under consideration in this chapter is in the form as:

x  f ( x )  g ( x)u
(1)
u U  Rm

where x  R n is state vector, u  R m is input vector, f(*) and g(*) are nonlinear smooth
functions with f(0) = 0. U is the control constraint.

Definition I:
For system (1), if there exists a C1 function V(x): x  RnR+  {0}, such that
1) V(0) = 0, V(x) > 0 if x ≠0;
2) a1(||x||) < V(x) < a2(||x||), where a1(*) and a2(*) are class K∞ functions;
3) inf [Vx ( x) f ( x)  Vx ( x) g ( x)u ]  0, x  c  {0} ,where c  {x  R n : V ( x)  c} .
uU  R m
then V(x) is called a CLF of system (1). Moreover, if x can be chosen as Rn and V(x) satisfies
the following condition,
A new kind of nonlinear model predictive control
algorithm enhanced by control lyapunov functions 61

V(x)∞ ==> ||x||∞

then V(x) is called a global CLF of system (1). █

If system (1) has uncertainty terms, i.e.,

x  f ( x)  g ( x)u  l ( x)
y  h( x ) (2)
m
u U  R

where ω  Rq is external disturbance; l(*) and h(*) are pre-defined nonlinear smooth
functions; y is the interested output. We have the following concept of robust version CLF –
called H∞CLF,

Definition II,
For system (2), if there exists a C1 function V(x): x  RnR+  {0}, such that
1) V(0) = 0, V(x) > 0 if x ≠0;
2) a1(||x||) < V(x) < a2(||x||), where a1(*) and a2(*) are class K∞ functions;
1 1
3) infm {Vx ( x )[ f ( x )  g ( x )u ]  h ( x ) h ( x )}  0, x   c   c2 , where c1>c2.
T T T
Vx ( x )l ( x )l ( x )Vx 
2
2
uR 2 1

then V(x) is called a local H∞CLF of system (2) in c1  c2 . Furthermore, V(x) is called a
global H∞CLF if c1 can be chosen +∞ with V(x)∞ as |x|∞. █

Definition I and II indicate that if we can obtain a CLF or H∞CLF of system (1) or (2), a
‘permitted’ control set can be found at every ‘feasible’ state, and the control action inside the
set can guarantee the closed loop stability of system (1) or input output finite gain L2
stability of system (2). Subsequently, in order to complete the controller design, what one
needs to do is just to find an approach to select a sequence of control actions from the
‘permitted control set’, see Fig. 1.

Input

State

Fig. 1. Sketch of CLF, the shadow indicates the ‘permitted’ set of (x, u) V ( x, u ) along system (1)
62 Model Predictive Control

CLF based nonlinear controller design method is also called direct method of Lyapunov
function based controller design, and its difficulty is how to ensure the controller’s
continuousness. Thus, most recently, researchers mainly pay their attentions to designing
continuous CLF based controller, and several universal formulas have been revealed.
Sontag’s formula (Sontag, 1989), for example, originated from the root calculation of 2nd-
order equation, can be written as Eq. (3) through slightly modification by Freeman (Freeman
& Kokotovic, 1996b),

 V ( x ) f  (V ( x ) f ( x )) 2  q( x)(V ( x ) g ( x ) g T ( x )V T ( x )) 
  x x x x
 Vx g  0
u  Vx ( x ) g ( x ) g T ( x )VxT ( x )  (3)


 0 Vx g  0

where q(x) is a pre-designed positive definite function.

Pointwise Min-Norm (PMN) control is another well known CLF-based approach proposed
by Freeman (Freeman & Kokotovic, 1996a),

min u
u

s.t. Vx ( x)[ f ( x)  g ( x)u ]   ( x) (4)

u U

where σ(x) is a pre-selected positive definite function. Controller (4) can also be explicitly
denoted as (5) if the constraint set U can be selected big enough.

 [Vx ( x) f ( x)   ( x)] g T ( x)VxT ( x)

 Vx ( x) f ( x)   ( x)  0
u Vx ( x) g ( x) g T ( x)VxT ( x) (5)
 0 Vx ( x) f ( x)   ( x)  0


(3) and (5) provide two different methods on how to design continuous and stable controller
based on CLF with respect to system (1). H∞CLF with respect to system (2) is a new given
concept, and there are no methods can be used to designed robust controller based on it.
Although the closed loop stability can be guaranteed using controller (3) or controller (5),
selection of parameters q(x) or σ(x) is too difficult to be used in real applications. This is
mainly because these parameters heavily influence some inconsistent closed loop
performance simultaneously. Furthermore, if the known CLF is not global, the selection of
q(x) and σ(x) will also influence stability margin of the closed loop systems, which makes
them more difficult to be selected (Sontag, 1989; Freeman & Kokotovic, 1996a). In this
chapter, we will firstly give a new CLF based controller design strategy, which is superior
compared to the existing CLF based controller design methods referred to above.
Furthermore, the most important is that this new strategy can be used in designing robustly
stable and fast NMPC algorithm.
A new kind of nonlinear model predictive control
algorithm enhanced by control lyapunov functions 63

3. GPMN-ENMPC
3.1 CLF based GPMN controller
Since q(x) and σ(x) in controller (3) and controller (5) are difficult to select, a guide function is
proposed in this subsection into the PMN controller to obtain a new CLF based nonlinear
controller with respect to system (1), in the following section, this controller will be
generated with respect to system (2). In the new controller, σ(x) is only used to ensure the
stability of the closed loop, while the other desired performance of the controller, for
example tracking performance, can be guaranteed by the guide function, which, as new
controller parameters, can be designed without deteriorating the stability. The following
proposition is the main result of this subsection.

Proposition I:
If V(x) is a CLF of system (1) in Ωc and ξ(x): RnRm is a continuous guide function such that
ξ(0) = 0, then, the following controller can stabilize system (1),

u ( x)  arg min { u   ( x) }
uKV ( x )
(6)
KV ( x)  { y | Vx ( x) f ( x)  Vx ( x) g ( x) y   ( x), y  U }

where σ(x) is a positive definite function of state, and ξ(x), called guide function, is a
continuous state function.
Proof of Proposition I:
Let V(x) be a Lyapunov function candidate for system (1), then we have

V ( x)  Vx ( x) f ( x)  Vx ( x) g ( x)u (7)

Substitute Eq. (6) into (7), it is not difficult to obtain the following inequality,

V ( x)  Vx ( x) f ( x)  Vx ( x) g ( x)u   ( x)

Because σ(x) is a positive definite function, proposition I is proved. █

Controller (6) is called Generalized Pointwise Min-Norm (GPMN) controller. The difference
between the proposed GPMN controller and the normal PMN controller of Eq. (4) can be
illustrated in Fig.2: for the normal PMN algorithm (Fig. 2a), the controller output in each
state point has the minimum ‘permitted’ norm (close to the state-axis as much as possible),
while the GPMN controller’s output has nearest distance from the guide function ξ(x) (Fig.
2b). Thus, ξ(x) in GPMN controller is actual a performance criterion which the controller is
expected to pursue, while σ(x) dedicates only on providing the ‘permitted’ stable control
input sets.
Up to now, the design of new GPMN controller has been completed. However, in order to
use a GPMN controller in reality or in NMPC algorithm, analytical form of the solution of
Eq. (6) is necessary to be studied.
Firstly, if there are no input constraints (or the input constraint sets are big enough), the
analytical form of controller (6) can be obtained as follows, based on the projection theory,
64 Model Predictive Control

 [Vx f    Vx g ( x)]g T VxT

   ( x) , Vx f    Vx g ( x)  0
u ( x ) ( x)   Vx gg T VxT (8)
  ( x) , Vx f    Vx g ( x)  0

Secondly, if there exist input constraints, the analytical expression of controller (6) might be
very complicated or even inexistent. Thus in this subsection, only analytical form of
controller (6) with a typical super ball input constraint is researched, i.e., input constraints is
as
U  {(u1 , , um ) | u12    um2  r 2 } (9)

where (u1, … ,um) is the input vector, and r is the radius of the super ball.
In order to obtain the analytical expression of Eq. (6) with input constraint as Eq. (9), we
propose the following 4 steps (For a general control input constraint U, one can always find
a maximal inscribed super ball B of it, and then use B replacing U before continuing the
following processes):

Input

State

Fig. 2a. the sketch of PMN

Input

State

Fig. 2b. the sketch of GPMN

* the dashed line is the PMN controller in a) and the GPMN control in b); the solid line
denotes the guide function of ξ(x).
A new kind of nonlinear model predictive control
algorithm enhanced by control lyapunov functions 65

Step1: For each state x, the following equation denotes a super plane in Rm (u  Rm).

Vx f ( x)   ( x )  Vx g ( x)u  0 (10)

Let d be the distance from zero to the super plane (10), we have,

Vx f ( x)   ( x)
d (11)
Vx g ( x) g T ( x)VxT

Fig. 3a

Vx f ( x)   ( x) T
 g ( x)VxT
Vx g ( x) g T ( x)VxT

r
 g T ( x)VxT
T T
Vx g ( x ) g ( x )V
x

Fig. 3b.
* Sketch of the process to build the analytic GPMN controller

Step2: From Eq. (11), the ‘permitted’ stable control input set KV(x) in controller (6) can be
denoted as Fig. 3a, where the right figure (left figure) is the case that the super plane of (10)
intersects (does not intersect) with the super ball (9), and the region filled by the dotted line
is the ‘permitted’ stable control input set. For the case denoted by the left figure of Fig. 3a, it
is easy to obtain a minimal distance from any point p to KV(x), and the corresponding point,
i.e., the controller’s output, in KV(x) with minimal distance from p can also be obtained (the
66 Model Predictive Control

point of intersection of the super ball (9) and the beeline connecting the centre of it and p).
With respect to the case of the right figure, the maximally inscribed super ball B’ is used to
replace KV(x) (see Fig. 3b). Thus, the same processes as above can be used to obtain the
output of controller (6).
Step 3: A new ‘permitted’ stable control input sets KV ( x) is defined,

 Vx f ( x)   ( x)
U r
 Vx g ( x) g T ( x)VxT
KV ( x)   (12)
{u | u   ( x)  R 2 ( x)} Vx f ( x)   ( x)
 r
 Vx g ( x ) g T ( x)VxT
where
Vx f ( x)   ( x) r
 ( x )  (  ) g T ( x)VxT
2Vx g ( x) g T ( x)VxT 2 Vx g ( x) g T ( x)VxT

[Vx f ( x )   ( x)]
r
Vx g ( x) g T ( x)VxT
R( x) 
2

It is obvious that KV ( x)  KV ( x ) , thus the stability of the closed loop can be ensured from
Proposition I.
Step 4: The analytical expression of GPMN controller with super-ball input constraint can
thus be described as

 ( x)  ( x)   ( x)  R( x)

u ( x ) ( x)   R( x) (13)
  ( x)   ( x) [ ( x)   ( x )]   ( x) else


where ξ(x) is the guide function of controller (6). █

From the preceding procedure, it is evidently that Eq. (13) is the solution of Eq. (6) with
KV(x) being placed by KV ( x) .

3.2 GPMN-ENMPC
In order to achieve a stable NMPC with reduced computational burden, we propose to use
the GPMN to parameterize the control input sequence in NMPC. Assuming that  ( x,  ) is a
function of state x, where θ is the vector of unknown parameters, the following NMPC can
be formulated,
A new kind of nonlinear model predictive control
algorithm enhanced by control lyapunov functions 67

u *   ( x,  * )
 *  arg min J ( x,  )
  IRl
t T
J ( x,  )   l ( x,  ( x, ))d (14)
t

s.t. x  f ( x )  g ( x) ( x,  )
 ( x,  )  U , t  [t , t  T ]

NMPC algorithm of (14) is different from the normal NMPC in the following aspect: in
normal NMPC algorithm, one tries to optimize the continuous control profile of u (Mayne et
al., 2000), while controller (14) tries to achieve good performance by optimizing the
parameter vector θ. Thus, the computational cost of controller (14) dependents mainly on
dimension of θ instead of that of control input profile in normal NMPC algorithm. The most
important problem of the latter algorithm is that its computational cost increases rapidly
with the control horizon. Based on (14), our new designed NMPC controller is introduced in
the following proposition.

Proposition II:
Assuming V(x) is a known CLF of system (1), Ωc is the stability region of V(x), then
controller (14) with the following GPMN controller  ( x,  ) ,

 ( x,  )  u ( x,  )  arg min { u   ( x,  ) } (15)

uKV ( x )

(u(x,θ) is the GPMN control and ξ(x,θ) the guide function in Eq. (6)), is stable in Ωc.
Furthermore, if V(x) is a global CLF, controller of (14) combined with (15) is stable over Rn.
(14), combined with (15), is called GPMN-Enhanced NMPC (GPMN-ENMPC).
Proof of Proposition II:
At any time instant t, by assuming that θ* is the optimal parameters at t, control input at t
can be represented as u(x,θ*). From Proposition I, we can conclude that the control inputs
u(x,θ*) can guarantee a negative definite V ( x) . Due to the randomicity of t, GPMN-ENMPC
actually makes the V ( x) negative in any time instant, which means that the closed loop
stability of controller (14) and (15) is guaranteed. █

3.3 Selection of ξ(x,θ)

Theoretically, ξ(x,θ) in (15) can be selected in any forms since it does not influence the closed
loop stability, which is guaranteed by GPMN. However, it is natural that ξ(x,θ) will
influence other closed loop performances of the GPMN-ENMPC except the stability.
Since optimality is the main concern in designing NMPC algorithm, the Bellman’s
Optimization Principle (BOP, Lewis & Syrmos, 1995) is used to design ξ(x,θ) in this sub-
section.
In BOP, with the following quadratic cost function,
 t T
J ( x,  )   ( xT Px  u T Qu )d (16)
t
68 Model Predictive Control

and J*(x0,θ) denoting the optimal value function of J(x0,θ) in state x0, the following controller
of system (1) is optimal,
1 J *
u *   (Q 1 )T g T ( x) (17)
2 x

Unfortunately, in most applications, it is impossible to obtain J*(x*,θ).

Based on the Stone-Weierstrass theorem (Brinkhuis & Tikhomirov, 2005), any continuous
function defined in a bounded set can be uniformly approximated by a polynomial function,

v1 v

*
BkJ ( x1 , , xn )  J*( , , n ) pk ;v1 ,, vn ( x1 , , xn ) (18)
v1 ,, vn  0 k k
v1  vn  k

where
 k  v1 vn k  v1  vn
pk ;v1 ,,vn ( x1 , , xn )    x1  xn (1  x1    xn ) ,
v
 1 ,  , vn 
(19)
 k  k!
  ,
v
 1 ,  , vn  v1 !v2 ! vn !(k  v1    vn )!
and
*
lim BkJ1 ,, kn ( x1 , , xn )  J * ( x1 , , xn ) (20)
ki 
( i 1,, k )

Thus, take the coefficients of the Bernstein polynomial as the parameters θ, and select θ
optimally using the NMPC algorithm, a ‘quasi-optimal’ function closed to J*(x*,θ) can be
obtained. That means we can complete the design of GPMN-ENMPC algorithm by taking

 ( x,  )  
v1 ,, vn  0
v ,,v pk ;v ,,v ( x1 , , xn )
1 n 1 n
(21)
v1  vn  k

where v1 ,..., vn , v1 ,..., vn ≥ 0 and v1  ...  vn ≤ k are the parameters to be optimized, k is the
order of the Bernstein polynomial, and

  [k , k 1 2 ,..., kn
]nk 1 (22)

It should be noted that the order of the Bernstein polynomial determines the consequent
optimization cost, i.e., the higher the order is, the higher the computational cost is. About
the GPMN-ENMPC, we have the following remarks:
Remark-1: Selection of ξ(x,θ) as Eq. (21) provides a feasible way to complete the GPMN-
ENMPC of (20) and (21). By this way, the computation cost is controllable, namely, one can
select the order of k to meet the CPU capability of a specific real system. This makes the
GPMN-ENMPC feasible to be implemented.
A new kind of nonlinear model predictive control
algorithm enhanced by control lyapunov functions 69

Remark-2: The selection of k does not influence the closed loop stability, which has already
guaranteed by the GPMN scheme. But there still exist trade-offs between computation cost
and the optimal performance which is determined by ξ(x,θ).

Remark-3: Compared to nominal NMPC algorithm, the huge computational burden

problem of GPMN-ENMPC algorithm is improved due to the following two reasons: 1) the
dimension of optimizing variables is one of key elements which increase the computational
burden of NMPC, while that of GPMN-ENMPC algorithm is independent of the predictive
horizon; 2) online considerations of control input constraints are not necessary in GPMN-
ENMPC algorithm since it can be dealt with offline during designing GPMN controller.

3.4 The Feasibility of GPMN-ENMPC

Another important problem, normally called the feasibility problem of NMPC, is that
general NMPC algorithm may not guarantee that a control set always exists to meet all of
the input and state constraints, while the proposed GPMN-ENMPC can guarantee such a
control sequence always exists. This is because for any θ, from the proposition-I, one can
always obtain a stable GPMN controller, i.e., u(x,θ) of (6) meeting all input and state
constraints. Therefore, by Eq. (14) and (15), there will always exist a feasible control
u =  ( x, ) , and the task left is just to find an optimal parameter set of θ to minimize the cost
function of J(x,θ) in Eq. (14).

4. H∞ GPMN-ENMPC
In section 3, GPMN-ENMPC algorithm is introduced with respect to system (1). In this
section, it will be generalized to deal with the disturbed system as Eq. (2). Firstly, an H∞
controller with partially known disturbances is given, and then it is used to design
H∞GPMN controller, which followed by the designing process of H∞GPMN-ENMPC.

4.1 H∞ Control With Partially Known Disturbances

Suppose the following two assumptions are satisfied with respect to system (2),

Assumption I:
System (2) is static feedback linearizable, i.e., there exists a state feedback controller u = k(x)
such that (2) can be transformed into a linear system without considering ω.

Assumption II:
The disturbances of system (2) are partially obtainable, i.e., the variables ω can be used to
construct controller.

Assumption II is reasonable because the uncertainty information ω can often be measured or

estimated in reality (He & Han, 2007; Chen, 2004). Moreover, the tracking problem of
general nonlinear system, where ω is composed of known desired trajectory, can also be
modeled as Eq. (2). However, the higher order derivative of the disturbances with respective
to time is often difficult to be obtained due to the heavy additive noise. Thus, the
disturbances are often ‘partially obtainable’.
70 Model Predictive Control

Based on assumption I, system (2) can be changed into the following equations through
some coordination transformation,

z1  z2  F1 ( z )

(23)
zn  f1 ( z )  g1 ( z )u  Fn ( z )
y  z1
where z = [z1,z2,…,zn]T is the new state variable.
An H∞ robust controller for system (23) can be designed based on the following Theorem,

Theorem I:
Consider system (23), if there exists a control u = u1(z) and a radially unbounded function
V(x) to satisfy the following inequality,

n 1
2 T
V zi zi 1  Vzn [ f1 ( z )  g1 ( z )u1 ( z )] 

Vz  F1T ( z ) F2T ( z )  FnT ( z )  
2
i 1 (24)
 F ( z ) F ( z )  F ( z )  V  z  0
1
T
2
T
n
T
z
T 2
1

Then, controller

n
u  g11 ( z )[ f1 ( z )  g1 ( z )u1  f1 ( z )   F2 ( z ,  , ,  (i 1) )  ( n  i ) ]  [Vzn g1 ( z )]1 
i 1
T
 F1 ( z )  F1 ( z ,  )  F1 ( z )  F1 ( z ,  ) 
   (25)
2 F ( z )  F ( z ,  , 
 ) F ( z )  F ( z ,  , 
 )
{ 2 Vz  2 2  2 2  VT}
      z
   
 Fn ( z )  Fn ( z ,  , ,  )   Fn ( z )  Fn ( z ,  , ,  ( n 1) ) 
( n 1)

can make the system (23) finite gain L2 stable from Δ+ρ to y, and the gain is less than or
equal to γ. ρ is a new defined signal to further attenuate the disturbances.
Proof of Theorem I:
Define new variables,

z1  z1
z2  z2  F1 ( z ) 
 (26)
n 1
zn  zn   Fi ( z )  ( n  i )
i 1

Then, system (23) can be written as

A new kind of nonlinear model predictive control
algorithm enhanced by control lyapunov functions 71

z1  z2  F1 ( z ,  )(   )
z2  z3  F2 ( z ,  ,  )(   )
 (27)
n
zn  f1 ( z )  g1 ( z )u   Fi ( z ,  , ,  (i 1) )  ( n  i )  Fn ( z ,  , ,  ( n 1) )(   )
i 1

y  z1

where

Fj ( z ,  , ,  ( j 1) )  F j ( z ) i 1
zi  z j   Fi ( z )  ( ni )
j 1

Let
V ( z )  V ( z ) |z  z (28)

where z =  z1 z2  zn  . Computing the HJI equation (Khalil, 2002) of system (27)

T

with respect to V ( z ) , we have,

n 1 n

V
i 1
zi zi 1  Vzn [ f1 ( z )  g1 ( z )u   F2 ( z ,  , ,  (i 1) )  ( n i ) ] 
i 1

2 T
V  F1T ( z ,  ) F2T ( z ,  ,  )  Fn ( z ,  , ,  ( n 1) )   (29)
 z2

 F1T ( z ,  ) F2T ( z ,  ,  )  Fn ( z ,  , ,  ( n 1) )  VzT  z12

Thus, combine controller (25) and Eq. (29), we have,

n 1 n
(29)   Vzi zi 1  Vzn [ f1 ( z )  g1 ( z )u   F2 ( z ,  , ,  (i 1) )  ( n  i ) ] 
i 1 i 1

2 T
Vz  F1T ( z ,  ) F2T ( z ,  ,  )  FnT ( z ,  , ,  ( n 1) )  
 2

 F1 ( z ,  ) F2 ( z ,  ,  )  Fn ( z ,  , ,  )  VzT  z12
T T T ( n 1)
(30)
n 1
2 T
 { Vzi zi 1  Vzn [ f1 ( z )  g1 ( z )u ]  2 Vz  F1T ( z ) F2T ( z )  FnT ( z )  
i 1 
 F1T ( z ) F2T ( z )  FnT ( z )  VzT  z12 } |z  z
0
Based on theorem 5.5 in reference (Khalil, 2002), controller (25) can make system (23) finite
gain L2 stable from Δ+ρ to y, and the L2 gain is less than or equal to γ. █
72 Model Predictive Control

Furthermore, ρ can be used to further attenuate the disturbances which are partially
obtainable from assumption II by the following equation,

B( s)
 ( s)  ( s) (31)
A( s )

where s is the Laplace operator. Thus, the new external disturbances Δ+ρ can be denoted as,

A( s )  B( s)
( s)   ( s)  ( s) (32)
A( s )

From Eq. (32), proper A(s) and B(s) is effective for attenuating the influence of external
disturbances on the closed loop system. Thus, we have designed an H∞ controller (25) and
(31) with partially known uncertainty information.

4.2 H∞ GPMN Controller Based on Control Lyapunov Functions

In this sub-section, by using the concept of H∞CLF, H∞ GPMN controller is designed as
following proposition,

Proposition III:
If V(x) is a local H∞CLF of system (23), and ξ(x): RnRm is a continuous guide function such
that ξ(0) = 0, then, the following controller, called H∞GPMN, can make system (23) finite
gain L2 stable from  to output y,

u H  ( x)  arg min{ u   ( x) } (33)

uKVH  ( x )

where
1 1
KVH  ( x)  {u  U ( x) : Vx [ f ( x)  g ( x)u ]  V l ( x)l T ( x)VxT  hT ( x)h( x )   ( x )} (34)
2 x
2 2
█
Proof of Proposition III can be easily done based on the definition of finite gain L2 stability
and H∞CLF. The analytical form of controller (33) can also be obtained as steps in section 3.
Here only the analytical form of controller without input constraints is given,
 1 T T 1 T
 [Vx ( f  2 2 ll Vx  g )  2 h h   ]g Vx
T T


u H  ( x)     0 (35)
 Vx gg T VxT
  0
where
  Vx f    Vx g ; f  f ( x); g  g ( x);    ( x);
   ( x); Vx  Vx ( x); h  h( x); l  l ( x)
A new kind of nonlinear model predictive control
algorithm enhanced by control lyapunov functions 73

It is not difficult to show that H∞GPMN satisfies inequality (24) of Theorem I, thus, it can be
used as u1(z) in controller (25) to bring the advantages of H∞GPMN controller to the robust
controller in section 4.1.

4.3 H∞GPMN-ENMPC
As far as the external disturbances are concerned, nominal model based NMPC, where the
prediction is made through a nominal certain system model, is an often used strategy in
reality. And the formulation of it is very similar to non-robust NMPC, so dose the GPMN-
ENMPC.

RGPMN

Feedback lineari-
zation z=T(x)
x

Robust control-
er with partially
H∞GPMN controller Uncertain No-
obtainable distu-
(35) nlinear System
rbances (25)
u1(z)
uH(x , * ) ( z )

uH(x , ) ( z )
θ*

x
z
Feedback lineari-
GPMN-ENMPC
zation z=T(x)

Fig. 4. Structure of new designed RNRHC controller

However, for disturbed nonlinear system like Eq. (23), GPMN-ENMPC algorithm can
hardly be used in real applications due to weak robustness. Thus, in this subsection, we will
combine it to the robust controller from sub-section 4.1 and sub-section 4.2 to overcome the
drawbacks originated from both GPMN-ENMPC algorithm and the robust controller (25)
and (35). The structure of the new parameterized H∞GPMN-ENMPC algorithm based on
controller (25) and (35) is as Fig. 4.
Eq. (36) is the new designed H∞GPMN-ENMPC algorithm. Compared to Eq. (14), it is easy
to find out that the control input in the H∞GPMN-ENMPC algorithm has a pre-defined
structure given in section 4.1 and 4.2.
74 Model Predictive Control

u *  u H  ( x,  * )

 *  arg min J ( x, u )
uU
 t T
J ( x, u )   l ( x( ), u ( ))d (36)
t

s.t. x  f ( x)  g ( x)u
u (t )  u H  ( x,  )

5. Practical Considering
Both GPMN-ENMPC algorithm and H∞GPMN-ENMPC algorithm can be divided into two
processes, including the implementation process and the optimization process as Fig.5.

Optimized parameter θ*

Implementation process Optimization process

Computing control input Computing the optimal

based on (H∞)GPMN scheme parameter θ* by solving an
optimal control problem

Current state xt

Fig. 5. The process of (H∞)GPMN-ENMPC

The implementation process and the optimization process in Fig. 5 are independent. In
implementation process, the (H∞)GPMN scheme is used to ensure the closed loop (L2)
stability, and in the optimization process, the optimization algorithm is responsible to
improving the optimality of the controller. And the interaction of the two processes is
realized through the optimized parameter θ* (from optimization process to implementation
process) and the measured states (form implementation process to optimization process).

5.1 Time Interval Between Two Neighboring Optimizing Process

Sample time in controller implemented using computer is often very short, especially in
mechatronic system. This is very challenging to implement complicated algorithm, such as
GPMN-ENMPC in this chapter. Fortunately, the optimization process of the new designed
controller will end up with a group of parameters which are used to form a stable
(H∞)GPMN controller, and the optimization process itself does not influence the closed loop
stability at all. Thus, theoretically, any group of optimized parameters can be used for
several sample intervals without destroying the closed loop stability.
A new kind of nonlinear model predictive control
algorithm enhanced by control lyapunov functions 75

Fig.6 denotes the scheduling of (H∞)GPMN-ENMPC algorithm. In Fig.6, t is the current time
instant; T is the prediction horizon; TS is the sample time of the (H∞)GPMN controller; and TI
is the duration of every optimal parameter θ*(t), i.e., the same parameter θ* is used to
implement the (H∞)GPMN controller from time t to time t+TI.

     
TS
TI
t

Fig. 6. Scheduling of ERNRHC

5.2 Numerical Integrator

How to predict the future’s behavior is very important during the implementation of any
kind of MPC algorithms. In most applications, the NMPC algorithm is realized by
computers. Thus, for the continuous systems, it will be difficult and time consuming if some
accurate but complicated numerical integration methods are used, such as Newton-Cotes
integration and Gaussian quadratures, etc. In this chapter, we will discretize the continuous
system (1) as follows (take system (1) as an example),

x(kTO  TO )  f ( x(kTO ))TO  g ( x(kTO ))u (kTO )TO (37)

where To is the discrete sample time. Thus, the numerical integrator can be approached by
the operation of cumulative addition.

5.3 Index Function

Replace x(kTo) with x(k), the index function can be designed as follows,

 k0  N
J ( x(k0 ),  c )   lT *  l*   J ( x(i), )
i  k0
c (38)

where k0 denotes the current time instant; N is the predictive horizon with N=Int(T/To) (here
Int(*) is the operator to obtain a integer nearest to *); θc is the parameter vector to be
optimized at current time instant; and θl* is the last optimization result; Q, Z, R are constant
matrix with Q>0, Z>0, and R≥0.
The new designed item θlT*Zθl* is used to reduce the difference between two neighboring
optimized parameter vector, and improve the smoothness of the optimized control inputs u.
76 Model Predictive Control

6. Numerical Examples
6.1 Example1 (GPMN-ENMPC without control input constrains)
Consider the following pendulum equation (Costa & do Va, 2003),
 x1  x2

 19.6sin x1  0.2 x22 sin 2 x1 0.2 cos x1 (39)
 x2  2
4 / 3  0.2 cos x1

4 / 3  0.2 cos 2 x1
u

A local CLF of system (39) can be given as,

151.57 42.36   x1 
V ( x)  xT Px   x1 x2     (40)
 42.36 12.96   x2 
Select
 ( x)  0.1( x12  x22 ) (41)

The normal PMN control can be designed according to (5) as,

  ( x)(4 / 3  0.2 cos 2 x1 )
   ( x)  0
u   0.4 cos x1 (42.36 x1  12.96 x2 )
 0  ( x)  0

19.6sin x1  0.2 x22 sin 2 x1
 ( x )  2[(151.57 x1  42.36 x2 ) x2  (42.36 x1  12.96 x2 ) ]  (42)
4 / 3  0.2 cos 2 x1
(10.54 x2  1.27 x1 ) 2  x22

Given initial state x0 = [x1,x2]T = [-1,2]T, and desired state xd = [0,0]T, time response of the
closed loop for PMN controller is shown in Fig. 7 in solid line. It can be seen that the closed
loop with PMN controller (42) has a very low convergence rate for state x1. This is mainly
because the only regulable parameter to change the closed loop performance is σ(x), which is
difficult to be properly selected due to its great influence on the stability region.
To design GPMN-ENMPC, two different guide functions are selected based on Eq. (21),

 ( x,  )   0,0 (1  x1  x2 )  1,0 x1   0,1 x2 (43)

 ( x,  )   0,0 (1  x1  x2 )  2( 0,1 x2  1,0 x1 )(1  x1  x2 )  21,1 x1 x2   x   x
2 2
0,2 2
2
2,0 1 (44)

CLF V(x) and σ(x) are given in Eq. (40) and Eq. (41), and others conditions in GPMN-
ENMPC are designed as follows,

T  20 0 
J   ( xT  2
 x  0.01u )dt (45)
0
 0 1
A new kind of nonlinear model predictive control
algorithm enhanced by control lyapunov functions 77

 x2 
T  20 0  2  
l ( x, u )  x   x  0.01u ; f ( x)  19.6sin x1  0.2 x2 sin 2 x1  ;
2

 0 1  
4 / 3  0.2 cos 2 x1
(46)
 0 
g ( x)   0.2 cos x1  ; z  0.1I
 
 4 / 3  0.2 cos 2 x1 

Integral time interval To in Eq. (37) is 0.1s. Genetic algorithm (GA) in MATLAB toolbox is
used to solve the online optimization problem. Time response of GPMN-ENMPC algorithm
with different predictive horizon T and approaching order are presented in Fig. 7, where the
dotted line denotes the case of T = 0.6s with guide function (43), and the dashed line is the
case of T = 1.5s with guide function (44). From Fig. 7, it can be seen that the convergence
performance of the proposed NMPC algorithm is better than PMN controller, and both the
prediction horizon and the guide function will result in the change of the closed loop
performance.
The improvement of the optimality is the main advantage of MPC compared with others
controller. In view of this, we propose to estimate the optimality by the following index
function,
  20 0 
J  lim  ( xT  2
 x  0.01u )dt (47)
 0
 0 1 

-0.5
x1

-1
0 1 2 3 4 5

2
PMN
ENMPC (1,0.6)
1 ENMPC (2,1.5)
x2

0
0 1 2 3 4 5
time (s)
Fig. 7. Time response of different controller, where the (a,b) indicates that the order of
 ( x,  ) is a, and the predictive horizon b

The comparison results are summarized in Table 1, from which the following conclusions
can be obtained, 1) GPMN-ENMPC has better optimizing performance than PMN controller
in terms of optimization. 2) In most cases, GPMN-ENMPC with higher order ξ(x,θ) will
usually result in a smaller cost than that with lower order ξ(x,θ). This is mainly because
78 Model Predictive Control

higher order ξ(x,θ) indicates larger inherent optimizing parameter space. 3) A longer
prediction horizon will usually be followed by a better optimal performance.

ENMPC PMN
x0 = (- x0 =
J x0 = (-1,2) x0 = (0.5,1)
1,2) (0.5,1)
K= k=
k=1 k=2 ----
1 2
T=0.6 29.39 28.87 6.54 6.26 +∞ +∞
T=0.8 23.97 23.83 5.02 4.96 +∞ +∞
T=1.0 24.08 24.07 4.96 4.90 +∞ +∞
T=1.5 26.31 24.79 5.11 5.28 +∞ +∞
Table 1. the cost value of different controller
* k is the order of Bernstein polynomial used to approach the optimal value function; T is the
predictive horizon; x0 is the initial state

Another advantage of the GPMN-ENMPC algorithm is the flexibility of the trade offs
between the optimality and the computational time. The computational time is influenced
by the dimension of optimizing parameters and the parameters of the optimizing algorithm,
such as the maximum number of iterations and the size of the population (the smaller these
values are selected, the less the computational cost is). However, it will be natural that the
optimality maybe deteriorated to some extent with the decreasing of the computational
burden. In preceding paragraphs, we have researched the optimality of GPMN-ENMPC
algorithm with different optimizing parameters, and now the optimality comparisons
among the closed loop systems with different GA parameters will be done. And the results
are listed in Table 2, from which the certain of the optimality loss with the changing of the
optimizing algorithm’s parameters can be observed. This can be used as the criterion to
determine the trade-off between the closed loop performance and the computational
efficiency of the algorithm.

G=100 G=50 G=50 G=50 G=50

OP
PS=50 PS=50 PS=30 PS=20 PS=10
cost 26.2 28.1 30.8 43.5 45.7
Table 2. The relation between the computational cost and the optimality
*x0 = (-1,2), T=1.5, k = 1, OP means Optimization Prameters, G means Generations, PS means
Population Size

Finally, in order to verify that the new designed algorithm is improved in the computational
burden, simulations comparing the performance of the new designed algorithm and
algorithm in (Primbs, 1999) are conducted with the same optimizing algorithm. Time
interval of two neighboured optimization (TI in Table 3) in Primbs’ algorithm is important
since control input is assumed to be constant at every time slice. Generally, large time
interval will result in poor stability.
While our new GPMN-ENMPC results in a group of controller parameter, and the closed loop
stability is independent of TI. Thus different TI is considered in these simulations of Primbs’
A new kind of nonlinear model predictive control
algorithm enhanced by control lyapunov functions 79

algorithm and Table 3 lists the results. From Table 3, the following items can be concluded: 1)
with same GA parameters, Primbs’ algorithm is more time-consuming and poorer in optimality
than GPMN-ENMPC. This is easy to be obtained through comparing results of Ex-2 and Ex-5; 2)
in order to obtain similar optimality, GPMN-ENMPC takes much less time than Primbs’
algorithm. This can be obtained by comparing results of Ex-1/Ex-4 and Ex-6, as well as Ex-3 and
Ex-5. The reasons for these phenomena have been introduced in Remark 3.

Algorithm in (Primbs, 1999) GPMN-ENMPC

Ex-1 Ex-2 Ex-3 Ex-4 Ex-5 Ex-6
TI 0.1 0.05 0.1
G=100 G=50 G=100 G=50 G=50 G=50
OP
PS=50 PS=50 PS=50 PS=50 PS=50 PS=30
Average Time
2.2075 1.8027 2.9910 2.2463 1.3961 0.8557
Consumption
Cost 31.2896 35.7534 27.7303 31.8055 28.1 31.1043
Table 3. Performance comparison of GPMN-ENMPC and Primbs’ algorithm
*x0 = (-1,2), TI means time interval of two neighbored optimization; OP means Optimization
Prameters; G means Generations, PS means Population Size. Other parameters of GPMN-
ENMPC are T=1.5, k = 1

6.2 Example 2 (GPMN-ENMPC with control input constraint)

In order to show the performance of the GPMN-ENMPC in handling input constraints, we
give another simulation using the dynamics of a mobile robot with orthogonal wheel
assemblies (Song, 2007). The dynamics can be denoted as Eq. (48),

x  f ( x )  g ( x)u (48)
where
 x2 
 2.3684x x  0.5921x 
 4 6 2

 x4 
f ( x)   
 2.3684x2 x6  0.5921x4 
 x6 
 
 0.2602x6 
 0 0 0 
 
0.8772( 3 sin x5  cos x5 ) 0.8772*2 cos x5 0.8772( 3 sin x5  cos x5 ) 
 0 0 0 
g ( x)   
0.8772( 3 cos x5  sin x5 ) 0.8772*2sin x5 0.8772( 3 cos x5  sin x5 ) 
 
 0 0 0 
 -1.4113 -1.4113 -1.4113 
80 Model Predictive Control

x1  xw ; x2  xw ; x3  yw ; x4  y w ; x5   w ; x6  w ; xw, yw, φw are respective the x-y positions

and yaw angle; u1, u2, u3 are motor torques.
Suppose that control input is limited in the following closed set,

U = {( u1, u2, u3)|( u12+ u22+ u32)1/2≤20} (49)

System (48) is feedback linearizable, and by which we can obtain a CLF of system (48) as
follows,
V ( x)  xT Px (50)
where
1.125 0.125 0 0 0 0 
0.125 0.156 0 0 0 0 

 0 0 1.125 0.125 0 0 
P 
 0 0 0.125 0.156 0 0 
 0 0 0 0 1.125 0.125
 
 0 0 0 0 0.125 1.156 

The cost function J(x) and σ(x) are designed as,

t0  T
J ( x)   (3x12  3x32  3x52  x22  x42  x62  5u12  5u22  5u32 )dt   T (k  1) Z (k  1);
t0 (51)
 ( x)  0.1( x12  x22  x32  x42  x52  x62 ); Z =0.1I

System (48) has 6 states and 3 inputs, which will introduce large computational burden if
using the GPMN-ENMPC method. Fortunately, one of the advantages of GPMN-ENMPC is
that the optimization does not destroy the closed loop stability. Thus, in order to reduce the
computation burden, we reduce the frequency of the optimization in this simulation, i.e.,
one optimization process is conducted every 0.1s while the controller of (13) is calculated
every 0.002s, i.e., TI = 0.1s, Ts = 0.002s.
15
5
2
10
5 0
1

0
x

0
-2
-5 -5
2 4 6 8 10 2 4 6 8 10
-4
5 5
0 -6
-5 0
u3
3

4
x

-10 -8
-5
-15
2 4 6 8 10 2 4 6 8 10 -10
1
1.5 -12
1
0
5

0.5 -14
x

0
-0.5
2 4 6 8 10
-1
2 4 6 8 10
-16
2 4 6 8 10 12 14 16 18 20
time(s) time(s)
time(s)

a) states response b) control input u1

A new kind of nonlinear model predictive control
algorithm enhanced by control lyapunov functions 81

4 14

12
3
10

8
2

6
u2

1
u
1 4

2
0
0

-1 -2
2 4 6 8 10 12 14 16 18 20 2 4 6 8 10 12 14 16 18 20
time(s) time(s)

c) control input u2 d) control input u3

Fig. 8. GPMN-ENMPC controller simulation results on the mobile robot with input
constraints

Feedback
Initial States
linearization GPMN-NMPC
(x1; x2; x3; x4; x5; x6)
controller
(10; 5; 10; 5; 1; 0) 2661.7 1377.0
(10; 5; 10; 5; -1; 0) 3619.5 1345.5
(-10; -5; 10; 5; 1; 0) 2784.9 1388.5
(-10; -5; 10; 5; -1; 0) 8429.2 1412.0
(-10; -5; -10; -5; 1; 0) 394970.0 1349.9
(-10; -5; -10; -5; -1; 0) 4181.6 1370.9
(10; 5; -10; -5; 1; 0) 3322 1406
(10; 5; -10; -5; -1; 0) 1574500000 1452.1
(-5; -2; -10; -5; 1; 0) 1411.2 856.1
(-10; -5; -5; -2; 1; 0) 1547.5 850.9
Table 4. The comparison of the optimality

Simulation results are shown in Fig.8 with the initial state (10; 5; -10; -5; 1; 0), From Fig.8, it is
clear that GPMN-ENMPC controller has the ability to handling input constraints.
In order to evaluate the optimal performance of the GPMN-ENMPC, we proposed the
following cost function according to Eq. (51),


cos t  lim
 0 (3 x12 3x32  3x52  x22  x42  x62  5u12  5u22  5u32 )dt (52)

Table 4 lists the costs by feedback linearization controller and GPMN-ENMPC for several
different initial states, from which it can be seen that the cost of GPMN-ENMPC is less than
the half of the cost of feedback linearization controller when the initial is (10; 5; -10; -5; 1; 0).
And in most cases listed in Table 4, the cost of GPMN-ENMPC is about one second of that of
feedback linearization controller. Actually, in some special cases, such as the initial of (10; 5;
-10; -5; -1; 0), the cost ratio of feedback linearization controller to GPMN-ENMPC is more
than 1000000.
82 Model Predictive Control

6.3 Example 3 (H∞GPMN-ENMPC)

In this section, a simulation will be given to verify the feasibility of the proposed H∞GPMN-
ENMPC algorithm with respect to the following planar dynamic model of helicopter,

 x = 9.8cos  sin   1

 
 y  9.8sin    2
 
  0.05 sin  cos      tan  (0.5+0.05cos 2 ) 
2
(53)
 0.07 cos  tan   L  M sin  tan    3


  0.05   sin  cos   0.07 sin   M cos   
4

where Δ1, Δ2, Δ3, Δ4 are all the external disturbances, and are selected as following values,

1  3;  2  3

3  10sin(0.5t )
  10sin(0.5t )
 4

Firstly, design an H∞CLF of system (53) by using the feedback linearization method,

V  X T PX (54)

where,

X  [ x, x , 
x, 
x , y, y ,  y ]T
y, 
14.48 11.45 3.99 0.74 0 0 0 0 
11.45 9.77 3.44 0.66 0 0 0 0 
 
 3.99 3.44 1.28 0.24 0 0 0 0 
 
 0.74 0.66 0.24 0.05 0 0 0 0 
P
 0 0 0 0 14.48 11.45 3.99 0.74 
 
 0 0 0 0 11.45 9.77 3.44 0.66 
 0 0 0 0 3.99 3.44 1.28 0.24 
 
 0 0 0 0 0.74 0.66 0.24 0.05 

Thus, the robust predictive controller can be designed as Eq. (25), (35) and (36) with the
following parameters,
A new kind of nonlinear model predictive control
algorithm enhanced by control lyapunov functions 83

 ( x)  X T X
N
J   lT * I l*   [ x T (iTo ) Px(iTo )  u T (iTo )Qu (iTo )]To
i 1

 1   2 x   3 x   4 y   5 y   6   7  8   9 

 ( x,  ) = 
 
 10  11 x  12 x  13 y  14 y  15  16  17  18 
50000 0 0 0 0 0 0 0
 0 1 0 0 0 0 0 0 

 0 0 50000 0 0 0 0 0 
 
 0 0 0 1 0 0 0 0
P
 0 0 0 0 0 0 0 0
 
 0 0 0 0 0 1 0 0
 0 0 0 0 0 0 0 0
 
 0 0 0 0 0 0 0 1 
1 0 
Q  ; To  0.1s; Ts  0.02s; TI  1s; N  20; Z = I
0 1 

Time response of the H∞GPMN-ENMPC is as solid line of Fig.9 and Fig.10. Furthermore, the
comparisons between the performance of the closed loop controlled by the proposed
H∞GPMN-ENMPC and some other controller design method are done. The dashed line in
Fig.9 and Fig.10 is the time response of the feedback linearization controller. From Fig.9 and
Fig.10, the disturbance attenuation performance of the H∞GPMN-ENMPC is apparently
better than that of feedback linearization controller, because the penalty gain of position
signals, being much larger than other terms, can be used to further improve the ability.

0.05 0.05
xdot

0 0
x

-0.05 -0.05
0 4 8 12 16 20 24 0 4 8 12 16 20 24
0.05 0.05
ydot

0 0
y

-0.05 -0.05
0 4 8 12 16 20 24 0 4 8 12 16 20 24
-0.3 0.1
dot

-0.31 0


-0.32 -0.1
0 4 8 12 16 20 24 0 4 8 12 16 20 24
0.34 0.1
dot

0.33 0


0.32 -0.1
0 4 8 12 16 20 24 0 4 8 12 16 20 24
time(s) time(s)

Fig. 9. Time response of states

84 Model Predictive Control

20

10

L 0

-10

-20
0 4 8 12 16 20 24

20

10

0
M

-10

-20
0 4 8 12 16 20 24
time(s)

Fig. 10. Control inputs

Simultaneously, the following index is used to compare the optimality of the two different
controllers,

0 
J  lim  [ x T (t ) Px(t )  u T (t )Qu (t )]dt (55)
 0

The optimality performance of H∞GPMN-ENMPC, computed from Eq. (55), is about 3280,
and the feedback linearization controller is about 5741, i.e., the H∞GPMN-ENMPC has better
optimality than the feedback linearization controller.

7. Conclusion
In this paper, nonlinear model predictive control (NMPC) is researched and a new NMPC
algorithm is proposed. The new designed NMPC algorithm, called GPMN-enhancement
NMPC (GPMN-ENMPC), has the following three advantages: 1) closed loop stability can be
always guaranteed; 2) performance other than optimality and stability can be considered in
the new algorithm through selecting proper guide function; 3) computational cost of the
new NMPC algorithm is regulable according to the performance requirement and available
CPU capabilities. Also, the new GPMN-ENMPC is generalized to a robust version with
respect to input-output feedback linearizable nonlinear system with partially known
uncertainties. Finally, extensive simulations have been conducted, and the results have
shown the feasibility and validity of the new designed method.

8. References
Brinkhuis, J. & Tikhomirov, V., Optimization : insights and applications, Princeton University
Press, ISBN : 978-0-691-10287-0, Oxfordshire, United Kingdom
Chen, C. & Shaw, L., On receding horizon feedback control, Automatica, Vol. 18, No. 3, May,
1982, 349-352, ISSN : 0005-1098
A new kind of nonlinear model predictive control
algorithm enhanced by control lyapunov functions 85

Chen, H. & Allgower, F., A quasi-infinite horizon nonlinear model predictive control
scheme with guaranteed stability. Automatica, Vol. 34, No. 10, Oct, 1998, 1205-1217,
ISSN : 0005-1098
Chen, W., Disturbance observer based control for nonlinear systems, IEEE/ASME
Transactions on mechatronics, Vol. 9, No. 4, 2004, 706-710, ISSN : 1083-4435
Costa, E. & do Val, J., Stability of receding horizon control of nonlinear systems, Proceedings
of The 42nd IEEE Conference on Decision and Control, pp. 2077-2801, ISSN : 0191-2216,
HI, USA, Dec, 2003, IEEE, Maui
Freeman, R. & Kokotovic, P, Inverse optimality in robust stabilization, SIAM Journal on
Control and Optimization, Vol. 34, No. 4, Aug, 1996a, 1365-1391, ISSN : 0363-0129
Freeman, R. & Kokotovic, P.(1996b), Robust nonlinear control design: state-space and Lyapunov
techniques, Birkhauser, ISBN: 978-0-8176-4758-2, Boston
Henson, M., Nonlinear model predictive control: current status and future directions.
Computers & Chemical Engineering, Vol. 23, No. 2, Dec, 1998, 187-202, ISSN : 0098-
1354
He, Y., and Han, J., Acceleration feedback enhanced H∞ disturbance attenuation control,
Proceedings of The 33rd Annual Conference of the IEEE Industrial Electronics Society
(IECON), pp. 839-844, ISSN : 1553-572X, Taiwan, Nov, 2007, IEEE, Taipei
Khalil, H. (2002), Nonlinear systems, 3rd edition, Printice Hall, ISBN: 0-13-067389-7, NJ, USA
Lewis, F. & Syrmos, V. (1995), Optimal control, John Wiley & Sons, ISBN : 0-471-03378-2,
Bangalore, India
Magni, L., De Nicolao, G., Magnani, L. & Scattolini, R., A stabilizing model-based predictive
control algorithm for nonlinear systems. Automatica, Vol. 37, No. 9, Sep, 2001, 1351-
1362, ISSN : 0098-1354
Mayne, D., Rawlings, J., Rao, C. & Scokaert, P., Constrained model predictive control:
stability and optimality. Automatica, Vol. 36, No. 6, Jun, 2000, 789-814, ISSN : 0005-
1098
Pothin, R., Disturbance decoupling for a class of nonlinear MIMO systems by static
measurement feedback, Systems & Control Letters, Vol. 43, No. 2, Jun, 2001, 111-116,
ISSN : 0167-6911
Primbs, J. A., Nevistic, V. & Doyle, J. C., Nonlinear optimal control: a control
Lyapunov function and receding horizon perspective, Asian Journal of
Control, Vol. 1, No. 1, Jan, 1999, 14-24, ISSN: 1561-8625
Primbs, J. & Nevistic, V., Feasibility and stability of constrained finite receding horizon
control, Automatica, Vol. 36, No. 7, Jul, 2000, 965-971, ISSN : 0005-1098
Qin, S., & Badgwell, T., A survey of industrial model predictive control technology. Control
Engineering Practice, Vol. 11, No. 7, Jul, 2003, 733-764, ISSN :0967-0661
Rawlings, J., Tutorial overview of model predictive control, IEEE Control System Magazine,
Vol. 20, No. 3, Jun, 2000, 38-52, ISSN : 0272-1708
Scokaert, P., Mayne, D. & Rawlings, J., Suboptimal model predictive control (feasibility
implies stability), IEEE Transactions on Automatica Control, Vol. 44, No. 3, Mar, 1999,
648-654, ISSN : 0018-9286
Song, Q., Jiang, Z. & Han, J., Noise covariance identification-based adaptive UKF with
application to mobile robot system, Proceedings of IEEE International Conference on
Robotics and Automation (ICRA 2007), pp. 4164-4169, ISSN: 1050-4729, Italy, May,
2007, Roma
86 Model Predictive Control

Sontag, E., A ‘universal’ construction of Artstein’s theorem on nonlinear stabilization,

Systems & Control Letters, Vol. 13, No. 2, Aug, 1989, 117-123, ISSN : 0167-6911
Zou, T., Li, S. & Ding, B., A dual-mode nonlinear model predictive control with the enlarged
terminal constraint sets, Acta Automatica Sinica, Vol. 32, No. 1, Jan, 2006, 21-27,
ISSN : 0254-4156
Wesselowske, K. & Fierro, R., A dual-mode model predictive controller for robot
formations, Proceedings of the 42nd IEEE Conference on Decision and Control, pp.
3615-3620, ISSN : 0191-2216, HI, USA, Dec, 2003, IEEE, Maui
Robust Model Predictive Control Algorithms for
Nonlinear Systems: an Input-to-State Stability Approach 87

0
4

Robust Model Predictive Control

Algorithms for Nonlinear Systems:
an Input-to-State Stability Approach
D. M. Raimondo
Automatic Control Laboratory, Electrical Engineering, ETH Zurich, Physikstrasse 3, 8092
Zurich
Switzerland

D. Limon, T. Alamo
Departamento de Ingeniería de Sistemas y Automática, Universidad de Sevilla, Escuela
Superior de Ingenieros, Camino de los Descubrimientos s/n 41092 Sevilla
Spain

L. Magni
Dipartimento di Informatica e Sistemistica, Università di Pavia, via Ferrata 1, 27100 Pavia
Italy

This paper presents and compares two robust MPC controllers for constrained nonlinear systems based
on the minimization of a nominal performance index. Under suitable modifications of the constraints
of the Finite Horizon Optimization Control Problems (FHOCP), the derived controllers ensure that the
closed loop system is Input-to-State Stable (ISS) with a robust invariant region, with relation to addi-
tive uncertainty/disturbance. Assuming smoothness of the model function and of the ingredients of the
FHOCP, the effect of each admissible disturbance in the predictions is considered and taken into account
by the inclusion in the problem formulation of tighter state and terminal constraints. A simulation exam-
ple shows the potentiality of both the algorithms and highlights their complementary aspects.
Keywords: Robust MPC, Input to State Stability, Constraints, Robust design.

1. Introduction
Model predictive control (MPC) is an optimal control technique which deals with constraints
on the states and the inputs. This strategy is based on the solution of a finite horizon optimiza-
tion problem (FHOCP), which can be posed as a mathematical programming problem. The
control law is obtained by means of the receding horizon strategy that requires the solution of
the optimization problem at each sample time Camacho & Bordons (2004); Magni et al. (2009);
Rawlings & Mayne (2009).
It is well known that considering a terminal cost and a terminal constraint in the optimization
problem, the MPC stabilizes asymptotically a constrained system in absence of disturbances
or uncertainties. If there exist uncertainties in the process model, then the stabilizing proper-
ties may be lost Magni & Scattolini (2007); Mayne et al. (2000) and these must be taken into
account in the controller design. Recent results have revealed that nominal MPC may have
88 Model Predictive Control

zero robustness, i.e. stability or feasibility may be lost if there exist model mismatches Grimm
et al. (2004). Therefore it is quite important to analyze when this situation occurs and to find
design procedures to guarantee certain degree of robustness. In Limon et al. (2002b); Scokaert
et al. (1997) it has been proved that under some regularity condition on the optimal cost, the
MPC is able to stabilize the uncertain system; however, this regularity condition may be not
ensured due to constraints, for instance.
The synthesis of N MPC algorithms with robustness properties for uncertain systems has been
developed by minimizing a nominal performance index while imposing the fulfillment of con-
straints for each admissible disturbance, see e.g. Limon et al. (2002a) or by solving a min-max
optimization problem, see e.g. Chen et al. (1997); Fontes & Magni (2003); Magni et al. (2003);
Magni, Nijmeijer & van der Schaft (2001); Magni & Scattolini (2005). The first solution calls for
the inclusion in the problem formulation of tighter state, control and terminal constraints. The
main advantage is that the on-line computational burden is substantially equal to the compu-
tational burden of the nominal N MPC. In fact, nominal prediction based robust predictive
controllers can be thought as a nominal MPC designed in such a way that a certain degree
of robustness is achieved. The main limitation is that it can lead to very conservative solu-
tions. With a significant increase of the computational burden, less conservative results can be
achieved by solving a min-max optimization problem.
Input-to-State Stability (ISS) is one of the most important tools to study the dependence of
state trajectories of nonlinear continuous and discrete time systems on the magnitude of in-
puts, which can represent control variables or disturbances. The concept of ISS was first
introduced in Sontag (1989) and then further exploited by many authors in view of its equiv-
alent characterization in terms of robust stability, dissipativity and input-output stability, see
e.g. Jiang & Wang (2001), Huang et al. (2005), Angeli et al. (2000), Jiang et al. (1994), Nešić &
Laila (2002). Now, several variants of ISS equivalent to the original one have been developed
and applied in different contexts (see e.g. Sontag & Wang (1996), Gao & Lin (2000), Sontag &
Wang (1995), Huang et al. (2005)). The ISS property has been recently introduced also in the
study of nonlinear perturbed discrete-time systems controlled with Model Predictive Control
(MPC), see e.g. Limon et al. (2009), Raimondo et al. (2009), Limon et al. (2002a), Magni &
Scattolini (2007), Limon et al. (2006), Franco et al. (2008), Magni et al. (2006). In fact, the devel-
opment of MPC synthesis methods with enhanced robustness characteristics is motivated by
the widespread success of MPC and by the availability of many MPC algorithms for nonlinear
systems guaranteeing stability in nominal conditions and under state and control constraints.
In this paper two algorithms based on the solution of a minimization problem with respect to
a nominal performance index are proposed. The first one, following the algorithm presented
in Limon et al. (2002a), proves that if the terminal cost is a Lyapunov function which ensures
a nominal convergence rate (and hence some degree of robustness), then the derived nominal
MPC is an Input-to-State stabilizing controller. The size of allowable disturbances depends
on the one step decreasing rate of the terminal cost.
The second algorithm, first proposed in a preliminary version in Raimondo & Magni (2006),
shares with de Oliveira Kothare & Morari (2000) the idea to update the state of the nominal
system with the value of the real one only each M step to check the terminal constraint. The
use of a prediction horizon larger than a time varying control horizon is aimed to provide
more robust results by means of considering the decreasing rate in a number of steps.
Both controllers are based on the Lipschitz continuity of the prediction model and of some
of the ingredients of the MPC functional such as stage cost function and the terminal cost
Robust Model Predictive Control Algorithms for
Nonlinear Systems: an Input-to-State Stability Approach 89

function. Under the same assumptions they ensure that the closed loop system is Input-to-
State-Stable (ISS) with relation to the additive uncertainty.
A simulation example shows the potentiality of both the algorithms and highlights their com-
plementary aspects.
The paper is organized as follows: first some notations and definitions are presented. In
Section 3 the problem is stated. In Section 4 the Regional Input-to-State Stability is introduced.
In Section 5 the proposed MPC controllers are presented. In Section 6 the benefits of the
proposed controllers are illustrated with several examples. Section 7 contains the conclusions.
All the proofs are gathered in an Appendix in order to improve the readability.

2. Notations and basic definitions

Let R, R≥ 0 , Z and Z≥0 denote the real, the non-negative real, the integer and the non-
negative integer numbers, respectively. For a given M ∈ Z≥0 , the following set is defined
T M  {kM, k ∈ Z≥0 }. Euclidean norm is denoted as | · |. Given a signal w, the signal’s
sequence is denoted by w  {w(0), w(1), · · ·} where the cardinality of the sequence is in-
ferred from the context. The set of sequences of w, whose values belong to a compact set
W ⊆ Rm is denoted by MW , while W sup  supw∈W {|w|}, W in f  infw∈W {|w|}. More-
over w  supk≥0 {|w(k)|} and w[τ ]   sup0≤k≤τ {|w(k)|}. The symbol id represents the
identity function from R to R, while γ1 ◦ γ2 is the composition of two functions γ1 and γ2
from R to R. Given a set A ⊆ Rn , |ζ |A  inf {|η − ζ | , η ∈ A} is the point-to-set distance
from ζ ∈ Rn to A. The difference between two given sets A ⊆ Rn and B ⊆ Rn with B ⊆ A,
is denoted by A\ B  { x : x ∈ A, x ∈ / B}. Given two sets A ⊆ Rn and B ⊆ Rn , then the
Pontryagin difference set C is defined as C = A ∼ B  { x ∈ Rn : x + ξ ∈ A, ∀ξ ∈ B}.
Given a closed set A ⊆ Rn , ∂A denotes the border of A. A function γ : R≥ 0 → R≥ 0 is of class
K (or a ”K-function”) if it is continuous, positive definite and strictly increasing. A function
γ : R≥ 0 → R≥ 0 is of class K ∞ if it is a K-function and γ(s) → +∞ as s → +∞. A function
β : R≥0 × Z≥0 → R≥0 is of class KL if, for each fixed t ≥ 0, β(·, t) is of class K, for each fixed
s ≥ 0, β(s, ·) is decreasing and β(s, t) → 0 as t → ∞.

3. Problem statement
In this paper it is assumed that the plant to be controlled is described by discrete-time nonlin-
ear model:

x (k + 1) = f ( x (k), u(k)) + w(k ), k ≥ t, x (t) = x̄ (1)

where x (k) ∈ Rn is the state of the system, u(k) ∈ Rm is the control variable, and w(k) ∈ Rn
is the additive uncertainty. Notice that the additive uncertainty can model perturbed systems
and a wide class of model mismatches. Take into account that these ones might depend on
the state and on the input of the system, consider a real plant xk+1 = f˜( x (k), u(k)). Then the
additive uncertainty can be taken as w(k ) = [ f˜( x (k), u(k)) − f ( x (k), u(k))]. Note that if, as
it will be assumed, x and u are bounded and f is Lipschitz, then w can be modeled as a
bounded uncertainty. This kind of model uncertainty has been used in previous papers about
robustness in MPC , as in Michalska & Mayne (1993) and Mayne (2000).
In the following assumption, the considered structure of such a model is formally presented.

Assumption 1.
90 Model Predictive Control

1. The uncertainty belongs to a compact set W ⊂ Rn containing the origin, defined as

W  {w ∈ Rn : |w| ≤ γ} (2)

where γ ∈ R≥0 .
2. The system has an equilibrium point at the origin, that is f (0, 0) = 0.
3. The control and state of the plant must fulfill the following constraints on the state and the input:

x (k) ∈ X (3)

u(k) ∈ U (4)
where X is and U are compact sets, both of them containing the origin.
4. The state of the plant x (k) can be measured at each sample time. 

The control objective consists in designing a control law u = κ ( x ) such that it steers the system
to (a neighborhood of) the origin fulfilling the constraints on the input and the state along the
system evolution for any possible uncertainty and yielding an optimal closed performance
according to certain performance index.

4. Regional Input-to-State Stability

In this section the ISS framework for discrete-time autonomous nonlinear systems is pre-
sented and Lyapunov-like sufficient conditions are provided. This will be employed in the
paper to study the behavior of perturbed nonlinear systems in closed-loop with MPC con-
trollers. Consider a nonlinear discrete-time system described by

x (k + 1) = F (k, x (k), w(k)), k ≥ t, x (t) = x̄ (5)

where F : Z≥0 × Rn × Rr → Rn is locally Lipschitz continuous, F (k, 0, 0) = 0, x (k) ∈ Rn

is the state, w(k) ∈ R p is the input (disturbance), limited in a compact set W containing the
origin w(k) ∈ W . The solution to the difference equation (5) at time k, starting from state
x (0) = x̄ and for inputs w is denoted by x (k, x̄, w). Consider the following definitions.

Definition 1 (Robust positively invariant set). A set Ξ(k) ⊆ Rn is a robust positively invariant
set for the system (5), if x (k, x̄, w) ∈ Ξ(k), ∀k ≥ t, ∀ x̄ ∈ Ξ(t) and ∀w ∈ MW . 

Definition 2 (Magni et al. (2006) Regional ISS in Ξ(k)). Given a compact set Ξ(k) ⊂ Rn contain-
ing the origin as an interior point, the system (5) with w ∈ MW , is said to be ISS (Input-to-State
Stable) in Ξ(k), if Ξ(k) is robust positively invariant for (5) and if there exist a KL-function β and a
K-function γ such that

| x (k, x̄, w)| ≤ β(| x̄ |, k) + γ(w[k−1] ), ∀k ≥ t, ∀ x̄ ∈ Ξ(t). (6)

Definition 3 (Magni et al. (2006) ISS-Lyapunov function in Ξ). A function V: Rn → R≥ 0 is

called an ISS-Lyapunov function in Ξ(k) ⊂ Rn for system (5) with respect to w, if:
1) Ξ(k) is a closed robust positively invariant set containing the origin as an interior point.
Robust Model Predictive Control Algorithms for
Nonlinear Systems: an Input-to-State Stability Approach 91

2) there exist a compact set Ω ⊆ Ξ(k) , ∀k ≥ t (containing the origin as an interior point), a
pair of suitable K∞ -functions α1 , α2 such that:

V ( x ) ≥ α1 (| x |), ∀ x ∈ Ξ(k), ∀k ≥ t (7)

V ( x ) ≤ α2 (| x |), ∀ x ∈ Ω (8)

3) there exist a suitable K∞ -function α3 , a K-function σ such that:

∆V ( x )  V ( F (k, x, w)) − V ( x )
(9)
≤ −α3 (| x |) + σ(|w|), ∀ x ∈ Ξ(k), ∀k ≥ t, ∀w ∈ W

4) there exist a suitable K∞ -functions ρ (with ρ such that (id − ρ) is a K∞ -function) and a suitable
constant cθ > 0, such that there exists a nonempty compact set Θ ⊂ { x : x ∈ Ω, d( x, δΩ) >
cθ } (containing the origin as an interior point) defined as follows:

Θ  { x : V ( x ) ≤ b(W sup )} (10)

where b  α4−1 ◦ ρ−1 ◦ σ, with α4  α3 ◦ α2−1 .



The following sufficient condition for regional ISS of system (5) can be stated.

Theorem 1. If system (5) admits an ISS-Lyapunov function in Ξ(k) with respect to w, then it is ISS
in Ξ(k) with respect to w and limk→∞ | x (k, x̄, w)|Θ = 0.

Remark 1. In order to analyse the control algorithm reported in Section 5.2, a time-varying system
has been considered. However, because all the bounds introduced in the ISS Lyapunov function are
time-invariant, Theorem 1 can be easily derived by the theorem reported in Magni et al. (2006) for
time-invariant systems. 

5. Nonlinear Model Predictive Control

In this section, the results derived in Theorem 1, are used to analyze the ISS property of two
open-loop formulations of stabilizing MPC algorithms for nonlinear systems. The idea on the
base of the two algorithms is the same one. However, there are important differences that,
based on the dynamic system under consideration, give advantages to an algorithm rather
than to the other in terms of domain of attraction and robustness. Notably, in the following
it is not necessary to assume the regularity of the value function and of the resulting control
law.

5.1 MPC with constant optimization horizon

The system (1) with w(k) = 0, k ≥ t, is called nominal model. Let denote ut1 ,t2  {u(t1 ), u(t1 +
1), . . . , u(t2 )}, t2 ≥ t1 , a sequence of vectors and ut1 ,t2 (t3 ) the vector ut1 ,t2 at time t3 . If it is
clear on the context the subscript will be omitted. The vector x̂ (k|t) is the predicted state of
the system at time k (k ≥ t) obtained applying the sequence of inputs ut,k−1 to the nominal
model, starting from the real state x (t) at time t, i.e. x̂ (k |t) = f ( x̂ (k − 1|t), u(k − 1)), k >
t, x̂ (t|t) = x (t).
92 Model Predictive Control

Assumption 2. The function f (·, ·) is Lispchitz with respect to x and u in X × U, with Lipschitz
constants L f and L f u respectively.

Remark 2. Note that the following results could be easily extended to the more general case of f (·, ·)
uniformly continuous with respect to x and u in X × U. Moreover, note that in virtue of the Heine-
Cantor, if X and U are compact, as assumed, then continuity is sufficient to guarantee uniform conti-
nuity Limon (2002); Limon et al. (2009).

Definition 4 (Robust invariant region). Given a control law u = κ ( x ), X̄ ⊆ X is a robust invariant

region for the closed-loop system (1) with u(k) = κ ( x (k)), if x̄ ∈ X̄ implies x (k) ∈ X̄ and κ ( x (k)) ∈
U, ∀w(k) ∈ W , k ≥ t. 

Since there are mismatches between real system and nominal model, the predicted evolution
using nominal model might differ from the real evolution of the system. In order to consider
this effect in the controller synthesis, a bound on the difference between the predicted and the
real evolution is given in the following lemma:

Lemma 1. Limon et al. (2002a) Consider the system (1) satisfying Assumption 2. Then, for a given
sequence of inputs, the difference between the nominal prediction of the state x̂ (k|t) and the real state
of the system x (k) is bounded by

Lkf −t − 1
| x̂ (k|t) − x (k)| ≤ γ, k ≥ t.
Lf − 1

To define the N MPC algorithms first let

Lkf −t −1
Bγk−t  {z ∈ Rn : |z| ≤ L f −1 γ }
Xk − t  X ∼ Bγk−t
= { x ∈ Rn : x + y ∈ X, ∀y ∈ Bγk−t }

then define the following Finite Horizon Optimal Control Problem.

Definition 5 (FHOCP1 ). Given the positive integer N, the stage cost l, the terminal penalty Vf and
the terminal set X f , the Finite Horizon Optimal Control Problem (FHOCP1 ) consists in minimizing,
with respect to ut,t+ N −1 , the performance index

t + N −1
J ( x̄, ut,t+ N −1 , N )  ∑ l ( x̂ (k|t), u(k)) + Vf ( x̂ (t + N |t))
k=t

subject to
(i) the nominal state dynamics (1) with w(k) = 0 and x (t) = x̄;
(ii) the state constraints x̂ (k |t) ∈ Xk−t , k ∈ [t, t + N − 1];
(iii) the control constraints (4), k ∈ [t, t + N − 1];
(iv) the terminal state constraint x̂ (t + N |t) ∈ X f . 
Robust Model Predictive Control Algorithms for
Nonlinear Systems: an Input-to-State Stability Approach 93

It is now possible to define a “prototype” of the first one of two nonlinear MPC algorithms: at
every time instant t, define x̄ = x (t) and find the optimal control sequence uot,t+ N −1 by solving
the FHOCP1 . Then, according to the Receding Horizon (RH) strategy, define κ MPC ( x̄ ) =
uot,t ( x̄ ) where uot,t ( x̄ ) is the first column of uot,t+ N −1 , and apply the control law

u = κ MPC ( x ). (11)

Although the FHOCP1 has been stated for nominal conditions, under suitable assumptions
and by choosing appropriately the terminal cost function Vf and the terminal constraint X f ,
it is possible to guarantee the ISS property of the closed-loop system formed by (1) and (11),
subject to constraints (2)-(4).
Assumption 3. The function l ( x, u) is such that l (0, 0) = 0, l ( x, u) ≥ αl (| x |) where αl is a K∞ -
function. Moreover, l ( x, u) is Lipschitz with respect to x and u, in X × U, with constant Ll and Llu
respectively.
Remark 3. Notice that if the stage cost l ( x, u) is a piece-wise differentiable function in X and U (as
for instance the standard quadratic cost l ( x, u) = x  Qx + u Ru) and X and U are bounded sets, then
the previous assumption is satisfied.
Assumption 4. The design parameter Vf and the set Φ  { x : Vf ( x ) ≤ α}, α > 0, are such that,
given an auxiliary control law κ f ,
1. Φ ⊆ X N −1 ;
2. κ f ( x ) ∈ U, ∀ x ∈ Φ;
3. f ( x, κ f ( x )) ∈ Φ, ∀ x ∈ Φ;
4. αVf (| x |) ≤ Vf ( x ) < β Vf (| x |), ∀ x ∈ Φ, where αVf and β Vf are K∞ -functions;
5. Vf ( f ( x, κ f ( x ))) − Vf ( x ) ≤ −l ( x, κ f ( x )), ∀ x ∈ Φ;
6. Vf is Lipschitz in Φ with a Lipschitz constant Lv .
Remark 4. The assumption above can appear quite difficult to be satisfied, but it is standard in the
development of nonlinear stabilizing MPC algorithms. Moreover, many methods have been proposed in
the literature to compute Vf , Φ satisfying the Assumption 4 (see for example Chen & Allgöwer (1998);
De Nicolao et al. (1998); Keerthi & Gilbert (1988); Magni, De Nicolao, Magnani & Scattolini (2001);
Mayne & Michalska (1990)).
Assumption 5. The design parameter X f  { x ∈ Rn : Vf ( x ) ≤ αv }, αv > 0, is such that for all
x ∈ Φ, f ( x, k f ( x )) ∈ X f .
Remark 5. If Assumption 4 is satisfied, then, a value of αv satisfying Assumption 5 is the following
1 −1
αv = (id + αl ◦ β−
Vf ) ( α ).

For each x (k) ∈ Φ there could be two cases. If Vf ( x (k)) ≤ αv , then, by Assumption 4, Vf ( x (k + 1)) ≤
αv . If V ( x (k)) > αv , then, by point 4 of Assumption 4, β Vf (| x (k)|) ≥ Vf ( x (k)) > αv , that means
1
| x (k)| > β−
Vf ( αv ). Therefore, by Assumption 3 and point 4 of Assumption 4, one has

Vf ( x (k + 1)) ≤ Vf ( x (k)) − l ( x (k), κ f ( x (k))) ≤ Vf ( x (k)) − αl (| x (k)|)

1
≤ α − αl ◦ β−
Vf ( αv )
94 Model Predictive Control

1
for all Vf ( x (k + 1)) ≤ αv . Then, αv = α − αl ◦ β−
Vf ( αv ) satisfy the previous equation. After some
1 −1
manipulations one has αv = (id + αl ◦ β−
Vf ) ( α ). 

Let X MPC ( N ) be the set of states of the system where an admissible solution of the FHOCP1
optimization problem exists.

Definition 6. Let α1 = α3 = αl , α2 = β Vf , Ξ = X MPC ( N ), Ω = Φ, σ = L J , where L J 

−1 L Nf −1 −1
LVf L N
f + Ll L f −1 .

Assumption 6. The values w are such that point 4 of Definition 2 is satisfied with V ( x ) 
J ( x, uot,t+ N −1 , N ).
Remark 6. From this assumption it is inferred that the allowable size of disturbances is related with
the size of the local region Ω where the upper bound of the terminal cost is found. This region can
be enlarged following the way suggested in Limon et al. (2006). However, this might not produce an
enlargement of the allowable size since the new obtained bound is more conservative. 
The main peculiarities of this N MPC algorithm are the use in the FHOCP1 of: (i) tightened
state constraints along the optimization horizon; (ii) terminal set that is only a subset of
the region where the auxiliary control law satisfies Assumption 4 in order to guarantee
robustness (see Assumptions 4 and 5).

Let introduce now following theorem.

Theorem 2. Let a system be described by a model given by (1). Assume that Assumptions 1-6 are
satisfied. Then the closed loop system (1), (11) is ISS with robust invariant region X MPC ( N ) if the
uncertainty is such that
α − αv
γ≤ −1
(12)
Lv L N
f

5.2 MPC with time-varying control horizon

In this sub-section the second algorithm will be shown. It is based on the same ideas of the
first one and it is motivated by the attempt to reduce its intrinsic conservativity.
The second Finite Horizon Optimal Control Problem (FHOCP2 ) to be introduced is character-
ized by using a time varying control horizon Nc (t) and a (time invariant) prediction horizon
Np . The control horizon is given by
  
t
Nc (t)  +1 M−t
M
where · indicates the integer part operator and M is a parameter which determines its
maximum value, i.e. Nc (t) ∈ [1, M ] .
Definition 7 (FHOCP2 ). Given a stabilizing control law κ f the maximum control horizon M, the
prediction horizon Np , the stage cost l, and the terminal penalty Vf , the Finite Horizon Optimal Con-
trol Problem (FHOCP2 ) consists in minimizing, with respect to ut,t+ Nc (t)−1 , the performance index
t + Np −1
J ( x̄, ut,t+ Nc (t)−1 , Nc (t), Np )  ∑ l ( x̂ (k|t), u(k)) + Vf ( x̂ (t + Np |t))
k=t
Robust Model Predictive Control Algorithms for
Nonlinear Systems: an Input-to-State Stability Approach 95

subject to
(i) the nominal state dynamics (1) with w(k) = 0 and x̄ = x (t);
(ii) the state constraints x̂ (k|t) ∈ Xk−t , k ∈ [t, ..., t + Nc (t) − 1];
(iii) the control constraints (4), k ∈ [t, ..., t + Nc (t) − 1];
(iv) the terminal state constraint x̃ (t + Nc (t)|t + Nc (t) − M) ∈ X f where x̃ denotes the nom-
inal prediction of the system considering as initial condition x (t + Nc (t) − M) and ap-
plying the sequence of control inputs ũt+ Nc (t)− M,t+ Nc (t)−1 defined as

uok,k if k < t
ũt+ Nc (t)− M,t+ Nc (t)−1 (k) =
ut,t+ Nc (t)−1 (k) if k ≥ t

(v) the control signal


ut,t+ Nc (t)−1 (k), k ∈ [t, t + Nc (t) − 1]
u(k) = (13)
κ f ( x̂ (k|t)), k ∈ [t + Nc (t), t + Np − 1]


It is now possible to introduce the second N MPC algorithm in the following way: at every
time instant t, define x̄ = x (t) and find the optimal control sequence uot,t+ N (t)−1 by solving
c
the FHOCP2 . Then, according to the RH strategy, define κ MPC (t, x̄, x̃ (t|t + Nc (t) − M)) =
uot,t ( x̄, x̃ (t|t + Nc (t) − M )) where uot,t ( x̄, x̃ (t|t + Nc (t) − M)) is the first column of uot,t+ N (t)−1 ,
c
and apply the control law

u(t) = κ MPC (t, x (t), x̃ (t|t + Nc (t) − M)). (14)

Note that the control law is time variant (periodic) due to the time variance of the control
horizon Nc (t) and depends also on x̃ (t|t + Nc (t) − M ).
Therefore, defining
   
x (t) ξ 1 (t)
ξ (t) = = ∈ R2n ,
x̃ (t|t + Nc (t) − M)) ξ 2 (t)

the closed-loop system formed by (1) and (14) is given by

ξ (k + 1) = F̃ (k, ξ (k), w(k)), k ≥ t, ξ (t) = ξ̄ (15)

where
 
 f (ξ 1 (k), κ MPC (k, ξ 1 (k), ξ 2 (k))) + w(k )
F̃ (k, ξ (k), w(k)) =  f (ξ 2 (k), κ MPC (k, ξ 1 (k), ξ 2 (k))), ∀(k + 1) ∈ / TM 
f (ξ 1 (k), κ MPC (k, ξ 1 (k), ξ 2 (k))) + w(k), ∀(k + 1) ∈ T M

Definition 8. Let X MPC (t, Np ) ∈ R2n be the set of states ξ (t) where an admissible solution of the
FHOCP2 exists. 
96 Model Predictive Control

Noting that x (t) = x̃ (t|t + Nc (t) − M )), ∀t ∈ T M since Nc (t) = M, the closed-loop system (1),
(14) for k ∈ T M is time invariant since the control law is time invariant and

x (k + M) = F̄ ( x (k), wk,k+ M−1 ), ∀k ∈ T M , k ≥ t, x (t) = x̄. (16)

MPC ( N ) ∈ Rn be the set x of states of the system (1) where an admissible

Definition 9. Let X M p
solution of the FHOCP2 exists ∀t ∈ T M . 

As in the previous algorithm, although the FHOCP2 has been stated for nominal conditions,
under suitable assumptions and by choosing accurately the terminal cost function Vf and the
terminal constraint X f , it is possible to guarantee the ISS property of the closed-loop system
formed by (1) and (14), subject to constraints (2)-(4).

Assumption 7. The auxiliary control law κ f is Lipschitz in Φ with a Lipschitz constant Lκ where
Φ  { x ∈ X M−1 : Vf ( x ) ≤ α}, α > 0.

Remark 7. Note that, an easy way to satisfy Assumption 7 is to choose κ f linear, e.g. the solution of
the infinite horizon optimal control problem for the unconstrained linear system.

Assumption 8. The design parameter X f  { x ∈ Rn : Vf ( x ) ≤ αv } is such that, considering

the system (1), with u = κ f ( x ) and w(k) = 0, for all x (t) ∈ Φ results x̂ (t + M|t) ∈ X f and
x̂ (k|t) ∈ Xk−t , k ∈ [t, t + M − 1].
MPC ( N ), Ω = Φ, σ = L M , where
Definition 10. Let α1 = α3 = αl , α2 = β Vf , Ξ = X M p J
 
N (k )−1 Np − Nc (k)+1
t + M −1  Lf c −1 N (k )−1 Lx −1 
N (k)−1 Np − Nc (k)+1
LM
J  ∑ Ll + Llx L f c + Lv L f c Lx
k=t
 Lf − 1 Lx − 1 

with L x  ( L f + L f u Lκ ) and Llx  ( Ll + Llu Lκ ).

Assumption 9. The values w are such that point 4 of Definition 2 is satisfied with V ( x ) 
J ( x, uot,t+ M−1 , M, Np ).

The main peculiarities of this N MPC algorithm, with respect to the one previously presented,
are the use in the FHOCP2 of: (i) a time varying control horizon; (ii) a control horizon that
is different from prediction horizon; (iii) the fact that the real value of the state is updated
only each M step to check the terminal constraint while it is updated at each step for the
computation of cost. These modifications allows to relax Assumption 5 with Assumption 8.
In this way it could be possible to enhance the robustness. The idea to use the measure of
the state only each M step has been already used in an other context in contractive MPC de
Oliveira Kothare & Morari (2000).

Theorem 3. Let a system be described by a model given by (1). Assume that Assumptions 1-4, 7-9
are satisfied. Then the closed loop system (15) is ISS with robust invariant region X MPC (t, Np ) if the
uncertainty is such that
α − αv
γ≤ (17)
L M −1
Lv Lff −1
Robust Model Predictive Control Algorithms for
Nonlinear Systems: an Input-to-State Stability Approach 97

Different from Magni, De Nicolao, Magnani & Scattolini (2001) the use of a prediction horizon
longer than the control horizon does not affect the size of the robust invariant region because
the terminal inequality constraint has been imposed at the end of the control horizon. How-
ever the following theorem proves that this choice has positive effect on the performance.

Theorem 4. Magni, De Nicolao, Magnani & Scattolini (2001) Letting l ( x, u) = x  Qx + u Ru,

Q> 0, R > 0, u = −K LQ x the solution of the infinite horizon optimal control problem for the
unconstrained linear system
x (k + 1) = Ax (k) + Bu(k)
with A = ∂ f ( x, u)/∂x | x=0,u=0 , B = ∂ f ( x, u)/∂u| x=0,u=0 , for each given Nc , if κ f ( x ) = −K LQ x,
then lim Np →∞ ∂κ MPC ( x )/∂x | x=0 = K LQ .

In conclusion, Theorems 2 and 3 proven that both the algorithm guarantee the ISS of the
closed-loop system. However a priori it is not possible to establish which of the two algo-
rithms give more robustness. This because of the dependance from the values of L f , M, Np of
the bounded on the maximum disturbance allowed. Therefore, based on the dynamic system
in object, it will be used an algorithm rather than the other.

6. Examples
The objective of the examples is to show that, based on the values of certain parameters, one
algorithm can be better than the other. In particular two examples are shown: in the first
one the algorithm based on FHOCP1 is better than the one based on FHOCP2 in terms of
robustness; in the second one the contrary happens.

6.1 Example 1
Consider the uncertain nonlinear system given by

x1 ( k + 1) = 0.55x1 (k) + 0.12x2 (k) + (0.01 − 0.6x1 (k ) + x2 (k) + Λ1 )u(k)

x2 ( k + 1) = 0.67x2 (k) + (0.15 + x1 (k) − 0.8x2 (k) + Λ2 )u(k)

where Λ1 and Λ2 are the parameters of the system model uncertainty. The control is con-
strained to be |u| ≤ umax = 0.2. Defining w = [Λ1 u T Λ2 u T ] T the disturbance is in the form (1)
and the nominal system is in the form x (k + 1) = Ax + Bu + Cxu. Considering the ∞-norm,
the Lipschitz constant of the system is

Lf = maxu (| A + Cu|∞ ) = max{| A + 3C |∞ , | A − 3C |∞ } = 1.03.

In theformulation
 of the FHOCP1 and FHOCP2 the stage is l ( x, u) = x  Qx + u Ru with
1 0
Q = 0 1
, R = 1 and the auxiliary control law u = −K LQ x is derived by solving an
Infinite Horizon optimal control problem for the linearized system around the origin

x1 ( k + 1) = 0.55x1 (k) + 0.12x2 (k) + 0.01u(k)

x2 ( k + 1) = 0.67x2 (k) + 0.15u(k)

with
 the same stage
 cost. The solution of the associated Riccati Equation is P =
1.4332 0.1441 LQ LQ
 
so that the value of K is K = −0.0190 −0.1818 . The value of the
0.1441 1.8316
98 Model Predictive Control

Lipschitz constant Lκ of the auxiliary control law is Lκ = |K LQ |∞ = 0.1818. The terminal

penalty Vf ( x ) = βx  Px, where β = 1.2 satisfies

λmax ( Q + K LQ RK LQ ) < βλmin ( Q + K LQ RK LQ )

in order to verify Assumption 7. Therefore, considering the presence of the constraint on the
control, the linear controller u = −K LQ x stabilizes the system only in the invariant set Φ, Φ =
{ x : 1.2x  Px ≤ α = 0.2} The value of the Lipschitz constant Lv is Lv = maxx∈Φ |2βPx |∞ =
2.4| Px |∞ = 1.3222. For the algorithm based on FHOCP2 the final constraint X f depends on
the value M while for the algorithm based on FHOCP1 it results X f = { x : 3x  Px ≤ 0.0966}.
In Figure 1.a the maximum value of γ that satisfies (12) (solid line) and the one that satisfies
the (17) (dotted line) for different values of M, are reported. In this example the algorithm
based on the FHOCP1 guarantees major robustness than the one based on FHOCP2 .

0.09 0.2
First algorithm Firts algorithm
Second algorithm 0.18 Second algorithm
0.08
0.16
0.07
0.14
0.06
0.12

0.05 0.1
γ
γ

0.08
0.04
0.06
0.03
0.04

0.02 0.02

0.01 0
1 2 3 4 5 6 7 8 1 2 3 4 5 6 7 8
M=Np M=Np

(a) Example 1: comparison of γ between the two (b) Example 2: comparison of γ between the two al-
algorithms. gorithms.

LMPC 1 LMPC
NMPC NMPC
2 MPC MPC
0.5
φ

0
1 X
f
−0.5

0 −1
x2

x2

−1.5

−1
−2

−2.5
−2

−3

−3 −3.5

−6 −4 −2 0 2 4 6 −5.8 −5.6 −5.4 −5.2 −5 −4.8 −4.6 −4.4 −4.2 −4

x1 x1

(c) Example 2: closed loop state evolution. (d) Example 2: detail of the closed-loop state evo-
lution with initial state (-4.1;-3).
Robust Model Predictive Control Algorithms for
Nonlinear Systems: an Input-to-State Stability Approach 99

6.2 Example 2
This example shows a case in which the algorithm based on FHOCP2 gives a better solution.
Consider the uncertain nonlinear system

x1 ( k + 1) = x2 (k) + (0.3x2 (k) + Λ1 )u

x2 ( k + 1) = −0.32x1 (k) + 1.8x2 (k) + (1 − 0.2x2 (k) + Λ2 )u

where Λ1 and Λ2 are the parameters of the system model uncertainty. The control is con-
strained to be |u| ≤ umax = 3 and the state x1 is constrained to be x1 ≥ −4.8. Considering the
∞-norm, the Lipschitz constant of the system is

Lf = maxu (| A + Cu|∞ ) = max{| A + 3C |∞ , | A − 3C |∞ } = 2.72.

In the formulation
 of the FHOCP1 and FHOCP2 the stage is l ( x, u) = x  Qx + u Ru with
1 0
Q = , R = 1 and the auxiliary control law u = −K LQ x is derived by solving an
0 1
Infinite Horizon optimal control problem for the linearized system around the origin

x1 ( k + 1) = x2 ( k )
x2 ( k + 1) = −0.32x1 (k) + 1.8x2 (k) + u

with
 the same stage
 cost. The solution of the associated Riccati Equation is P =
1.0834 −0.4428 LQ LQ
 
−0.4428 4.3902
so that the value of K is K = −0.2606 1.3839 . The value of
the Lipschitz constant Lκ of the auxiliary control law is Lκ = |K LQ |∞ = 1.3839. The terminal
penalty Vf ( x ) = βx  Px, where β = 3, satisfies

λmax ( Q + K LQ RK LQ ) < βλmin ( Q + K LQ RK LQ )

in order to verify Assumption 7. Therefore, considering the presence of the constraint on the
control, the linear controller u = −K LQ x stabilizes the system only in the invariant set Φ, Φ =
{ x : 3x  Px ≤ α = 40.18}. The value of the Lipschitz constant Lv is Lv = maxx∈Φ |2βPx |∞ =
6| Px |∞ = 45.9926. For the algorithm based on FHOCP2 the final constraint X f depends on the
value M while for the algorithm based on FHOCP1 it results X f = { x : 3x  Px ≤ 31.2683}. In
Figure 1.b the maximum value of γ that satisfies (12) (solid line) and the one that satisfies the
(17) (dotted line) for different values of M, are reported. In this example, the advantage of the
algorithm based on the FHOCP2 with respect to first one is due to the fact that the auxiliary
control law can lead the state of the nominal system from Φ to X f in M steps rather than in
only one. Hence, since the difference between Φ and X f is bigger, then a bigger perturbation
can be tolerated. In Figure 1.c the state evolutions of the nonlinear system obtained with
different control strategies with initial condition

x01 6 −4.1 7 6 −4.6

x02 −2.5 −3 1.5 −1 1

and γ = 0.0581 are reported: in solid line, using the new algorithm (N MPC), with Np = 10
and M = 3, in dashed line, using the new algorithm but with the linearized system in the
solution of the FHOCP (LMPC) and in dash-dot line the results of a nominal MPC (MPC)
with Np = 10 and Nc = 3. It is clear that, since the model used for the FHOCP differs from the
nonlinear model, using LMPC feasibility is not guaranteed along the trajectory as shown with
100 Model Predictive Control

initial states [−4.6; 1], [−4.1; −3], [6; −1]. Also with the nominal MPC, as shown with initial
states [−4.1; −3], [6; −2.5], since uncertainty is not considered, feasibility is not guaranteed.
Figure 1.d shows a detail of the unfeasibility phenomenon from the first to the second time
instant with initial state [−4.1; −3]. The state constraint infact is robustly fulfilled only with
the N MPC algorithm. For the other initial states, the evolutions of the three strategies are
close.

7. Conclusions
In this paper two design procedures of nominal MPC controllers are presented. The objec-
tive of these algorithms is to provide some degree of robustness when model mismatches are
present. Regional Input-to-State Stability (ISS) has been used as theoretical framework of the
closed loop analysis. Both controllers assume the Lipschitz continuity of the model and of
the stage cost and terminal cost functions. Robust constraint satisfaction is ensured by in-
troducing restricted constraints in the optimization problem based on the estimation of the
maximum effect of the uncertainty. The main differences between the proposed algorithms
are that the second one uses a time varying control horizon and, in order to check the terminal
constraints, it updates the state with the real one just only each M steps. Theorem 2 and The-
orem 3 give sufficient condition on the maximum uncertainty in order to guarantee regional
ISS. The bounds depend on both system parameters and control algorithm parameters. These
conditions, even if only sufficient, give an idea on the algorithm that it is better to use for a
particular system.

8. Appendix

Lemma 2. Let x ∈ Xk−t and y ∈ Rn such that |y − x | ≤ Lkf −t−1 γ. Then y ∈ Xk−t−1 .

Proof : Consider ek−t−1 ∈ Bγk−t−1 , and let denote z = y − x + ek−t−1 . It is clear that

Lkf −t−1 − 1 Lkf −t − 1

|z| ≤ |y − x | + |ek−t−1 | ≤ Lkf −t−1 γ + γ= γ
Lf − 1 Lf − 1

thus, z ∈ Bγk−t . Taking into account that x ∈ Xk−t , for all ek−t−1 ∈ Bγk−t−1 , it results that
y + ek−t−1 = ( x + z) ∈ X. This yields that y ∈ Xk−t−1 . 

Proof of Theorem 2: Firstly, it will be shown that region X MPC ( N ) is robust positively invariant
for the closed loop system: if x (t) ∈ X MPC ( N ), then x (t + 1) = f ( x (t), uo (t)) + w(t) ∈
X MPC ( N ) for all w(t) ∈ W. This is achieved by proving that for all x (t) ∈ X MPC ( N ), there
exists an admissible solution of the optimization problem in t + 1, based on the optimal
solution in t, i.e. ūt+1,t+ N = [uot+1,t+ N −1 , k f ( x̂ (t + N |t + 1))]. Let denote x̄ (k|t + 1) the state
obtained applying the input sequence ūt+1,k−1 to the nominal model with initial condition
x (t + 1). In order to prove that the sequence ūt+1,t+ N is admissible, it is necessary that
a) ū(k) ∈ U, k ∈ [t + 1, t + N ]: it follows from the feasibility of uot,t+ N −1 and the fact that
κ f ( x ) ∈ U, ∀ x ∈ X f ⊆ Φ.
Robust Model Predictive Control Algorithms for
Nonlinear Systems: an Input-to-State Stability Approach 101

b) x̄ (t + N + 1|t + 1) ∈ X f : first, it is going to be shown that x̄ (t + N |t + 1) ∈ Φ. Taking

−1
into account that | x̄ (t + N |t + 1) − x̂ (t + N |t)| ≤ L N
f γ then

−1 −1
Vf ( x̄ (t + N |t + 1)) ≤ Vf ( x̂ (t + N |t)) + Lv L N
f γ ≤ αv + Lv L N
f γ ≤ α.

Therefore x̄ (t + N |t + 1) ∈ Φ and hence, applying the auxiliary control law, x̄ (t + N +

1| t + 1) ∈ X f .
c) x̄ (k|t + 1) ∈ Xk−t−1 , k ∈ [t + 1, t + N ]: considering that | x (t + 1) − x̂ (t + 1|t)| ≤ γ by
recursion | x̄ (k|t + 1) − x̂ (k|t)| ≤ Lkf −t−1 γ for k ∈ [t + 1, t + N ]. Since x̂ (k|t) ∈ Xk−t ,
then, by Lemma 2, x̄ (k|t + 1) ∈ Xk−t−1 . Moreover, since x̄ (t + N |t + 1) ∈ Φ ⊆ X N −1 ,
the proof is completed.
Now, in order to show that the closed loop system (1), (11) is ISS in X MPC ( N ), let verify
that V ( x̄, N )  J ( x̄, uot,t+ N −1 , N ) is an ISS-Lyapunov function in X MPC ( N ). First note that by
Assumption 3
V ( x̄, N ) ≥ αl (| x̄ |), ∀ x̄ ∈ X MPC ( N ). (18)
Moreover, in view of Assumption 4, ũt,t+ N = [uot,t+ N −1 , k f ( x̂ (t + N |t))] is an admissible, pos-
sible suboptimal, control sequence for the FHOCP1 with horizon N + 1 at time t with cost

J ( x̄, ũt,t+ N , N + 1) = V ( x̄, N ) − Vf ( x̂ (t + N |t)) + Vf ( x̂ (t + N + 1|t))

+l ( x̂ (t + N |t), k f ( x̂ (t + N |t))).

Since ũt,t+ N is a suboptimal sequence, V ( x̄, N + 1) ≤ J ( x̄, ũt,t+ N , N + 1) and, using point 5 of
Assumption 4, it follows that J ( x̄, ũt,t+ N , N + 1) ≤ V ( x̄, N ). Then

V ( x̄, N + 1) ≤ V ( x̄, N ), ∀ x̄ ∈ X MPC ( N )

with V ( x̄, 0) = Vf ( x̄ ), ∀ x̄ ∈ Φ. Therefore

V ( x̄, N ) ≤ V ( x̄, N − 1) ≤ Vf ( x̄ ) < β Vf (| x̄ |), ∀ x̄ ∈ Φ. (19)

Moreover, let define ∆J as

∆J  J ( x (t + 1), ūt+1,t+ N , N ) − J ( x (t), uot,t+ N −1 , N )
k = t + N −1
= −l ( x (t), uo (t)) + ∑ {l ( x̄ (k|t + 1), ū(k)) − l ( x̂ (k|t), uo (k))}
k = t +1
+l ( x̄ (t + N |t + 1), ū(t + N )) + Vf ( x̄ (t + N + 1|t + 1) − Vf ( x̂ (t + N |t)). (20)

From the definition of ū, ū(k) = uo (k), for k

∈ [t + 1, t + N − 1], and hence l ( x̄ (k|t + 1), ū(k)) −
l ( x̂ (k |t), uo (k)) ≤ Ll Lkf −t−1 γ and analogously

−1
Vf ( x̄ (t + N |t + 1) − Vf ( x̂ (t + N |t)) ≤ Lv L N
f γ.

Substituting these expressions in (20) and considering that x̄ (t + N |t + 1) ∈ Φ, from Assump-

tion 4, there is

∆J ≤ [l ( x̄ (t + N |t + 1), ū(t + N )) + Vf ( x̄ (t + N + 1|t + 1) − Vf ( x̄ (t + N |t + 1)]

−l ( x (t), uo (t)) + L J γ ≤ −l ( x (t), uo (t)) + L J γ
102 Model Predictive Control

−1 L N −1 − 1
where L J  Lv L N
f + Ll Lf f −1 . Considering that by Assumption 3, l ( x, u) ≥ αl (| x |) and the
optimality of the solution, then
V ( x (t + 1), N ) − V ( x (t), N ) ≤ ∆J ≤ −αl (| x (t)|) + L J γ, ∀ x ∈ X MPC ( N ) (21)
Therefore, by (18), (19) and (21), V ( x̄, N ) is an ISS-Lyapunov function of the closed loop
system (1), (11), and hence, the closed-loop system is ISS with robust invariant region
X MPC ( N ). 

Proof of Theorem 3: Firstly, it will be shown that region X MPC (t, Np ) is robust positively invari-
ant for the closed-loop system. This is achieved by proving that for all ξ (t) ∈ X MPC (t, Np ),
there exists an admissible solution ūt+1,t+1+ Nc (t+1)−1 of the optimization problem in t + 1,
based on the optimal solution in t. This sequence is given by

uot,t+ N (t)−1 (k) if t + 1 ∈ T M
ūt+1,t+1+ Nc (t+1)−1 (k) = c
κ f ( x̂ (k|t + 1)) if t + 1 ∈ T M

for k ∈ [t + 1, · · · , t + 1 + Nc (t + 1) − 1]. Notice that if t + 1 ∈ T M , Nc (t + 1) = Nc (t) − 1 and

hence the sequence is well defined.
Moreover, since necessary for the ISS proof, it will be shown that, starting from the (nominal)
state x̂ (t + 1|t), the sequence ūt+1,t+1+ N (t+1)−1 is admissible. This is given by
c

uot,t+ N (t)−1 (k) if t + 1 ∈ T M
ūt+1,t+1+ Nc (t+1)−1 (k) = c
κ f ( x̂ (k|t)) if t + 1 ∈ T M

for k ∈ [t + 1, · · · , t + 1 + Nc (t + 1) − 1].

In order to prove that the two sequences are admissible, it is necessary that
1) x̃ (t + 1 + Nc (t + 1)|t + 1 + Nc (t + 1) − M) ∈ X f with ũt+1+ Nc (t+1)− M,t+1+ Nc (t+1)−1 derived
from both ū and ū ;
2) x̂ (k|t + 1) ∈ Xk−t−1 , k ∈ [t + 1, t + 1 + Nc (t + 1) − 1] with input ū;
3) x̂ (k|t) ∈ Xk−t , k ∈ [t + 1, t + 1 + Nc (t + 1) − 1] with input ū ;
4) ū(k) ∈ U, ū (k) ∈ U, k ∈ [t + 1, t + 1 + Nc (t + 1) − 1].

1) First note that if t + 1 ∈ T M , then ū(k) = ū (k) = uo (k), k ∈ [t + 1, t + 1 + Nc (t + 1) − 1].

This yields to x̃ (k|t + Nc (t) − M) = x̃ (k|t + 1 + Nc (t + 1) − M) for all k ∈ [t + 1 + Nc (t + 1) − M, t +
1 + Nc (t + 1)] and hence
x̃ (t + 1 + Nc (t + 1)|t + 1 + Nc (t + 1) − M ) = x̃ (t + Nc (t)|t + Nc (t) − M ) ∈ X f .
On the contrary, if t + 1 ∈ T M then ūt+1,t+1+ Nc (t+1)−1 (k) = κ f ( x̂ (k|t + 1)) and
ūt+1,t+1+ N (t+1)−1 (k) = κ f ( x̂ (k|t)). We are going to prove that both sequence satisfies the
c
terminal constraint:
• Consider the sequence ū and let denote ũ and x̃ the sequence and predictions derived
from ū. In virtue of Lemma 1 and the fact that Nc (t) = 1, the following inequality holds

LM
f −1
| x (t + 1) − x̃ (t + 1|t + Nc (t) − M)| ≤ γ (22)
Lf − 1
Robust Model Predictive Control Algorithms for
Nonlinear Systems: an Input-to-State Stability Approach 103

and by point 5 of Assumption 4 it follows that

Vf ( x (t + 1)) − Vf ( x̃ (t + 1|t + Nc (t) − M))

LM
f −1
≤ Lv | x (t + 1) − x̃ (t + 1|t + Nc (t) − M)| ≤ Lv γ
Lf − 1

Hence, considering that x̃ (t + 1|t + Nc (t) − M) ∈ X f and the uncertainty satisfies (17),
then
LM −1 LM −1
Vf ( x (t + 1)) ≤ Vf ( x̃ (t + 1|t + Nc (t) − M )) + Lv f
L f −1 γ ≤ αv + Lv f
L f −1 γ≤α (23)

and therefore x (t + 1) ∈ Φ. Hence, from Assumption 8, κ f ( x̂ (k|t + 1)) steers the nomi-
nal state in X f in M steps. Then ūt+1,t+ Nc (t+1)−1 satisfies the constraint.
• Let consider now ū and let denote ũ and x̃  the sequence and predictions derived from
ū . Since x̂ (t + 1|t) = f ( x (t), uot,t ) we have that
| x̂ (t + 1|t) − x̃  (t + 1|t + Nc (t) − M)|
= | f ( x (t), uo (t)) − f ( x̃  (t|t + Nc (t) − M), uo (t))|
≤ L f | x (t) − x̃  (t|t + Nc (t) − M )|
LM
f
−1 −1
and from (22) | x̂ (t + 1|t) − x̃ (t + 1|t + Nc (t) − M)| ≤ L f L f −1 γ. Finally, following the same
idea used to derive (23)
−1
LM
f −1
Vf ( x̂ (t + 1|t)) ≤ Vf ( x̃  (t + 1|t + Nc (t) − M )) + Lv L f γ
Lf − 1
LM
f −1
< αv + Lv γ ≤ α. (24)
Lf − 1

Therefore Vf ( x̂ (t + 1|t)) < α and consequently x̂ (t + 1|t) ∈ Φ. Hence κ f ( x̂ (k|t)) steers

the nominal state in X f in M steps. Then ūt+1,t+ N (t+1)−1 satisfies the constraint.
c

2) Consider the sequence of inputs ū and assume that t + 1 ∈ T M , then, since by optimality of
solution at time t, x̂ (k|t) ∈ Xk−t and

| x̂ (k|t + 1) − x̂ (k|t)| ≤ Lkf −t−1 γ, k ∈ [t + 1, t + 1 + Nc (t + 1) − 1]

from Lemma 2, it follows that x̂ (k|t + 1) ∈ Xk−t−1 . If t ∈ T M then x (t + 1) ∈ Φ as shown in

(23), and from Assumptions 4, 7, the constraints satisfaction is directly derived.
3) Consider that the sequence ūt+1,t+1+ N (t+1)−1 is applied from the state x̂ (t + 1|t). If
c
t + 1 ∈ T M then the constraints are satisfied since x̂ (k|t) ∈ Xk−t . If t + 1 ∈ T M , as shown in
(24), x̂ (t + 1|t) ∈ Φ and then, by Assumptions 4, 7, constraints satisfaction is directly derived.
4) From the admissibility of uot,t+ N (t)−1 and the fact that for all x ∈ Φ, κ f ( x ) ∈ U, it follows
c
that ū(k) ∈ U, ū (k) ∈ U, k ∈ [t + 1, t + 1 + Nc (t + 1) − 1].

Now, in order to show that the closed loop system (15) is ISS in X MPC (t, Np ), it is first
MPC ( N ).
proven that the closed-loop system (16), defined for each t ∈ T M , is ISS in X M p
104 Model Predictive Control

In order to prove the first part let verify that V ( x̄, M, Np )  J ( x̄, uot,t+ M−1 , M, Np ), is
an ISS-Lyapunov function for the system (16).

Let denote x̄ (k|t + 1) and x̄  (k|t) the state evolution obtained with input ū(k) and ini-
tial state x (t + 1) and with input ū (k) and initial state x (t + 1|t) respectively. Let call J ∗ (t, x ),
J̄ ( x ) and J̄  ( x ) the optimal cost and the cost relative to the admissible sequences ū and ū
respectively. First note that by Assumption 3
MPC
V ( x̄, M, Np ) ≥ αl (| x̄ |), ∀ x̄ ∈ X M ( Np ). (25)

Moreover, ũt,t+ M−1 = uot,t+ M−1 , where uot,t+ M−1 is the optimal control sequence for the
FHOCP2 with prediction horizon Np , is an admissible, possible suboptimal, control sequence
for the FHOCP2 with control horizon M and prediction horizon Np + 1 at time t with cost
J ( x̄, ũt,t+ M−1 , M, Np + 1) = V ( x̄, M, Np ) − Vf ( x̂ (t + Np |t)) + Vf ( x̂ (t + Np + 1|t))
+l ( x̂ (t + Np |t), k f ( x̂ (t + Np |t))).

Since ũt,M−1 is a suboptimal sequence V ( x̄, M, Np + 1) ≤ J ( x̄, ũt,t+ M−1 , M, Np + 1) and, us-
ing point 5 of Assumption 4, it follows that J ( x̄, ũt,t+ M−1 , Np + 1) ≤ V ( x̄, M, Np ). Then
V ( x̄, M, Np + 1) ≤ V ( x̄, M, Np ), ∀ x̄ ∈ X M MPC ( N ), N ≥ M. In particular, it is true that
p p
MPC
V ( x̄, M, Np ) ≤ V ( x̄, M, M ), ∀ x̄ ∈ X M ( M). Now, in view of Assumption 4, ũt,t+ M =
[uot,t+ M−1 , k f ( x̂ (t + M|t))] is an admissible, possible suboptimal, control sequence for the
FHOCP2 with horizon M + 1 with cost
J ( x̄, ũt,t+ M , M + 1, M + 1) = V ( x̄, M, M ) − Vf ( x̂ (t + M|t)) + Vf ( x̂ (t + M + 1|t))
+l ( x̂ (t + M|t), k f ( x̂ (t + M|t))).

Since ũt,t+ M is a suboptimal sequence V ( x̄, M + 1, M + 1) ≤ J ( x̄, ũt,t+ M , M + 1, M + 1) and, using

point 5 of Assumption 4, it follows that J ( x̄, ũt,t+ M , M + 1) ≤ V ( x̄, M, M). Then V ( x̄, M + 1, M +
MPC ( M ) with V ( x̄, 0, 0) = V ( x̄ ), ∀ x̄ ∈ Φ. Therefore
1) ≤ V ( x̄, M, M ), ∀ x̄ ∈ X M f

V ( x̄, M, M ) ≤ V ( x̄, M − 1, M − 1) ≤ Vf ( x̄ ) < β Vf (| x̄ |), ∀ x̄ ∈ Φ. (26)

Moreover, let calculate

J̄  ( x̂ (t + 1|t)) − J ∗ (t, x (t))
t+1+ Nc (t+1)−1 t + Np
= ∑ l ( x̄  (k |t), ū (k)) + ∑ l ( x̄  (k |t), κ f ( x̄  (k |t)))
k = t +1 k =t+1+ Nc (t+1)

t+ Nc (t)−1 t + N p −1
− ∑ l ( x̂ (k |t), uo (k )) − ∑ l ( x̂ (k |t), κ f ( x̂ (k |t)))
k=t k=t+ Nc (t)

+Vf ( x̄  (t + 1 + Np |t)) − Vf ( x̂ (t + Np |t)).

Since, both the state evolutions are obtained with initial condition x̂ (t + 1|t) and the same
input sequence from time t + 1 and until t + Np − 1, there is x̄  (k|t) = x̂ (k|t), k ∈ [t + 1, t + Np ]
so that

J̄  ( x̂ (t + 1|t)) − J ∗ (t, x (t)) = l ( x̂ (t + Np |t), κ f ( x̂ (t + Np |t)) − l ( x (t), uo (t))

+Vf ( x̂ (t + 1 + Np |t)) − Vf ( x̂ (t + Np |t)).
Robust Model Predictive Control Algorithms for
Nonlinear Systems: an Input-to-State Stability Approach 105

Using point 5 of the Assumption 4

J̄  ( x̂ (t + 1|t)) − J ∗ (t, x (t)) ≤ −l ( x (t), κ MPC (t, x (t))). (27)

Let consider now the difference

t+ Nc (t)−1
J̄ ( x (t + 1)) − J̄  ( x̂ (t + 1|t)) = ∑ {l ( x̄ (k|t + 1), ū(k)) − l ( x̄  (k|t), ū (k))}
k = t +1
t + Np
+ ∑ {l ( x̄ (k|t + 1), κ f ( x̄ (k|t + 1)))
k =t+ Nc (t)

−l ( x̄  (k|t), κ f ( x̄  (k|t)))}
+Vf ( x̄ (t + 1 + Np |t + 1)) − Vf ( x̄  (t + 1 + Np |t)).

Note that ū(k) = ū (k), k ∈ [t + 1, t + Nc (t) − 1], while the signals are different for k >
t + Nc (t) − 1. Since | x̄ (k|t + 1) − x̄  (k|t)| ≤ Lkf −t−1 γ from Assumption 3 it is derived that

|l ( x̄ (k|t + 1), ū(k)) − l ( x̄  (k|t), ū (k))| ≤ Ll Lkf −t−1 γ, k ∈ [t + 1, ...., t + Nc (t) − 1]

Therefore, an upper bound for the first part of the summation is given by
N (t)−1
t+ Nc (t)−1 Lf c −1
∑ 
{l ( x̄ (k|t + 1), ū(k)) − l ( x̄ (k|t), ū (k))} 
≤ Ll
Lf − 1
γ. (28)
k = t +1

For k > t + Nc (t), where ū and ū are obtained applying the auxiliary control law to x̄ (k|t + 1)
and x̄  (k|t) respectively, the upper bound is obtained using Assumptions 3 and 7, l ( x̄ (k|t +
1), ū(k |t + 1)) − l ( x̄  (k |t), ū (k|t)) ≤ ( Ll + Llu Lκ )| x̄ (k |t + 1) − x̄  (k |t)| and Assumption 2, | x̄ (k + 1|t +
1) − x̄  (k + 1|t)| ≤ ( L f + L f u Lκ )| x̄ (k |t + 1) − x̄  (k |t)|.Moreover | x̄ (t + Nc (t)|t + 1) − x̄  (t + Nc (t)|t)| ≤
N (t)−1
Lf c γ and defining L x  ( L f + L f u Lκ ) and Llx  ( Ll + Llu Lκ ), the following upper bound
is obtained
t + Np t + Np
∑ {l ( x̄ (k|t + 1), ū(k|t + 1)) − l ( x̄  (k|t), ū (k|t))} ≤ Llx ∑ | x̄ (k|t + 1) − x̄  (k|t)|
k=t+ Nc (t) k=t+ Nc (t)

t + Np
k−t− Nc (t)
≤ Llx ∑ Lx | x̄ (t + Nc (t)|t + 1) − x̄  (t + Nc (t)|t)|
k=t+ Nc (t)
Np − Nc (t)+1
N (t)−1 Lx −1
≤ Llx L f c γ.
Lx − 1
Finally in order to compute an upper bound for the difference of terminal penalties note that
N (t)−1 Np − Nc (t)+1
| x̄ (t + Np + 1|t + 1) − x̄  (t + Np + 1|t)| ≤ L f c
Lx γ and using point 6 of Assumption
N − Nc (t)+1
4, Vf ( x̄ (t + Np + 1|t + 1)) − Vf ( x̄ (t + Np + 1|t)) ≤ Lv L Nf c (t)−1 L x p γ. Therefore the following
bound is obtained
N (t)−1 Np − Nc (t)+1
Lf c −1 N (t)−1 Lx −1
J̄ ( x (t + 1)) − J̄  ( x̂ (t + 1|t)) ≤ Ll γ + Llx L f c γ
Lf − 1 Lx − 1
N (t)−1 Np − Nc (t)+1
+ Lv L f c Lx γ.
106 Model Predictive Control

Defining
N (t)−1 Np − Nc (t)+1
Lf c −1 N (t)−1 Lx −1 N (t)−1 Np − Nc (t)+1
L J (t)  Ll + Llx L f c + Lv L f c Lx
Lf − 1 Lx − 1

it follows that J̄ ( x (t + 1)) ≤ J̄  ( x̂ (t + 1|t)) + L J (t)γ. Considering that J ∗ (t + 1, x (t + 1)) is the

optimal solution at time t + 1, J ∗ (t + 1, x (t + 1)) ≤ J̄ ( x (t + 1)) ≤ J̄  ( x̂ (t + 1|t)) + L J (t)γ. From (27)
it is possible to conclude J ∗ (t + 1, x (t + 1)) − J ∗ (t, x (t)) ≤ −l ( x (t), κ MPC (t, x (t))) + L J (t)γ. and by
Assumption 3
J ∗ (t + 1, x (t + 1)) − J ∗ (t, x (t)) ≤ −αl (| x (t)|) + L J (t)γ. (29)
Now, since V ( x (t), M, Np ) = J ∗ (t, x (t)), ∀t ∈ T M , using (29), there is
t + M −1
V ( x (t + M ), M, Np ) − V ( x (t), M, Np ) ≤ ∑ −αl (| x (k)|) + L J (k)γ
k=t
t + M −1
≤ −αl (| x (t)|) + ∑ L J (k )γ. (30)
k=t

Therefore, by (25), (26) and (30), V ( x, M, Np ) is an ISS-Lyapunov function for the closed-loop
system (16) and hence, the closed-loop system is ISS with robust invariant region X M MPC ( N ).
p
Now, to conclude the proof, it is necessary to demonstrate that, for t ∈ / T M , the system (15) is
ISS in X MPC (t, Np ). Since the model predictive control law (14) is admissible for the FHOCP2 ,
the closed-loop system (15) is such that ξ 1 (t + nM ) ∈ Φ, ∀t ∈ T M , ∀n ∈ Z>0 . Hence, in order
to prove that the system (15) is ISS in X MPC (t, Np ), it is sufficient to prove that the system (15)
is ISS in Φ.
Noting that
t + i −1
αl (| x (t + i )|) ≤ J ∗ (t + i, x (t + i )) ≤ V ( x (t), M, Np ) − αl (| x (t)|) + ∑ L J (k)γ
k=t
t + i −1
≤ β Vf (| x (t)|) + ∑ L J (k )γ, ∀ x (t) ∈ Φ
k=t

considering that for any K∞ -function γ, γ(r + s) ≤ γ(2r ) + γ(2s), there is

t + i −1
1 −1
| x (t + Mn + i )| ≤ α−
l (2β Vf (| x ( t + Mn )|)) + αl (2 ∑ L J ( k ) γ ), ∀ x ( t ) ∈ Φ (31)
k=t

for all n ∈ Z>0 and i ∈ [0, · · · , M − 1]. Since the closed-loop system (16) is ISS with robust
invariant region X M MPC ( N ), there exist a KL-function β (·, ·), and a K -function λ such that
p ∞
| x (t + Mn)| ≤ β(| x (t)|, n) + λ(γ), ∀n ∈ Z≥0 , ∀ x (t) ∈ X M MPC ( N ). Applying this to (31), there is
p
| x (t + Mn + i )| ≤ β̃(| x (t)|, n) + λ̃(γ), ∀ x (t) ∈ Φ. Hence, in conclusion, the system (15) is ISS in
X MPC (t, Np ). 

9. References
Angeli, D., Sontag, E. D. & Wang, Y. (2000). A characterization of integral input-to-state sta-
bility, IEEE Trans. on Automatic Control 45: 1082–1097.
Camacho, E. F. & Bordons, C. (2004). Model Predictive Control, Springer.
Chen, H. & Allgöwer, F. (1998). A quasi-infinite horizon nonlinear model predictive control
scheme with guaranteed stability, Automatica 34: 1205–1217.
Robust Model Predictive Control Algorithms for
Nonlinear Systems: an Input-to-State Stability Approach 107

Chen, H., Scherer, C. W. & Allgöwer, F. (1997). A game theoretical approach to nonlinear
robust receding horizon control of constrained systems, American Control Conference
’97.
De Nicolao, G., Magni, L. & Scattolini, R. (1998). Stabilizing receding-horizon control of non-
linear time-varying systems, IEEE Trans. on Automatic Control AC-43: 1030–1036.
de Oliveira Kothare, S. L. & Morari, M. (2000). Contractive model predictive control for con-
strained nonlinear systems, IEEE Trans. on Automatic Control pp. 1053–1071.
Fontes, F. A. C. C. & Magni, L. (2003). Min-max model predictive control of nonlinear systems
using discontinuous feedbacks, IEEE Trans. on Automatic Control 48: 1750–1755.
Franco, E., Magni, L., Parisini, T., Polycarpou, M. & Raimondo, D. M. (2008). Cooperative
constrained control of distributed agents with nonlinear dynamics and delayed infor-
mation exchange: A stabilizing receding-horizon approach, IEEE Trans. on Automatic
Control 53(1): 324–338.
Gao, K. & Lin, Y. (2000). On equivalent notions of input-to-state stability for nonlinear discrete
time systems, IASTED International Conference on Control and Applications, Cancun,
Mexico, pp. 81–86.
Grimm, G., Messina, M. J., Tuna, S. E. & Teel, A. R. (2004). Examples when nonlinear model
predictive control is nonrobust, Automatica 40: 1729–1738.
Huang, S., James, M. R., Nesic, D. & Dower, P. M. (2005). A unified approach to controller
design for achieving ISS and related properties, IEEE Trans. on Automatic Control
50: 1681–1697.
Jiang, Z.-P., Teel, A. R. & Praly, L. (1994). Small-gain theorem for ISS systems and applications,
Mathematics of Control, Signals, and Systems 7: 95–120.
Jiang, Z.-P. & Wang, Y. (2001). Input-to-state stability for discrete-time nonlinear systems,
Automatica 37: 857–869.
Keerthi, S. S. & Gilbert, E. G. (1988). Optimal, infinite-horizon feedback laws for a general
class of constrained discrete-time systems, J. Optimiz. Th. Appl. 57: 265–293.
Limon, D. (2002). Predictive control of constrained nonlinear systems: stability and robustness, PhD
thesis, Universidad de Sevilla. In spanish.
Limon, D., Alamo, T. & Camacho, E. F. (2002a). Input-to-state stable MPC for constrained
discrete-time nonlinear systems with bounded additive uncertainties, IEEE CDC,
pp. 4619–4624.
Limon, D., Alamo, T. & Camacho, E. F. (2002b). Stability analysis of systems with bounded ad-
ditive uncertainties based on invariant sets: Stability and feasibility of MPC, ACC02,
pp. 364–369.
Limon, D., Alamo, T., Raimondo, D. M., Muñoz de la Peña, D., Bravo, J. M., Ferramosca, A. &
F. Camacho, E. (2009). Input-to-state stability: a unifying framework for robust model
predictive control, in L. Magni, D. M. Raimondo & F. Allgöwer (eds), Nonlinear Model
Predictive Control: Towards New Challenging Applications, Springer-Verlag, pp. 1–26.
Limon, D., Alamo, T., Salas, F. & Camacho, E. F. (2006). Input to state stability of min-max MPC
controllers for nonlinear systems with bounded uncertainties, Automatica 42: 797–
803.
Magni, L., De Nicolao, G., Magnani, L. & Scattolini, R. (2001). A stabilizing model-based
predictive control for nonlinear systems, Automatica 37: 1351–1362.
Magni, L., De Nicolao, G., Scattolini, R. & F.Allgöwer (2003). Robust model predictive control
of nonlinear discrete-time systems, International Journal of Robust and Nonlinear Control
13: 229–246.
108 Model Predictive Control

Magni, L., Nijmeijer, H. & van der Schaft, A. J. (2001). A receding-horizon approach to the
nonlinear H∞ control problem, Automatica 37: 429–435.
Magni, L., Raimondo, D. M. & Allgöwer, F. (eds) (2009). Nonlinear Model Predictive Control:
Towards New Challenging Applications, Springer-Verlag.
Magni, L., Raimondo, D. M. & Scattolini, R. (2006). Regional input-to-state stability for non-
linear model predictive control, IEEE Trans. on Automatic Control 51: 1548–1553.
Magni, L. & Scattolini, R. (2005). Control design for nonlinear systems: Trading robustness and
performance with the model predictive control approach, IEE Proceedings - Control
Theory & Application pp. 333–339.
Magni, L. & Scattolini, R. (2007). Robustness and robust design of MPC for nonlinear discrete-
time systems, R. Findeisen et al. (Eds.): Assessment and Future Directions, Vol. LNCIS
358, Springer-Verlag Berlin Heidelberg, pp. 239–254.
Mayne, D. Q. (2000). Nonlinear model predictive control: Challenges and opportunities, in
F. Allgoẅer & A. Zheng (eds), Nonlinear Model Predictive Control, Progress in Systems
and Control Theory, Birkhauser Verlag, pp. 23–44.
Mayne, D. Q. & Michalska, H. (1990). Receding horizon control of nonlinear systems, IEEE
Trans. on Automatic Control 35: 814–824.
Mayne, D. Q., Rawlings, J. B., Rao, C. V. & Scokaert, P. O. M. (2000). Constrained model
predictive control: Stability and optimality, Automatica 36: 789–814.
Michalska, H. & Mayne, D. Q. (1993). Robust receding horizon control of constrained nonlin-
ear systems, IEEE Trans. on Automatic Control 38: 1623–1633.
Nešić, D. & Laila, D. S. (2002). A note on input-to-state stabilization for nonlinear sampled-
data systems, IEEE Transactions on Automatic Control 47: 1153–1158.
Raimondo, D. M., Limon, D., Lazar, M., Magni, L. & Camacho, E. F. (2009). Min-max model
predictive control of nonlinear systems: a unifying overview on stability, European
Journal of Control 15(1): 5–21.
Raimondo, D. M. & Magni, L. (2006). A robust model predictive control algorithm for non-
linear systems with low computational burden, IFAC Workshop on Nonlinear Model
Predictive Control for Fast Systems, Grenoble, France.
Rawlings, J. B. & Mayne, D. Q. (2009). Model Predictive Control: Theory and Design, Nob Hill
Publishing.
Scokaert, P. O. M., Rawlings, J. B. & Meadows, E. S. (1997). Discrete-time stability with per-
turbations: Application to model predictive control, Automatica 33: 463–470.
Sontag, E. D. (1989). Smooth stabilization implies coprime factorization, IEEE Trans. on Auto-
matic Control 34: 435–443.
Sontag, E. D. & Wang, Y. (1995). On characterizations of the input-to-state stability property,
System & Control Letters 24: 351–359.
Sontag, E. D. & Wang, Y. (1996). New characterizations of input-to state-stability, IEEE Trans.
on Automatic Control 41: 1283–1294.
Model predictive control of nonlinear processes 109

x5

Model predictive control

of nonlinear processes

Ch. Venkateswarlu
Indian Institute of Chemical Technology

1. Historical background
Process control has become an integral part of process plants. An automatic controller must
be able to facilitate the plant operation over a wide range of operating conditions. The
proportional-integral (PI) or proportional-integral-derivative (PID) controllers are
commonly used in many industrial control systems. These controllers are tuned with
different tuning techniques to deliver satisfactory plant performance.

Fig. 1. MPC multi-step prediction scheme.

However, specific control problems associated with the plant operations severely limit the
performance of conventional controllers. The increasing complexity of plant operations
110 Model Predictive Control

together with tougher environmental regulations, rigorous safety codes and rapidly
changing economic situations demand the need for more sophisticated process controllers.
Model predictive control (MPC) is an important branch of automatic control theory. MPC
refers to a class of control algorithms in which a process model is used to predict and
optimize the process performance. MPC has been widely applied in industry (Qin and
Badgwell, 1997). The idea of MPC is to calculate a control function for the future time in
order to force the controlled system response to reach the reference value. Therefore, the
future reference values are to be known and the system behavior must be predictable by an
appropriate model. The controller determines a manipulated variable profile that optimizes
some open-loop performance objective over a finite horizon extending from the current time
into the future. This manipulated variable profile is implemented until a plant measurement
becomes available. Feedback is incorporated by using the measurement to update the
optimization problem for the next time step. Figure 1 explains the basic idea of MPC
showing how the past input-output information is used to predict the future process
behavior at the current time and how this information is extended to future to track the
desired setpoint trajectory. The notation y, u and Ts refer process output, control action and
sample time, respectively.

2. Model predictive control scheme

Model predictive control (MPC) refers to a wide class of control algorithms that use an
explicit process model to predict the behavior of a plant. The most significant feature that
distinguishes MPC from other controllers is its long range prediction concept. This concept
enables MPC to perform current computations to account the future dynamics, thus
facilitating it to overcome the limitations of process dead time, non-minimum phase
behavior and slow dynamics. In addition, MPC exhibits superior performance by
systematically handling constraints violation.

Fig. 2. MPC block diagram.

Model predictive control of nonlinear processes 111

The fundamental framework of MPC algorithms is common for any kind of MPC schemes.
The main differences in many MPC algorithms are the types models used to represent the
plant dynamics and the cost function to be minimized. The multi-step model predictive
control scheme shown in Figure 1 can be realized from the block diagram represented in
Figure 2.
The basic elements in the block diagram are defined as follows. An appropriate model is
used to predict the process outputs, y(t  i ), i  1,...., N over a future time interval known
as prediction horizon, N. A sequence of control actions, u(t+j), j=1,…., m over the control
horizon m are calculated by minimizing some specified objective which is a function of
predicted outputs, y(t+i), set-point values, w(t+i) and control actions, u(t). The first control
move, u(t) of the sequence is implemented and the calculations are repeated for the
subsequent sampling instants. In order to account the plant-model mismatch, a prediction
error, d(t), that is calculated based on plant measurement, y(t) and model prediction, ym(t) is
used to update the future predictions.
In MPC, the control law generates a control sequence, which forces the future system
response to be equal to the reference values. The system response is based on future control
actions, model parameters and the actual system states. Many methods for updating the
optimization problem are possible, such as estimating model parameters and/or states,
inferring about disturbances etc. MPC design considers different types of process models.
These include first principle models, auto regressive moving average models, polynomial
models, neural network models, fuzzy models etc. The attraction for MPC is due to its
capability of handling various constraints directly in the formulation through on-line
optimization. A variety of model predictive control techniques have been reported for
controlling the processes of various complexities.
This chapter presents different linear and nonlinear model predictive controllers with case
studies illustrating their application to real processes.

3. Linear model predictive control

Linear MPC (LMPC) algorithms employ linear or linearized models to obtain the predictive
response of the controlled process. These algorithms include the Model Algorithmic Control
(MAC) (Richalet et al., 1978), the Dynamic Matrix Control (DMC) (Cutler and Ramaker,
1980) and the Generalized Predictive Control (GPC) (Clarke et al., 1987). These algorithms
are all similar in the sense that they rely on process models to predict the behavior of the
process over some future time interval, and the control calculations are based on these
model predictions. The models used for these predictions have usually been derived from
linear approximations of the process or experimentally obtained step response data. A
survey of theory and applications of such algorithms have been reported by Garcia et al.
(1989).

3.1 LMPC design

A classical autoregressive moving average (ARX) model structure that relates the plant
output with the present and past plant input-output can be used to formulate a predictive
model. The model parameters can be determined a priori by using the known input-output
data to form a fixed predictive model or these parameters are updated at each sampling
112 Model Predictive Control

time by an adaptive mechanism. The one step ahead predictive model can be recursively
extended to obtain future predictions for the plant output. The minimization of a cost
function based on future plant predictions and desired plant outputs generates an optimal
control input sequence to act on the plant. The strategy is described as follows.

Predictive model
The relation between the past input-output data and the predicted output can be expressed
by an ARX model of the form

y(t+1) = a1y(t) + . . . + anyy(t-ny+1) + b1u(t) +. . . . . . . + bnuu(t-nu+1) (1)

where y(t) and u(t) are the process and controller outputs at time t, y(t+1) is the one-step
ahead model prediction at time t, a’s and b’s represent the model coefficients and the nu and
ny are input and output orders of the system.

Model identification

The model output prediction can be expressed as

ym(t+1) =  xm(t) (2)

where
 = [1 . . . ny 1 . . . nu] (3)
and
xm(t) = [y(t) . . . y(t-ny+1) u(t) . . . u(t-nu+1)]T (4)

with  and  as identified model parameters.

One of the most widely used estimators for model parameters and covariance is the popular
recursive least squares (RLS) algorithm (Goodwin and Sin, 1984). The RLS algorithm
provides the updated parameters of the ARX model in the operating space at each sampling
instant or these parameters can be determined a priori using the known data of inputs and
outputs for different operating conditions. The RLS algorithm is expressed as

(t+1) =  (t) + K(t) [y(t+1) - ym(t+1)]

K(t) = P(t) xm(t+1) / [1 + xm(t+1)T P(t) xm(t+1)] (5)
P(t+1) = 1/ [P(t) - {( P(t) xm(t+1) xm(t+1)T P(t)) / (1 + xm(t+1)T P(t) xm(t+1))}]

where (t) represents the estimated parameter vector,  is the forgetting factor, K(t) is the
gain matrix and P(t) is the covariance matrix.

Controller formulation

The N time steps ahead output prediction over a prediction horizon is given by
y p (t  N )   1 y(t+N-1)+...+nyy(t-ny+N)+1u(t+N-1)+...+nuu(t-nu+N)+err(t) (6)

where yp(t+N) represent the model predictions for N steps and err(t) is an estimate of the
modeling error which is assumed as constant for the entire prediction horizon. If the control
horizon is m, then the controller output, u after m time steps can be assumed to be constant.
Model predictive control of nonlinear processes 113

An internal model is used to eliminate the discrepancy between model and process outputs,
error(t), at each sampling instant
error(t) = y(t) - ym(t) (7)

where ym(t) is the one-step ahead model prediction at time (t-1). The estimate of the error is
then filtered to produce err(t) which minimizes the instability introduced by the modeling
error feedback. The filter error is given by
err(t) = (1-Kf) err(t-1) + Kf error(t) (8)
where Kf is the feedback filter gain which has to be tuned heuristically.
Back substitutions transform the prediction model equations into the following form
y p (t  N )  f N ,1 y(t )  .... 
f N ,ny y(t  ny  1)  f N ,ny  1 u(t  1) 
.....  f N ,ny  nu 1 u(t  nu  1)  (9)
gN ,1 u(t )  .....  gN ,m u(t  m  1)
 eN err (t )
The elements f, g and e are recursively calculated using the parameters  and  of
Eq. (3). The above equations can be written in a condensed form as
Y(t) = F X(t) + G U(t) + E err(t) (10)
where
Y(t) = [yp(t+1) . . . yp(t+N)]T (11)
X(t) = [y(t) y(t-1) . . . y(t-ny+1) u(t-1) . . . u(t-nu+1)]T (12)
U(t) = [u(t) . . . u(t+m-1)]T (13)

 f 11 f 12 ..... f 1( ny  nu  1) 
 
 f 11 f 12 ..... f 2( ny  nu 1) 
F : 
 
 : 
 
 f N 1 f N 12 ..... f N ( ny  nu 1) 
 g11 0 0 ... 0 
g g 0 ... 0 
 21 21
 . . . . . 
 
 . . . . . 
 . . . . . 
G 
 m1g g m2 g m3 ... g mm 
 . . . ... . 
 
 . . . ... . 
 
 . . . ... . 
 gN 1 gN 2 gN 3 ... gNm 
E = [e1 . . . eN]T
114 Model Predictive Control

In the above, Y(t) represents the model predictions over the prediction horizon, X(t) is a
vector of past plant and controller outputs and U(t) is a vector of future controller outputs. If
the coefficients of F, G and E are determined then the transformation can be completed. The
number of columns in F is determined by the ARX model structure used to represent the
system, where as the number of columns in G is determined by the length of the control
horizon. The number of rows is fixed by the length of the prediction horizon.

Consider a cost function of the form

N m
2
  y (t  i)  w(t  i)    u(t  i  1
2
J p
i 1 i 1
N m

   U (t ) U (t )
T
 Y (t )  W (t ) Y (t )  W (t )  T
(14)
i 1 i 1
where W(t) is a setpoint vector over the prediction horizon

W(t) = [ w(t+1) . . . . w(t+N)]T (15)

The minimization of the cost function, J gives optimal controller output sequence

U(t) = [GTG + I ]-1GT[W(t) - FX(t) - Eerr(t)] (16)

The vector U(t) generates control sequence over the entire control horizon. But, the first
component of U(t) is actually implemented and the whole procedure is repeated again at the
next sampling instant using latest measured information.
Linear model predictive control involving input-output models in classical, adaptive or
fuzzy forms is proved useful for controlling processes that exhibit even some degree of
nonlinear behavior (Eaton and Rawlings, 1992; Venkateswarlu and Gangiah, 1997 ;
Venkateswarlu and Naidu, 2001).

3.2 Case study: linear model predictive control of a reactive distillation column
In this study, a multistep linear model predictive control (LMPC) strategy based on
autoregressive moving average (ARX) model structure is presented for the control of a
reactive distillation column. Although MPC has been proved useful for a variety of chemical
and biochemical processes (Garcia et al., 1989 ; Eaton and Rawlings, 1992), its application to
a complex dynamic system like reactive distillation is more interesting.

The process and the model

Ethyl acetate is produced through an esterfication reaction between acetic acid and ethyl
alcohol
H
CH 3COOH  C2 H 5OH  H 2O  CH 3COOC2 H 5 (17)

The achievable conversion in this reversible reaction is limited by the equilibrium

conversion. This quaternary system is highly non-ideal and forms binary and ternary
Model predictive control of nonlinear processes 115

azeotropes, which introduce complexity to the separation by conventional distillation.

Reactive distillation can provide a means of breaking the azeotropes by altering or
eliminating the conditions for azeotrope formation. Thus reactive distillation becomes
attractive alternative for the production of ethyl acetate.
The rate equation of this reversible reaction in the presence of a homogeneous acid catalyst
is given by (Alejski and Duprat, 1996)

k1
r1  k1C1C 2  C 3C 4
Kc
k1  (4.195C k  0.08815)exp(  6500.1 ) (18)
T
Kc  7.558  0.012T

Vora and Daoutidis (2001) have presented a two feed column configuration for ethyl acetate
reactive distillation and found that by feeding the two reactants, ethanol and acetic acid, on
different trays counter currently allows to enhance the forward reaction on trays and results
in higher conversion and purity over the conventional column configuration of feeding the
reactants on a single tray. All plates in the column are considered to be reactive. The column
consists of 13 stages including the reboiler and the condenser. The less volatile acetic acid
enters the 3 rd tray and the more volatile ethanol enters the 10 th tray. The steady state
operating conditions of the column are shown in Table 1.

Acetic acid feed flow rate, FAc 6.9 mol/s

Ethanol flow rate, FEth 6.865 mol/s
Reflux flow rate, Lo 13.51 mol/s
Distillate flow rate, D 6.68 mol/s
Bottoms flow rate, B 7.085 mol/s
Reboiler heat duty, Qr 5.88 x 105 J/mol
Boiling points, K
o 391.05, 351.45, 373.15, 350.25
(Acetic acid, ethanol, water, ethyl acetate)
Distillate composition 0.0842, 0.1349, 0.0982, 0.6827
(Acetic acid, ethanol, water, ethyl acetate)
Bottoms composition 0.1650, 0.1575, 0.5470, 0.1306
(Acetic acid, ethanol, water, ethyl acetate)

Table 1. Design conditions for ethyl acetate reactive distillation column

The dynamic model representing the process operation involves mass and component
balance equations with reaction terms, along with energy equations supported by vapor-
liquid equilibrium and physical properties (Alejski & Duprat, 1996). The assumptions made
in the formulation of the model include adiabatic column operation, negligible heat of
reaction, negligible vapor holdup, liquid phase reaction, physical equilibrium in streams
leaving each stage, negligible down comer dynamics and negligible weeping of liquid
through the openings on the tray surface. The equations representing the process are given
as follows.
116 Model Predictive Control

Total mass balance

Total condenser:
dM 1
 V2 ( L1  D)  R1 (19)
dt
Plate j:
dM j
 FV j  V j 1  FL j  L j 1  V j  L j  R j (20)
dt
Reboiler :
dM n
 Ln 1  Vn  Ln  Rn (21)
dt
Component mass balance
Total condenser :
d ( M 1 xi ,1 )
 V2 yi , 2  ( L1  D) xi ,1  Ri ,1 (22)
dt
Plate j:
d(Mj xi , j )
 FVj yfi, j Vj1yi , j1  FLj xfi , j  Lj1xi, j1 Vj yi , j  Lj xi, j  Ri , j (23)
dt
Reboiler:
d( M n x i , n )
 Ln  1 xi ,n 1  Vn yi ,n  Ln xi ,n  Ri ,n (24)
dt

Total energy balance

Total condenser :
dE1
 V2 H 2  ( L1  D)h1  Q1 (25)
dt
Plate j:
dE j
 FV j Hf j  V j 1 H j 1  FL j hf j  L j 1 h j 1  V j H j  L j h j  Q j (26)
dt
Reboiler:
dE n
 Ln 1 hn 1  Vn H n  Ln hn  Qn (27)
dt
Level of liquid on the tray
M n MWav
Lliq  (28)
Atray  av
Flow of liquid over the weir
If ( Lliq<hweir ) then Ln = 0 (29)
Model predictive control of nonlinear processes 117

else
Lweir  av liq 3
Ln  1.84 ( L  hweir ) 2 (30)
MWav
Mole fraction normalization

NC NC

 xi  y i  1
i 1 i 1
(31)

VLE calculations

For the column operation under moderate pressures, the VLE equation assumes the ideal
gas model for the vapor phase, thus making the vapor phase activity coefficient equal to
unity. The VLE relation is given by

yi P = xi i Pisat (i = 1,2,….,NC) (32)

The liquid phase activity coefficients are calculated using UNIFAC method (Smith et al.,
1996).

Enthalpies Calculation
The relations for the liquid enthalpy h, the vapor enthalpy H and the liquid density  are:

h  h(P,T , x)
H  H (P,T , y) (33)

 liq
  liq
(P,T , x)

Control scheme

The design and implementation of the control strategy is studied for the single input-single
output (SISO) control of the ethyl acetate reactive distillation column with its double feed
configuration. The objective is to control the desired product purity in the distillate stream
inspite disturbances in column operation. This becomes the main control loop. Since reboiler
and condenser holdups act as pure integrators, they also need to be controlled. These
become the auxiliary control loops. Reflux flow rate is used as a manipulated variable to
control the purity of the ethyl acetate. Distillate flow rate is used as a manipulated variable
to control the condenser holdup, while bottom flow rate is used to control the reboiler
holdup. In this work, it is proposed to apply a multistep model predictive controller for the
main loop and conventional PI controllers for the auxiliary control loops. This control
scheme is shown in the Figure 3.
118 Model Predictive Control

Fig. 3. Control structure of two feed ethyl acetate reactive distillation column.

Analysis of Results
The performance of the multistep linear model predictive controller (LMPC) is evaluated
through simulation. The product composition measurements are obtained by solving the
model equations using Euler’s integration with sampling time of 0.01 s. The input and
output orders of the predictive model are considered as nu = 2 and ny = 2. The diagonal
elements of the initial covariance matrix, P(0) in the RLS algorithm are selected as 10.0, 1.0,
0.01, 0.01, respectively. The forgetting factor,  used in recursive least squares is chosen as
5.0. The feedback controller gain Kf is assigned as 0.65. The tuning parameter  in the
control law is set as 0.115 x 10-6. The PI controller parameters of ethyl acetate composition
are evaluated by using the continuous cycling method of Ziegler and Nichols. The tuned
 I = 1.61 x 104 s. The PI controller parameters used for
controller settings are kc = 11.15 and
reflux drum and reboiler holdups are kc = - 0.001 and  I = 5.5 h, and kc = - 0.001 and
 I = 5.5 h, respectively (Vora and Daoutidis, 2001).
The LMPC is implemented by adaptively updating the prediction model using recursive
least squares. On evaluating the effect of different prediction and control horizons, it is
observed that the LMPC with a prediction horizon of around 5 and a control horizon of 2
has shown reasonably better control performance. The LMPC is also referred here as MPC.
Figure 4 shows the results of MPC and PI controller when they are applied for tracking
series of step changes in ethyl acetate composition. The regulatory control performance of
MPC and PI controller for 20% decrease in feed rate of acetic acid is shown in Figure 5. The
results thus show the effectiveness of the multistep linear model predictive control strategy
for the control of highly nonlinear reactive distillation column.
Model predictive control of nonlinear processes 119

Fig. 4. Performance of MPC and PI controller for tracking series of step changes in distillate
composition.

Fig.5. Output and input profiles for MPC and PI controller for 20% decrease in the feed rate
of acetic acid.
120 Model Predictive Control

4. Generalized predictive control

The generalized predictive control (GPC) is a general purpose multi-step predictive control
algorithm (Clarke et al., 1987) for stable control of processes with variable parameters,
variable dead time and a model order which changes instantaneously. GPC adopts an
integrator as a natural consequence of its assumption about the basic plant model. Although
GPC is capable of controlling such systems, the control performance of GPC needs to be
ascertained if the process constraints are to be encountered in nonlinear processes. Camacho
(1993) proposed a constrained generalized predictive controller (CGPC) for linear systems
with constrained input and output signals. By this strategy, the optimum values of the
future control signals are obtained by transforming the quadratic optimization problem into
a linear complementarity problem. Camacho demonstrated the results of the CGPC strategy
by carrying out a simulation study on a linear system with pure delay. Clarke et al. (1987)
have applied the GPC to open-loop stable unconstrained linear systems. Camacho applied
the CGPC to constrained open-loop stable linear system. However, most of the real
processes are nonlinear and some processes change behavior over a period of time.
Exploring the application of GPC to nonlinear process control will be more interesting.
In this study, a constrained generalized predictive control (CGPC) strategy is presented and
applied for the control of highly nonlinear and open-loop unstable processes with multiple
steady states. Model parameters are updated at each sampling time by an adaptive
mechanism.

4.1 GPC design

A nonlinear plant generally admits a local-linearized model when considering regulation
about a particular operating point. A single-input single-output (SISO) plant on linearization
can be described by a Controlled Autoregressive Integrated Moving Average (CARIMA)
model of the form
A(q-1)y(t) = B(q-1)q-d u(t) + C (q-1)e(t )/ (34)

where A, B and C are polynomials in the backward shift operator q-1. The y(t) is the
measured plant output, u(t) is the controller output, e(t) is the zero mean random Gaussian
noise, d is the delay time of the system and  is the differencing operator 1-q-1.
The control law of GPC is based on the minimization of a multi-step quadratic cost function
defined in terms of the sum of squares of the errors between predicted and desired output
trajectories with an additional term weighted by projected control increments as given by

 N 2 N3 

J(N1 , N 2 , N 3 )  E  
 j  N 1
[ y(t  j|t )  w(t  j )]2  
j 1
[u(t  j  1)]2 

(35)

where E{.} is the expectation operator, y(t + j| t ) is a sequence of predicted outputs, w(t + j)
is a sequence of future setpoints, u(t + j -1) is a sequence of predicted control increments
and  is the control weighting factor. The N1 , N2 and N3 are the minimum costing horizon,
the maximum costing horizon and the control horizon, respectively. The values of N1 , N2
and N3 of Eq. (35) can be defined by N1 = d + 1, N2 = d + N, and N3 = N, respectively.
Model predictive control of nonlinear processes 121

Predicting the output response over a finite horizon beyond the dead-time of the process
enables the controller to compensate for constant or variable time delays. The recursion of
the Diophantine equation is a computationally efficient approach for modifying the
predicted output trajectory. An optimum j-step a head prediction output is given by

y(t + j| t) = Gj (q-1 ) u(t + j - d - 1) + Fj (q-1 )y(t) (36)

where Gj (q-1 ) = Ej (q-1 )B(q-1), and Ej and Fj are polynomials obtained recursively solving the
Diophantine equation,

1  E j (q 1 ) A  q  j F j (q 1 ) (37)

The j-step ahead optimal predictions of y for j = 1, . . . , N2 can be written in condensed form

Y =Gu + f (38)

where f contains predictions based on present and past outputs up to time t and past inputs
and referred to free response of the system, i.e., f = [f1, f2, ….., fN]. The vector u corresponds
to the present and future increments of the control signal, i.e., u = [u(t), u(t+1), …….,
u(t+N-1)]T. Eq. (35) can be written as

J  Gu  f  w Gu  f  w   u T u
T
(39)

The minimization of J gives unconstrained solution to the projected control vector

u  (G T G  I ) 1 G T ( w  f ) (40)

The first component of the vector u is considered as the current control increment u(t),
which is applied to the process and the calculations are repeated at the next sampling
instant. The schematic of GPC control law is shown in Figure 6, where K is the first row of
the matrix (GT G   I )1 GT .

w u(t) y(t)
+ K Process
-
f

Free response
of system

Fig. 6. The GPC control law

122 Model Predictive Control

4.2 Constrained GPC design

In practice, all processes are subject to constraints. Control valves are limited by fully closed
and fully open positions and maximum slew rates. Constructive and safety reasons as well
as sensor ranges cause limits in process variables. Moreover, the operating points of plants
are determined in order to satisfy economic goals and usually lie at the intersection of
certain constraints. Thus, the constraints acting on a process can be manipulated variable
limits (umin, umax ), slew rate limits of the actuator (dumin, dumax), and limits on the output
signal (ymin, ymax ) as given by

umin  u (t )  umax
dumin  u (t )  u (t  1)  dumax (41)

ymin  y (t )  ymax

These constraints can be expressed as

lumin  Tu  u (t  1)l  lumax

ldumin  u  ldumax (42)

lymin  Gu  f  lymax

where l is an N vector containing ones, and T is an N x N lower triangular matrix containing

ones. By defining a new vector x = u - ldumin, the constrained equations can be transformed
as

x0
(43)
Rx  c
with

 I NxN   l( dumax  dumin ) 

   
T  lu
 max  Tldu min  u( t  1)l 
R   T  ; c   lumin  Tldumin  u(t  1)l  (44)
   
G  lymax  f  Gldumin 
   
 G   lymin  f  Gldumin 

Eq. (39) can be expressed as

1 T
J  u Hu  bu  f o (45)
2
where
Model predictive control of nonlinear processes 123

H  2(G T G  I )
b  GT ( f  w)  ( f  w)T G
f o  ( f  w)T ( f  w)

The minimization of J with no constraints on the control signal gives

u   H 1b (46)
Eq. (45) in terms of the newly defined vector x becomes

1 T
J  x Hx  ax  f1 (47)
2
where
a  b  umin l T H
1 2 T
f1  f o  dumin l Hl  bldumin
2
The solution of the problem can be obtained by minimization of Eq. (47) subject to the
constraints of Eq. (43). By using the Lagrangian multiplier vectors v1 and v for the
constraints, x ≥ 0 and Rx ≤ c, respectively, and introducing the slack variable vector v2, the
Kuhn-Tucker conditions can be expressed as

Rx  v 2  c
 Hx  R T v  v1  a
x T v1  0 (48)
T
v v2  0
x, v, v1 , v 2  0

Camacho (1993) has proposed the solution of this problem with the help of Lemke’s
algorithm (Bazaraa and Shetty, 1979) by expressing the Kuhn-Tucker conditions as a linear
complementarity problem starting with the following tableau

v2 x v v1 z0
1 T 1
v2 I mxn OmxN  RH R RH 1 v2 min (49)
1 T 1
x ONxm I NxN H R H 1 xmin

Here, z0 is the artificial variable which will be driven to zero iteratively.

In this study, the above stated constrained generalized predictive linear control of Camacho
(1993) is extended to open-loop unstable constrained control of nonlinear processes. In this
124 Model Predictive Control

strategy, process nonlinearities are accounted through adaptation of model parameters

while taking care of input and output constraints acting on the process. The following
recursive least squares formula (Hsia, 1977) is used for on-line estimation of parameters and
the covariance matrix after each new sample:

 (k  1)   (k )   (k  1) P (k ) (k  1)y (k  1)   T (k  1) (k )
P (k  1) 
1

P(k )   (k  1) P(k ) (k  1) T
(k  1) P (k ) 
(50)
 (k  1)  1
1   T
(k  1) P (k ) (k  1) 
where θ is the parameter vector, γ is the intermediate estimation variable, P is the covariance
matrix, v is the vector of input-output variables, y is the output variable, and 0 <  < 1 is the
forgetting factor. The initial covariance matrix and exponential forgetting factor are selected
based on various trials so as to provide reasonable convergence in parameter estimates.

The CGPC strategy of nonlinear processes is described in the following steps:

1. Specify the controller design parameters N1, N2, N3 and also the initial parameter
estimates and covariance matrix for recursive identification of model parameters.
2. Update the model parameters using recursive least squares method.
3. Initialize the polynomials E1 and F1 of Diophantine identity, Eq. (37), using the estimated
parameters. Further initialize G1 as E1 B.
4. Compute the polynomials Ej , Fj and Gj over the prediction horizon and control horizon
using the recursion of Diophantine.
5. Compute matrices H, R, and G, and vectors f and c using the polynomials determined in
step 4.
6. Compute the unconstrained solution xmin = - H-1 a.
7. Compute v2min = c - Rxmin . If xmin and v2min are nonnegative, then go to step 10.
8. Start Lemke’s algorithm with x and v2 in the basis with the tableau, Eq. (49).
9. If x1 is not in the first column of the tableau, make it zero; otherwise, assign it the
corresponding value.
10. Compute u(t) = x1 + dumin + u(t - 1).
11. Implement the control action, then shift to the next sampling instant and go to step 2.

4.3 Case study: constrained generalized predictive control (CGPC) of open-loop

unstable CSTR
The design and implementation of the CGPC strategy is studied by applying it for the
control of a nonlinear open-loop unstable chemical reactor (Venkateswarlu and Gangiah,
1997).

Reactor
A continuous stirred tank reactor (CSTR) in which a first order exothermic irreversible
reaction occurs is considered as an example of an unstable nonlinear process. The dynamic
equations describing the process can be written as
Model predictive control of nonlinear processes 125

dC A
V  F (C Af  C A )  Vk0 exp  E / R gTr  C A (51)
dt 
dTr
V C p  C p F(T f  Tr )  V ( H )k0 exp  E / R gTr  C A  UAh (Tr  Tc ) (52)
dt 

where CA and Tr are reactant concentration and temperature, respectively. The coolant
temperature Tc is assumed to be the manipulated variable. Following the analysis of Uppal
et al. (1974), the model is made dimensionless by introducing the parameters as

E (   H )C Afo k e  V UAh
 , Bh  , Da  o ,   (53)
R g T fo  C pT fo Fo Fo  C p

where Fo, CAfo and Tfo are the nominal characteristic values of volumetric flow rate, feed
composition and feed temperature, respectively. The corresponding dimensionless variables
are defined by

t Fo C Afo  C A Tr  T fo Tc  Tco

t , x1  , x2  , u  (54)
V C Afo T fo T fo

where Tco is some reference value for the coolant temperature.

The modeling equations can be written in dimensionless form (Calvet and Arkun, 1988;
Hernandez and Arkun, 1992) as

 x2 
x1   x1  Da (1  x1 ) exp 
 1  x 2 /  
 x2 
x2   x2  Bh Da(1  x1 ) exp    (u  x2 ) (55)
 1  x2 /  
y = x1

where x1 and x2 are the dimensionless reactant concentration and temperature, respectively.
The input u is the cooling jacket temperature, Da is the Damkohler number,  is the
dimensionless activation energy, Bh is the heat of reaction and  is the heat transfer coeffi-
cient. If the physical parameters chosen are Da = 0.072,  = 20.0, Bh = 8.0, and  = 0.3, then the
system can exhibit up to three steady states, one of which is unstable as shown in Figure 7.
Here the task is to control the reactor at and around the unstable operating point. The
cooling water temperature is the input u, which is the manipulated variable to control the
reactant concentration, x1.
126 Model Predictive Control

Fig. 7. Steady state output vs. steady state input for CSTR system.

Analysis of Results
Simulation studies are carried out in order to evaluate the performance of the Constrained
Generalized Predictive Control (CGPC) strategy. The results of unconstrained Generalized
Predictive Control (GPC) are also presented as a reference. The CGPC strategy considers an
adaptation mechanism for model parameters.

Na 2
Nb 2
N1 2
N2 7
N3 6
 0.2
umin -1.0
umax 1.0
dumin -0.5
dumax 0.5
ymin 0.1
ymax 0.5
Forgetting factor 0.95
Initial covariance matrix 1.0x109
Sample time 0.5

Table 2. Constraints and parameters of CSTR system.

The controller and design parameters as well as the constraints employed for the CSTR
system are given in Table 2. The same controller and design parameters are used for both
the CGPC and GPC. Two set-point changes are introduced for the output concentration of
the system and the corresponding results of CGPC and GPC are analyzed. A step change is
Model predictive control of nonlinear processes 127

introduced in the output concentration of CSTR from a stable equilibrium point (x1 = 0.2, x2
= 1.33, u = 0.42) to an unstable operating point (x1 = 0.5, x2 = 3.303, u = - 0.2). The input and
output responses of both CGPC and GPC are shown in Figure 8. Another step change is
introduced for the set-point from a stable operating point (x1 = 0.144, x2 = 0.886, u = 0.0) to
an unstable operating point (x1 = 0.445, x2 = 2.75, u = 0.0). The input and output responses of
CGPC and GPC for this case are shown in Figure 9. The results show that for the specified
controller and design parameters, CGPC provides better performance over GPC.

Fig. 8. Cooling water temperature and concentration plots of CSTR for a step change in
concentration from 0.20 to 0.50.

Fig. 9. Cooling water temperature and concentration plots of CSTR for a step change in
concentration from 0.144 to 0.445.

The results illustrate the better performance of CGPC for SISO control of nonlinear systems
that exhibit multiple steady states and unstable behavior.

5. Nonlinear model predictive control

Linear MPC employs linear or linearized models to obtain the predictive response of the
controlled process. Linear MPC is proved useful for controlling processes that exhibit even
some degree of nonlinear behavior (Eaton and Rawlings, 1992; Venkateswarlu and Gangiah,
1997). However, the greater the mismatch between the actual process and the representative
128 Model Predictive Control

model, the degree of deterioration in the control performance increases. Thus the control of
a highly nonlinear process by MPC requires a suitable model that represents the salient
nonlinearities of the process. Basically, two different approaches are used to develop
nonlinear dynamic models. These approaches are developing a first principle model using
available process knowledge and developing an empirical model from input-output data.
The first principle modeling approach results models in the form of coupled nonlinear
ordinary differential equations and various model predictive controllers based on this
approach have been reported for nonlinear processes (Wright and Edgar, 1994 ; Ricker and
Lee, 1995). The first principle models will be larger in size for high dimensional systems thus
limiting their usage for MPC design. On the other hand, the input-output modeling
approach can be conveniently used to identify nonlinear empirical models from plant data,
and there has been a growing interest in the development of different types of MPCs based
on this approach (Hernandez and Arkun, 1994; Venkateswarlu and Venkat Rao, 2005). The
other important aspect in model predictive control of highly nonlinear systems is the
optimization algorithm. Efficient optimization algorithms exist for convex optimization
problems. However, the optimization problem often becomes nonconvex in the presence of
nonlinear characteristics/constraints and is usually more complex than convex
optimization. Thus, the practical usefulness of nonlinear predictive control is hampered by
the unavailability of suitable optimization tools (Camacho and Bordons, 1995). Sequential
quadratic programming (SQP) is widely used classical optimization algorithm to solve
nonlinear optimization problems. However, for the solution of large problems, it has been
reported that gradient based methods like SQP requires more computational efforts (Ahn et
al., 1999). More over, classical optimization methods are more sensitive to the initialization
of the algorithm and usually leads to unacceptable solutions due to convergence to local
optima. Consequently, efficient optimization techniques are being used to achieve the
improved performance of NMPC.
This work presents a NMPC based on stochastic optimization technique. Stochastic
approach based genetic algorithms (GA) and simulated annealing (SA) are potential
optimization tools because of their ability to handle constrained, nonlinear and nonconvex
optimization problems. These methods have the capacity to escape local optima and find
solutions in the vicinity of the global optimum. They have the ability to use the values from
the model in a black box optimization approach with out requiring the derivatives. Various
studies have been reported to demonstrate the ability of these methods in providing
efficient optimization solutions (Hanke and Li, 2000 ; Shopova and Vaklieva-Bancheva,
2006).

5.1 NMPC design

In NMPC design, the identified input-output nonlinear process model is explicitly used to
predict the process output at future time instants over a specified prediction horizon. A
sequence of future control actions over a specified control horizon is calculated using a
stochastic optimizer which minimizes the objective function under given operating
constraints. In this receding horizon approach, only the first control action in the sequence is
implemented. The horizons are moved towards the future. The structure of the stochastic
optimizer based NMPC is shown in Figure 10.
Model predictive control of nonlinear processes 129

Fig.10. Structure of stochastic optimization based NMPC.

Simulated Annealing
Simulated annealing (SA) is analogous to the process of atomic rearrangement of a
substance into a highly ordered crystalline structure by way of slowly cooling-annealing the
substance through successive stages. This method is found to be a potential tool to solve a
variety of optimization problems (Kirkpatrick et al., 1983 ; Dolan et al., 1989). Crystalline
structure with a high degree of atomic order is the purest form of the substance, indicating
the minimum energy state. The principle of SA mimics the annealing process of slow
cooling of molten metal to achieve the minimum function value. The cooling phenomena is
simulated by controlling a temperature like parameter introduced with the concept of the
Bolzmann probability distribution, which determines the energy distribution probability, P
of the system at the temperature, T according to the equation:

P( E )  exp( E / k BTA ) (56)

where kB is the Bolzmann constant. The Bolzmann distribution concentrates around the state
with the lowest energy. For T  0, P(E)  0 and only the state with the lowest energy can
have a probability greater than zero. However, cooling the system too fast could result in a
higher state of energy and may lead to frozen the system to a metastable state.
130 Model Predictive Control

The SA is a point by point method based on Monte Carlo approach. The algorithm begins at
an initial random point called u and a high temperature T, and the function value at this
point is evaluated as E(k). A second point is created in the vicinity of the initial point u and
the function value corresponding to this point is obtained as E(k+1). The difference in
function values at these points E is obtained as

E = E(k+1) – E(k) (57)

If E  0, the second point is accepted, otherwise the point is accepted probabilistically,

governed by the temperature dependent Bolzmann probability factor

Pr  exp(E / k BTA ) (58)

The annealing temperature, TA is a parameter in the optimization algorithm and is set by a

predefined annealing schedule which starts at a relatively high temperature and steps
slowly downward at a prescribed rate in accordance with the equation

TAm 1  TAm , 0  1 (59)

As the temperature decreases, the probability of the acceptance of the point u will be
decreased according to Eq. (58). The parameter  is set such that at the point of convergence,
the temperature TA reaches a small value. The procedure is iteratively repeated at each
temperature with the generation of new points and the search is terminated when the
convergence criterion set for the objective is met.

Nonlinear modeling and model identification

Various model structures such as Volterra series models, Hammerstein and Wiener models,
bilinear models, state affine models and neural network models have been reported in
literature for identification of nonlinear systems. Haber and Unbehauen (1990) presented a
comprehensive review on these model structures. The model considered in this study for
identification of a nonlinear process has a polynomial ARMA structure of the form
ny nu ny  1

yˆ ( k )   0  
i 1
1,i
y( k  i )  
i 1
2 ,i
u( k  i )  
i 1
3, i y( k  i )u( k  i ) 
(60)
nu i 1

 i  1 j 1
4,i j u( k  i )u( k  j )  .............

or simply

yˆ (k )  f  y ( k  1),......, y ( k  n y ), u (k  1),........ , u (k  nu )  (61)

Here k refers the sampling time, y and u are the output and input variables, and ny and nu
refer the number of output and input lags, respectively. This type of polynomial model
structures have been used by various researchers for process control (Morningred et al.,
Model predictive control of nonlinear processes 131

1992 ; Hernandez and Arkun, 1993). The main advantage of this model is that it represents
the process nonlinearities in a structure with linear model parameters, which can be
estimated by using efficient parameter estimation methods such as recursive least squares.
Thus the model in (61) can be rearranged in a linear regression form as

yˆ (k )   T (k  1) (k  1)   (k ) (62)

where  is a parameter vector,  represents input-output process information and  is the

estimation error. The parameters in the model can be estimated by using recursive least
squares based on a priori process knowledge representing the process characteristics over a
wide range of operating conditions.

Predictive Model Formulation

The primary purpose of NMPC is to deal with complex dynamics over an extended horizon.
Thus, the model must predict the process dynamics over a prediction horizon enabling the
controller to incorporate future set point changes or disturbances. The polynomial input-
output model provides one step ahead prediction for process output. By feeding back the
model outputs and control inputs, the one step-a head predictive model can be recurrently
cascaded to itself to generate future predictions for process output. The N step predictions
can be obtained as follows

yˆ (k  1 k )  f  y (k ),......., y (k  1  n y ), u (k ),......., u (k  1  nu ) 

yˆ ( k  2 k )  f yˆ ( k  1 k ) ,......., y( k  2  ny ), u( k  1),......., u( k  M  nu )  (63)

 yˆ ( k  N  1 k ),......., yˆ ( k  N  ny k ), u( k  M  1),......., 
yˆ ( k  N k )  f  
 u( k  M  n )
 u 

where N is the prediction horizon and M is the control horizon.

Objective function
The optimal control input sequence in NMPC is computed by minimizing an objective
function based on a desired output trajectory over a prediction horizon:

N M

    u(k  i  1) 
2 T
Min J   w(k  i)  yˆ p (k  i)  u(k  i  1)
i 1 i 1 (64)
u(k), u(k  1),......., u(k  M  1)

subject to constraints:

ymin  yˆ p (k  i)  ymax i  1, ........,N 

umin  u(k  i)  umax i  0, ......., M  1
132 Model Predictive Control

umin  u(k  i)  umax i  0, ......., M 1

where yˆ p (k  i ) , i=1, …., N, are the future process outputs predicted over the prediction
horizon, wk+i , i=1, …., N, are the setpoints and u(k+i), i=0, .…., M-1, are the future control
signals. The  and  represent the output and input weightings, respectively. The umin and
umax are the minimum and maximum values of the manipulated inputs, and umin and umax
represent their corresponding changes, respectively. Computation of future control signals
involves the minimization of the objective function so as to bring and keep the process
output as close as possible to the given reference trajectory, even in the presence of load
disturbances. The control actions are computed at every sampling time by solving an
optimization problem while taking into consideration of constraints on the output and
inputs. The control signal, u is manipulated only with in the control horizon, and remains
constant afterwards, i.e., u(k+i) = u(k+M-1) for i = M, …., N-1. Only the first control move
of the optimized control sequence is implemented on the process and the output
measurements are obtained. At the next sampling instant, the prediction and control
horizons are moved ahead by one step, and the optimization problem is solved again using
the updated measurements from the process. The mismatch dk between the process y(k)
and the model yˆ (k ) is computed as
d k  b( y (k )  yˆ (k )) (65)

where b is a tunable parameter lying between 0 and 1. This mismatch is used to compensate
the model predictions in Eq. (62):

yˆ p (k  i)  yˆ (k  i )  d k (for all i  1 to N ) (66)

These predictions are incorporated in the objective function defined by Eq. (64) along with
the corresponding setpoint values.

NMPC based on stochastic optimization

NMPC design based on simulated annealing (SA) requires to specify the energy function
and random number selection for control input calculation. The control input is normalized
and constrained with in the specified limits. The random numbers used for the control
input, u equals the length of the control horizon, and these numbers are generated so that
they satisfy the constraints. A penalty function approach is considered to satisfy the
constraints on the input variables. In this approach, a penalty term corresponding to the
penalty violation is added to the objective function defined in Eq. (64). Thus the violation of
the constraints on the variables is accounted by defining a penalty function of the form
N
P     u ( k  i ) 2 (67)
i 1

where the penalty parameter,  is selected as a high value. The penalized objective function
is then given by
f(x) = J + P (68)
Model predictive control of nonlinear processes 133

where J is defined by Eq. (64). At any instant, the current control signal, uk and the
prediction output based on this control input, yˆ (k  i ) are used to compute the objective
function f(x) in Eq. (68) as the energy function, E(k+i). The E(k+i) and the previously
evaluated E(k) provides the E as

E(k) = E(k+i) – E(k) (69)

The comparison of the E with the random numbers generated between 0 and 1 determines
the probability of acceptance of u(k). If E  0, all u(k) are accepted. If E  0, u(k) are
accepted with a probability of exp(-E/TA). If nm be the number of variables, nk be the
number of function evaluations and nT be the number of temperature reductions, then the
total number of function evaluations required for every sampling condition are (nT x nk x nm).
Further details of NMPC based on stochastic optimization can be referred elsewhere
(Venkateswarlu and Damodar Reddy, 2008).

Implementation procedure
The implementation of NMPC based on SA proceeds with the following steps.

1. Set TA as a sufficiently high value and let nk be the number of function evaluations to be
performed at a particular TA. Specify the termination criterion, . Choose the initial
control vector, u and obtain the process output predictions using Eq. (63). Evaluate the
objective function, Eq. (68) as the energy function E(k).
2. Compute the incremental input vector uk stochastically and update the control vector as

u(k+i) = u(k) + u(k) (70)

Calculate the objective function, E(k+i) as the energy function based on this vector.
3. Accept u(k+i) unconditionally if the energy function satisfies the condition

E(k+i)  E(k) (71)

Otherwise, accept u(k+i) with the probability according to the Metropolis criterion
 E ( k  i )  E ( k )  
exp  '
r
 (72)
 T A 
'
where TA is the current annealing temperature and r represents random number. This step
proceeds until the specified function evaluations, nk are completed.
4. Carry out the temperature reduction in the outer loop according to the decrement
function
TA/  TA (73)
where  is temperature reduction factor. Terminate the algorithm if all the differences are
less than the prespecified .
5. Go to step 2 and repeat the procedure for every measurement condition based on the
updated control vector and its corresponding process output.
134 Model Predictive Control

5.2 Case study: nonlinear model predictive control of reactive distillation column
The performance of NMPC based on stochastic optimization is evaluated through
simulation by applying it to a ethyl acetate reactive distillation column.

Analysis of Results
The process, the column details, the mathematical model and the control scheme of ethyl
acetate reactive distillation column given in Section 3.2 is used for NMPC implementation.
In this operation, since the ethyl acetate produced is withdrawn as a product in the distillate
stream, controlling the purity of this main product is important in spite of disturbances in
the column operation. This becomes the main control loop for NMPC in which reflux flow
rate is used as a manipulated variable to control the purity of ethyl acetate. Since reboiler
and condenser holdups act as pure integrators, they also need to be controlled. These
become the auxiliary control loops and are controlled by conventional PI controllers in
which the distillate flow rate is considered as a manipulated variable to control the
condenser molar holdup and the bottom flow rate is used to control the reboiler molar
holdup. The tuning parameters used for both the PI controllers of reflux drum and reboiler
holdups are kc = - 0.001 and  I = 1.99 x 104 (Vora and Dauotidis, 2001). The SISO control
scheme for the column with the double feed configuration used in this study is shown in the
Fig. 3.
The input-output data to construct the nonlinear empirical model is obtained by solving the
model equations using Euler's integration with a step size of 2.0 s. A PI controller with a
series of step changes in the set point of ethyl acetate composition is used for data
generation. The input data (reflux flow) is normalized and used along with the outputs
(ethyl acetate composition) in model building. The reflux flow rate is constrained with in the
limits of 20 mol/s and 5 mol/s. A total number of 25000 data sets is considered to develop
the model. The model parameters are determined by using the well known recursive least
squares algorithm (Goodwin and Sin, 1984), the application of which has been shown
elsewhere (Venkateswarlu and Naidu, 2001). After evaluating model structure in Eq. (60) for
different orders of ny and nu , the model with the order ny=2 and nu=2 is found to be more
appropriate to design and implement the NMPC with stochastic optimization. The structure
of the model is in the form

yˆ k   0  1 yk 1   2uk 1   3 yk 1uk 1   4 yk  2uk  2   5uk 1uk  2 (74)

The parameters of this model are determined as θ0=-0.000774, θ1=1.000553, θ2=0.002943, θ3=-
0.003828, θ4=0.000766 and θ5=-0.000117. This identified model is then used to derive the
future predictions for the process output by cascading the model to it self as in Eq. (63).
These model predictions are added with the modeling error, d(k) defined by Eq. (65), which
is considered to be constant for the entire prediction horizon. The weightings  and  in the
objective function, Eq. (64) are set as 1.0 x 107 and 7.5 x 104, respectively. The penalty
parameter,  in Eq. (67) is assigned as 1.0 x 105. The cost function used in NMPC is the
penalized objective function, eq. (68), based on which the SA search is computed. The
incremental input, u in SA search is constrained with in the limits -0.0025 and 0.0025,
respectively. The actual input, u involved with the optimization scheme is a normalized
value and is constrained between 0 and 1. The objective function in Eq. (68) is evaluated as
the energy function at each instant. The initial temperature T is chosen as 500 and the
Model predictive control of nonlinear processes 135

number of iterations at each temperature is set as 250. The temperature reduction factor, 
in Eq. (73) is set as 0.5. The control input determined by the stochastic optimizer is
denormalized and implemented on the process. A sample time of 2 s is considered for the
implementation of the controller.
The performance of NMPC based on SA is evaluated by applying it for the servo and
regulatory control of ethyl acetate reactive distillation column. On evaluating the results
with different prediction and control horizons, the NMPC with a prediction horizon of
around 10 and a control horizon of around 1 to 3 is observed to provide better performance.
The results of NMPC are also compared with those of LMPC presented in Section 3 and a PI
controller. The tuning parameters of the PI controller are set as kC = 10.0 and I = 1.99 x 104
(Vora and Dauotidis, 2001). The servo and regulatory results of NMPC along with the
results of LMPC and PI controller are shown in Figures 11-14. Figure 11 compares the input
and output profiles of NMPC with LMPC and PI controller for step change in ethyl acetate
composition from 0.6827 to 0.75. The responses in Figure 12 represent 20% step decrease in
ethanol feed flow rate, and the responses in Figure 13 correspond to 20% step increase in
reboiler heat load. These responses show the better performance of NMPC over LMPC and
PI controller. Figure 14 compares the performance of NMPC and LMPC in tracking multiple
step changes in setpoint of the controlled variable. The results thus show the stability and
robustness of NMPC towards load disturbances and setpoint changes.

Fig.11. Output and input profiles for step increase in ethyl acetate composition setpoint.
136 Model Predictive Control

Fig.12. Output and input profiles for step decrease in ethanol feed flow rate.

Fig.13. Output and input profiles for step increase in reboiler heat load.
Model predictive control of nonlinear processes 137

Fig. 14. Output responses for multiple setpoint changes in ethyl acetate composition

6. Conclusions
Model predictive control (MPC) is known to be a powerful control strategy for a variety of
processes. In this study, the capabilities of linear and nonlinear model predictive controllers
are explored by designing and applying them to different nonlinear processes. A linear
model predictive controller (LMPC) is presented for the control of an ethyl acetate reactive
distillation. A generalized predictive control (GPC) and a constrained generalized predictive
control (CGPC) are presented for the control of an unstable chemical reactor. Further, a
nonlinear model predictive controller (NMPC) based on simulated annealing is presented
for the control of a highly complex nonlinear ethyl acetate reactive distillation column. The
results of these controllers are evaluated under different disturbance conditions for their
servo and regulatory performance and compared with the conventional controllers. From
these results, it is observed that though linear model predictive controllers offer better
control performance for nonlinear processes over conventional controllers, the nonlinear
model predictive controller provides effective control performance for highly complex
nonlinear processes.

Nomenclature
ARX autoregressive moving average
Ah heat transfer area, m2
Atray tray area, m2
B bottom flow rate, mol s-1
Bh dimensionless heat of reaction
C concentration, mol m-3
CA reactant concentration, mol m-3
CAf feed concentration, mol m-3
Ck catalyst concentration, % vol
Cp specific heat capacity, J kg-1 K-1
D distillate flow rate, mol s-1
Da Damkohler number
dumin lower limit of slew rate
138 Model Predictive Control

dumax upper limit of slew rate

E total enthalpy of liquid on plate, kJ
FL liquid feed flow rate on plate, mol s-1
FV vapor feed on plate, mol s-1
FAc acetic acid feed flow rate, mol s-1
FEth ethanol feed flow rate, mol s-1
Fo volumetric feed rate, m3 s-1
H molar enthalpy of vapor stream, kJ mol-1
h molar enthalpy of liquid stream, kJ mol-1
k1 reaction rate constant, m3 mol-1 s-1
hweir weir height, m
KC constant of reaction equilibrium
L molar liquid flow rate, mol s-1
Lweir weir length, m
Lliquid liquid level on tray, m
M molar holdup on plate, m
MWav average molecular weight, g mol-1
N1 minimum costing horizon
N2 maximum costing horizon
N3 control horizon
P pressure on plate, pascal
Q heat exchange, kJ
R number of moles reacted, mol s-1
Rg gas constant, J mol-1 K-1
RLS recursive least squares
r rate of reaction, mol s-1 m-3
av average density, g m-3
T temperature, K
Tc coolant temperature, K
Tf feed temperature, K
Tr reactor temperature, K
U heat transfer coefficient, J m-2 s-1 K-1
u controller output
umin lower limit of manipulated variable
umax upper limit of manipulated variable
VLE vapor-liquid equilibrium
V molar vapor flow rate, mol s-1
x mole fraction in liquid phase
x1 dimensionless reactant concentration
x2 dimensionless reactant temperature
y mole fraction in vapor phase
ymin lower limit of output variable
ymax upper limit of output variable
av average density, g m-3
Model predictive control of nonlinear processes 139

7. References
Ahn, S.M., Park, M.J., Rhee, H.K. Extended Kalman filter based nonlinear model predictive
control of a continuous polymerization reactor. Industrial &. Engineering Chemistry
Research, 38: 3942-3949, 1999.
Alejski, K., Duprat, F. Dynamic simulation of the multicomponent reactive distillation.
Chemical Engineering Science, 51: 4237-4252, 1996.
Bazaraa, M.S., Shetty, C.M. Nonlinear Programming, 437-443 (John Wiley & Sons, New York),
1979.
Calvet, J P., Arkun, Y. Feedforward and feedback linearization of nonlinear systems and its
implementation using internal model control (IMC). Industrial &. Engineering
Chemistry Research, 27: 1822-1831, 1988.
Camacho, E. F. Constrained generalized predictive control. IEEE Trans Aut Contr, 38: 327-
332, 1993.
Camacho, E. F., Bordons, C. Model Predictive Control in the Process Industry; Springer Verlag:
Berlin, Germany, 1995.
Clarke, D.W., Mohtadi, C and Tuffs, P.S. Generalized predictive control – Part I. The basic
algorithm. Automatica, 23: 137-148, 1987.
Cutler, C.R. and Ramker, B.L. Dynamic matrix control – a computer control algorithm,
Proceedings Joint Automatic Control Conference, Sanfrancisco, CA.,1980.
Dolan, W.B., Cummings, P.T., Le Van, M.D. Process optimization via simulated annealing:
application to network design. AIChE Journal. 35: 725-736, 1989.
Garcia, C.E., Prett, D.M., and Morari, M. Model predictive control: Theory and Practice - A
survey. Automatica, 25: 335-348, 1989.
Eaton, J.W., Rawlings, J.B. Model predictive control of chemical processes. Chemical
Engineering Science, 47: 705-720, 1992.
Goodwin, G.C., Sin, K.S. Adaptive Filtering Prediction and Control (Printice Hall,
Englewood Cliffs, New Jersey), 1984.
Haber, R., Unbehauen, H. Structure identification of nonlinear dynamical systems -a
survey on input/output approaches. Automatica, 26: 651-677, 1990.
Hanke, M., Li, P. Simulated annealing for the optimization of batch distillation process.
Computers and Chemical Engineering, 24: 1-8, 2000.
Hernandez, E., Arkun, Y., Study of the control relevant properties of backpropagation
neural network models of nonlinear dynamical systems. Computers & Chemical
Engineering, 16: 227-240, 1992.
Hernandez, E., Arkun, Y. Control of nonlinear systems using polynomial ARMA models.
AIChE Journal, 39: 446-460, 1993.
Hernandez, E., Arkun, Y. On the global solution of nonlinear model predictive control
algorithms that use polynomial models. Computers and Chemical Engineering, 18:
533-536, 1994.
Hsia, T.C. System Identification: Least Square Methods (Lexington Books, Lexington, MA),
1977.
Kirkpatrick, S., Gelatt Jr, C.D., Veccchi, M.P. Optimization by simulated annealing. Scienc,
220: 671-680, 1983.
Morningred, J.D., Paden, B.E., Seborg D.E., Mellichamp, D.A., An adaptive nonlinear
predictive controller. Chemical Engineering Science, 47: 755-762, 1992.
140 Model Predictive Control

Qin, J., Badgwell, T. An overview of industrial model predictive control technology; In: V th
International Conference on Chemical Process Control (Kantor, J.C., Garcia, C.E.,
Carnhan, B., Eds.): AIChE Symposium Series, 93: 232-256, 1997.
Richalet, J., Rault, A., Testud, J. L. and Papon, J. Model predictive heuristic control:
Application to industrial processes. Automatica, 14: 413-428, 1978.
Ricker, N.L., Lee, J.H. Nonlinear model predictive control of the Tennessee Eastman
challenging process. Computers and Chemical Engineering, 19: 961-981, 1995.
Smith, J.M., Van Ness, H.C. Abbot, M.M., A Text Book on Introduction to Chemical
Engineering Thermodynamics, 5 th Ed., Mc-Graw Gill International. 1996.
Shopova, E.G., Vaklieva-Bancheva, N.G. BASIC-A genetic algorithm for engineering
problems solution. Computers and Chemical Engineering, 30: 1293-1309, 2006.
Venkateswarlu, Ch., Gangiah, K. Constrained generalized predictive control of unstable
nonlinear processes. Transactions of Insitution of Chemical Engineers, 75: 371-376,
1997.
Venkateswarlu, Ch., Naidu, K.V.S. Adaptive fuzzy model predictive control of an
exothermic batch chemical reactor. Chemical Engineering Communications, 186: 1-23,
2001.
Venkateswarlu, Ch., Venkat Rao, K. Dynamic recurrent radial basis function network model
predictive control of unstable nonlinear processes. Chemical Engineering Science, 60:
6718-6732, 2005.
Venkateswarlu, Ch., Damodar Reddy, D. Nonlinear model predictive control of reactive
distillation based on stochastic optimization. Industrial Engineering & Chemistry
Research, 47: 6949-6960, 2008.
Vora, N., Daoutidis, P. Dynamics and control of ethyl acetate reactive distillation column.
Industrial &. Engineering Chemistry Research, 40: 833-849, 2001.
Uppal, A., Ray, W.H., Poore, A. B. On the dynamic behavior of continuous stirred tank
reactors. Chemical Engineering Science, 29: 967- 985,1974.
Wright, G. T., Edgar, T. F. Nonlinear model predictive control of a fixed-bed water-gas shift
reactor: an experimental study. Computers and Chemical Engineering, 18: 83-102,
1994.
Approximate Model Predictive Control for Nonlinear Multivariable Systems 141

0
6

Approximate Model Predictive Control for

Nonlinear Multivariable Systems
Jonas Witt and Herbert Werner
Hamburg University of Technology
Germany

1. Introduction
The control of multi-input multi-output (MIMO) systems is a common problem in practical
control scenarios. However in the last two decades, of the advanced control schemes, only
linear model predictive control (MPC) was widely used in industrial process control (Ma-
ciejowski, 2002). The fundamental common idea behind all MPC techniques is to rely on
predictions of a plant model to compute the optimal future control sequence by minimiza-
tion of an objective function. In the predictive control domain, Generalized Predictive Control
(GPC) and its derivatives have received special attention. Particularly the ability of GPC to
be applied to unstable or time-delayed MIMO systems in a straight forward manner and the
low computational demands for static models make it interesting for many different kinds of
tasks. However, this method is limited to linear models.

Travel-Axis Pitch-Axis

Elevation-Axis

Engines
Counterweight
Fig. 1. Quanser 3-DOF Helicopter

If nonlinear dynamics are present in the plant a linear model might not yield sufficient pre-
dictions for MPC techniques to function adequately. A related technique that can be applied
to nonlinear plants is Approximate (Model) Predictive Control (APC). It uses an instantaneous
linearization of a nonlinear model based on a neural network in each sampling instant. It is
142 Model Predictive Control

similar to GPC in most aspects except that the instantaneous linearization of the neural net-
work yields an adaptive linear model. Previously this technique has already successfully been
applied to a pneumatic servomechanism (Nørgaard et al., 2000) and gas turbine engines (Mu
& Rees, 2004), however both only in simulation.
The main challenges in this work were the nonlinear, unstable and comparably fast dynamics
of the 3-DOF helicopter by Quanser Inc. (2005) (see figure 1). APC as proposed by Nørgaard
et al. (2000) had to be extended to the MIMO case and model parameter filtering was proposed
to achieve the desired control and disturbance rejection performance.
This chapter covers the whole design process from nonlinear MIMO system identification
based on an artificial neural network (ANN) in section 2 to controller design and presentation
of enhancements in section 3. Finally the results with the real 3-DOF helicopter system are
presented in section 4. On the way pitfalls are analyzed and practical application hints are
given.

2. System Identification
The correct identification of a model is of high importance for any MPC method, so special
attention has to be paid to this part of controller design. The success of the identification will
determine the performance of the final controlled system directly or even whether the system
is stable at all.
Basically there are a few points one has to bear in mind during the experiment design (Ljung,
1999):
• The sampling rate should be chosen appropriately.
• The experimental conditions should be close to the situation for which the model is
going to be used. Especially for MIMO systems this plays an important role as this may
be nontrivial.
• The identification signal should be sufficiently rich to excite all modes of the system. For
nonlinear systems not only the frequency spectrum but also the excitation of different
amplitudes should be sufficient.
• Periodic inputs have the advantage that they reduce the influence of noise on the output
signal but increase the experiment length.
The following sections guide through the full process of the MIMO identification by means of
the practical experiences with the helicopter model.

2.1 Excitation Signal

The type of the excitation signal plays an important role as it should exhibit a few properties
which affect the outcome essentially. Generally the input signal should be persistently exciting
of at least twice the system order. There are many different types of input signals which are not
covered here (see Ljung (1999) for further reading). Despite the desirable optimal Crest factor,
for nonlinear system identification binary signals are not an option due to the lack of excitation
of different amplitudes. For this work an excitation signal comprised of independent multi-
sine signals as described in (Evan et al., 2000) was designed. This is explored in the following
section.
Approximate Model Predictive Control for Nonlinear Multivariable Systems 143

2.1.1 Assembling of Multisine Signals

A multisine is basically a sum of sinusoids:
ns
u(t) = ∑ Ak cos(ωk t + φk )
k =1

where ns is the number of present frequencies. This parameter should be large enough to
guarantee persistent excitation.
A favourable attribute of multisine signals is that the spectrum can be determined directly. By
this property it is possible to just include the frequency ranges that excite the system which
is done by splitting the spectrum in a low (or main) and a high frequency band. As a rule of
thumb one should choose the upper limit of the main frequency band ωc around the system
bandwidth ωb , since choosing ωc too low may result in unexcited modes, while ωc  ωb does
not yield additional information (Ljung, 1999). In a relay feedback experiment the bandwidth
of the helicopter’s pitch axis was measured to be f b ≈ 0.67Hz. As one can see in figure 2 the
upper limit of the main frequency band f c = ωc /2π = 1.5Hz was chosen about twice as large
but the higher frequencies from ωc up to the Nyquist frequency ωn are not entirely absent.
This serves the purpose of making the mathematical model resistant to high frequency noise
as the real system will typically not react to this high frequency band.

40
20
0
Amplitude (dB)

−20
−40
−60
−80
−100
0 0.5 1 1.5
2 2.5 3 3.5 4 4.5 5
Frequency (Hz)
Fig. 2. Spectrum of the multisine excitation signal for the helicopter

2.1.2 Periodic Signals

To reduce the influence of noise present in the output signal of the plant, taking an integer
number of periods of the input signal can be considered. If K periods of the input signal are
taken, the signal to noise ratio is improved by this factor K. A drawback of periodic inputs is
that they generally can not inject as much excitation into the system over a given time span as
non-periodic inputs, since a signal of length N can at most excite a system of order N (Ljung,
1999). But as a periodic signal of length N = KM consists of K periods of length M it has the
same level of excitation as one period.
In the case of the helicopter three signal periods were chosen, as this proved to give consistent
results for the present noise level.
144 Model Predictive Control

2.1.3 MIMO Considerations

For MIMO systems the design of the input signal is a bit more complex, as there may be cross
couplings between the different inputs which drive the outputs out of desired limits or the
signal does not excite all modes sufficiently. The design process of the identification signal
involves the consideration of system specifics and can not be generalized.

2 2

1 1
Amplitude

0 0

−1 −1

−2 −2
0 500 1000 1500 2000 0 500 1000 1500 2000
Samples Samples
(a) Signal 1 (b) Signal 2

40
20
0
Amplitude (dB)

−20
−40
−60
−80
−100
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
Frequency (Hz)
(c) Spectrum of signal 1
Amplitude (dB)

−50

−100
0 1 2 3 4 5
Frequency (Hz)
(d) Spectrum of signal 2

Fig. 3. MIMO signals with appropriate setpoints

In the case of the helicopter three axes need to be excited in all modes. A first attempt was
to directly apply multisine signals to both inputs. For this attempt both inputs were limited
Approximate Model Predictive Control for Nonlinear Multivariable Systems 145

to low amplitudes though, as coincidental add-up effects quickly drove the system out of the
operating bounds. Naturally this yielded bad models that did not resemble the actual plant
very well.
A way to drive a MIMO system to different operating states is the use of setpoints that are
added to the multisine signal. This enables the selective identification of certain modes of the
system. At the same time this can be used as a means to keep the outputs inside of valid op-
erating bounds, since the amplitude of the multisine signal can be chosen to be much lower
than without setpoints. This enables much safer operation during the experiment, as the en-
ergy of the random signal can be reduced. Of course one has to keep in mind that the actual
excitation signals amplitude has to be as large as possible to assure maximal excitation around
each setpoint.
The spectrum of a multisine signal with additive setpoints does not differ much from the
original multisine (figure 2) as can be seen in figure 3. The only difference is a peak in the low
frequency band and a general small lifting in the upper band. Both signals are composed of
multisines of same spectrum with unit variance and additive setpoints in the range of [−1, 1].
This assures overlapping amplitude ranges, which is desirable for a consistent model.

2.2 Closed Loop Identification

In recent years the interest in closed loop identification has generally risen, due to its impor-
tance for practical system identification. In many industrial processes existing control loops
cannot be switched off during system identification for safety reasons or process restrictions.
Likewise when dealing with unstable systems every experiment setup must involve stabiliz-
ing control loops to keep the output in a valid operating range. Another advantage of models
computed from closed loop data is their better approximation of the behavior of a process
under feedback which is important for successful controller design (Pico & Martinez, 2002).
There are a few different approaches to closed loop identification of which the two most gen-
eral are covered here:
• Direct Approach: ignore the presence of the feedback and directly identify the plant by
plant input and output data. This has the advantage that no knowledge about the type
of control feedback or even linearity of the controller is required.
• Indirect Approach: identify the closed loop and obtain the open loop model by decon-
volution if possible. Obtaining the open loop model is only possible if the controller is
known and both the closed loop plant model and the controller are linear.

r (t) u(t) y(t)

G0 (z)
+
−

K F (z)

Fig. 4. Closed loop setup for identification

146 Model Predictive Control

This section is limited to the techniques and practical experiences of the identification of the
helicopter model and does not cover the whole theory of closed loop identification. For further
information on this matter see Pico & Martinez (2002) or Ljung (1999).

2.2.1 Direct Identification

This is the natural approach to closed loop identification as it is similar to open loop identifi-
cation if one keeps some fallacies in mind. In general this method applies a straightforward
identification by taking the raw data from plant input u(t) and plant output y(t) and thus
computes the model as if in open loop. Figure 4 shows the basic setup of the experiment.
A few general statements can be made about the direct identification in closed loop (Pico &
Martinez, 2002):
• The experiment is informative if r (t) is persistently exciting.
• Even if r (t) is not a rich signal, the experiment can be informative if the system is driven
by the output-noise and the feedback mechanism is of adequate structure avoiding a
linear dependency between y(t) and u(t).
• Problems can arise if the amplitude of r (t) is small in comparison to u(t) and the feed-
back mechanism is approximately linear, i.e. u(t) ≈ −K F (z)y(t).
• The direct approach can be problematic if the open loop plant is unstable, since the
spectrum of u(t) may be altered in a suboptimal way.
So if it is possible to use a rich signal for r (t), the experiment is informative. But as this may
not always be the case in a practical scenario, informative experiments can still be achieved by
choosing an appropriate controller. Generally if r (t) has a very low amplitude in comparison
to u(t) or if r (t) is not a rich signal, problems arise if a dependency like u(t) ≈ −K F (z)y(t) ex-
ists. But this can be avoided by choosing high order, time varying or nonlinear feedback mech-
anisms. For further information about appropriate controllers see (Pico & Martinez, 2002).

Issues with Unstable Systems

In the case of the helicopter the multisine from section 2.1 was used which exhibits a high level
of excitation, thus a simple proportional derivative controller (PD controller) could be used for
stabilization since the main source of exitation is the reference signal r (t) in this case. Figure
5 shows the spectrum of the input signal u(t) that was recorded during a SISO experiment
to identify the helicopters pitch axis. The elevation axis was controlled separately by a PID
controller. As one can see, the spectrum has changed significantly if compared to the original
spectrum in figure 2. The low frequencies are heavily damped which comes naturally for
an unstable plant like the helicopter’s pitch axis, as a constant input would drive the system
to infinity. For identification though this is unfavourable, since the low frequencies are not
sufficiently excited although these frequencies lie in the normal operating band.
More problems arise during validation, since with the direct approach unstable plants yield
unstable models directly and validating these is tedious. The problem with validating un-
stable models is that the validation is usually open loop, as the recorded input sequence is
applied to the model and the output is compared to the recorded actual output. Errors are not
compensated and thus add up as there are no control loops during validation so the model
response may ascend to infinity - even with a decent model. A better way to verify the quality
of an unstable model is to look at a k-step-ahead prediction, because errors do not have as
much time to add up.
Approximate Model Predictive Control for Nonlinear Multivariable Systems 147

40
20
0
Amplitude (dB)

−20
−40
−60
−80
−100
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
Frequency (Hz)

60
40
Amplitude (dB)

20
0
−20
−40
−60
−80
0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
Frequency (Hz)
Fig. 5. Spectrum of the plant input u(t) in closed loop excited by multisine signal (lower
picture). Note that this signal is scaled in comparison to the original spectrum of the reference
r (t) (upper picture).

A comparison of the direct and the indirect approach by example of the helicopter’s unstable
pitch axis is presented in section 2.2.3.

2.2.2 Indirect Identification

The basic concept of indirect identification is to identify the closed loop as a whole and vali-
date this model of the closed loop in the first place. Consecutive steps may include a decon-
volution of the plant and controller to obtain the open loop model. This requires the regulator
and the plant model to be both linear, and of course the regulator transfer function must be
known. So in the case of nonlinear system identification these steps are not possible, even if
the controller is linear and known, because in this case the equation can usually not be solved
for G0 (z) analytically.
In contrast to the direct approach the indirect approach avoids any alteration of the input sig-
nal’s spectrum since the input to the closed loop directly is the reference signal r (t) which has
no dependence on other signals. This has the advantage, that even if r (t) is small compared
to u(t) and the feedback mechanism is of a simple linear kind, an informative experiment is
still achieved as long as r (t) is persistently exciting (which the multisine signal from section
2.1 ensures). Another advantage is that unstable systems can be handled intuitively as the re-
sulting model is stable. This eliminates the problems discussed in 2.2.1 in the validation phase
148 Model Predictive Control

since the closed loop model does not have unstable poles as the open loop model would. A
drawback of this method is, that the feedback mechanism increases the model order since it is
identified along with the actual open loop model.
For the case of a known linear controller and a linear closed loop model, the open loop model
can be obtained by deconvolution as was already mentioned. The closed loop transfer func-
tion corresponding to figure 4 is:
G0 (z)
Gcl (z) = . (1)
1 + G0 (z)K F (z)
Solving for G0 (z) yields:
Gcl (z)
G0 (z) = , (2)
1 − Gcl (z)K F (z)
which is the final formula for obtaining the open loop model G0 (z). So if either Gcl (z) or K F (z)
are nonlinear both formulas cannot be applied and a deconvolution is not possible. However
for the linear case it will be shown that a controller design for the closed loop model Gcl (z)
can yield exactly the same overall system dynamics as a controller design for the open loop
model G0 (z). For control strategies that utilize linearizations of a nonlinear model like APC,
this similarly implies that the direct use of a (in this case nonlinear) closed loop model has no
adverse effects on the final performance.

r (t) y(t) r (t) y(t)

K1 ( z ) G0 (z) K2 ( z ) G0 (z)
+ + +
− − −

⇔
K F (z)

Fig. 6. Controller Setup for open loop and closed loop models

Theorem. Given the closed loop system Gcl (z) consisting of the open loop plant G0 (z) and the con-
troller K F (z), a controller K2 (z) can be found that transforms the system to an equivalent system
consisting of an arbitrary controller K1 (z) that is applied to the plant G0 (z) directly.

Proof. The two system setups are depicted in figure 6. The transfer function of the left system
is:
K1 (z) G0 (z)
G1 (z) =
1 + K1 (z) G0 (z)
while the right system has the transfer-function:
K2 (z) Gcl (z)
G2 (z) =
1 + K2 (z) Gcl (z)
G (z)
K2 (z) 1+G (0z)K (z)
0 F
= G (z)
1 + K2 (z) 1+G (0z)K (z)
0 F

K2 (z) G0 (z)
= .
1 + (K F (z) + K2 (z)) G0 (z)
Approximate Model Predictive Control for Nonlinear Multivariable Systems 149

Now it has to be proved that there exists a controller K2 (z) that transforms G2 (z) to G1 (z)
for any given K1 (z). It is clear that the system G2 (z) with the K F (z) feedback controller can
achieve exactly the same performance as the G1 (z) system if this is the case.

K1 (z) G0 (z) K2 (z) G0 (z)

=
1 + K1 (z) G0 (z) 1 + (K F (z) + K2 (z)) G0 (z)
K1 (z) G0 (z) + K1 (z)K F (z) G0 (z) G0 (z) = K2 (z) G0 (z)
K2 ( z ) = K1 (z) + K1 (z)K F (z) G0 (z)
= K1 (z)(1 + K F (z) G0 (z))

2.2.3 Indirect vs. Direct Approach

The quality of the models heavily depends on the experiment setup and the identification
approach chosen. This is valid even more for the identification of unstable systems. This will
be shown in the following with the example of the SISO identification of the helicopter’s pitch
axis. Consider the experiment setup from figure 4. As discussed in section 2.1 three periods
of a multisine signal with spectrum as in figure 2 are applied to the reference input r (t). The
feedback controller K F (z) is a hand tuned PD controller. The input data for the identification
process are r (t) for the indirect approach and u(t) for the direct approach respectively, the
output data is y(t) in both cases. The spectrum of u(t) is shown in figure 5 (the spectrum of
r (t) is just the one of the multisine in figure 2). To obtain a fair comparison of the approaches
the open loop model is computed for both (for the indirect approach (2) is used to compute
G0 (z)). Since direct validation of unstable models is not very meaningful in most cases, a
stabilizing controller is added for the simulation. Here, the natural choice is the same PD
feedback controller as used with the real helicopter. The comparison of the model outputs to
the original helicopter output is shown in figure 7.

40
indirect identification
direct identification
20 original output
Pitch (deg)

−20

−40
5 10 15 20 25
Time (sec)
Fig. 7. Simulated closed loop response of models from direct and indirect approach compared
to experimental measurement

Judging from the predicted outputs of both models they seem almost identical, as it is even
difficult to distinguish between both model outputs. Both are not perfectly tracking the real
output but it seems that decent models have been acquired. In figure 8 the bode plots of both
150 Model Predictive Control

open loop models are shown and this illustrates that the models are not as similar as it had
seemed in the closed loop validation, since the static gain differs in a few orders of magnitude.
The high frequency part of the plot is comparable, though.

Bode Diagram

100
indirect identification
direct identification
Magnitude (dB)

50

−50
−180

−225
Phase (deg)

−270

−315

−360
−4 −3 −2 −1 0 1 2
10 10 10 10 10 10 10
Frequency (rad/sec)

Fig. 8. Bode plot of models from direct and indirect approach

This correlates with the spectra of r (t) and u(t) since they are similar for higher frequencies,
too. Taking a close look at the step responses in figure 9 of both stabilized open loop models,
although looking similar over all, the model of the direct approach shows a small oscillation
for a long time period which seems negligible at first.
To see the consequences of the differences in the models they have to be used in a controller de-
sign process and tested on the actual plant. Figure 10 shows the responses of the helicopter’s
pitch axis to a rectangular reference stabilized by two LQG controllers. Both controllers were
designed with the same parameters differing only in the employed plant models.
The controller designed with the model of the indirect approach performs well and is also
very robust to manual disturbances. In contrast the LQG controller designed with the model
of the direct approach even establishes a static oscillation indicating that the model is not a
good representation of the real plant. During all identification approaches the indirect method
performed superiorly, which led to the conclusion that the direct approach is not ideal for our
setup.

2.3 Linear Identification Results

With the bad results for the direct identification in the SISO case the MIMO identification was
attempted with the indirect approach only.
The output of a MIMO model computed from a data set with usage of setpoints as described
in 2.1.3 is shown in figure 11. The model used for the output in figure 11 is a state space model
of order 16 computed with the prediction error/maximum likelihood (PEM) method of the
Approximate Model Predictive Control for Nonlinear Multivariable Systems 151

9 7.2 indirect identification

indirect identification
8 direct identification direct identification
7
7
6.8
6
Pitch (deg)

Pitch (deg)
6.6
5

4 6.4

3 6.2

2 6
1
5.8
0
2 4 6 8 2 4 6 8 10 12 14 16
Time (sec) Time (sec)

Fig. 9. Simulated step response of models from direct and indirect approach (plotted at differ-
ent scales)

20

10
Pitch (deg)

−10 LQG from indirect

LQG from direct
−20 reference

30 40 50 60 70 80 90 100 110 120

Time (sec)
Fig. 10. Experimental results of controllers tracking a rectangular reference on the pitch axis.
The LQG controller design was done with models from the direct and indirect approach re-
spectively.

identification toolbox in Matlab. This method uses an iterative search starting at the result of
the subspace-method. Other methods like MIMO ARX or directly the sub-space method also
yielded good results.
From the model output it can be seen that the characteristics of the model seem to resemble the
real ones correctly. During more dynamic maneuvers a discrepancy between the measurement
and the prediction becomes visible, though.

2.4 Neural Networks for System Identification

Opposed to the common linear plant models with widely spread structures like ARX (AutoRe-
gressive with eXogenous input), ARMAX (AutoRegressive Moving Average with eXogenous
input) or state space models the use of neural networks for system identification is a relatively
new approach. Traditionally the identification of models with neural networks falls in the
category of black box modelling as typically very few information about the system can be
incorporated in the process of identification. Neural networks as they are used most often
are very general approximators which can be trained to resemble any given function (given
that the network complexity is sufficient). Many different approaches and network structures
152 Model Predictive Control

Measured and 20 step predicted output

300
original output
model output
200

100
Travelspeed

−100

−200

320 340 360 380 400 420 440

Time (sec)

original output
10
model output

5
Elevation (deg)

−5

−10

−15
320 340 360 380 400 420 440
Time (sec)

30
original output
model output
20

10
Pitch (deg)

−10

−20

−30

−40
320 340 360 380 400 420 440
Time (sec)
Fig. 11. 20-step ahead prediction output of the linear model for a validation data set
Approximate Model Predictive Control for Nonlinear Multivariable Systems 153

exist. For an introduction to the field of neural networks the reader is referred to Engelbrecht
(2002). The common structures and specifics of neural networks for system identification are
examined in Nørgaard et al. (2000).

2.4.1 Network Structure

The network that was chosen as nonlinear identification structure in this work is of NNARX
format (Neural Network ARX, corresponding to the linear ARX structure), as depicted by
figure 12. It is comprised of a multilayer perceptron network with one hidden layer of sigmoid
units (or tanh units which are similar) and linear output units. In particular this network
structure has been proven to have a universal approximation capability (Hornik et al., 1989).
In practice this is not very relevant knowledge though, since no statement about the required
number of hidden layer units is made. Concerning the total number of neurons it may still
be advantageous to introduce more network layers or to introduce higher order neurons like
product units than having one big hidden layer.

y k −1
.. ..
. . NN
yk−n
ŷk
uk−d
.. .. θ
. .
uk−d−m
Fig. 12. SISO NNARX model structure

The prediction function of a general two-layer network with tanh hidden layer and linear
output units at time k of output l is
1
 
s r
ŷl (k) = ∑ w2lj tanh ∑ w1ji ϕi (k) + w1j0 + w2l0 (3)
j =1 i =1

where w1ji and w1j0 are the weights and biases of the hidden layer, w2lj and w2l0 are the weights
and biases of the output layer respectively and ϕi (k) is the ith entry of the network input
vector (regression vector) at time k which contains past inputs and outputs in the case of the
NNARX structure. The choice of an appropriate hidden layer structure and input vector are of
great importance for satisfactory prediction performance. Usually this decision is not obvious
and has to be determined empirically. For this work a brute-force approach was chosen, to
systematically explore different lag space and hidden layer setups, as illustrated in figure 13.
From the linear system identification can be concluded that significant parts of the dynamics
can be described by linear equations approximately. This knowledge can pay off during the
identification using neural networks. If only sigmoid units are used in the hidden layer the
network is not able to learn linear dynamics directly. It can merely approximate the linear
behavior which would be wasteful. Consequently in this case it is beneficial to introduce linear
neurons to the hidden layer. The benefits are twofold as training speed is greatly improved
when using linear units (faster convergence) and the linear behavior can be learned "natively".
Since one linear neuron in the hidden layer can represent a whole difference equation for an
output the number of linear neurons should not exceed the number of system outputs.
154 Model Predictive Control

300
MSE of 10−step ahead prediction

250

200

150

100

50

0
0
2 0
4 4 2
6 8 6
8 12 10
Number of sigmoid Units
Lagspace
Fig. 13. Comparison of network structures according to their MSE of the 10-step ahead pre-
diction using a validation data set (all networks include three linear units in the hidden layer).
Each data point reresents the best candidate network of 10 independent trainings.

The final structure that was chosen according to the results depicted by figure 13 includes three
linear and twelve sigmoid units in the hidden layer with a lag space of six for both inputs and
the three outputs. For this network accordingly ((2 + 3)· 6 + 1)· (12 + 3) + (12 + 3 + 1)· 3 =
513 weights had to be optimized.

2.4.2 Instantaneous Linearization

To implement APC, linearized MIMO-ARX models have to be extracted from the nonlinear
NNARX model in each sampling instant. The coefficients of a linearized model can be ob-
tained by the partial derivative of each output with respect to each input (Nørgaard et al.,
2000). Applying the chain rule to (3) yields
  
s1 r
∂ŷl (k)
= ∑ w2lj w1ji 1 − tanh2 ∑ w1ji ϕi (k) + w1j0 (4)
∂ϕi (k) j =1 i =1

for tanh units in the hidden layer. For linear hidden layer units in both the input and the
output layer one yields
1
s
∂ŷl (k)
= ∑ w2lj w1ji . (5)
∂ϕi (k) j =1
Approximate Model Predictive Control for Nonlinear Multivariable Systems 155

2.4.3 Network Training

All networks were trained with Levenberg Marquardt Backpropagation (Hagan & Menhaj, 1994).
Due to the monotonic properties of linear and sigmoid units, networks using only these unit
types have the inherent tendency to have only few local minima, which is beneficial for local
optimization algorithms like backpropagation. The size of the final network (513 weights)
that was used in this work even makes global optimization techniques like Particle Swarm Op-
timization or Genetic Algorithms infeasible. Consequently for a network of the presented size,
higher order units such as product units cannot be incorporated due to the increased amount
of local minima, requiring global optimization techniques (Ismail & Engelbrecht, 2000).
But also with only sigmoid units, based on the possibility of backpropagation getting stuck in
local minima, always a set of at least 10 networks with random initial parameters were trained.
To minimize overfitting a weight decay of D = 0.07 was used. The concept of regularization
to avoid overfitting using a weight decay term in the cost function is thoroughly explored by
Nørgaard et al. (2000).

2.5 Nonlinear Identification Results

For the nonlinear identification the same excitation signal and indirect measurement setup
was used as for the linear identification. Thus a stabilized closed-loop model was acquired.
The controller that was inevitably identified along with the unstable plant model cannot be
removed from the model analytically. In section 2.2.2 we showed that the stabilizing controller
will not hinder the final control performance in the case of APC, though.
The prediction of the finally chosen network with a validation data set is depicted in figure
14. If one compares the neural network prediction with the prediction of the linear model
in figure 11 it is obvious that the introduction of nonlinear neurons benefited the prediction
accuracy. This is underlined by figure 13 also visualizing a declining prediction error for
increasing sigmoid unit numbers. Whether the improvements in the model can be transferred
to an improved controller remains to be seen, though.

2.6 Conclusion
This section demonstrated successful experiment design for an unstable nonlinear MIMO sys-
tem and showed some pitfalls that may impede effective identification. The main approaches
to closed loop identification have been presented and compared by means of the helicopters
unstable pitch axis. It was shown that the identification of unstable systems can be just as suc-
cessful as for stable systems if the presented issues are kept in mind. Both linear and nonlinear
identifications can be regarded as successful, although the nonlinear predictions outperform
the linear ones.
156 Model Predictive Control

Measured and 20 step predicted output

400
Measured output
200 Network output
Travelspeed

−200

−400
0 10 20 30 40 50 60 70 80 90 100
Time (sec)

10
Elevation (deg)

−10

−20
0 10 20 30 40 50 60 70 80 90 100
Time (sec)

40

20
Pitch (deg)

−20

−40

0 10 20 30 40 50 60 70 80 90 100
Time (sec)

Fig. 14. 20-step ahead prediction output of the best network for a validation data set
Approximate Model Predictive Control for Nonlinear Multivariable Systems 157

3. Approximate Model Predictive Control

The predictive controller that is discussed in this chapter is a nonlinear adaptation of the
popular Generalized Predictive Control (GPC), proposed in (Clarke et al., 1987a;b). Approximate
(Model) Predictive Control (APC) as proposed by Nørgaard et al. (2000) uses the GPC principle
on instantaneous linearizations of a neural network model. Although presented as a single-
input single-output (SISO) algorithm, its extension to the multi-input multi-output (MIMO)
case with MIMO-GPC (Camacho & Borbons, 1999) is straightforward. The scheme is visual-
ized in figure 15.

A( z −1 ), B( z −1 )

Tuning parameters GPC NN

Nu , N1 , N2 , Qr , Qu synthesis linearization

r u
GPC Plant y

Fig. 15. Approximate predictive control scheme

The linearized model that is extracted from the neural network at each time step (as described
in section 2.4.2) is used for the computation of the optimal future control sequence according
to the objective function:

N2  T  
J (k) = ∑ r (k + i ) − ŷ(k + i ) Qr r (k + i ) − ŷ(k + i )
i = N1
Nu
+ ∑ ∆uT (k + i − 1) Qu ∆u(k + i − 1) (6)
i =1
where N1 and N2 are the two prediction horizons which determine how many future samples
the objective function considers for minimization and Nu denotes the length of the control
sequence that is computed. As common in most MPC methods, a receding horizon strategy is
used and thus only the first control signal that is computed is actually applied to the plant to
achieve loop closure.
A favourable property of quadratic cost functions is that a closed-form solution exists, en-
abling its application to fast processes under hard realtime constraints (since the execution
time remains constant). If constraints are added, an iterative optimization method has to be
used in either way, though. The derivation of MIMO-GPC is given in the following section for
the sake of completeness.

3.1 Generalized Predictive Control for MIMO Systems

In GPC, usually a modified ARX (AutoRegressive with eXogenous input) or ARMAX (Au-
toRegressive Moving Average with eXogenous input) structure is used. In this work a struc-
ture like
158 Model Predictive Control

1
A( z −1 ) y ( k ) = B( z −1 ) u ( k ) +
e(k) (7)
∆
is used for simplicity, with ∆ = 1 − z−1 where y(k) and u(k) are the output and control
sequence of the plant and e(k) is zero mean white noise. This structure is called ARIX and
basically extends the ARX structure by integrated noise. It has a high relevance for practical
applications as the coloring polynomials for an integrated ARMAX structure are very difficult
to estimate with sufficient accuracy, especially for MIMO systems (Camacho & Borbons, 1999).
The integrated noise term is introduced to eliminate the effects of step disturbances.
For an n-output, m-input MIMO system A(z−1 ) is an n × n monic polynomial matrix and
B(z−1 ) is an n × m polynomial matrix defined as:

A( z −1 ) = In×n + A1 z−1 + A2 z−2 + ... + Ana z−na

B( z −1 ) = B0 + B1 z−1 + B2 z−2 + ... + Bnb z−nb

The output y(k) and noise e(k) are n × 1-vectors and the input u(k) is an m × 1-vector for the
MIMO case. Looking at the cost function from (6) one can see that it is already in a MIMO
compatible form if the weighting matrices Qr and Qu are of dimensions n × n and m × m
respectively. The SISO case can easily be deduced from the MIMO equations by inserting
n = m = 1 where A(z−1 ) and B(z−1 ) degenerate to polynomials and y(k), u(k) and e(k ) be-
come scalars.

To predict future outputs the following Diophantine equation needs to be solved:

In×n = E j (z−1 )(A(z−1 )∆) + z− j F j (z−1 ) (8)

where E j (z−1 ) and F j (z−1 ) are both unique polynomial matrices of order j − 1 and n a re-
spectively. This special Diophantine equation with In×n on the left hand side is called Bizout
identity, which is usually solved by recursion (see Camacho & Borbons (1999) for the recur-
sive solution). The solution to the Bizout identity needs to be found for every future sampling
point that is to be evaluated by the cost function. Thus N2 − N1 + 1 polynomial matrices
E j (z−1 ) and F j (z−1 ) have to be computed. To yield the j step ahead predictor, (7) is multiplied
by E j (z−1 )∆z j :

E j (z−1 )∆A(z−1 )y(k + j) = E j (z−1 )B(z−1 )∆u(k + j − 1) + E j (z−1 )e(k + j) (9)

which by using equation 8 can be transformed into:

y(k + j) = E j (z−1 )B(z−1 )∆u(k + j − 1) + F j (z−1 )y(k) + E j (z−1 )e(k + j) (10)

        
past and f uture inputs f ree response f uture noise

Since the future noise term is unknown the best prediction is yielded by the expectation value
of the noise which is zero for zero mean white noise. Thus the expected value for y(k + j) is:

ŷ(k + j|k) = E j (z−1 )B(z−1 )∆u(k + j − 1) + F j (z−1 )y(k) (11)

The term E j (z−1 )B(z−1 ) can be merged into the new polynomial matrix G j (z−1 ):

G j (z−1 ) = G0 + G1 z−1 + . . . + Gj−1 z−( j−1) + ( Gj ) j z− j + . . . + ( Gj−1+nb ) j z−( j−1+nb )

Approximate Model Predictive Control for Nonlinear Multivariable Systems 159

where ( Gj+1 ) j is the ( j + 1)th coefficient of G j (z−1 ) and nb is the order of B(z−1 ). So the
coefficients up to ( j − 1) are the same for all G j (z−1 ) which stems from the recursive properties
of E j (z−1 ) (see Camacho & Borbons (1999)). With this new matrix it is possible to separate the
first term of (10) into past and future inputs:

G j (z−1 )∆u(k + j − 1) = G0 ∆u(k + j − 1) + G1 ∆u(k + j − 2) + . . . + Gj−1 ∆u(k)

  
f uture inputs
+ ( Gj ) j ∆u(k − 1) + ( Gj+1 ) j ∆u(k − 2) + . . . + ( Gj−1+nb ) j ∆u(k − nb )
  
past inputs

Now it is possible to separate all past inputs and outputs from the future ones and write this
in matrix form:

     
ŷ(k + 1|k) G0 0 ··· 0 f1
 ŷ(k + 2|k)   G1 G0 ··· 0    f2 


 
 

 ∆u(k) 



 ..   .. .. .. ..  ∆u ( k + 1)   .. 
 .   . . . .    . 
 =  .. + 
 ŷ(k + Nu |k)   GNu −1 GNu −2 ··· G0  .   f Nu 
     
 ..   .. .. ..  ∆u(k + N − 1)  .. 
 .   . . ··· .  u  . 
  
ŷ(k + N2 |k) GN2 −1 GN2 −2 · · · GN2 − Nu ũ f N2
        
ŷ G f
(12)
which can be condensed to :

ŷ = Gũ + f (13)
where f represents the influence of all past inputs and outputs and the columns of G are the
step responses to future ũ (for further reading, see (Camacho & Borbons, 1999)). Since each Gi
is an n × m matrix G has block matrix structure.

Now that we obtained a j-step ahead predictor form of a linear model this can be used to
compute the optimal control sequence with respect to a given cost function (like (6)). If (6) is
written in vector form and with (13) one yields:

J (k) = (r − ŷ) T Qr (r − ŷ) + ũT Qu ũ

= (r − Gũ − f) T Qr (r − Gũ − f) + ũT Qu ũ

where
r = [r (k + 1), r (k + 2), . . . , r (k + N2 )] T
In order to minimize the cost function J (k) for the future control sequence ũ the derivative
dJ (k)/dũ is computed and set to zero:
160 Model Predictive Control

dJ (k)
= 0
dũ
= 2GT Qr Gũ − 2GT Qr (r − f) + 2Qu ũ

(GT Qr G + Qu )ũ = G T Qr (r − f) (14)

ũ = (G T Q r G + Q u ) −1 G T Q r (r − f) (15)
  
K

Thus the optimization problem can be solved analytically without any iterations which is true
for all quadratic cost functions in absence of constraints. This is a great advantage of GPC
since the computation effort can be very low for time-invariant plant models as the main
computation of the matrix K can be carried out off-line. Actually just the first m rows of K
must be saved because of the receding horizon strategy using only the first input of the whole
sequence ũ. Therefore the resulting control law is linear, each element of K weighting the
predicted error between the reference and the free response of the plant.
Finally for a practical implementation of APC one has to bear in mind that the matrix (GT Qr G +
Qu ) can be singular in some instances. In the case of GPC this is not a problem since the so-
lution is not computed online. For APC in this work a special Gauss solver was used which
assumes zero control input where no unambiguous solution can be found.

3.2 Reducing Overshoot with Reference Filters

With the classic quadratic cost function it is not possible to control the overshoot of the result-
ing controller in a satisfying manner. If the overshoot needs to be influenced one can choose
three possible ways. The obvious and most elaborate way is to introduce constraints, however
the solution to the optimization problems becomes computationally more expensive. Another
possible solution is to change the cost function, introducing more tuning polynomials, as men-
tioned by Nørgaard et al. (2000) referring to Unified Predictive Control.
A simple but yet effective way to reduce the overshoot for any algorithm that minimizes the
standard quadratic cost function (like LQG, GPC or APC) is to introduce a reference prefilter
which smoothes the steep areas like steps in the reference. For the helicopter, the introduction
of prefilters made it possible to eliminate overshoot completely, retaining comparably fast rise
times. The utilized reference prefilters are of first order low-pass kind

1−l
GRF =
1 − lz−1
which have a steady-state gain of one and can be tuned by the parameter l to control the
smoothing.

3.3 Improving APC Performance by Parameter Filtering

A problem with APC is that a network that has a good prediction capability does not neces-
sarily translate into a good controller, as for APC the network dynamics need to be smooth for
consistent linear models which is not a criterion the standard Levenberg-Marquardt backprop-
agation algorithm trains the network for. A good way to test whether the network dynamics
are sufficiently smooth is to start a simulation with the same neural network as the plant and
Approximate Model Predictive Control for Nonlinear Multivariable Systems 161

as the predictive controllers system model. If one sees unnecessary oscillation this is good ev-
idence that the network dynamics are not as smooth as APC desires for optimal performance.
The first solution to this is simply training more networks and test whether they provide a
better performance in the simulation.

0 Reference
Elevation (deg)

d=0
−10 d=0.9

−20

0 1 2 3 4 5 6 7 8 9 10
Time (sec)

20 Reference
Pitch (deg)

0 d=0
d=0.9
−20

−40
0 1 2 3 4 5 6 7 8 9 10
Time (sec)
10
Disturbance
5
d=0
Torque

0 d=0.9
−5

−10
0 1 2 3 4 5 6 7 8 9 10
Time (sec)

5
d=0
0
d=0.9
Thrust

−5
−10
−15
0 1 2 3 4 5 6 7 8 9 10
Time (sec)

Fig. 16. Simulation results of disturbance rejection with parameter filtering. Top two plots:
Control outputs. Bottom two plots : Control inputs

In the case of the helicopter a neural network with no unnecessary oscillation in the simu-
lation could not be found, though. If one assumes sufficiently smooth nonlinearities in the
real system, one can try to manually smooth linearizations of the neural network from sample
to sample, as proposed in (Witt et al., 2007). Since APC is not able to control systems with
nonlinearities that are not reasonably smooth within the prediction horizon anyway, the idea
of smoothing the linearizations of the network does not interfere with the basic idea of APC
being able to control nonlinear systems. It is merely a means to flatten out local network areas
where the linearized coefficients start to jitter within the prediction horizon.
162 Model Predictive Control

This idea has been realized by a first order low-pass filter:

1−d
GPF =
1 − dz−1

with tuning parameter d. When applied to the polynomial matrix A(z−1 ), (3.3) results in the
following formula:
Âk (z−1 ) = (1 − d)Ak (z−1 ) + dÂk−1 (z−1 )
where Âk (z−1 ) contains the filtered polynomial coefficients Ak (z−1 ). For prediction horizons
around N2 = 10...20 a good starting value for the tuning parameter d was found to be 0.9,
however this parameter depends on the sampling rate.
If the filtering parameter d is increased, the adaptivity of the model decreases and shifts to-
wards a linear model (in the case of d = 1). The importance of parameter filtering in the case
of the helicopter is displayed in figure 16 where an input disturbance acts on the torque input
of a standard APC controller and the parameter filtered version.

4. Experimental Results
During the practical experiments the setup shown in figure 17 was used. It necessarily in-
corporates the stabilizing proportional derivative controller that is included in our nonlinear
model from section 2. The sampling time was 0.1 seconds and the experiments were run on
a 1 GHz Intel Celeron CPU. All APC related algorithms were implemented in C++ to achieve
the computational performance that was necessary to be able to compute the equations in
realtime on this system at the given sampling rate.

d(t)

r (t)
u(t) y(t)
Controller Helicopter
Prefilter

PD-Stabilizers

Fig. 17. Control setup for helicopter with inner stabilizing control loop and reference prefilter.

For our experiments only the control of the pitch and elevation axis was considered as the
travelspeed axis has significantly longer rise times (about factor 15) than the other two axes,
making predictive control with the same sampling rate and prediction horizons impractical.
To control the travelspeed axis in this setup one could design an outer cascaded control loop
with a slower sampling rate, but this is beyond the scope of this work.
APC as well as GPC were tuned with the same 5 parameters, being the horizons N1 , N2 , Nu
and the weighting matrices Qr and Qu . The tuning was done as suggested in (Clarke et al.,
1987a;b) and resulted in N1 = 1, N2 = 10, Nu = 10 and the weighting matrices Qr =
Approximate Model Predictive Control for Nonlinear Multivariable Systems 163

diag(0, 1, 1) and Qu = diag(20, 10). The choice of Qr disables weighting for the first output
which is the uncontrolled travelspeed-axis.
The computational limits of the test platform were found at horizons of N2 = Nu = 20 which
does not leave too much headroom.

4.1 Tracking Performance

APC has been benchmarked with both tracking and disturbance rejection experiments. We
also designed a linear GPC and an integrator augmented LQG controller for comparison. The
benchmark reference signals are designed to cover all operating ranges for all outputs. All
controllers were benchmarked with identically parameterized reference prefilters to eliminate
overshoot.
In figure 18 it can be seen that LQG achieves a suitable performance only for the pitch axis
while performance on the elevation axis is much poorer than both APC and GPC. For both
outputs, APC yields slightly better performance than linear GPC which is most visible for
the large reference steps on the more nonlinear elevation axis. However looking at the plant
input signals one can see that the APC signals have less high frequency oscillation than for
GPC which is also an important issue because of actuator stress in practical use. Parameter
filtering does not change the response to the benchmark sequence up to about d = 0.9 but
significantly improves the performance for disturbance rejection as will be shown in the next
section.

4.2 Disturbance Rejection

The performance of the benchmarked controllers becomes more diverse when disturbance
rejection is considered. In figure 19 one can see the response to disturbances applied to the
two inputs. Again LQG can be tuned to satisfactory performance only for the pitch axis, but
also the standard APC and GPC do not give satisfying results. Considering input disturbance
rejection the standard APC even shows a lower stability margin than GPC. The introduction
of parameter filtering however changes this aspect significantly. With parameter filtering of
d = 0.9 the stability margin of APC becomes much larger than the one of GPC and it can be
seen in the plot that it shows the best disturbance response of all tested controllers - especially
note the low input signal amplitude, while superiorly managing the disturbance.

4.3 Conclusion
With this work it has been shown that MIMO APC for a fast process is indeed feasible with
mid-range embedded hardware. It was found that standard APC can be problematic if the
network dynamics are unsmooth. For this purpose, parameter filtering was presented as an
improvement to the standard APC implementation with which it was possible to enhance
the stability margin and overall performance of APC in the face of disturbances significantly.
Still the acquisition of a decent model should be the first step before one should tune the
performance with parameter filtering, since it remains the most important constituent to good
control performance.
Finally although the helicopter is not a highly nonlinear system, APC with parameter filtering
was able to outperform the linear GPC while being the more generally applicable control
scheme.
164 Model Predictive Control

10

5
Elevation (deg)

-5 Reference
LQG
-10 GPC
APC
-15
0 10 20 30 40 50 60 70
Time (sec)
20

10
Pitch (deg)

0
Reference
LQG
-10
GPC
APC
-20
0 10 20 30 40 50 60 70
Time (sec)
5
Torque

0 LQG
GPC
APC
-5
0 10 20 30 40 50 60 70
Time (sec)
10

5
Thrust

0
LQG
-5 GPC
APC
-10
0 10 20 30 40 50 60 70
Time (sec)

Fig. 18. Experimental results for tracking performance of APC compared to GPC and LQG.
Top two plots: Control outputs. Bottom two plots: Control inputs.
Approximate Model Predictive Control for Nonlinear Multivariable Systems 165

5
Reference
LQG
Elevation (deg)

0 GPC
APC, d=0
APC, d=0.9
-5

0 2 4 6 8 10 12 14 16 18 20
Time (sec)

10
Reference
5 LQG
Pitch (deg)

GPC
0 APC, d=0
APC, d=0.9
-5

-10
0 2 4 6 8 10 12 14 16 18 20
Time (sec)

6
Disturbance
4
LQG
2 GPC
Torque

0 APC, d=0
-2 APC, d=0.9
-4
-6
0 2 4 6 8 10 12 14 16 18 20
Time (sec)

6
Disturbance
4
LQG
2 GPC
Thrust

0 APC, d=0
-2 APC, d=0.9

-4
-6
0 2 4 6 8 10 12 14 16 18 20
Time (sec)

Fig. 19. Experimental results for disturbance rejection performance. Top two plots: Control
outputs. Bottom two plots: Control inputs.
166 Model Predictive Control

5. References
Camacho, E. & Borbons, C. (1999). Model Predictive Control, Springer-Verlag, London.
Clarke, D., Mohtadi, C. & Tuffs, P. (1987a). Generalized Predictive Control – Part I. The basic
algorithm, Automatica 23(2): 137–148.
Clarke, D., Mohtadi, C. & Tuffs, P. (1987b). Generalized Predictive Control – Part II. Extension
and Interpretations, Automatica 23(2): 149–160.
Engelbrecht, A. (2002). Computational Intelligence: An Introduction, Halsted Press New York,
NY, USA.
Evan, C., Rees, D. & Borrell, A. (2000). Identification of aircraft gas turbine dynamics using
frequency-domain techniques, Control Engineering Practice 8: 457–467.
Hagan, M. & Menhaj, M. (1994). Training feedforward networks with the Marquardt algo-
rithm, IEEE transactions on Neural Networks 5(6): 989–993.
Hornik, K., Stinchcombe, M. & White, H. (1989). Multilayer Feedforward Networks are Uni-
versal Approximators, Neural networks 2(5): 359–366.
Ismail, A. & Engelbrecht, A. (2000). Global Optimization Algorithms for Training Product Unit
Neural Networks, Proceedings of the IEEE-INNS-ENNS International Joint Conference on
Neural Networks, Vol. 1, Como, Italy, pp. 132–137.
Ljung, L. (1999). System Identification Theory for the User, 2nd edn, Prentice Hall PTR, Upper
Saddle River, NJ.
Maciejowski, J. (2002). Predictive Control with Constraints, Prentice Hall.
Mu, J. & Rees, D. (2004). Approximate model predictive control for gas turbine engines,
Proceedings of the 2004 American Control Conference, Boston, Massachusetts, USA,
pp. 5704–2709.
Nørgaard, M., Ravn, O., Poulsen, N. K. & Hansen, L. K. (2000). Neural Networks for Modelling
and Control of Dynamic Systems, Springer-Verlag, London, UK.
Pico, J. & Martinez, M. (2002). Iterative Identification and Control, Springer-Verlag London,
chapter System Identification. Performance and Closed-loop Issues.
Quanser Inc. (2005). 3 DOF Helicopter System, www.quanser.com.
Witt, J., Boonto, S. & Werner, H. (2007). Approximate model predictive control of a 3-dof he-
licopter, Proceedings of the 46th IEEE Conference on Decision and Control, New Orleans,
LA, USA, pp. 4501–4506.
Multi-objective Nonlinear Model Predictive Control: Lexicographic Method 167

x7

Multi-objective Nonlinear Model Predictive

Control: Lexicographic Method
Tao ZHENG, Gang WU, Guang-Hong LIU and Qing LING
University of Science and Technology of China
China

1. Introduction
The design of most process control is essentially a dynamic multi-objective optimization
problem (Meadowcroft et al., 1992), sometimes with nonlinear characters, and in which both
economic benefit and social benefit should be considered. Commonly speaking, there are
contradictory objectives such as quantity of products, quality of products, safety of
manufacturing, cost of manufacturing, environment protection and so on. Since the different
relative importance of these objectives cannot be ignored in the process of the controller
design, we should manage the different priority of each objective correctly and exactly.
Therefore, multivariable process control could be formulated as a complicated dynamic
multi-objective optimization problem.
Traditionally, a multi-objective control problem could be transformed into a single-objective
dynamic optimization with the quadratic objective function, where the weights denote the
different relative importance of different objectives. This method is easy to understand, but
the value of the weight coefficients usually could be only decided by try-and-error method,
based on engineering experiences, repeating simulations and other information, while there
is no accurate theoretical analysis of these weight coefficients yet. So it can be seen that, the
design process of the traditional method is complicated and time-consuming indeed.
Especially, when the situation of manufacturing changes (such as sudden load increasing of
a power supplier and so on), it is very hard for operators to renew the weights rapidly.
Therefore, a new framework of multi-objective controller is desired, it should be driven by
the relative importance of different objectives, which reflect the practical requirement of
control problems, and it also should be convenient to redesign for engineers and operators,
when the values or priorities of the objectives are changed.
Using lexicographic method, which also called completely stratified method, Meadowcroft
et al. proposed a priority-driven framework of controller: Modular Multivariable Controller
(MMC), and analyzed its steady-state properties (Meadowcroft et al., 1992). It sorts
objectives sequentially according to their relative importance, and then satisfies them as
many as possible in the corresponding control modules by the order as Fig. 1., where one
module handles with only one objective. Later, because of its advantages, researchers have
extended MMC to the dynamic optimization of linear systems with model predictive control
(MPC) and other controllers in past years (Ocampo-Martinez et al., 2008, Wu et al., 2000).
168 Model Predictive Control

Initialize
Feasible
Control Set

U0

Module 1

Control U1
Objectives System’s
& Information
Constraints Module 2

U2
...
U n 1
Module n

Un

Process

Fig. 1. Lexicographic structure of Modular Multivariable Controller

While has the mentioned advantages, the lexicographic structure still has some serious
problems. First, in this structure, the priorities of objectives are absolute and rigid, if an
objective cannot be completely satisfied (usually a objective with a setpoint form or an
extremum form), the objectives with lower priorities than it will not be considered any
more, even if they can be satisfied without any bad influence on other objectives. Second, in
some practical cases, it is hard to distinguish the difference on priorities between some
“parallel objectives”, and it is also not necessary indeed. In practical need, the number of
priorities is no need to equals to the number of objectives, it can be smaller, that means a
certain priority may have several objectives. So sometimes, the partially stratified structure
is more flexible than completely stratified structure (lexicographic structure), the number of
priorities could be determined by the essential control problem, and the objectives with
relatively lower importance can be handled in a same priority together for simplicity.
Besides the structure of the controller, the control algorithm is also important in multi-
objective control nowadays. Since the control demand of modern process industry is
heightening continuously, nonlinearity of systems cannot be ignored in controller design, to
utilize the advantages of MPC in process control, nonlinear model predictive control
(NMPC) now are developing rapidly (Alessio & Bemporad, 2009, Cannon, 2004). Naturally,
for multi-objective NMPC in many industrial cases, the priority-driven method is also
necessary. We have tried to combine lexicographic structure (or partially stratified structure)
and NMPC directly, as dynamic MMC of linear systems (Ocampo-Martinez et al., 2008, Wu
et al., 2000). But the nonlinear character makes it difficult to obtain analytic solution of
control problem, and the modular form for stratified structure seems to be too complex for
Multi-objective Nonlinear Model Predictive Control: Lexicographic Method 169

nonlinear systems in some extent. Both these facts lead us to find a new way for the
nonlinear multi-objective control problem. Genetic algorithm (GA) now is recognized as an
efficient computing means for single-objective NMPC already (Yuzgec et al., 2006), and it
also can be used to solve lexicographic optimization (Coello, 2000). So, in this chapter, a
series of dynamic coefficients are used to make up a combined fitness function of GA, which
makes GA be able to handle lexicographic optimization or partially stratified optimization
in multi-objective NMPC. It can solve the nonlinear multi-objective control problem in the
same way as MMC, but with a simple structure and much little computational load.
Since the partially stratified structure could be modified from lexicographic structure easily
(or lexicographic structure can be seen as a special case of partially stratified structure), in
this chapter, we will introduce lexicographic method as the main content, then the
corresponding content of partially stratified method can be obtained directly. The rest of this
chapter is organized as follow, Section 2 will introduce the basic theory of lexicographic
optimization and partially stratified optimization, then the modified GA for them will be
proposed in Section 3, lexicographic NMPC and partially stratified NMPC based on the
proposed GA will be studied in Section 4, using the control problem of a two-tank system as
a case study. At last, conclusions and acknowledgements will be done in Section 5.

2. Lexicographic optimization and partially stratified optimization

2.1 Lexicographic optimization
Lexicographic optimization is a strategy of multivariable optimization derived from
priority-driven thought, without loss of generality, we just considers the minimization of
multi-objective problem in this chapter.
Suppose a complex goal g  {g 1 , g 2 , , g n } contains n objectives, and the subscript also
describe the relative importance of each objective, where g 1 is the most important one and
g i 1 is always more important than g i . The solution g ( 1)  {g (11 ) , g (21) , , g (n1 ) } is better than
the solution g ( 2 )  {g (12 ) , g (22 ) , , g (n2 ) } , if and only if g (k1)  g (k2 ) and g (i1)  g (i 2 )  min g i hold
for certain k  n and all i  k . It means that, before priority k , all objectives are satisfied in
both g ( 1 ) and g ( 2 ) , but on priority k, g ( 1) is preferred to g ( 2 ) , so it is a better solution for the
whole multi-objective optimization, no matter what will be on the objectives of lower
priorities than g k . Thus the formulation of the lexicographic minimization problem can be
written as follow (Meadowcroft et al., 1992):
min g k , k  1,2 , n
s. t. g i  min g i , i  k (1)
Therefore, lexicographic optimization would be defined as the computing process of a
lexicographic minimum solution of a multi-objective problem (or sometimes maybe a
maximum solution). This solution usually is not the optimal solution of any quadratic
objective function and vice versa. As mentioned in Section 1, in lexicographic optimization,
one priority can have only one objective, so it also called completely stratified optimization.
If needed, the readers can find more about the definition of lexicographic optimization from
other references.
170 Model Predictive Control

2.2 Partially stratified optimization

Still suppose the complex goal g  {g 1 , g 2 , , g n } contains n objectives, and all objectives
need to be minimized. If these n objectives can be divided into m priorities ( m  n ), the
complex goal can be rewritten as G  {G 1 , G 2 , , G m } , where Gi    ijg ij is a combined
j

goal of a certain priority i that contains j goals, and the goals in the same priority still could
be combined with weight coefficients.
Because the relation between priorities is still lexicographic, the subscript of G i also
describes the relative importance, where G 1 is the most important and G i 1 is always more
important than G i . The solution G ( 1)  {G (11) , G (21) , , G (m1 ) } is better than the solution
G ( 2 )  {G (12 ) , G (22 ) , , G (m2 ) } , if and only if G(k1)  G(k2 ) and G(i1)  G(i 2 )  min Gi hold for certain
k  m and all i  k . Similar to de definition of lexicographic minimization in (1), the
partially stratified optimization now can be defined as the computing process of a partially
stratified minimum solution:
min G k , k  1,2 ,m
s. t. G i  min G i , i  k (2)

Simply speaking, partially stratified multi-objective optimization here means lexicographic

method between priorities and traditional weight coefficients method on goals in the same
priority. Specially, if the number of priorities equals to the number of goals ( m  n ),
partially stratified multi-objective optimization will equal to lexicographic multi-objective
optimization.

3. GA for lexicographic optimization and partially stratified optimization

3.1 GA for lexicographic optimization
In GA, the survival opportunity and competitiveness of individuals are only determined by
fitness function. So the key to a lexicographic genetic algorithm is a special fitness function,
which is suitable for lexicographic optimization for multi-objective control.
Still suppose a complex goal g  {g 1 , g 2 ,  , g n } contains n objectives, and the fitness
function of each objective is Fi  [0 ,1],1  i  n , while Fi  1 means objective i has been
completely satisfied. Since lexicographic optimization can only deal with a certain objective
when all the objectives with higher priority have been achieved already, a series of dynamic
coefficients is introduced to describe this decision procedure:
1 i1

 i  1 F1  F2   Fi 1  1,2  i  n (3)
0 otherwise

Here, since F1 is the fitness function of the most important objective, which has no objective
with higher priority than itself, so  1  1 should be held all the time. Then the combined
fitness function could be:
n
F   1F1   2 F2     n Fn    i Fi (4)
i 1
Multi-objective Nonlinear Model Predictive Control: Lexicographic Method 171

Using this lexicographic combined fitness function in multi-objective GA, the lexicographic
optimum solution can be obtained directly, and there are no special rules on coding method,
crossover operator, mutation operator or any other parameters of GA. Constraints on the
value of individuals can be matched by lethal penalty or other kinds of penalties in GA, and
to ensure the solution’s convergence to the optimal solution, the best individual should be
remained in every evolution. For the convenience to readers, we will describe the steps of
this modified GA for lexicographic optimization briefly as follow:
Step 1: create M initial parent individuals randomly.
Step 2: create M offspring individuals by crossover operator, mutation operator
with proper operation on constraints.
Step 3: compute the fitness of all the 2M individuals (parents and offspring)
respectively by (4).
Step 4: choose M individuals with higher fitness among the 2M individuals as new
parent individuals.
Step 5: if the ending condition for evolution computation is matched, output the
individual of the highest fitness, or return to Step 2.

3.2 GA for partially stratified optimization and some discussion

The only difference between GA for lexicographic optimization (LMGA) and partially
stratified GA (PSMGA) is the definition of Fi , in LMGA it is the fitness function for a single
objective, but in PSMGA it is the fitness function for all the objectives in a same priority.
Since the form of fitness fuction is depend on the problem will be solved, the fitness fuction
of NMPC based on proposed GA will be introduced in detail later in the Section 4.
Both LMGA and PSMGA are quite different from many other multi-objective GA.
Traditional multi-objective GA usually need to find out Pareto Surface (Coello, 2000), which
contains a set of Pareto optimal solutions, then choose a best solution by the given criterion.
But LMGA and PSMGA don’t need this additional selection after evolutionary computing,
since the optimal solution of the multi-objective optimization can be obtained directly. For
controller design, what we need is just an optimal solution, no matter what the Pareto
Surface is, so PSMGA and LMGA’s disposal is quite suitable and time-saving.
In LMGA and PSMGA, if the priority order of objectives changes, we only need to modify
the logical descriptions of the priorities in the combined fitness function, and if the value of
objectives changes, we only need to modify the numerical description of the combined
fitness function, while there is no parameters need to be tuned.

4. Multi-objective NMPC based on GA: a case study

4.1 The model of the two-tank system
To be used in this chapter to carry out simulations, the nonlinear model of a two-tank
system in Fig. 2. would be introduced here as (5), which is obtained by mechanism
modelling, and the sample time of this discrete system is 1 second. Here outputs
y 1 ( k ), y 2 (k ) denote the height of water in two tanks T1 and T2 respectively, and control
input u( k ) is the water fed into tank T1 from the valve V3 . The manual valve V1 and V2
are kept open at the maximal position all the time, and magnetic valve V3 is controlled by
PC to be the actuator of the system, to control the fluid speed of water from pulp P1 . (5-1)
172 Model Predictive Control

and (5-2) is the fluid mechanical character of T1 and T2 and (5-3) is the constraints on
outputs, input, and the increment of input respectively. For convenience, all the variables in
the model are normalized to the scale 0%-100%.

Fig. 2. Structure of the two-tank system

y 1 (k  1)  y 1 ( k )  0.01573  u(k )  0.2232  sign( y 1 (k )  y 2 (k ))  y 1 (k )  y 2 (k ) (5-1)

y 2 (k  1)  y 2 (k )  0.1191 y 2 (k )  0.2232  sign( y 1 ( k )  y 2 (k )) y 1 (k )  y 2 (k ) (5-2)

s. t. y 1 (k ), y 2 (k )  [ 0%,100%]
u(k )  [ 20%,80%]
u(k )  u(k  1)  u(k )  [ 5%,5%] (5-3)
where the sign function is
 1 x0
sign( x )   .
 1 x  0

4.2 The basic control problem of the two-tank system

The NMPC of the two-tank system would have two forms of objective functions, according
to two forms of practical goals in control problem: setpoint and restricted range.
For goals in the form of restricted range g : y i (k )  [ y i low , y i high ], i  1,2 , suppose the
predictive horizon contains p sample time, k is the current time and the predictive value at
time k of future output is denoted by ŷ i (|k ) , the objective function can be chosen as:
p
J(k )   [pos( ŷ i (k  j|k )  y i high )  neg( ŷ i (k  j|k )  y i low )]2 , i  1,2 (6)
j 1

where the positive function and negative function are

x x  0  0 x0
pos( x )   , neg( x )   .
 0 x  0  x x  0
In (6), if the output is in the given restricted range, the value of objective function J( k ) is
zero, which means this objective is completely satisfied.
For goals in the form of setpoint g : y i (k )  y i set , i  1,2 , since the output cannot reach the
setpoint from recent value immediately, we can use the concept of reference trajectories, and
Multi-objective Nonlinear Model Predictive Control: Lexicographic Method 173

the output will reach the set point along it. Suppose the future reference trajectories of
output y i ( k ) are w i (k ), i  1,2 , in most MPC (NMPC), these trajectories often can be set as
exponential curves as (7) and Fig. 3. (Zheng et al., 2008)
w i (k  j)   i  w i (k  j  1)  (1   i )  y i set ,1  j  p , i  1,2 (7)
where
w i (k )  y i (k ) and 0   i  1 .
Then the objective function of a setpoint goal would be:
p
J(k )   ( ŷ i ( k  j|k )  w i (k  j))2 , i  1,2 (8)
j 1

Setpoint
1

0.8
Output

0.6

alfa=0
0.4 =0.8
=0.8
=0.9
=0.9
0.2 =0.95
=0.95

0
0 50 100 150
Time
Fig. 3. Description of exponential reference trajectory

4.3 The stair-like control strategy

To enhance the control quality and lighten the computational load of dynamic optimization
of NMPC, especially the computational load of GA in this chapter, stair-like control strategy
(Wu et al., 2000) could be used here. Suppose the first unknown increment of instant control
input is u(k )  u(k )  u(k  1) , and the stair constant  is a positive real number, in stair-
like control strategy, the future control inputs could be decided as follow (Wu et al., 2000,
Zheng et al., 2008):
u( k  j)    u(k  j  1)   j  u(k ),1  j  p  1 (9)

18

16

14 data1
beta=2
data2
=1
12
data3
=0.5
Control Input

10

8
6

0
0 1 2 3 4 5
Time
Fig. 4. Description of stair-like control strategy
174 Model Predictive Control

With this disposal, the elements in the future sequence of control input
u( k ) u( k  1)  u(k  p  1) are not independent as before, and the only unknown
variable here in NMPC is the increment of instant control input u(k ) , which can determine
all the later control input. The dimension of unknown variable in NMPC now decreases
from i  p to i remarkably, where i is only the dimension of control input, thus the
computational load is no longer depend on the length of the predictive horizon like many
other MPC (NMPC). So, it is very convenient to use long predictive horizon to obtain better
control quality without additional computational load under this strategy. Because MPC
(NMPC) will repeat the dynamic optimization at every sample time, and only
u( k )  u(k )  u(k  1) will be carried out actually in MPC (NMPC), this strategy is surely
efficient here. At last, in stair-like control strategy, it also supposes the future increment of
control input will change in the same direction, which can prevent the frequent oscillation of
control input’s increment, while this kind of oscillation is very harmful to the actuators of
practical control plants. A visible description of this control strategy is shown in Fig. 4.

4.4 Multi-objective NMPC based on GA

Based on the proposed LMGA and PSMGA, the NMPC now can be established directly.
Because NMPC is an online dynamic optimal algorithm, the following steps of NMPC will
be executed repeatedly at every sample time to calculate the instant control input.
Step 1: the LMGA (PSMGA) initialize individuals as different u(k ) (with
population M) under the constraints in (5-3) with historic data u( k  1) .
Step 2: create M offspring individuals by evolutionary operations as mentioned in
the end of Section 3.1. In control problem, we usually can use real number coding, linear
crossover, stochastic mutation and the lethal penalty in GA for NMPC. Suppose P1 , P2 are
parents and O1 , O2 are offspring, linear crossover operator 0    1 and stochastic
mutation operator  is Gaussian white noise with zero mean, the operations can be
described briefly as bellow:
O 1    P1  (1   )  P2  1
,01 (10)
O 2    P2  (1   )  P1   2

Step 3: predictions of future outputs ( ŷ i (k  1|k ) ŷ i (k  2|k )  ŷ i (k  p|k ) ,

i=1,2) are carried out by (5-1) and (5-2) on all the 2M individuals (M parents and M
offspring), and their fitness will be calculated. In this control problem, the fitness function F
of each objective is transformed from its objective function J easily as follow, to meet the
value demand of F  [0 ,1] , in which J is described by (6) or (8):
F  1 ( J  1) (11)
To obtain the robustness to model mismatch, feedback compensation can be used in
prediction, thus the latest predictive errors e i ( k )  y i ( k )  ŷ i (k|k  1), i  1,2 should be added
into every predictive output ŷ i (k  j|k ),1  1,2, 1  j  p .
Step 4: the M individuals with higher fitness in the 2M individuals will be
remained as new parents.
Multi-objective Nonlinear Model Predictive Control: Lexicographic Method 175

Step 5: if the condition of ending evaluation is met, the best individual will be the
increment of instant control input u(k ) of NMPC, which is taken into practice by the
actuator. Else, the process should go back to Step 2, to resume dynamic optimization of
NMPC based on LMGA (PSMGA).

4.5 Simulations and analysis of lexicographic multi-objective NMPC

First, the simulation about lexicographic Multi-objective NMPC will be carried out. Choose
control objectives as: g 1 : y 1 ( k )  [ 40%,60%] , g 2 : y 2 (k )  [ 20%, 40%] , g 3 : y 2  30% . Consider
the physical character of the system, two different order of priorities can be chosen as: [A]:
g 1  g 2  g 3 , [B]: g 2  g 1  g 3 , and they will have the same initial state as y 1 (0 )  80% ,
y 2 (0 )  0% and u( 0 )  20% . Parameters of NMPC are   0.95 ,   0.85 for both y 1 and y 2 ,
and parameters of GA are   0.9 , while  is a zero mean Gaussian white noise, whose variance
is 5. Since the feasible control input set is relatively small in our problem according to constraints
(5-3), it is enough to have only 10 individuals in our simulation, and they will evolve for 20
generations. While in process control practice, because the sample time is often has a time scale of
minutes, even hours, we can have much more individuals and they can evolve much more
generations to get a satisfactory solution. (In following figures, dash-dot lines denote g 1 , g 2 , dot
line denote g 3 and solid lines denote y 1 , y 2 , u )
Compare Fig. 5. and Fig. 6. with Fig. 7. and Fig. 8., although the steady states are the same in
these figures, the dynamic responses of them are with much difference, and the objectives
are satisfied as the order appointed before respectively under all the constraints. The reason
of these results is the special initial state: y 1 (0 ) is higher than g 1 (the most important
objective in order [A]: g 1  g 2  g 3 ), while y 2 (0 ) is lower than g 2 (the most important
objective in order [B]: g 2  g 1  g 3 ). So the most important objective of the two orders must
be satisfied with different control input at first respectively. Thus the difference can be seen
from the different decision-making of the choice in control input more obviously: in Fig. 5.
and Fig. 6. the input stays at the lower limit of the constraints at first to meet g 1 , while in
Fig. 7. and Fig. 8. the input increase as fast as it can to satisfy g 2 at first. The lexicographic
character of LMGA is verified by these comparisons.

80%
60%
Y1

40%
20%
0 20 40 60 80 100
60%
40%
Y2

20%
0%
0 20 40 60 80 100
100%

50%
U

0%
0 20 40 60 80 100
Time (second)
Fig. 5. Control simulation: priority order [A] and p=1
176 Model Predictive Control

80%
60%

Y1
40%
20%
0 20 40 60 80 100
60%
40%

Y2
20%
0%
0 20 40 60 80 100
100%

50%
U

0%
0 20 40 60 80 100
Time (second)
Fig. 6. Control simulation: priority order [A] and p=20

80%
60%
Y1

40%
20%
0 20 40 60 80 100
60%
40%
Y2

20%
0%
0 20 40 60 80 100
100%

50%
U

0%
0 20 40 60 80 100
Time (second)
Fig. 7. Control simulation: priority order [B] and p=1
80%
60%
Y1

40%
20%
0 20 40 60 80 100
60%
40%
Y2

20%
0%
0 20 40 60 80 100
100%

50%
U

0%
0 20 40 60 80 100
Time (second)
Fig. 8. Control simulation: priority order [B] and p=20

And the difference in control input with different predictive horizon can also be observed
from above figures: the control input is much smoother when the predictive horizon
Multi-objective Nonlinear Model Predictive Control: Lexicographic Method 177

becomes longer, while the output is similar with the control result of shorter predictive
horizon. It is the common character of NMPC.

100%
80%

Y1
60%
40%
0 20 40 60 80 100
60%
40%
Y2

20%
0%
0 20 40 60 80 100
100%

50%
U

0%
0 20 40 60 80 100
Time (second)
Fig. 9. Control simulation: when an objective cannot be satisfied

In Fig. 9., g 1 is changed as y 1  [60%,80%] , while other objectives and parameters are kept
the same as those of Fig. 6., so that g 3 can’t be satisfied at steady state. The result shows that
y 1 will stay at lower limit of g 1 to reach set-point of g 3 as close as possible, when g 1 must
be satisfied first in order [A]. This result also shows the lexicographic character of LMGA
obviously.

80%
60%
Y1

40%
20%
0 20 40 60 80 100
60%
40%
Y2

20%
0%
0 20 40 60 80 100
100%

50%
U

0%
0 20 40 60 80 100
Time (second)
Fig. 10. Control simulation: when model mismatch

Finally, we would consider about of the model mismatch, here the simulative plant is
changed, by increasing the flux coefficient 0.2232 to 0.25 in (5-1) and (5-2), while all the
objectives, parameters and predictive model are kept the same as those of Fig. 6. The result
in Fig. 10. shows the robustness to model mismatch of the controller with error
compensation in prediction, as mentioned in Section 4.4.
178 Model Predictive Control

4.6 Simulations and analysis of partially stratified multi-objective NMPC

To obtain evident comparison to Section 4.5, simulations are carried out with the same
parameters (   0.95,   0.85 for both y 1 and y 2 , predictive horizon p=20 and the same
GA parameters), and the only difference is an additional objective on y 1 in the form of a
setpoint.
The four control objectives now are g 1 : y 1 (k )  [ 40%,60%] , g 2 : y 2 (k )  [ 20%, 40%] ,
g 3 : y 2  30% , g 4 : y 1  50% , and then choose the new different order of priorities as: [A]:
g 1  g 2  g 3  g 4 , [B]: g 2  g 1  g 3  g 4 , if we still use lexicographic multi-objective NMPC
as Section 4.5, the control result in Fig. 11. and Fig. 12. is completely the same as Fig. 6. and
Fig. 8., when there are only three objectives g 1，g 2，g 3 . That means, the additional
objective g 4 (setpoint of y 1 ) could not be considered by the controller in both situations
above, because the solution of g 3 (setpoint of y 2 ) is already a single-point set of u . (In
following figures, dash-dot lines denote g 1 , g 2 , dot line denote g 3 , g 4 and solid lines denote
y1 , y2 , u )

80%
60%
Y1

40%
20%
0 20 40 60 80 100
60%
40%
Y2

20%
0%
0 20 40 60 80 100
100%

50%
U

0%
0 20 40 60 80 100
Time (second)
Fig. 11. Control simulation: priority order [A] of four objectives, NMPC based on LMGA

80%
60%
Y1

40%
20%
0 20 40 60 80 100
60%
40%
Y2

20%
0%
0 20 40 60 80 100
100%

50%
U

0%
0 20 40 60 80 100
Time (second)
Fig. 12. Control simulation: priority order [B] of four objectives, NMPC based on LMGA
Multi-objective Nonlinear Model Predictive Control: Lexicographic Method 179

In another word, in lexicographic multi-objective NMPC based on LMGA, if optimization of

an objective uses out all the degree of freedom on control inputs (often an objective in the
form of setpoint), or an objective cannot be completely satisfied (often an objective in the
form of extremum, such as minimization of cost that can not be zero), the objectives with
lower priorities will not be take into account at all. But this is not rational in most practice
cases, for complex process industrial manufacturing, there are often many objectives in the
form of setpoint in a multi-objective control problem, if we handle them with the
lexicographic method, usually, we can only satisfy only one of them. Take the proposed
two-tank system as example, g 3 and g 4 are both in the form of setpoint, seeing about the
steady-state control result in Fig. 13. and Fig. 14., if we want to satisfy g 3 : y 2  30% , then
y 1 will stay at 51.99%, else if we want to satisfy g 4 : y 1  50% , then y 2 will stay at 28.92%,
the errors of the dissatisfied output are both more than 1%.

70%
60%
Y1

50%
40%
30%
100 110 120 130 140 150

50%
40%
30%
Y2

20%
10%
100 110 120 130 140 150
Time (second)
Fig. 13. Steady-state control result when g 3 is completely satisfied
70%
60%
50%
Y1

40%
30%
100 110 120 130 140 150

50%
40%
30%
Y2

20%
10%
100 110 120 130 140 150
Time (second)
Fig. 14. Steady-state control result when g 4 is completely satisfied

In the above analysis, the mentioned disadvantage comes from the absolute, rigid
management of lexicographic method, if we don’t develop it, NMPC based on LMGA can
only be used in very few control practical problem. Actually, in industrial practice,
objectives in the form of setpoint or extremum are often with lower importance, they are
usually objectives for higher demand on product quality, manufacturing cost and so on,
180 Model Predictive Control

which is much less important than the objectives about safety and other basic
manufacturing demand. Especially, for objectives in the form of setpoint, under many kinds
of disturbances, it always can not be accurately satisfied while it is also not necessary to
satisfy them accurately.
A traditional way to improve it is to add slack variables into objectives in the form of setpoint or
extremum. Setpoint may be changed into a narrow range around it, and instead of an extremum,
the satisfaction of a certain threshold value will be required. For example, in the two-tank
system’s control problem, setpoint g 3 : y 2  30% could be redefined as g 3 : y 2  30%  1% .
Another way is modified LMGA into PSMGA as mentioned in Section 3, because sometimes
there is no need to divide these objectives with into different priorities respectively, and they are
indeed parallel. Take order [A] for example, we now can reform the multi-objective control
problem of the two-tank system as: G 1  G 2  G 3  g 1  g 2   3g 3   4g 4 . Choose weight
coefficients as  3  30 ,  4  1 and other parameters the same as those of Fig. 6., while NMPC
base on PSMGA has the similar dynamic state control result to that of NMPC based on LMGA,
the steady state control result is evidently developed as in Fig. 15. and Fig. 16., y 1 stays at
50.70% and y 2 stays at 29.27%, both of them are in the 0.8% neighborhood of setpoint in g 3 , g 4 .

80%
60%
Y1

40%
20%
0 20 40 60 80 100
60%
40%
Y2

20%
0%
0 20 40 60 80 100
100%

50%
U

0%
0 20 40 60 80 100
Time (second)
Fig. 15. NMPC based on PSMGA: priority order [A]

70%
60%
50%
Y1

40%
30%
100 110 120 130 140 150

50%
40%
30%
Y2

20%
10%
100 110 120 130 140 150
Time (second)
Fig. 16. Steady-state control result of NMPC base on PSMGA
Multi-objective Nonlinear Model Predictive Control: Lexicographic Method 181

4.7 Some discussions

In the above simulative examples, there is only one control input, but for many practical
systems, coordinated control of multi-input is also a serious problem. The brief discussions
on multi-input proposed NMPC can be achieved if we still use priorities for inputs. If all the
inputs have the same priority, in another word, it is no obvious difference among them in
economic cost or other factors, we can just increase the dimension of GA’s individual. But,
in many cases, the inputs actually also have different priorities: for certain output, different
input often has different gain on it with different economic cost. The cheap ones with large
gain are always preferred by manufacturers. In this case, we can form anther priority list,
and then inputs will be used to solve the control problem one by one, using single input
NMPC as the example in Section 4, that can divide an MIMO control problem into some
SIMO control problems.
We should point out that, the two kinds of stratified structures proposed in this paper are
basic structures for multi-objective controllers, though we use NMPC to realize them in this
chapter, they are independent with control algorithms indeed. For certain multi-objective
control problem, other proper controllers and computational method can be used.
Another point must be mentioned is that, NMPC proposed in this paper is based on LMGA
and PSMGA, because it is hard for most NMPC to get an online analytic solution. But the
LMGA and PSMGA are also suitable for other control algorithms, the only task is to modify
the fitness function, by introducing the information from the control algorithm which will
be used.
At last, all the above simulations could been done in 40-200ms by PC (with 2.7 GHz CPU,
2.0G Memory and programmed by Matlab 6.5), which is much less than the sample time of
the system (1 second), that means controllers proposed in this chapter are actually
applicable online.

5. Conclusion
In this chapter, to avoid the disadvantages of weight coefficients in multi-objective dynamic
optimization, lexicographic (completely stratified) and partially stratified frameworks of
multi-objective controller are proposed. The lexicographic framework is absolutely priority-
driven and the partially stratified framework is a modification of it, they both can solve the
multi-objective control problem with the concept of priority for objective’s relative
importance, while the latter one is more flexible, without the rigidity of lexicographic
method.
Then, nonlinear model predictive controllers based on these frameworks are realized based
on the modified genetic algorithm, in which a series of dynamic coefficients is introduced to
construct the combined fitness function. With stair–like control strategy, the online
computational load is reduced and the performance is developed. The simulative study of a
two-tank system indicates the efficiency of the proposed controllers and some deeper
discussions are given briefly at last.
The work of this chapter is supported by Fund for Excellent Post Doctoral Fellows (K. C.
Wong Education Foundation, Hong Kong, China and Chinese Academy of Sciences),
Science and Technological Fund of Anhui Province for Outstanding Youth (08040106910),
and the authors also thank for the constructive advices from Dr. De-Feng HE, College of
Information Engineering, Zhejiang University of Technology, China.
182 Model Predictive Control

6. References
Alessio A. & Bemporad A. (2009). A survey on explicit model predictive control. Lecture
Notes in Control and Information Sciences (Nonlinear Model Predictive Control:
Towards New Challenging Applications), Vol. 384, pp 345-369, ISSN 0170-8643.
Cannon M. (2004). Efficient nonlinear model predictive control algorithms. Annual Reviews
in Control. Vol.28, No.2, pp229-237, ISSN 1367-5788.
Coello C. A. C. (2000). An Updated Survey of GA-Based Multiobjective Optimization
Techniques, ACM Computing Surveys, Vol.32, No.2, pp109-143, ISSN 0360-0300.
Meadowcroft T. A.; Stephanopoulos G. & Brosilow C. (1992). The modular multivariable
controller: I: steady-state properties. AIChE Journal, Vol.38, No.8, pp1254-1278,
ISSN 0001-1541.
Ocampo-Martinez C.; Ingimundarson A.; Vicenç P. & J. Quevedo. (2008). Objective
prioritization using lexicographic minimizers for MPC of sewer networks. IEEE
Transactions on Control Systems Technology, Vol. 16, No.1, pp113-121, ISSN 1063-
6536.
Wu G.; Lu X. D.; Ying A. G.; Xue M. S.; Zhang Z. G. & Sun D. M. (2000). Modular
Multivariable Self-tuning Regulator. Acta Automatica Sinica, Vol.26, No.6, pp811-
815, ISSN 0254-4156.
Yuzgec U.; Becerikli Y. & Turker M. (2006). Nonlinear Predictive Control of a Drying Process
Using Genetic Algorithms, ISA Transactions, Vol.45, No.4, pp589-602, ISSN 0019-
0578.
Zheng T.; Wu G.; He D. F.; Yue D. Z. (2008). An Efficient Model Nonlinear Predictive
Control Algorithm Based on Stair-like Control Strategy, Proceedings of the 27th
Chinese Control Conference, Vol.3, pp557-561, ISBN 9787811243901, Kunming,
China, July, 2008, Beihang University Press, Beijing, China.
Model Predictive Trajectory Control for High-Speed Rack Feeders 183

0
8

Model Predictive Trajectory Control

for High-Speed Rack Feeders
Harald Aschemann and Dominik Schindele
Chair of Mechatronics, University of Rostock
18059 Rostock, Germany

1. Introduction
Rack feeders represent the commonly used handling systems for the automated operation of
high-bay rackings. To further increase the handling capacity by shorter transport times, con-
trol measures are necessary for the reduction of excited structural oscillations, see also Asche-
mann & Ritzke (2009). One possible approach is given by flatness-based feedforward control,
where the desired control inputs are determined by dynamic system inversion using the de-
sired trajectories for the flat outputs as in Bachmayer et al. (2008) and M. Bachmayer & Ulbrich
(2008). However, both publications consider only a constant mass position in vertical direc-
tion on an elastic beam without any feedback control. A variational approach is presented in
Kostin & Saurin (2006) to compute an optimal feedforward control for an elastic beam. Unfor-
tunately, feedforward control alone is not sufficient to guarantee small tracking errors when
model uncertainty is present or disturbances act on the system. For this reason in this con-
tribution a model predictive control (MPC) design is presented for fast trajectory control. In
general, at model predictive control the optimal input vector is mostly calculated by minimis-
ing a quadratic cost function as, e.g., in Wang & Boyd (2010) or Magni & Scattolini (2004). In
contrast, the here considered MPC approach aims at reducing future state errors, see Jung &
Wen (2004), and allows for a relatively small computational effort as required in a real-time
implementation. Hence, the proposed MPC algorithm is well suited for systems with fast
dynamics, e.g., a high-speed linear axis with pneumatic muscles as presented in Schindele &
Aschemann (2008) or high-speed rack feeders as in the given case. A further attractive char-
acteristic of this MPC approach is its applicability to linear as well as nonlinear systems.
For the experimental investigation of modern control approaches to active oscillation damp-
ing as well as tracking control, a test rig of a high-speed rack feeder has been build up at the
Chair of Mechatronics at the University of Rostock, see Figure 1. The experimental set-up
consists of a carriage driven by an electric DC servo motor via a toothed belt, on which an
elastic beam as the vertical supporting structure is mounted. On this beam structure, a cage
with variable load mass is guided relocatably in vertical direction. This cage with the coor-
dinate yK (t) in horizontal direction and xK (t) in vertical direction represents the tool center
point (TCP) of the rack feeder that should track desired trajectories as accurate as possible.
The movable cage is driven by a tooth belt and an electric DC servo motor as well. The angles
of the actuators are measured by internal angular transducers, respectively. Additionally, the
horizontal position of the carriage is detected by a magnetostrictive transducer. Both axes are
operated with a fast underlying velocity control on the current converter. Consequently, the
184 Model Predictive Control

Fig. 1. Experimental set-up of the high-speed rack feeder (left) and the corresponding elastic
multibody model (right).

corresponding velocities deal as new control input, and the implementational effort is tremen-
dously reduced as compared to the commonly used force or torque input, like in Staudecker
et al. (2008), where passivity techniques were employed for feedback control of a similar set-
up. Two strain gauges are used to determine the bending deformation of the elastic beam.
Basis of the control design for the rack feeder is a planar elastic multibody system, where
for the mathematical description of the bending deflection of the elastic beam a Ritz ansatz
is introduced, covering for instance the first bending mode. The decentralised feedforward
and feedback control design for both axes is performed employing a linearised state space
representation, respectively. Given couplings between both axes are taken into account by the
gain-scheduling technique with the normalised vertical cage position as scheduling param-
eter, see also Aschemann & Ritzke (2010). This leads to an adaptation of the whole control
structure for the horizontal axis. The capability of the proposed control concept is shown by
experimental results from the test set-up with regard to tracking behaviour and damping of
bending oscillations. Especially the artificial damping introduced by the closed control loop
represents a main improvement. The maximum velocity of the TCP during the tracking ex-
periments is approx. 2.5 m/s.

2. Control-oriented modelling of the mechatronic system

Elastic multibody models have proven advantageously for the control-oriented modelling of
flexible mechanical systems. For the feedforward and feedback control design of the rack
feeder a multibody model with three rigid bodies - the carriage (mass mS ), the cage movable
on the beam structure (mass mK , mass moment of inertia θK ), and the end mass at the tip of
the beam (mass m E ) - and an elastic Bernoulli beam (density ρ, cross sectional area A, Youngs
modulus E, second moment of area IzB , and length ) is chosen. The varying vertical position
Model Predictive Trajectory Control for High-Speed Rack Feeders 185

xK (t) of the cage on the beam is denoted by the dimensionless system parameter
xκ ( t )
κ (t) = . (1)
l
The elastic degrees of freedom of the beam concerning the bending deflection can be described
by the following Ritz ansatz
   
3 x 2 1  x 3
v ( x, t) = v̄¯1 ( x ) v1 (t) = − v1 ( t ) , (2)
2 l 2 l
which takes into account only the first bending mode. The vector of generalised coordinates
results in  
yS (t)
q (t) = . (3)
v1 ( t )
The nonlinear equations of motion can be derived either by Lagrange’s equations or, advan-
tageously, by the Newton-Euler approach, cf. Shabana (2005). After a linearisation for small
bending deflections, the equations of motion can be stated in M-D-K form
M q̈ (t) + D q̇ (t) + Kq (t) = h · [ FSM (t) − FSR (ẏS (t))] . (4)
The symmetric mass matrix is given by
 
3 mK κ 2
mS + ρAl + mK + m E 8 ρAl + 2 [3 − κ ] + m E
M = 2 , (5)
3
8 ρAl + mK2κ [3 − κ ] + m E m22
 
6ρI 2 2
ρAl + 5lzB + mK4κ [3 − κ ]2 + 9θlK2κ 1 − κ + κ4 + m E . The damping matrix,
2
33
with m22 = 140
which is specified with stiffness-proportional damping properties, and the stiffness matrix
become
 
0 0
D = , (6)
0 3kdlEI 3
zB

 
0 0 
K = 3EIzB 3mK gκ 3 2 6m E g . (7)
0 l3
− 38 ρAg − l 1+ 3κ
20 − 3κ
4 − 5l

The input vector of the generalised forces, which accounts for the control input as well as the
disturbance input, reads
 T
h = 1 0 . (8)
The electric drive for the carriage is operated with a fast underlying velocity control on the
current converter. The resulting dynamic behaviour is characterised by a first-order lag system
with a time constant T1y
T1y ÿS (t) + ẏS (t) = vS (t) . (9)
This differential equation replaces now the equation of motion for the carriage in the mechan-
ical system model, which leads to a modified mass matrix as well as a modified damping
matrix
 
T1y 0
My = 3 mK κ 2 , (10)
8 ρAl + 2 [3 − κ ] + m E m22
186 Model Predictive Control

 
1 0
Dy = 3k d EIzB . (11)
0 l3

The stiffness matrix K = Ky and the input vector for the generalised forces h = hy , however,
remain unchanged. Hence, the equations of motion are given by

q̈ = −My−1 Ky q − My−1 Dy q̇ + My−1 hy vS . (12)

For control design, the system representation is reformulated in state space form
      
q̇ 0 I q 0
ẋy = = + vS . (13)
q̈ −My−1 Ky −My−1 Dy q̇ My−1 hy 
         uy
Ay xy by

The design model for the vertical movement of the cage can be directly stated in state space
representation. Here, an underlying velocity control is employed on the current converter,
which is also described by a first-order lag system

T1x ẍK (t) + ẋK (t) = vK (t) . (14)

The corresponding state space description follows immediately in the form

      
ẋK 0 1 xK 0
ẋx = = + vK . (15)
ẍK 0 − T11x ẋK 1
T1x 
         ux
Ax xx bx

Whereas the state space respresentation for the horizontal y-axis depends on the varying sys-
tem parameter κ (t), the description of the x-axis is invariant. A gain-scheduling, hence, is
necessary only for the horizontal axis in y-direction.

3. Decentralised control design

As for control, a decentralised approach is followed, at which the coupling of the vertical
cage motion with the horizontal axis is taken into account by gain-scheduling techniques. For
the control of the cage position xK (t) a simple proportional feedback in combination with
feedforward control, which is based on the inverse transfer function of this axis, is sufficient

vK (t) = K R ( xKd (t) − xK (t)) + ẋKd (t) + T1x ẍKd (t) . (16)

For this purpose, the desired trajectory xKd (t) and its first two time derivatives are available
from trajectory planning. The design of the state feedback for the horizontal motion is carried
out by the MPC approach, which is explained in the following chapter.
Model Predictive Trajectory Control for High-Speed Rack Feeders 187

Horizontal Axis
[]
v1
v̇ 1
Real
Differentiation

[]
y Sd
ẏ Sd Real

[]
y Kd v 1d Differentiation
ẏ Kd
w y = ÿ Kd Inverse
v̇ 1d
[]
yS
ẏ S

Hight-Speed Rack Feeder

y Kd Kinematics
Gain-Scheduled
y 4 Kd
v Sd
Model
Inverse Predictive
Dynamics v S t
Control
d

x K t 
Vertical Axis

[] []
x Kd d Proportional
wx = ẋ Kd ̇ d Feedback
ẍ Kd ̈ d
v K t 
Feedforward
Normalisation
Control

Fig. 2. Implementation of the control structure.

4. Model Predictive Control

The main idea of the control approach consists in a minimisation of a future tracking error
in terms of the predicted state vector based on the actual state and the desired state vector
resulting from trajectory planning, see Lizarralde et al. (1999), Jung & Wen (2004). The min-
imisation is achieved by repeated approximate numerical optimisation in each time step, in
the given case using the Newton-Raphson technique. The optimisation is initialised in each
time step with the optimisation result of the preceding time step in form of the input vector.
The MPC-algorithm is based on the following discrete-time state space representation

xk +1 = Axk + buk , (17)

T
yk = c xk , (18)

with the state vector xk ∈ R n , the control input uk ∈ R and the output vector yk ∈ R.
The constant M specifies the prediction horizon TP as a multiple of the sampling time ts , i.e.
TP = M · ts . The predicted input vector at time k becomes
 
(k) (k) T
uk,M = u1 , ..., u M , (19)
188 Model Predictive Control

with uk,M ∈ R M . The predicted state vector at the end of the prediction horizon φ M (xk , uk,M )
is obtained by repeated substitution of k by k + 1 in the discrete-time state equation (17)

xk+2 = Axk+1 + buk+1 = A2 xk + Abuk + buk+1

..
. (20)
xk+ M = A M xk + A M−1 buk + A M−2 buk−1 + . . . + buk+ M−1
= φ M (xk , uk,M ) .
The difference of φ M (xk , uk,M ) and the desired state vector xd yields the final control error

e M,k = φ M (xk , uk,M ) − xd , (21)

i.e. to the control error at the end of the prediction horizon. The cost function to be minimised
follows as
1
J MPC = · eTM,k e M,k , (22)
2
and, hence, the necessary condition for an extremum can be stated as
∂J MPC !
= e M,k = 0 . (23)
∂e M,k
A Taylor-series expansion of (23) at uk,M in the neighbourhood of the optimal solution leads
to the following system of equations
∂φ M
0 = e M,k + ∆uk,M + T.h.O.. (24)
∂uk,M
The vector ∆uk,M denotes the difference which has to be added to the input vector uk,M to
obtain the optimal solution. The n equations (24) represent an under-determined set of equa-
tions with m · M unknowns having an infinite number of solutions. An unique solution for
∆uk,M can be determined by solving the following L2 -optimisation problem with (24) as side
condition  
1 T ∂φ M
J = · ∆uk,M ∆uk,M + λT e M,k + ∆uk,M . (25)
2 ∂uk,M
Consequently, the necessary conditions can be stated as
 
∂J ! ∂φ M T
= 0 = ∆uk,M + λ,
∂∆uk,M ∂uk,M
(26)
∂J ! ∂φ M
= 0 = e M,k + ∆uk,M ,
∂λ ∂uk,M
which results in e M,k
 
∂φ M ∂φ M T
e M,k = λ. (27)
∂uk,M uk,M
  
S (φ M ,uk,M )
 
If the matrix S φ M , uk,M is invertible, the vector λ can be calculated as follows
 
λ = S −1 φ M , uk,M e M,k . (28)
Model Predictive Trajectory Control for High-Speed Rack Feeders 189

 
An almost singular matrix S φ M , uk,M can be treated by a modification of (28)
  −1
λ = µI + S φ M , uk,M e M,k , (29)

where I denotes the unity matrix. The regularisation parameter µ > 0 in (29) may be chosen
constant or may be calculated by a sophisticated algorithm. The latter solution improves
the convergence of the optimisation but increases, however, the computational complexity.
Solving (26) for ∆uk,M and inserting λ according to (28) or (29), directly yields the L2 -optimal
solution
   
∂φ M T −1   ∂φ M †
∆uk,M = − S φ M , uk,M e M,k = − e M,k . (30)
∂uk,M ∂uk,M
 †
∂φ ∂φ
Here, ∂u M denotes the Moore-Penrose pseudo inverse of ∂u M . The overall MPC-
k,M k,M
algorithm can be described as follows:
Choice of the initial input vector u0,M at time k = 0, e.g. u0,M = 0, and repetition of steps a) -
c) at each sampling time k ≥ 0:
a) Calculation of an improved input vector vk,M according to
 †
∂φ M
vk,M = uk,M − ηk e M,k . (31)
∂uk,M

The step width ηk can be determined with, e.g., the Armijo-rule.

b) For the calculation of uk+1,M the elements of the vector vk,M have to be shifted by one
element and the steady-state input ud corresponding to the final state has to be inserted
at the end
   
0(( M−1)×1) 0(( M−1)×1) I( M −1)
uk+1,M = ud + vk,M . (32)
1 0 0(1×( M−1))

In general, the steady-state control input ud can be computed from

xd = Axd + bud . (33)

Alternatively, the desired input vector ud can be calculated by an inverse system model.
If the system is differentially flat, see Fliess et al. (1995) the desired input ud can be cal-
culated exactly by the flat system output and a finite number of its time derivative. For
non-flat outputs -as in the given case- the approach presented in chapter 4.4 is useful.
c) The first element of the improved input vector vk,M is applied as control input at time k
 
uk = 1 0(1×( M−1)) vk,M . (34)

In the proposed algorithm only one iteration is performed per time step. A similar approach
using several iteration steps is described in Weidemann et al. (2004). An improvement of
the trajectory tracking behaviour can be achieved if an input vector resulting from an inverse
system model is used as initial vector for the subsequent optimisation step instead of the last
input vector. The slightly modified algorithm can be stated as follows
190 Model Predictive Control

(d)
a) Calculation of the ideal input vector uk,M by evaluating an inverse system model with
the specified reference trajectory as well as a certain number β ∈ N of its time deriva-
tives  
(d) (d) ( β)
uk,M = uk,M yd , ẏd , ..., y d . (35)

b) Calculation of the improved input vector vk,M based on the equation

 †
(d) ∂φ M
vk,M = uk,M − ηk e M,k . (36)
∂uk,M

c) Application of the first element of vk,M to the process

 
uk = 1 0(1×( M−1)) vk,M . (37)

(d)
If the reference trajectory is known in advance, the according reference input vector uk,M can
be computed offline. Consequently, the online computational time remains unaffected.

4.1 Numerical calculations

∂φ
The analytical computation of the Jacobian ∂u M becomes increasingly complex for larger
k,M
values of M. Therefore, a numerical approach is preferred taking advantage of the chain rule
with i = 0, ..., M − 1
∂φ M ∂φ M ∂x ∂x ∂x i+1
(k)
= · k + M −1 · . . . · k + i +2 · k +
(k)
. (38)
∂ui+1 ∂xk+ M−1 ∂xk+ M−2 ∂xk+i+1 ∂ui+1

In this way, the Jacobian can be computed as follows

∂φ M
= [A M−1 b, A M−2 b, ..., Ab, b] . (39)
∂uk,M
 
For the inversion of the symmetric and positive definite matrix S φ M , uk,M =
 T
∂φ M ∂φ M
∂uk,M ∂uk,M the Cholesky-decomposition has proved advantageous in terms of compu-
tational effort.

4.2 Choice of the MPC design parameters

The most important MPC design parameter is the prediction horizon TP , which is given as the
product of the sampling time ts and the constant value M. Large values of TP lead to a slow
and smooth transient behaviour and result in a robust and stable control loop. For fast trajec-
tory tracking, however, a smaller value TP is desirable concerning a small tracking error. The
choice of the sampling time ts is crucial as well: a small sampling time is necessary regard-
ing both discretisation error and stability; however, the MPC-algorithm has to be evaluated
in real-time within the sampling inverval. Furthermore, the smaller ts , the larger becomes M
for a given prediction horizon, which in turn increases the computational complexity of the
optimisation step. Consequently, a system-specific trade-off has to be made for the choice of
M and ts . This paper follows the moving horizon approach with a constant prediction hori-
zon and, hence, a constant dimension m · M of the corresponding optimisation problem in
constrast to the shrinking horizon approach according to Weidemann et al. (2004).
Model Predictive Trajectory Control for High-Speed Rack Feeders 191

Desired Trajectory
e M,1
x d,1
φM(x1 , u1,M )
x d,0 e M,0
TP

φM(x0 , u 0,M )
Predicted State

x0 x1

ts M ts t
Fig. 3. Design parameters.

4.3 Input constraints

One major advantage of predictive control is the possibility to easily account for input con-
straints, which are present in almost all control applications. To this end, the cost function can
be extended with a corresponding term
 (k)

 0 umin ≤ u j ≤ umax

(k) (k) (k)
h(u j ) = g1 (u j ) for uj > umax , (40)


 (k) (k)
g2 ( u j ) uj < umin

which has to be evaluated componentwise, i.e. for each input variable at each sampling time.
Thus, the contribution of the additional input constraints depending on uk,M is given by

M
(k)
z(uk,M ) = ∑ h(u j ). (41)
j =1

 T
Instead of e M,k , the extended vector eTM,k , z has to be minimised in the MPC-algorithm.

4.4 MPC of the horizontal cage position

The state space representation for the cage position control in y-direction design is given by
(13). The discrete-time representation of the continous-time system is obtained by Euler dis-
cretisation  
xy,k+1 = I + ts · Ay (κ ) xy,k + ts · by (κ ) uy,k . (42)
192 Model Predictive Control

Using this simple discretisation method, the computational effort for the MPC-algorithm can
be kept acceptable. By the way, no significant improvement could be obtained for the given
system with the Heun discretisation method because of the small sampling time ts = 3 ms.
Only in the case of large sampling times, e.g. ts > 20 ms, the increased computational effort
caused by a sophisticated time discretisation method is advantageous. Then, the smaller dis-
cretisation error allows for less time integration steps for a specified prediction horizon, i.e. a
smaller number M. As a result, the smaller number of time steps can overcompensate the
larger effort necessary for a single time step.
The ideal input ud (t) can be obtained in continous time as function of the output variable
 
yK (t) = cyT xy (t) = 1 12 κ 2 (3 − κ ) 0 0 xy (t) , (43)

and a certain number of its time derivatives. For this purpose the corresponding transfer
function of the system under consideration is employed
 
YK (s)   −1 b0 + b1 · s + b2 · s2
= cyT sI − Ay by = . (44)
Ud (s) N (s)

Obviously, the numerator of the control transfer function contains a second degree polynomial
in s, leading to two transfer zeros. This shows that the considered output yK (t) represents a
non-flat output variable that makes computing of the feedforward term more difficult. A pos-
sible way for calculating the desired input variable is given by a modification of the numerator
of the control transfer function by introducing a polynomial ansatz for the feedforward action
according to  
Ud (s) = kV0 + kV1 · s + . . . + kV4 · s4 YKd (s) . (45)
For its realisation the desired trajectory yKd (t) as well as the first four time derivatives are
available from a trajectory planning module. The feedforward gains can be computed from
a comparison of the corresponding coefficients in the numerator as well as the denominator
polynomials of
  
YK (s) b0 + . . . + b2 · s2 kV0 + . . . + kV4 · s4
=
YKd (s) N (s)
     
bV0 kV j + bV1 kV j · s + . . . + bV6 kV j · s6
= (46)
a0 + a1 · s + . . . + s4
according to  
ai = bVi kV j , i = 0, . . . , n = 4 . (47)
This leads to parameter-dependent feedforward gains kV j = kV j (κ ). It is obvious that due
the higher numerator degree in the modified control transfer function a remaining dynamics
must be accepted. Lastly, the desired input variable in the time domain is represented by
 ... 
(4)
ud (t) = ud ẏKd (t), ÿKd (t), y Kd (t), yKd (t), κ . (48)

To obtain the desired system states as function of the output trajectory the output equation
Model Predictive Trajectory Control for High-Speed Rack Feeders 193

0.8
yKd
0.8
y
K
0.6
0.6
yK in m

x in m
0.4
0.4

K
0.2
0.2 xKd
xK
0 0
0 5 10 15 0 5 10 15
t in s t in s

2 1.5

1
1
0.5
yKpd in m/s

xKpd in m/s
0 0

−0.5
−1
−1

−2 −1.5
0 5 10 15 0 5 10 15
t in s t in s

Fig. 4. Desired trajectories for the cage motion: desired and actual position in horizontal
direction (upper left corner), desired and actual position in vertical direction (upper right
corner), actual velocity in horizontal direction (lower left corner) and actual velocity in vertical
direction (lower right corner).

and its first three time derivatives are considered. Including the equations of motion (12)
yields the following set of equations

yKd (t) = yS (t) + 12 κ 2 (3 − κ ) · v1 (t), (49)

ẏKd (t) = ẏS (t) + 12 κ 2 (3 − κ ) · v̇1 (t), (50)
1 2
ÿKd (t) = ÿS (t) + (3 − κ ) · v̈1 (t) = ÿK (v1 (t), ẏS (t), v̇1 (t), ud (t), κ ) ,
2κ (51)
... ...
y Kd (t) = y K (v1 (t), ẏS (t), v̇1 (t), ud (t), u̇d (t), κ ) . (52)

Solving equation (49) to (52) for the system states results in the desired state vector
 ... 
ySd (yKd (t), ẏKd (t), ÿKd (...t), y Kd (t), ud (t), u˙d (t), κ )
 y Kd (t), ud (t), u˙d (t), κ )
v1d (ẏKd (t), ÿKd (t), ... 
xd ( t ) = 
 y Kd (t), ud (t), u˙d (t), κ )
 .
 (53)
ẏSd (ẏKd (t), ÿKd (t), ...
v̇1d (ẏKd (t), ÿKd (t), y Kd (t), ud (t), u˙d (t), κ )

This equation still contains the inverse dynamics ud (t) and its time derivative u̇d . Substituting
ud for equation (48) and u̇d (t) for the time derivative of (48), which can be calculated analyti-
194 Model Predictive Control

cally, finally leads to

  ...  
(4) (5)
ySd yKd (t), ẏKd (t), ÿKd (t), y Kd (t), yKd (t), yKd (t), κ
  ...  
 (4) (5) 
 v1d yKd (t), ẏKd (t), ÿKd (t), y Kd (t), yKd (t), yKd (t), κ 
xd ( t ) = 

 ... (4) (5)
 .
 (54)
 ẏSd yKd (t), ẏKd (t), ÿKd (t), y Kd (t), yKd (t), yKd (t), κ 
  ...  
(4) (5)
v̇1d yKd (t), ẏKd (t), ÿKd (t), y Kd (t), yKd (t), yKd (t), κ

−3
x 10
8

4
ey in m

−2

−4

0 5 10 15
t in s
Fig. 5. Tracking error ey (t) for the cage motion in horizontal direction.

−3
x 10
5

1
ex in m

−1

−2

−3

−4
0 5 10 15
t in s
Fig. 6. Tracking error e x (t) for the cage motion in vertical direction.
Model Predictive Trajectory Control for High-Speed Rack Feeders 195

5. Experimental validation on the test rig

The benefits and the efficiency of the proposed control measures shall be pointed out by exper-
imental results obtained from the test set-up available at the Chair of Mechatronics, University
of Rostock. For this purpose, a synchronous four times continuously differentiable desired
trajectory is considered for the position of the cage in both x- and y-direction. The desired
trajectory is given by polynomial functions that comply with specified kinematic constraints,
which is achieved by taking advantage of time scaling techniques. The desired trajectory
shown in Figure 4 comprises a sequence of reciprocating motions with maximum velocities of
2 m/s in horizontal direction and 1.5 m/s in vertical direction. The resulting tracking errors

ey (t) = yKd (t) − yK (t) (55)

and
e x (t) = xKd (t) − xK (t) (56)
are depicted in Figure 5 and Figure 6. As can be seen, the maximum position error in y-
direction during the movements is about 6 mm and the steady-state position error is smaller
than 0.2 mm, whereas the maximum position error in x-direction is approx. 4 mm. Figure 7

0.015
v
1d
0.01 v1

0.005
v in m

0
1

−0.005

−0.01

−0.015
0 5 10 15
t in s
Fig. 7. Comparison of the desired values v1d (t) and the actual values v1 (t) for the bending
deflection.

shows the comparison of the bending deflection measured by strain gauges attached to the
flexible beam with desired values. During the acceleration as well as the deceleration inter-
vals, physically unavoidable bending deflections could be noticed. The achieved benefit is
given by the fact the remaining oscillatons are negligible when the rack feeder arrives at its
target position. This underlines both the high model accuracy and the quality of the active
damping of the first bending mode. Figure 8 depicts the disturbance rejection properties due
to an external excitation by hand. At the beginning, the control structure is deactivated, and
the excited bending oscillations decay only due to the very weak material damping. After
approx. 2.8 seconds, the control structure is activated and, hence, the first bending mode is
actively damped. The remaining oscillations are characterised by higher bending modes that
decay with material damping. In future work, the number of Ritz ansatz functions shall be
196 Model Predictive Control

0.03

0.02

0.01
v1 in m

−0.01 Manual
excitation
Control activated
−0.02

−0.03
0 1 2 3 4 5
t in s
Fig. 8. Transient response after a manual excitation of the bending deflection: at first without
feedback control, after approx. 2.8 seconds with active control.

increased to include the higher bending modes as well in the active damping. The correspond-
ing elastic coordinates and their time derivatives can be determined by observer techniques.

6. Conclusions
In this paper, a gain-scheduled fast model predictive control strategy for high-speed rack feed-
ers is presented. The control design is based on a control-oriented elastic multibody system.
The suggested control algorithm aims at reducing the future tracking error at the end of the
prediction horizon. Beneath an active oscillation damping of the first bending mode, an accu-
rate trajectory tracking for the cage position in x- and y-direction is achieved. Experimental
results from a prototypic test set-up point out the benefits of the proposed control structure.
Experimental results show maximum tracking errors of approx. 6 mm in transient phases,
whereas the steady-state tracking error is approx. 0.2 mm. Future work will address an active
oscillation damping of higher bending modes as well as an additional gain-scheduling with
respect to the varying payload.

7. References
Aschemann, H. & Ritzke, J. (2009). Adaptive aktive Schwingungsdämpfung und Trajektorien-
folgeregelung für hochdynamische Regalbediengeräte (in German), Schwingungen in
Antrieben, Vorträge der 6. VDI-Fachtagung in Leonberg, Germany. (in German).
Aschemann, H. & Ritzke, J. (2010). Gain-scheduled tracking control for high-speed rack feed-
ers, Proc. of the first joint international conference on multibody system dynamics (IMSD),
2010, Lappeenranta, Finland .
Bachmayer, M., Rudolph, J. & Ulbrich, H. (2008). Flatness based feed forward control for a
horizontally moving beam with a point mass, European Conference on Structural Con-
trol, St. Petersburg pp. 74–81.
Fliess, M., Levine, J., Martin, P. & Rouchon, P. (1995). Flatness and defect of nonlinear systems:
Introductory theory and examples, Int. J. Control 61: 1327–1361.
Model Predictive Trajectory Control for High-Speed Rack Feeders 197

Jung, S. & Wen, J. (2004). Nonlinear model predictive control for the swing-up of a rotary in-
verted pendulum, ASME J. of Dynamic Systems, Measurement and Control 126(3): 666–
673.
Kostin, G. V. & Saurin, V. V. (2006). The Optimization of the Motion of an Elastic Rod by
the Method of Integro-Differential Relations, Journal of computer and Systems Sciences
International, Vol. 45, Pleiades Publishing, Inc., pp. 217–225.
Lizarralde, F., Wen, J. & Hsu, L. (1999). A new model predictive control strategy for affine
nonlinear control systems, Proc of the American Control Conference (ACC ’99), San Diego
pp. 4263 – 4267.
M. Bachmayer, J. R. & Ulbrich, H. (2008). Acceleration of linearly actuated elastic robots avoid-
ing residual vibrations, Proceedings of the 9th International Conference on Motion and
Vibration Control, Munich, Germany.
Magni, L. & Scattolini, R. (2004). Model predictive control of continuous-time nonlin-
ear systems with piecewise constant control, IEEE Transactions on automatic control
49(6): 900–906.
Schindele, D. & Aschemann, H. (2008). Nonlinear model predictive control of a high-speed lin-
ear axis driven by pneumatic muscles, Proc. of the American Control Conference (ACC),
2008, Seattle, USA pp. 3017–3022.
Shabana, A. A. (2005). Dynamics of multibody systems, Cambridge University Press, Cambridge.
Staudecker, M., Schlacher, K. & Hansl, R. (2008). Passivity based control and time optimal tra-
jectory planning of a single mast stacker crane, Proc. of the 17th IFAC World Congress,
Seoul, Korea pp. 875–880.
Wang, Y. & Boyd, S. (2010). Fast model predictive control using online optimization, IEEE
Transactions on control systems technology 18(2): 267–278.
Weidemann, D., Scherm, N. & Heimann, B. (2004). Discrete-time control by nonlinear online
optimization on multiple shrinking horizons for underactuated manipulators, Pro-
ceedings of the 15th CISM-IFToMM Symposium on Robot Design, Dynamics and Control,
Montreal .
198 Model Predictive Control
Plasma stabilization system design on the base of model predictive control 199

0
9

Plasma stabilization system design

on the base of model predictive control
Evgeny Veremey and Margarita Sotnikova
Saint-Petersburg State University,
Faculty of Applied Mathematics and Control Processes
Russia

1. Introduction
Tokamaks, as future nuclear power plants, currently present exceptionally significant re-
search area. The basic problems are electromagnetic control of the plasma current, shape
and position. High-performance plasma control in a modern tokamak is the complex prob-
lem (Belyakov et al., 1999). This is mainly connected with the design requirements imposed
on magnetic control system and power supply physical constraints. Besides that, plasma is
an extremely complicated dynamical object from the modeling point of view and usually con-
trol system design is based on simplified linear system, representing plasma dynamics in the
vicinity of the operating point (Ovsyannikov et al., 2005). This chapter is focused on the con-
trol systems design on the base of Model Predictive Control (MPC) (Camacho & Bordons,
1999; Morari et al., 1994). Such systems provide high-performance control in the case when
accurate mathematical model of the plant to be controlled is unknown. In addition, these
systems allow to take into account constraints, imposed both on the controlled and manip-
ulated variables (Maciejowski, 2002). Furthermore, MPC algorithms can base on both linear
and nonlinear mathematical models of the plant. So MPC control scheme is quite suitable for
plasma stabilization problems.
In this chapter two different approaches to the plasma stabilization system design on the base
of model predictive control are considered. First of them is based on the traditional MPC
scheme. The most significant drawback of this variant is that it does not guarantee stability
of the closed-loop control circuit. In order to eliminate this problem, a new control algorithm
is proposed. This algorithm allows to stabilize control plant in neighborhood of the plasma
equilibrium position. Proposed approach is based on the ideas of MPC and modal paramet-
ric optimization. Within the suggested framework linear closed-loop system eigenvalues are
placed in the specific desired areas on the complex plane for each sample instant. Such areas
are located inside the unit circle and reflect specific requirements and constraints imposed on
closed-loop system stability and oscillations.
It is well known that the MPC algorithms are very time-consuming, since they require the
repeated on-line solution of the optimization problem at each sampling instant. In order to re-
duce computational load, algorithms parameters tuning are performed and a special method
is proposed in the case of modal parametric optimization based MPC algorithms.
200 Model Predictive Control

The working capacity and effectiveness of the MPC algorithms is demonstrated by the exam-
ple of ITER-FEAT plasma vertical stabilization problem. The comparison of the approaches is
done.

2. Control Problem Formulation

2.1 Mathematical model of the plasma vertical stabilization process in ITER-FEAT tokamak
The dynamics of plasma control process can be commonly described by the system of ordinary
differential equations (Misenov, 2000; Ovsyannikov et al., 2006)

dΨ
+ RI = V, (1)
dt
where Ψ is the poloidal flux vector, R is a diagonal resistance matrix, I is a vector of active and
passive currents, V is a vector of voltages applied to coils. The vector Ψ is given by nonlinear
relation
Ψ = Ψ( I, I p ), (2)
where I p is the plasma current. The vector of output variables is given by

y = y( I, I p ). (3)

Linearizing equations (1)–(3) in the vicinity of the operating point, we obtain a linear model of
the process in the state space form. In particular, the linear model describing plasma vertical
control in ITER-FEAT tokamak is presented below.
ITER-FEAT tokamak (Gribov et al., 2000) has a separate fast feedback loop for plasma vertical
stabilization. The Vertical Stabilization (VS) converter is applied in this loop. Its voltage is
evaluated in the feedback controller, which uses the vertical velocity of plasma current cen-
troid as an input. So the linear model can be written as follows

ẋ = Ax + bu,
(4)
y = cx + du,

where x ∈ E58 is a state space vector, u ∈ E1 is the voltage of the VS converter, y ∈ E1 is the
vertical velocity of the plasma current centroid.
Since the order of this linear model is very high, an order reduction is desirable to simplify
the controller synthesis problem. The standard Matlab function schmr was used to perform
model reduction from 58th to 3rd order. As a result, we obtain a transfer function of the
reduced SISO model (from input u to output y)

1.732 · 10−6 (s − 121.1)(s + 158.2)(s + 9.641)

P(s) = . (5)
(s + 29.21)(s + 8.348)(s − 12.21)
This transfer function has poles which dominate the dynamics of the initial plant. The un-
stable pole corresponds to vertical instability. It is natural to assume that two other poles
are determined by the virtual circuit dynamic related to the most significant elements in the
tokamak vessel construction. The quality of the model reduction can be illustrated by the
comparison of the Bode diagram for both initial and reduced models. Fig. 1 shows the Bode
diagrams for initial and reduced 3rd order models on the left and for initial and reduced 2nd
order model on the right. It is easy to see that the curves for initial model and reduced 3rd
order model are actually indistinguishable, contrary to the 2nd order model.
Plasma stabilization system design on the base of model predictive control 201

Bode Diagram Bode Diagram

−70 −70

−80 −80
Magnitude (dB)

Magnitude (dB)
−90 −90

−100 −100

−110 −110

−120 −120
20 20

15 15
Phase (deg)

Phase (deg)
10 10

5 5

0 0

−5 −5
0 2 4 0 2 4
10 10 10 10 10 10
Frequency (rad/sec) Frequency (rad/sec)

Fig. 1. Bode diagrams for initial (solid lines) and reduced (dotted lines) models.

In addition to plant model (5), we must take into account the following limits that are imposed
on the power supply system
VS VS
Vmax = 0.6kV, Imax = 20.7kA, (6)

where VS
Vmax is the maximum voltage, VS
is the maximum current in the VS converter. So,
Imax
the linear model (5) together with constraints (6) is considered in the following as the basis for
controller synthesis.

2.2 Optimal control problem formulation

The desired controller must stabilize vertical velocity of the plasma current centroid. One of
the approaches to control synthesis is based on the optimal control theory. In this framework,
plasma vertical stabilization problem can be stated as follows. One needs to find a feedback
control algorithm u = u(t, y) that provides a minimum of the quadratic cost functional
 ∞
J = J (u) = (y2 (t) + λu2 (t))dt, (7)
0

subject to plant model (5) and constraints (6), and guarantees closed-loop stability. Here λ is a
constant multiplier setting the trade-off between controller’s performance and control energy
costs.
Specifically, in order to find an optimal controller, LQG-synthesis can be performed. Such a
controller has high stabilization performance in the unconstrained case. However, it is per-
haps not the best choice in the presence of constraints.
Contrary to this, the MPC synthesis allows to take into account constraints. Its basic scheme
implies on-line optimization of the cost functional (7) over a finite horizon subject to plant
model (5) and imposed constraints (6).
202 Model Predictive Control

3. Model Predictive Control Algorithms

3.1 MPC Basic Principles
Suppose we have a mathematical model, which approximately describes control process dy-
namics
x̃˙ (τ ) = f(τ, x̃(τ ), ũ(τ )), x̃|τ =t = x(t). (8)
Here x̃(τ ) ∈ En is a state vector, ũ(τ ) ∈ Em is a control vector, τ ∈ [t, ∞), x(t) is the actual
state of the plant at the instant t or its estimation based on measurement output.
This model is used to predict future outputs of the process given the programmed control
ũ(τ ) over a finite time interval τ ∈ [t, t + Tp ]. Such a model is called prediction model and
the parameter Tp is named prediction horizon. Integrating system (8) we obtain x̃(τ ) =
x̃(τ, x(t), ũ(τ ))—predicted process evolution over time interval τ ∈ [t, t + Tp ].
The programmed control ũ(τ ) is chosen in order to minimize quadratic cost functional over
the prediction horizon
 t+ Tp
J = J (x(t), ũ(·), Tp ) = ((x̃ − r x ) R(x̃)(x̃ − r x ) + (ũ − ru ) Q(x̃)(ũ − ru ))dτ, (9)
t

where R (x̃) , Q (x̃) are positive definite symmetric weight matrices, r x , ru are state and con-
trol input reference signals. In addition, the programmed control ũ(τ ) should satisfy all of the
constraints imposed on the state and control variables. Therefore, the programmed control
ũ(τ ) over prediction horizon is chosen to provide minimum of the following optimization
problem
J (x (t) , ũ (·) , Tp ) → min , (10)
ũ(·)∈Ωu

where Ωu is the admissible set given by

 
Ωu = ũ(·) ∈ K0n [t, t + Tp ] : ũ(τ ) ∈ U, x̃(τ, x(t), ũ(τ )) ∈ X, ∀τ ∈ [t, t + Tp ] . (11)

Here, K0n [t, t + Tp ] is the set of piecewise continuous vector functions over the interval
[t, t + Tp ], U ⊂ Em is the set of feasible input values, X ⊂ En is the set of feasible state values.
Denote by ũ∗ (τ ) the solution of the optimization problem (10), (11). In order to implement
feedback loop, the obtained optimal programmed control ũ∗ (τ ) is used as the input only on
the time interval [t, t + δ], where δ << Tp . So, only a small part of ũ∗ (τ ) is implemented. At
time t + δ the whole procedure—prediction and optimization—is repeated again to find new
optimal programmed control over time interval [t + δ, t + δ + Tp ]. Summarizing, the basic
MPC scheme works as follows:
1. Obtain the state estimation x̂ on the base of measurements y.
2. Solve the optimization problem (10), (11) subject to prediction model (8) with initial
conditions x̃|τ =t = x̂(t) and cost functional (9).
3. Implement obtained optimal control ũ∗ (τ ) over time interval [t, t + δ].
4. Repeat the whole procedure 1–3 at time t + δ.
From the previous discussion, the most significant MPC features can be noted:
• Both linear and nonlinear model of the plant can be used as a prediction model.
• MPC allows taking into account constraints imposed both on the input and output vari-
ables.
Plasma stabilization system design on the base of model predictive control 203

• MPC is the feedback control with the discrete entering of the measurement information
at each sampling instant 0, δ, 2δ, . . . .
• MPC control algorithms imply the repeated (at each sampling instant with interval δ)
on-line solution of the optimization problems. It is especially important from the real-
time implementation point of view, because fast calculations are needed.

3.2 MPC real-time implementation

In order for real-time implementation, piece-wise constant functions are used as a pro-
grammed control over the prediction horizon. That is, the programmed control ũ(τ ) is pre-
sented by the sequence{ũ k , ũk+1 , ..., ũk+ P−1 }, where ũi ∈ Em is the control input at the time
interval [iδ, (i + 1)δ] , δ is the sampling interval. Note that, P is a number of sampling intervals
over the prediction horizon, that is Tp = Pδ. Likewise, general MPC formulation presented
above consider nonlinear prediction model in the discrete form
x̃i+1 = f(x̃i , ũi ), i = k + j, j = 0, 1, 2, ..., x̃k = xk ,
(12)
ỹi = Cx̃i .
Here ỹi ∈ Er is the vector of output variables, xk ∈ En is the actual state of the plant at time
instant k or its estimation on the base of measurement output. We shall say that the sequence
of vectors {ỹ k+1 , ỹk+2 , ..., ỹk+ P } represents the prediction of future plant behavior.
Similar to the cost functional (9), consider also its discrete analog given by
 y y
Jk = Jk (ȳ, ū) = ∑ Pj=1 (ỹk+ j − rk+ j ) T Rk+ j (ỹk+ j − rk+ j )
 (13)
+ (ũk+ j−1 − rk+ j−1 ) Qk+ j (ũk+ j−1 − ruk+ j−1 ) ,
u T

y
where Rk+ j and Qk+ j are the weight matrices as in the functional (9), ri and riu are the output
and input reference signals,
 T
ȳ = ỹk+1 ỹk+2 ... ỹk+ P ∈ ErP ,
 T
ū = ũk ũk+1 ... ũk+ P−1 ∈ EmP
are the auxiliary vectors.
The optimization problem (10), (11) can now be stated as follows
Jk (xk , ũk , ũk+1 , ...ũk+ P−1 ) → min , (14)
{ũk ,ũk+1 ,...,ũk+ P−1 }∈Ω∈EmP
 
where Ω = ū ∈ EmP : ũk+ j−1 ∈ U, x̃k+ j ∈ X, j = 1, 2, ..., P is the admissible set.
Generally, the function J (xk , ũk , ũk+1 , ...ũk+ P−1 ) is a nonlinear function of mP variables and Ω
is a non-convex set. Therefore, the optimization task (14) is a nonlinear programming prob-
lem.
Now real-time MPC algorithm can be presented as follows:
1. Obtain the state estimation x̂k based on measurements yk using the observer.
2. Solve the nonlinear programming problem (14) subject to prediction model (12) with
initial conditions x̃k = x̂k and cost functional (13). It should be noted, that the value
of the function Jk (xk , ũk , ũk+1 , ...ũk+ P−1 ) is obtained by numerically integrating the pre-
diction model (12) and then substituting the predicted behavior x̄ ∈ EnP into the cost
function (13) given the programmed control {ũ k , ũk+1 , ..., ũk+ P−1 } over the prediction
horizon and initial conditions x̂k .
204 Model Predictive Control

3. Let {ũ ∗k , ũ∗k+1 , ..., ũ∗k+ P−1 } be the solution of the problem (14). Implement only the first
component ũ∗k of the obtained optimal sequence over time interval [kδ, (k + 1)δ].
4. Repeat the whole procedure 1–3 at next time instant (k + 1)δ.
Note, that the algorithm stated above implies real-time solution of the nonlinear programming
problem at each sampling instant. The complexity of such a problem is determined by the
number of sampling intervals P.
The simplest way to reduce the optimization problem order is to decrease the prediction hori-
zon. But, it is necessary to keep in mind that the performance of the closed-loop system
depends strongly on the number P of samples. The quality of the processes is decreased if
the prediction horizon is reduced. Moreover, the system can lose stability if the quantity P is
sufficiently small.
So, the following approaches to reduce computational load can be proposed:
1. Using the control horizon. The positive integer number M < P is called the control
horizon if the following condition hold:
ũk+ M−1 = ũk+ M = ... = ũk+ P−1 .
Thus, the number of independent variables is decreased from mP to mM. This approach
allows to essentially reduce the optimization problem order. However, if the control
horizon M is too small, the closed-loop stability can be compromised and the quality of
the processes can decrease.
2. Increasing the sampling interval δ and reducing the number P of samples over the pre-
diction horizon. This also allows to decrease the optimization problem order while
preserving the value of the prediction horizon.
3. The computational consumption also depends on the prediction model used. So, one
needs to use as simple models as possible. But the prediction model should adequately
reflect the dynamics of the plant considered. The simplest case is using the linear pre-
diction model.

3.3 Linear MPC

In this particular case, MPC is based on the linear prediction model. These algorithms are
computationally efficient which is especially important from the real-time implementation
point of view.
Generally, linear prediction model is presented by
x̃i+1 = Ax̃i + Bũi , i = k + j, j = 0, 1, 2, ..., x̃k = xk ,
(15)
ỹi = Cx̃i .
 T
Suppose ū = ũk ũk+1 ... ũk+ P−1 is the programmed control over the prediction
horizon. Then, integrating (15) we obtain future outputs of the plant in the form
ȳ = Lxk + Mū, (16)
where  
  CB 0 ... 0
CA
 .. 

 CA2 


 CAB . 

L= .. , M= .. .
 .   .. 
 . . 
CA P CA P−1 B ... CAB CB
Plasma stabilization system design on the base of model predictive control 205

Substituting (16) into (13) we get

Jk = Jk (xk , ū) = ū T Hū + 2f T ū + g. (17)

Here we assumed that all weight matrices are equal, that is

Rk+1 = Rk+2 = ... = Rk+ P = R,

Qk+1 = Qk+2 = ... = Qk+ P = Q.

The matrix H and vector f in (17) are as follows

H = M RM + Q, f = M RLxk . (18)

It can easily be shown that in this case the optimization problem (14) is reduced to the
quadratic programming problem of the form

Jk (xk , ũk , ũk+1 , ..., ũk+ P−1 ) = ū T Hū + 2f T ū + g → min . (19)

ū∈Ω⊂EmP

Here H is a positive definite matrix and Ω is a convex set defined by the system of linear con-
straints. On-line solution of the optimization problem (19) at each sampling instant generally
leads to nonlinear feedback control law.
Note that the optimization problem (19) can be solved analytically for the unconstrained case.
The result is the linear controller
ũk = Kx̃k , (20)
which converges to the LQR-optimal one as P is increased. This convergence is obvious, be-
cause the discrete LQR controller minimizes the functional (13) with infinity prediction hori-
zon for linear model (15).

4. Model Predictive Control on the base of modal parametrical optimization

In this section a new approach to MPC control algorithm synthesis is considered. The key
feature of corresponding algorithms is that they guarantee linear closed-loop system stability
at each sampling period. It is necessary to remark that in the case of traditional MPC algorithm
implementation, described above, closed-loop system stability can be provided only for the
simplest case when we have a linear prediction model, quadratic cost functional and without
constraints.
Let us assume that the mathematical model of the plant to be controlled is described by the
following system of difference equations

x̂k+1 = F(x̂k , ûk , ϕ̂k ),

(21)
ŷk = Cx̂k .

Here ŷk ∈ Es is the vector of output variables, x̂k ∈ En is the state space vector, ûk ∈ Em is the
vector of controls, ϕ̂k ∈ El is the vector of external disturbances.
Equations (21) can be used as a basis for nonlinear prediction model construction. Suppose
that obtained prediction model is given by

x̃i+1 = f(x̃i , ũi ), i = k + j, j = 0, 1, 2, ..., x̃k = xk ,

(22)
ỹi = Cx̃i .
206 Model Predictive Control

Here xk ∈ En is the actual state of the plant at time instant k or its estimation on the base of
measurement output.
Let desired object dynamics is presented by the given vector sequences {rkx } and {ruk }, k =
0,1,2,... . The linear mathematical model of the plant, describing its behavior in the neighbour-
hood of the desired trajectory, can be obtained by performing the equations (21) linearization.
As a result of this action, we get the linear system of difference equations

x̄k+1 = Ax̄k + Būk + H ϕ̄k ,

(23)
ȳk = Cx̄k ,

where x̄k ∈ En , ūk ∈ Em , ȳk ∈ Es , ϕ̄k ∈ El are the vectors of the state, control input, measure-
ments and external disturbances respectively. These vectors represent the deviations from the
desired trajectory. Next we shall consider only such situations when all matrices in equations
(23) have constant elements. In the framework of proposed approach, the control input over
the prediction horizon is generated by the controller of the form

ūk = W(q, h)ȳk . (24)

Here q is the shift operator, W(q, h) is the controller transfer function with the fixed structure
(that is the degrees of the polynomials in the numerator and denominator of all its components
are given and fixed), h ∈ Er is the vector of tuned parameters, which must be chosen on the
stage of control design.
The prediction model equations (22), closed by the feedback (24), can be presented as follows

x̃i+1 = f(x̃i , ũi ), i = k + j, j = 0, 1, 2, ..., x̃k = xk ,

(25)
ũi = riu + W(q, h)C(x̃i − rix ).

Let us assume that parameters vector h is chosen and fixed. Then we can solve system of
difference equations (25) with a given initial conditions for the instants i = k, k + 1, ..., k + P −
1. As a result we obtain vectors sequence {x̃i }, (i = k + 1, ..., k + P), which represents the
prediction of future plant behavior over the prediction horizon P. It must be noted, that the
control sequence ũk , ũk+1 , ...ũk+ P−1 over this horizon is determined uniquely by the choice
of parameter vector h. So, in this case the problem of control is reduced to the problem of
parameters vector h tuning.
The controlled processes quality over the prediction horizon P can be presented by the fol-
lowing cost functional

Jk = Jk ({x̃i }, {ũi }) = Jk (W(q, h)) = Jk (h) ≥ 0, (26)

where {x̃i }, i = k + 1, ..., k + P, {ũi }, i = k, ..., k + P − 1 are the state and control vectors
sequences correspondently, which satisfies the system of equations (25). It is easy to see, that
the cost functional (26) is reduced to the function of parameter vector h.
Let us consider the following optimization problem

Jk = Jk (h) → inf , (27)

h∈Ω H

where Ω H is a set of parameter vectors providing that the eigenvalues of the closed-loop
system (23), (24) are placed in the desired area C∆ inside a unit circle.
It is necessary to remark that the problem (27) is a nonlinear programming problem with an
extremely complicated definition of the cost function, which, in generall, has no analytical
Plasma stabilization system design on the base of model predictive control 207

representation and is given only algorithmically. Besides that, the specific character of the
problem (27) is also defined by the complicated constraints imposed, which determines the
admissible areas of eigenvalues displacement. It must be noted, that the dimension of the
optimization problem (27) is defined only by the dimension of parameter vector h and it does
not depend on the prediction horizon P value.

Definition 1. We shall say that the controller (24) has a full structure if the degrees of polyno-
mials in the numerators and denominators of the matrix W(q, h) components and the struc-
ture of parameter vector h are such that it is possible to assign any given roots for closed-loop
system (23),(24) characteristic polynomial ∆(z, h) by appropriate selection of parameter vector
h.

In order to get another form of the presented definition, consider the equations of the closed-
loop system (23),(24). They can be represented in the normal form as follows

x̄k+1 = Ax̄k + Būk + H ϕ̄k ,

ȳk = Cx̄k ,
(28)
ξ k+1 = Ac (h)ξ k + Bc (h)ȳk ,
ūk = Cc (h)ξ k + Dc (h)ȳk ,

where ξ k ∈ Eν is a controller (24) state vector. After applying Z-transformation to the system
of equations (28) with zero initial conditions, obtain

(En z − A)x̄ = Bū + H ϕ̄,

(Eν z − Ac (h))ξ = Bc (h)Cx̄,
ū = Cc (h)ξ + Dc (h)Cx̄,
ȳ = Cx̄,
or     
En z − A − BDc (h)C −BCc (h) x̄ H
= ϕ̄.
−Bc (h)C Eν z − Ac (h) ξ 0
Therefore, the closed-loop system characteristic polynomial ∆(z, h) is given by
 
En z − A − BDc (h)C −BCc (h)
∆(z, h) = det .
−Bc (h)C Eν z − Ac (h)

Let us denote the degree of the polinomial ∆(z, h) by nd .

Let find parameter vector h, which provide a given roots for the system (28) characteristic
polynomial. In other words, it is nesessary to find such parameter vector h that provide the
following identity
∆(z, h) ≡ ∆ ˜ ( z ),
˜
where ∆(z) is a given polynomial with degree nd , having desired roots. In order to find vector
h, equate the correspondent coefficients for the same degrees of z-variable. As a result obtain
the system of (nd + 1) nonlinear equations with r unknown components of vector h in the
form
L(h) = γ. (29)
It is evident that the controller (24) has a full structure if and only if the system of equations
(33) has a solution for any vector γ.
208 Model Predictive Control

It can be easy shown that if the parameter vector h consists of the coefficients of numerator
and denominator polynomials of matrix W(q, h), then the system (29) reduced to the linear
system of the form
Lh = γ, (30)
where L is a constant matrix. Note that for any case, the controller (24) has a full structure
only if the system (23) is fully controllable and observable.
Let us now refine the optimization problem (27) statement in suppose that the controller (23)
has a full structure and that the following set Ω H is determined as admissible set of the form

Ω H = {h ∈ Er : δi (h) ∈ C∆ , i = 1, 2, ..., nd }. (31)

Here δi is the roots of the characteristic polynomial ∆(z, h), nd = deg∆(z, h).
Let consider two different variants of the desired areas C∆ , depicted in Fig. 2. This areas are
located inside a unit circle, i. e. r < 1.

(a) area C∆1 (b) area C∆2

Fig. 2. The areas C∆1 and C∆2 of the desired root displacement

The formalized description for the desired areas C∆ are as follows:

C∆ = C∆1 = {z ∈ C1 : |z| ≤ r }, where r ∈ (0, 1) is a given real number;
C∆ = C∆2 = {z ∈ C1 : z = ρexp(±iϕ), 0 ≤ ρ ≤ r, 0 ≤ ϕ ≤ ψ(ρ)}, where r ∈ (0, 1) is a
given real number, ψ(ξ ) is a real function of variable ξ ∈ (0, r ], which takes the values on the
interval [0, π ] and ψ(r ) = 0.
The reasons of these areas introduction is obvious. The first area C∆1 determines the lower
bound for the closed-loop system stability margin and, therefore, the settling time for transient
processes. Second area C∆2 determines stability bound and, in addition, constraints on the
closed-loop system oscillations.
In order to form the algorithm for the problem (27) solution on the admissible set (31), let us
firstly perform the parametrization of the considered areas C∆ with the n-dimensional real
vectors on the base of the following statement.

Theorem 1. For any real vector γ ∈ End the roots of the polynomial ∆∗ (z, γ), given by the formulas
presented below, are located inside the area C∆1 or on its bound. And reversly, if the roots of the some
Plasma stabilization system design on the base of model predictive control 209

polynomial ∆(z) are located inside the area C∆1 and, in addition, all its real roots are positive, then it
can be found such a vector γ ∈ End that the following identity holds ∆(z) ≡ ∆∗ (z, γ). Here
d
∆∗ (z, γ) = ∏(z2 + a1i (γ, r)z + a0i (γ, r)), (32)
i =1

if nd is even, d = nd /2;
d
∆∗ (z, γ) = (z − ad+1 (γ, r )) ∏ (z2 + a1i (γ, r )z + a0i (γ, r )), (33)
i =1

if nd is odd, d = [nd /2];

   4    4 
γ2 γi1 2
γi1 γi1
a1i (γ, r ) = −r exp − 2i1 + 4 − γ 2
i2 + exp − 2 − 4 − γ 2
i2 ,
 2  2  (34)
0 2
ai (γ, r ) = r exp −γi1 , i = 1, ..., d, ad+1 (γ, r ) = r exp −γd0 ,

γ = {γ11 , γ12 , γ21 , γ22 , ..., γd1 , γd2 , γd0 }. (35)

Proof If the nd is even, then the proof of the direct and reverse propositions arises from the
elementary properties of the quadratic trinomials in the formula (32). Really, for any given
pair of the real numbers γi1 , γi2 the roots of the trinomial ∆i∗ (z) in (32) are presented by the
expression   
2
γi1 4
γi1
i 2 .
z1,2 = r · exp − ± − γi2
2 4

From this expression it follows that |z1,2i | ≤ r and, therefore, the roots zi of the trinomial are
1,2
located inside the area C∆1 or on its bound, and this proves the direct proposition.
In order to prove reverse one, let consider some quadratic trinomial of the form ∆i (z) =
z2 + β 1 z + β 0 . By the conditions of the reverse proposition, the roots z1,2 of this trinomial are
located inside the area C∆1 and, if the roots are real numbers, then they are positive. In order
to locate the roots z1,2 inside the area C∆1 , it is necessary and sufficient that the following
relations holds
β β β β β
1 − 1 + 20 ≥ 0, 1 − 20 ≥ 0, 1 + 1 + 20 ≥ 0. (36)
r r r r r
Besides that, the roots product z1 z2 is positive in anycase if they are being complex conjugated
pair or positive real numbers. Therefore, the following inequality is true

β 0 > 0. (37)

Let find such numbers γi1 and γi2 that the identity ∆i∗ (z) ≡ ∆i (z) is satisfied. By equating the
correspondent coefficients for the same degrees of z-variable, obtain
      
2
γi1 4
γi1 γ 2 γ 4
−r exp − + 2  + exp  − i1 −
− γi2 i1 2  = β ,
− γi2 1
2 4 2 4

r2 exp(−γi1
2
) = β0 ,
210 Model Predictive Control

and consequently

γi1 = −ln ( β 0 /r2 ), 
  2   2
1 r β0 β21 β21 (38)
γi2 = − 4 ln w β0 ln w r2 , where w = 2β0 − 1 + 2β 0 − 1 − 1.

Now let verify that the γi1 and γi2 , given by the formulas (38), are the real numbers. 
Really, from the inequalities (36), (37) it follows that 0 < β 0 /r2 ≤ 1, therefore −ln β 0 /r2 ≥ 0
and γi1 is a real number.
Let show that the expression under radical in the formula for γi2 is nonnegative. For the
first, consider the case when the trinomial ∆i (z) has two real positive roots z1,2 . Then his
coefficients must satisfies to the condition β21 − 4β 0 ≥ 0, whence it follows that w ≥ 1 – is a
real number. As a result, taking into account (36), we obtain
 
ln w · r2 /β 0 ≥ 0. (39)

It could be noted that the inequalities (36) implies also the satisfaction of the inequality
β21 − 2β 0 ≤ r2 + β 0 /r2 . Hence, we have
 
wβ 0 ≤ r2 , and − ln wβ 0 /r2 ≥ 0. (40)

Thus from the inequalities (39) and (40) it is easy to see that the expression under radical in
the formula for γi2 is nonnegative and γi2 is a real number.
Consider now a case, when the trinomial ∆i (z) has a pair of complex-conjugate roots z1,2 .
Then the following inequality is hold β21 − 4β 0 < 0, and therefore w is a complex number,
  2
which can be presented in the form w = β21 /2β 0 − 1 + i 1 − β21 /2β 0 − 1 . It is not difficult
to see that |w| = 1, hence, the expression under the radical for γi2 has a form
   2        2 
1 r β0 1 2 r 2w ,
γi2 = − ln + i · argw ln + i · argw = ln + arg
4 β0 r2 4 β0
i.e. it is nonnegative and γi2 is a real number.
If the nd is odd, the polynomial ∆∗ has, in according to (33), an additional linear binomial, for
which the propositions of the theorem are evident.
Now consider more difficult second variant of the admissible set C∆ . Let us prove the analo-
gous theorem, which allows to perform parametrization of this area.
Theorem 2. For any real vector γ ∈ End the roots of the polynomial ∆∗ (z, γ) (32),(33) are located
inside the area C∆2 , and reversly, if the roots of the some polynomial ∆(z) are located inside the area
C∆2 and, in addition, all its real roots are positive, then it can be found such a vector γ ∈ End that the
following identity holds ∆(z) ≡ ∆∗ (z, γ). Here
  2   2 
a1i (γ, r ) = −r exp −γi1 + νi + exp −γi1 − νi ,
 (41)
a0i (γ, r ) = r2 exp −2γi1 2 , i = 1, ..., d, a 2
d+1 ( γ, r ) = r · exp (− γd0 ),
     
where νi = 4 − f ( γ ) ψ2 r · exp − γ2
γi1 + γi14 ,i = 1, 2, ..., d; γ =
i2 i1
{γ11 , γ12 , γ21 , γ22 , ..., γd1 , γd2 , γd0 }.
The function f is such that f (·) : (−∞, +∞) → (0, 1) and its inverse function exists in the whole
region of the definition; the function ψ(ξ ) is a real function from the variable ξ ∈ (0, r ], which takes
the values in the interval [0, π ] and ψ(r ) = 0.
Plasma stabilization system design on the base of model predictive control 211

Proof Similar to theorem 1, consider the properties of the quadratic trinomials in (32). Firstly,
let prove a direct proposition.
For any given pair of the real numbers γi1 , γi2 the roots of the trinomial ∆i∗ (z) in (32) is given
by the expression z1,2i = r · exp (− γ2 ± ν ). Here two different variants are possible. If ν is a
i1 i i
i are also real. Besides that, taking into account the properties
real number, then the roots z1,2
   
4 − f ( γ ) ψ2 r · exp (− γ2 ) + γ4 ≤ γ4 .
of the function f , the following inequality holds γi1 i2 i1 i1 i1
i | ≤ r, that is zi ∈ C .
Hence the roots are positive and |z1,2 1,2 ∆2
If νi is a complex number, then z1,2 i is the pair of complex-conjugated roots and
i 2
|z1,2 | = ρ = r · exp(−γi1 ) ≤ r. Taking into account the properties of the function f , the
following inequality is valid
       
ϕ = f (γi2 ) ψ2 r · exp(−γi1 2 ) + γ4 − γ4 ≤ 2 ) = ψ ( ρ ).
ψ2 r · exp(−γi1 (42)
i1 i1

i = ± ϕ and, accordingly to (42), 0 ≤ ϕ ≤ ψ ( ρ ), then the roots zi are located

Since the arg z1,2 1,2
inside the area C∆2 , so the direct proposition is proven.
Let consider the reverse proposition. The roots z1,2 of some trinomial ∆i (z) = z2 + β 1 z +
β 0 are located inside the area C∆2 in accordance with the reverse proposition if these roots
are positive real numbers. Notice that the coefficients of this trinomial must satisfy to the
inequalities (36),(37), because |z1,2 | ≤ r and the roots product z1 z2 is positive in any way.
Let find such numbers γi1 , γi2 that the identity ∆i∗ (z) ≡ ∆i (z) holds. By equating the corre-
spondent coefficients for the same degrees of z-variable, obtain
    
2 2
−r exp −γi1 + νi + exp −γi1 − νi = β i , r2 exp(−2γi1
2
) = β0 ,

hence 
−0.5 · ln( β 0 /r2 ),
γi1 =
   
   2
2  2
1  β0  β β1 
f (γi2 ) =  2  
2 ) + γ4 
 ln2 2
− ln2  1 − 1 +  − 1 − 1  .
4 ψ r · exp(−γi1 i1
r 2β 0 2β 0

Let us show that the γi2 is a real number. For γi1 the proof is equivalent to the such one in the
first theorem.
It is evident that the equation with respect to γi2 has a solution, if the expression in the right
part of it takes the values inside the interval (0,1). Let denote this expression by h. Notice that
the denominator for h is equal to zero only if z1 = z2 = r, but in this case γi2 can be chosen
as any real number. In general case, taking into account the proof of the theorem 1, it is not
difficult to see that h > 0. Besides that the following inequality holds
    
 2 2
h < 1 − ln2 β21 /2β 0 − 1 + β 1 /2β 0 − 1 − 1 /ln2 β 0 /r2 ,

hence the real number γi2 exists and this one is determined as a solution of the equation
f (γi2 ) = h.
If the nd is odd, the polynomial ∆∗ has, in accordance with (33), an additional linear binomial,
for which the propositions of the theorem are evident.
212 Model Predictive Control

Now let us show how introced areas C∆1 and C∆2 are related to the standart areas on the
complex plane, which are commonly used in the analysis and synthesis of the continuos time
systems.
Primarily, it may be noticed that the eigenvalues of the continues linear model and the discrete
linear model are connected by the following rule (Hendricks et al., 2008): if s is the eigenvalue
of the continuos time system matrix, then z = esT is the correspondent eigenvalue of the
discrete time system matrix, where T is the sampling period. Taking into account this relation,
let consider the examples of the mapping of some standart areas for continuous systems to the
areas for discrete systems.
Example 1 Let we have given area C = {s = x ± yj ∈ C1 : x ≤ −α}, depicted in Fig. 3. It is
evident that the points of the line x = −α are mapped to the points of the circle |z| = e−αT .
The area C itself is mapped on the disc |z| ≤ e−αT , as shown in Fig.3. This disc corresponds to
the area C∆1 , which defines the degree of stability for discrete system.

Fig. 3. The correspondence of the areas for continuous and discrete system

Example 2 Consider the area

C = {s = x ± yj ∈ C1 : x ≤ −α, 0 ≤ y ≤ (− x − α)tgβ},

depicted in Fig. 4, where 0 ≤ β < π2 and α > 0 is a given real numbers.

Let perform the mapping of the area C on the z-plane. It is evident that the vertex of the angle
(−α, 0) is mapped to the point with polar coordinates r = e−αT , ϕ = 0 on the plane z. Let now
map each segment from the set

Lγ = {s = x ± yj ∈ C1 : x = γ, γ ≤ −α, 0 ≤ y ≤ (−γ − α)tgβ}

to the z-plane. Each point s = γ ± yj of the segment Lγ is mapped to the point z = esT =
eγT ± jyT on the plane z. Therefore, the points of the segment Lγ are mapped to the arc of the
circle with radius eγT if the following condition holds −α − π/( Ttgβ) < γ ≤ −α, and to the
whole circle if γ ≤ −α − π/( Ttgβ). Therefore, the maximum radius of the circle, which is
fullfilled by the points of the segment, is equal to r  = e−αT −π/tgβ , corresponding with the
equality γ0 = −α − π/( Ttgβ). Notice that the rays, which constitutes the angle, mapped to
the logarithmic spirals. Moreover, the bound of the area on the plane z is formed by the arcs
of these spirals in accordace with the x varying from −α to γ0 .
Plasma stabilization system design on the base of model predictive control 213

Fig. 4. The correspondence of the areas for continuous and discrete time systems

Let introduce the notation ρ = e xT , and define the function ψ(ρ), which represents the con-
straints on the argument values while the radius ρ of the circle is fixed:

(−lnρ − αT )tgβ, i f ρ ∈ [r  , r ],
ψ(ρ) =
π, i f ρ ∈ [0, r  ].

The result of the mapping is shown on the Fig. 4. It can be noted that the obtained area reflects
the desired degree of the discrete time system stability and oscillations.
Let us use the results of the theorem 2 in order to formulate the computational algoritm for the
optimization problem (27) solution on the admissible set Ω H taking into account the condition
C∆ = C∆2 . It is evident that the first case, where C∆ = C∆2 , is a particular case of the second
one.
Consider a real vector γ ∈ End and form the polynomial ∆∗ (z, γ) with the help of formulas
(32),(33),(41). Let require that the tuned parameters of the controller (24), defined by the vector
h ∈ Er , provides the identity
∆(z, h) ≡ ∆∗ (z, γ), (43)
where ∆(z, h) is the characteristic polynomial of the closed-loop system with the degree nd .
By equating the correspondent coefficients for the same degrees of z-variable, we obtain the
following system of nonlinear equations
L(h) = χ(γ) (44)
with respect to unknown components of the parameters vector h. The last system has a solu-
tion for any given γ ∈ End due to the controller (24) has a full structure. Let consider that, in
general case, the system (44) has a nonunique solution. Then the vector h can be presented as
a set of two vectors h = {h̄, hc }, where hc ∈ Enc is a free component, h̄ is the vector that is
uniquely defined by the solution of the system (44) for the given vector hc .
Let introduce the following notation for the general solution of the system (44)
h = h∗ = {h̄∗ (hc , γ), hc } = h∗ (γ, hc ) = h∗ (),
where  = {γ, hc } is a vector of the independent parameters with the dimension λ given by
λ = dim  = dim γ + dim hc = nd + nc .
214 Model Predictive Control

Let form the equations of the prediction model, closed by the controller (24) with the obtained
parameter vector h∗

x̃i+1 = f(x̃i , ũi ), i = k + j, j = 0, 1, 2, ..., x̃k = xk ,

(45)
ũi = riu + W(q, h∗ ())C(x̃i − rix ).

Now the functional Jk , which is given by (26) and computed on the solutions of the system
(45), becomes the function of the vector :

Jk = Jk ({x̃i }, {ũi }) = Jk∗ (W (q, h∗ ())) = Jk∗ (). (46)

Theorem 3. Consider the optimization problem (27), where Ω H is the admissible set, given by (31),
and the desired area C∆ = C∆2 . If the extremum of this problem is achieved at the some point hk0 ∈
Ω H , then there exists a vector  ∈ Eλ such that

hk0 = h∗ (k0 ), with k0 = arg min Jk∗ (). (47)

∈Eλ

And reversly, if there exists such a vector k0 ∈ Eλ , that satisfies to the condition (47), then the
following vector hk0 = h∗ (k0 ) is the solution of the optimization problem (27). In other words, the
problem (27) is equivalent to the unconstrained optimization problem of the form

Jk∗ = Jk∗ () → inf . (48)

∈Eλ

Proof Assume that the following condition is hold

hk0 = arg min Jk (h), Jk0 = Jk (hk0 ). (49)

h∈Ω H

In this case, the characteristic polynomial ∆(z, hk0 ) of the closed-loop system (28) has the roots
that are located inside the area C∆2 . Then, accordingly to the theorem 2, it can be found such
a vector γ = γk0 ∈ End , that ∆(z, hk0 ) ≡ ∆∗ (z, γk0 ), where ∆∗ is a polynomial formed by the
formulas (32), (33). Hence, there exists such a vector  = {γk0 , hk0c }, for which the following
conditions is hold hk0 = h∗ (k0 ), Jk∗ (k0 ) = Jk0 . Here hk0c is the correspondent constituent
part of the vector hk0 .
Now it is only remain to show that there no exists a vector 01 ∈ Eλ that the condition
Jk∗ (01 ) < Jk0 is valid. Really, let suppose that such vector exists. But then for the vector
h∗ (01 ) the following inequality takes place Jk (h∗ (01 ) = Jk∗ (01 ) < Jk0 . But this is not possi-
ble due to the condition (49). The reverse proposition is proved analogously.
Let formulate the computational algorithm in order to get the solution of the optimization
problem (27) on the base of the theorems proved above.
The algorithm consists of the following operations:
1. Set any vector γ ∈ End and construct the polynomial ∆∗ (z, γ) by formulas (32),(33), (41).
2. In accordance with the identity ∆(z, h) ≡ ∆∗ (z, γ), form the system of nonlinear equa-
tions
L ( h ) = χ ( γ ), (50)
which has a solution for any vector γ. If the system (50) has a nonunique solution,
assign the vector of the free parameters hc ∈ Enc .
3. For a given vector  = {γ, hc } ∈ Eλ solve the system of equations (50). As a result,
obtain vector h∗ ().
Plasma stabilization system design on the base of model predictive control 215

4. Form the equations of the prediction model closed by the controller (24) with the pa-
rameter vector h∗ () and compute the value of the cost function Jk∗ () (46).
5. Solve the problem (48) by using any numerical method for unconstrained minimization
and repeating the steps 3–5.
6. When the optimal solution k0 = arg min Jk∗ () is found, compute the parameter vector
∈Eλ
hk0 = h∗ (k0 ) and accept them as a solution.
Now real-time MPC algorithm, which is based on the on-line solution of the problem (27), can
be formulated. This algorithm consists of the following steps:
• Obtain the state estimation x̂k on the base of measurements yk .
• Solve the optimization problem (27), using the algorithm stated above, subject to the
prediction model (22) with initial conditions x̃k = x̂k .
• Let hk0 be the solution of the problem (27). Implement controller (24) with the parame-
ter vector hk0 over time interval [kδ, (k + 1)δ], where δ is the sampling period.
• Repeat the whole procedure 1–3 at next time instant (k + 1)δ.
As a result, let notice the following important features of the proposed MPC-algorithm. For
the first, the linear closed-loop system stability is provided at each sampling interval. Sec-
ondly, the control is realised in the feedback loop. Thirdly, the dimension of the unconstrained
optimization problem is fixed and does not depend on the length of prediction horizon P.

5. Plasma Vertical Stabilization Based on the Model Predictive Control

Let us remember that SISO model (5) represents plasma dynamics in the vertical stabilization
process and limits (6) are imposed on the power supply system. It is necessary to transform
the system (5) to the state-space form for MPC algorithms implementation. Besides that, in
order to take into account the constraint imposed on the current, one more equation should
be added to the model (5). Finally, the linear model of the stabilization process is given by
ẋ = Ax + bu,
(51)
y = cx + du,

where x ∈ E4 and the last component of x corresponds to VS converter current, y = (y1 , y2 ) ∈

E2 , y1 is the vertical velocity and y2 is the current in the VS-converter. We shall assume that
the model (51) describes the process accurately.
We can obtain a linear prediction model in the form (15) by the system (51) discretization. As
a result, we get
x̃i+1 = Ad x̃i + bd ũi , x̃k = xk ,
(52)
ỹi = Cd x̃i .
The constraints (6) form the system of linear inequalities given by
VS ,
ũi ≤ Vmax i = k, ..., k + P − 1;
VS , (53)
ỹi2 ≤ Imax i = k + 1, ..., k + P.
These constraints define the admissible convex set Ω. The discrete analog of the cost func-
tional (7) with λ = 1 is given by
P  
Jk = Jk (ȳ, ū) = ∑ ỹ2k+ j,1 + ũ2k+ j−1 . (54)
j =1
216 Model Predictive Control

So, in this case MPC algorithm leads to real-time solution of the quadratic programming prob-
lem (19) with respect to the prediction model (52), constraints (53) and the cost functional (54).
From the experiments the following values for the sampling time and number of sampling
intervals over the horizon were obtained

δ = 0.004 sec, P = 250.

Hence, we have the following prediction horizon

Tp = Pδ = 1 sec .

Let us consider the MPC controller synthesis without taking into account the constraints im-
posed. Remember that in this case we obtain a linear controller (20) that is practically the
same as the LQR-optimal one. The transient response of the system closed by the controller is
presented in Fig. 5. The initial state vector x (0) = h is used, where h is a scaled eigenvector
of the matrix A corresponding to the only unstable eigenvalue. The eigenvector h is scaled to
provide the initial vertical velocity y1 = 0.03 m/sec. It can be seen from the figure that the
constraints (6) imposed on the voltage and current are violated.
4
x 10
0.06 700 3

0.05 600 2.5

500
0.04 2
y (m/sec)

400
u(Volt)

y (A)

0.03 1.5
2

300
1

0.02 1
200

0.01 100 0.5

0 0 0
0 0.5 1 0 0.5 1 0 0.5 1
sec sec sec

Fig. 5. Transient response of the closed-loop system with unconstrained MPC-controller

Now consider the MPC algorithm synthesis with constraints. Fig. 6 shows transient response
of the closed-loop system with constrained MPC-controller. It is not difficult to see that all
constraints imposed are satisfied. In order to reduce computational consumptions, the ap-
proaches proposed above in Section 3.2 can be implemented.
1. Experiments with using the control horizon were carried out. This experiments show
that the quality of stabilization remains approximately the same with control horizon
M = 50 and prediction horizon P = 250. So, optimization problem order can be signif-
icantly reduced.
2. Another approach is to increase the sampling interval up to δ = 0.005 sec and reduce
the number of samples down to P = 200. Hence, prediction horizon has the same
value Tp = Pδ = 1 sec. The optimization problem order is also reduced in this case
and consequently time consumptions at each sampling instant is decreased. However,
further increase of δ tends to compromise closed-loop system stability.
Now consider the processes of the plasma vertical stabilization on the base of new MPC-
scheme.
Plasma stabilization system design on the base of model predictive control 217

4
x 10
0.06 700 3

0.05 600 2.5

500
0.04 2
y (m/sec)

400

u(Volt)

y2(A)
0.03 1.5
300
1

0.02 1
200

0.01 100 0.5

0 0 0
0 0.5 1 0 0.5 1 0 0.5 1
sec sec sec

Fig. 6. Transient response of the closed-loop system with constrained MPC-controller

Let us, for the first, transform system (5) into the state space form. As a result, we get
ẋ = Ax + bu,
(55)
y = cx + du,

where x ∈ E3 , y is the vertical velocity, u is the voltage in the VS-converter. We shall assume
that this model describes the process accurately.
As early, we can obtain linear prediction model by the system (55) discretization. So, we have
the following prediction model
x̃i+1 = Ad x̃i + bd ũi , x̃k = xk ,
(56)
ỹi = Cd x̃i .
Let also form the discrete linear model of the process, describing its behavior in the neigh-
bourhood of the zero equilibrium position. Such a model is obtained by the system (55) dis-
cretization and can be presented as follows
x̄k+1 = Ad x̄k + bd ūk ,
(57)
ȳk = Cd x̄k ,

where x̄k ∈ E3 , ūk ∈ E1 , ȳk ∈ E1 . We shall form the control over the prediction horizon by the
linear proportional controller, that is given by
ūk = K x̄k , (58)
where K ∈ E3
is the parameter vector of the controller. In the real processes control input
(58) is computed on the base of the state estimation, obtained with the help of asymptotic
observer. It must be noted that the controller (58) has a full structure, because the matrices of
the controllability and observability for the system (57) have a full rank.
Now consider the equations of the prediction model (56), closed by the controller (58). As a
result, we get
x̃i+1 = (Ad + bd K)x̃i , x̃k = xk ,
(59)
ỹi = Cd x̃i .
The controlled processes quality over the prediction horizon P is presented by the cost func-
tional
P  
Jk = Jk (K) = ∑ ỹ2k+ j + ũ2k+ j−1 . (60)
j =1
218 Model Predictive Control

It is easy to see that the cost functional (60) becomes the function of three variables, which
are the components of the parameter vector K. It is important to note that the cost function
remains essentialy nonlinear for this variant of the MPC approach even in the case when the
prediction model is linear. It is a price for providing stability of the closed-loop linear system.
Consider the optimization problem (27) statement for the particular case of plasma vertical
stabilization processes

Jk = Jk (K) → min , where ΩK = {K ∈ E3 : δi (K) ∈ C∆ , i = 1, 2, 3}. (61)

K∈ΩK

Here δi are the roots of the closed-loop system (57), (58) characteristic polynomial ∆(z, K) with
the degree nd = 3. Let given desirable area be C∆ = C∆2 , where r = 0.97 and the function
ψ(ρ) is presented by the formula
  
ln ρr tgβ, re−π/tgβ ≤ ρ ≤ r,
ψ(ρ) =
π, i f 0 < ρ ≤ re−π/tgβ ,

where β = π/10. This area is presented on the Fig. 7.

1.5
C
∆
Unit Circle
1

0.5

−0.5

−1

−1.5
−1.5 −1 −0.5 0 0.5 1 1.5

Fig. 7. The area C∆ of the desired roots location

Let construct now the system of equations in accordance with the identity ∆(z, K) ≡ ∆∗ (z, γ),
where γ ∈ E3 and the polynomial ∆∗ (z, γ) is defined by the formulas (33), (41). As a result,
we obtain linear system with respect to unknown parameter vector K

L0 + L1 K = χ ( γ ). (62)

Here vector L0 and square matrix L1 are constant for any sampling instant k. These are fully
defined by the matrices of the system (57). Besides that, the matrix L1 is nonsingular, hence
we can find the unique solution for system (62)

K = L̃0 + L̃1 χ(γ), (63)

where L̃1 = L1−1 and L̃0 = −L1−1 L0 . Substituting (63) into the prediction model (59) and then
into the cost functional (60), we get Jk = Jk (K) = Jk∗ (γ). That is the functional Jk becomes
Plasma stabilization system design on the base of model predictive control 219

the function of three indepent variables. Then, accordingly to the theorem 3, optimization
problem (61) is equivalent to the unconstrained minimization

Jk∗ = Jk∗ (γ) → min . (64)

γ ∈ E3

Thus, in conformity with the algorithm of the MPC real-time implementation, presented in the
section 4 above, in order to form control input we must solve the unconstrained optimization
problem (64) at each sampling instant.
Consider now the processes of the plasma vertical stabilization. For the first, let us consider
the unconstrained case. Remember that the structure of the controller (58) is linear. So, if
the roots of the characteristic polynomial for the system (57) closed by the LQR-controller
are located inside the area C∆ then parameter vector K will be practically equivalent to the
matrix of the LQR-controller. The roots of the system closed by the discrete LQR are the
following z1 = 0.9591, z2 = 0.8661, z3 = 0.9408. This roots are located inside the area C∆ .
So, the transient responce of the system closed by the MPC-controller, which is based on the
optimization (64), is approximately the same as presented in Fig. 5.

4
x 10
0.06 700 3

0.05 600 2.5

500
0.04 2
y (m/sec)

400
u(Volt)

y2(A)

0.03 1.5
300
1

0.02 1
200

0.01 100 0.5

0 0 0
0 0.5 1 0 0.5 1 0 0.5 1
sec sec sec

Fig. 8. Transient response of the closed-loop system with constrained MPC-controller

Consider now the processes of plasma stabilization with the constraints (53) imposed. As
mentioned above, in order to take into account the constraint imposed on the current, the
additional equation should be added. It is necessary to remark that in the presence of the con-
straints, the optimization problem (64) becomes the nonlinear programming problem. Fig.8
shows transient responce of the closed-loop system with MPC-controller when the only con-
straint on the VS converter voltage is taked into account. It can be seen from the figure that
the constraint imposed on the voltage is satisfied, but the constraint on the current is violated.
Fig.9 shows transient responce of the closed-loop system with MPC-controller when both the
constraint on the VS converter voltage and current are taken into account. It is not difficult to
see that all the imposed constraints are satisfied.

6. Conclusion
The problem of plasma vertical stabilization based on the model predictive control has been
considered. It is shown that MPC algorithms are superior compared to the LQR-optimal con-
troller, because they allow taking constraints into account and provide high-performance con-
trol. It is also shown that in the case of the traditional MPC-scheme it is possible to reduce
220 Model Predictive Control

4
x 10
0.06 700 3

0.05 600 2.5

500
0.04 2
y (m/sec)

400

u(Volt)

y2(A)
0.03 1.5
300
1

0.02 1
200

0.01 100 0.5

0 0 0
0 0.5 1 0 0.5 1 0 0.5 1
sec sec sec

Fig. 9. Transient response of the closed-loop system with constrained MPC-controller

the computational load significantly using relatively small control horizon or by increasing
sample interval while preserving the processes quality in the closed-loop system.
New MPC approach was provided. This approach allows us to guarantee linear closed-loop
system stability. It’s implementation in real-time is connected with the on-line solution of the
unconstrained nonlinear optimization problem if there is not constraint imposed and with the
nonlinear programming problem in the presence of constraints. The significant feature of this
approach is that the dimension of the optimization problem is not depend on the prediction
horizon P. The algorithm for the real-time implementation of the suggested approach was
described. It allows us to use MPC algorithms to solve plasma vertical stabilization problem.

7. References
Belyakov, V., Zhabko, A., Kavin, A., Kharitonov, V., Misenov, B., Mitrishkin, Y., Ovsyannikov,
A. & Veremey, E. (1999). Linear quadratic Gaussian controller design for plasma cur-
rent, position and shape control system in ITER. Fusion Engineering and Design, Vol.
45, No. 1, pp. 55–64.
Camacho E.F. & Bordons C. (1999). Model Predictive Control, Springer-Verlag, London.
Gribov, Y., Albanese, R., Ambrosino, G., Ariola, M., Bulmer, R., Cavinato, M., Coccorese, E.,
Fujieda, H., Kavin A. et. al. (2000). ITER-FEAT scenarios and plasma position/shape
control, Proc. 18th IAEA Fusion Energy Conference, Sorrento, Italy, 2000, ITERP/02.
Hendricks, E., Jannerup, O. & Sorensen, P.H. (2008) Linear Systems Control: Deterministic and
Stochastic Methods, Springer-Verlag, Berlin.
Maciejowski, J. M. (2002). Predictive Control with Constraints, Prentice Hall.
Misenov, B.A., Ovsyannikov, D.A., Ovsyannikov, A.D., Veremey, E.I. & Zhabko, A.P. (2000).
Analysis and synthesis of plasma stabilization systems in tokamaks, Proc. 11th IFAC
Workshop. Control Applications of Optimization, Vol.1, pp. 255-260, New York.
Morari, M., Garcia, C.E., Lee, J.H. & Prett D.M. (1994). Model Predictive Control, Prentice Hall,
New York.
Ovsyannikov, D. A., Ovsyannikov, A. D., Zhabko, A. P., Veremey, E. I., Makeev I. V.,
Belyakov V. A., Kavin A. A. & McArdle G. J. (2005). Robust features analysis for the
MAST plasma vertical feedback control system.(2005). 2005 International Conference
on Physics and Control, PhysCon 2005, Proceedings, 2005, pp. 69–74.
Plasma stabilization system design on the base of model predictive control 221

Ovsyannikov D. A., Veremey E. I., Zhabko A. P., Ovsyannikov A. D., Makeev I. V., Belyakov
V. A., Kavin A. A., Gryaznevich M. P. & McArdle G. J.(2005) Mathematical methods
of plasma vertical stabilization in modern tokamaks, in Nuclear Fusion, Vol.46, pp.
652-657 (2006).
222 Model Predictive Control
Predictive Control of Tethered Satellite Systems 223

10
x

Predictive Control of
Tethered Satellite Systems
Paul Williams
Delft University of Technology
Australia

1. Introduction
Tethered satellite systems have many potential applications, ranging from upper
atmospheric research (Colombo et al., 1975) to momentum transfer (Nordley & Forward,
2001; Williams et al., 2004). The major dynamical features of the system have been studied
extensively (Misra & Modi, 1986), but there still remain open questions with regard to
control (Blanksby & Trivailo, 2000). Many of the open issues stem from the fact that there
have been limited flight tests. The most recent flight of the Young Engineers’ Satellite 2
(YES-2) highlighted from its results that tether dynamic modelling is relatively mature, but
that there is a need to provide fault tolerant design in the control and sensor subsystems
(Kruijff et al., 2009).
In applications such as momentum transfer and payload capture, it is imperative that
robust, accurate and efficient controllers can be designed. For example, although it is
conceivable to use onboard thrusters to manipulate the motion of the tethered satellite, this
negates some of the advantages of using tethers, i.e., little to no fuel expenditure in ideal
circumstances. The main source of control, therefore, has to be sought from manipulating
the length of deployed tether. This has two main aims: first, the length of tether directly
controls the distance of the tether tip from the main spacecraft, and second, changes in
tether length induce Coriolis-type forces on the system due to the orbital motion, which
allows indirect control over the swing motion of the tether (librations). Typically, control
over the tether length is achieved via manipulating the tension at the mother satellite (Rupp,
1975; Lorenzini et al., 1996). This can help to prevent the tether from becoming slack – a
situation that can lead to loss of control of the system.
A variety of different control strategies have been proposed in the literature on tethered
systems. Much of the earlier work focused on controlling the deployment and retrieval
processes (Xu et al., 1981; Misra & Modi, 1982; Fujii & Anazawa, 1994). This was usually
achieved by combining an open-loop length control scheme with feedback of the tether
states, either appearing linearly or nonlinearly. Other schemes were devoted to ensuring
nonlinear asymptotic stability through the use of Lyapunov’s second method (Fujii &
Ishijima, 1989; Fujii, 1991; Vadali & Kim, 1991). Most of these techniques do not ensure well-
behaved dynamics, and can be hard to tune to make the deployment and retrieval fast.
224 Model Predictive Control

Because deployment and retrieval is an inherent two-point boundary value problem, it

makes much more sense to approach the problem from the point-of-view of optimal control.
Several examples of the application of optimal control theory to tethered satellite systems
can be found (Fujii & Anazawa, 1994; Barkow, 2003; Lakso & Coverstone, 2000). However,
the direct application of the necessary conditions for optimality leads to an extremely
numerically sensitive two-point boundary value problem. The state-costate equations are
well-known to suffer from instability, but the tethered satellite problem is notorious because
of the instability of the state equations to small errors in the control tension. More recent
work has focused on the application of direct transcription methods to the tethered satellite
problem (Lakso & Coverstone, 2000; Williams, 2008; Williams & Blanksby, 2008). This
provides advantages with respect to robustness of convergence and is typically orders of
magnitude faster than other methods.
In recent work, the effect of the performance index used in solving the optimal control
problem for tethered satellites was examined in detail (Williams, 2008). The work in
(Williams, 2008) was prompted by the fact that bang-bang tension control trajectories have
been proposed (Barkow, 2003), which is extremely undesirable for controlling a flexible
tether. The conclusions reached in (Williams, 2008) suggest that an inelastic tether model
can be sufficient to design the open-loop trajectory, provided the cost function is suitably
selected. Suitable costs include the square of the tether length acceleration, tension rate or
tension acceleration. These trajectories lead to very smooth variations in the dynamics,
which ultimately improves the tracking capability of feedback controllers, and reduces the
probability of instabilities.
Much of the previous work on optimal control of tethered satellites has focused on
obtaining solutions, as opposed to obtaining rapid solutions. Some of the ideas that will be
explored in this paper have been discussed in (Williams, 2004), which presented two
approaches for implementing an optimal-based controller for tethered satellites. One of the
methods was based on quasilinearization of the necessary conditions for optimality
combined with a pseudospectral discretization, whereas the second was a direct
discretization of the continuous optimal control problem. In (Williams, 2004), NPSOL was
used as the nonlinear programming (NLP) solver, which implements methods based on
dense linear algebra, and is significantly slower than the sparse counterpart SNOPT (Gill et
al., 2002).
The aforementioned YES-2 mission had the aim of deploying a 32 km long tether in two
phases. The first phase had the objective of stabilizing the tether swinging motion
(librations) at the local vertical with the tether length at 3.5 km. The second phase had the
objective of inducing a sufficient swinging motion at the end of deployment to allow a
specially designed payload to re-enter the atmosphere and be recovered in Khazikstan. The
deployment controller consisted of using a reference trajectory computed offline via direct
transcription (Williams et al., 2008), in combination with a feedback controller to stabilize
the deployment dynamics. The feedback controller used a time-varying feedback gain
calculated via a receding horizon approach documented in (Williams, 2005). Flight results
showed that despite very large perturbations from nominal, the tether was deployed
successfully in the first phase. An issue with one of the sensors that measured the
deployment rate caused the feedback controller to believe that the tether was being
deployed too slowly. As a consequence, the tether was deployed too quickly. It has been
Predictive Control of Tethered Satellite Systems 225

shown that the tether was nonetheless fully deployed, making it the longest tether ever
deployed in space.
The aim of this Chapter is to explore the possibility of providing real-time optimal control
for a tethered satellite system. A realistic tether model is combined with a nonlinear Kalman
filter for estimating the tether state based on available measurements. A nonlinear model
predictive controller is implemented to satisfy the mission requirements.

2. System Model
In order to generate rapid optimal trajectories and test closed-loop performance for a real
system, it is necessary to introduce mathematical models of varying fidelity. In this chapter,
two models are distinguished: 1) a high fidelity truth model, 2) a low fidelity control model.
A truth model is required for testing the closed-loop performance of the controller in a
representative environment. Typically, the truth model will incorporate effects that are not
present in the model used by the controller. In the simplest case, these can be environmental
disturbances. Truth models are usually of higher fidelity than the control model, and as
such, they become difficult to use for real-time closed-loop control. For this reason, it is
necessary to employ a reduced order model in the controller. It should be pointed out that a
truth model will typically include a set of parameter perturbations that alter the
characteristics of the simulated system compared to the assumptions made in the control
model. Such perturbations are used in Monte Carlo simulations of the closed-loop system to
gather statistics on the controller performance.
For the particular case of a tethered satellite system, there are a number of important
dynamics that exist in the real system: 1) Rigid-body, librations of the tether in- and out-
plane, 2) Lateral string oscillations of the tether between the tether attachment points, 3)
Longitudinal spring-mass oscillations of the tether, 4) Rigid body motions of the end bodies,
5) Orbital perturbations caused by exchange of angular momentum from the tethered
system with orbital angular momentum. All of these dynamic modes are coupled to
varying degrees. However, the dominant dynamics are due to (1) and (2) as these directly
impact the short-term response of the system.
The following subsections derive the fundamental equations of motion for modeling the
tethered system taking into account the dominant dynamics. A simplified model suitable
for model predictive control is then developed.

2.1 Truth Model

The most sophisticated models for tethered satellite systems treat the full effects of tether
elasticity and flexibility. Examples include models based on discretization by assumed
modes (Xu et al., 1986) or discretization by lumped masses (Kim & Vadali, 1995). In a
typical lumped mass model, the tether is discretized into a series of point masses connected
by elastic springs. The tension in each element can be computed explicitly based on the
positions of the adjacent lumped masses. It is well known that the equations of motion for
the system are ‘stiff’, referring to the fact that the dynamics occur over very different
timescales, requiring small integration step sizes to capture the very high frequency modes.
For a tethered satellite system, the high frequency modes are the longitudinal elastic modes,
followed by the string modes of the tether, libration modes, and finally the orbital motion.
For short duration missions or analysis, the longitudinal modes are unlikely to have a
226 Model Predictive Control

significant effect on the overall motion (provided the tether remains taut). Thus, in this
model the effects of longitudinal vibrations are ignored, and the tether is divided into a
series of point masses connected via inelastic links. The geometric shortening of the
distance to the tether tip is accounted for due to the changes in geometry of the system, but
stretching of the tether is not. The degree of approximation is controlled by the number of
discretized elements that are used.
The tether is modeled as consisting of a series of n point masses connected via inelastic
links, as shown in Fig. 1. The ( x , y , z ) coordinate system rotates at the orbit angular velocity
and is assumed to be attached at the center of mass of the orbit (mother satellite). Although
not a necessary assumption in the model, it is assumed that the orbit of the mother satellite
is prescribed and remains Keplerian. In general, this coordinate system would orbit in a
plane defined by the classical orbital elements (argument of perigee, inclination, longitude
of ascending node). In the presence of a Newtonian gravitational field, the orientation of the
orbital plane does not affect the system dynamics. However, it does affect any aerodynamic
or electrodynamic forces due to the nature of the Earth’s rotating atmosphere and magnetic
field. These effects are not considered here.

l1 m1
f1
mj
q1

lj fj

Z z mn qj
Y y
ln fn

R qn
x
n

Fig. 1. Discretized multibody tether model.

The acceleration of a mass in the rotating frame is given by

r = ( x - w y - 2 w y - w 2 x )i + ( y + w x + 2 w x - w 2 y ) j + zk (1)

where w = k 2 m / p 3 is the orbital angular velocity, p = a(1 - e 2 ) is the semilatus rectum,

m is the Earth’s gravitational parameter, e is the orbit eccentricity, k = 1 + e cos n , and a is
the orbit semimajor axis. The contribution of forces due to the gravity-gradient is given by

mx j my j mz j w2x j w2y j w2zj

Fjg = 2 m j 3
i - mj 3
j - mj 3
k = 2m j i - mj j - mj k (2)
R R R k k k
Predictive Control of Tethered Satellite Systems 227

Note that in Equation (1), the contributions due to the center of mass motion R and
corresponding true anomaly n are cancelled with the Newtonian gravity terms for the
system center of mass. This is valid if the system is assumed to be in a Keplerian orbit.
Define the tension vector in the j th segment as

Tj = Tj ( cos q j cos f j i + sin q j cos f j j + sin f j k ) (3)

Also, define the j th mass in terms of the generalized coordinates,

n
x j = å lk cos qk cos fk (4)
k= j
n
y j = å lk sin qk cos fk (5)
k= j
n
z j = å lk sin fk (6)
k= j

The tension forces on the j th mass are given by

ìï Tj-1 - Tj , 1< j£n
Fjtension := ï í (7)
ïîï -Tj , j=1

The equations of motion can be expressed in component form as

w2 Fjx
x j - w y j - 2w y j - w 2 x j - 2 xj =
k mj

w2 Fjy
y j + w x j + 2 w x j - w 2 y j + yj = (8)
k mj
w2 Fjz
z j + zj =
k mj

where m j is the mass of the jth cable mass, and (Fjx , Fjy , Fjz ) is the vector of external forces
acting on the jth mass in the orbital frame. Substitution of Equations (4) through (6) into
Equation (8) gives the governing equations of motion in spherical coordinates. The
equations of motion may be decoupled by employing a matrix transformation and forward
substitution of the results. By multiplying the vector of Equation (8) by the matrix

é - sin q j cos q j 0 ù
ê ú
é C j ù = ê - cos q j sin f j - sin q j sin f j cos f j úú
ê (9)
ë û
ê ú
êë cos q j cos f j sin q j cos f j sin f j úû
the general decoupled equations of motion can be expressed as
228 Model Predictive Control

é lj ù w2 Tj-1 sin q j cos q j-1 cos f j-1

qj = -w + 2 ( q j + w ) êê f j tan f j - úú - 3 sin q j cos q j -
êë l j úû k m j l j cos f j
Fjx sin q j Tj-1 cos q j sin q j-1 cos f j-1 Tj +1 cos q j sin q j +1 cos f j +1
- + + (10)
m j l j cos f j m j l j cos f j m j +1l j cos f j
Tj +1 sin q j cos q j +1 cos f j +1 Fjy cos q j Fjx+1 sin q j Fjy+1 cos q j
- + + -
m j +1l j cos f j m j l j cos f j m j +1l j cos f j m j +1l j cos f j
lj é 2 3w 2 cos 2 q j ù
fj = -2 f j - êê ( q j + w ) + ú sin f j cos f j
ú
lj k
ëê ûú
Tj-1
+
mjlj
( cos fj sin fj-1 - cos q j sin f j cos q j-1 cos fj-1 - sin q j sin fj sin q j-1 cos f j-1 )
Tj +1
+
m j + 1l j
( cos fj sin fj +1 - cos q j sin fj cos q j +1 cos f j +1 - sin q j sin fj sin q j +1 cos fj +1 ) (11)
y y
Fjz cos f j Fjz+1 cos f j Fj +1 sin q j sin f j Fj sin q j sin f j Fjx cos q j sin f j
+ - + - -
mjlj m j + 1l j m j + 1l j m jl j mjlj
Fjx+1 cos q j sin f j
+
m j + 1l j

l = l éê ( q + w )2 cos 2 f + f 2 - w ( 1 - 3cos 2 q cos 2 f ) ùú - Tj - Tj

2
j j
êë j j j
k j j
úû m j m j +1
y y
Fjz sin f j Fjz+1 sin f j Fj sin q j cos f j Fj +1 sin q j cos f j
+ - + -
mj m j +1 mj m j +1
Fjx cos q j cos f j Fjx+1 cos q j cos f j
+ - (12)
mj m j +1
Tj-1
é cos q j cos f j cos q j-1 cos f j-1 + sin f j sin f j-1 + sin q j cos f j sin q j-1 cos f j-1 ù
mj ë û
T j +1
é cos q j cos f j cos q j +1 cos f j +1 + sin f j sin f j +1 + sin q j cos f j sin q j +1 cos f j +1 ù
m j +1 ë û
These equations may be nondimensionalized by utilizing the following relationships
d d d2 d d2
=w , = w + w2 2 (13)
dt dn dt 2 dn dn
Tj
l j = L jL , u j = (14)
m jw 2L
w2
e sin nw = -2 (15)
k
Thus, the following nondimensional equations of motion are obtained
Predictive Control of Tethered Satellite Systems 229

é e sin n L¢j ù 3 u j +1 cos q j sin q j +1 cos f j +1

q ¢¢j = 2 ( q ¢j + 1 ) ê + f ¢j tan f j - ú - sin q j cos q j +
ê k L ú k L j cos f j
ë jû

m j-1 u j-1 cos q j sin q j-1 cos f j-1 m j-1 u j-1 sin q j cos q j-1 cos f j-1
+ -
mj L j cos f j mj L j cos f j
(16)
u j +1 sin q j cos q j +1 cos f j +1 Fjy cos q j Fjx sin q j Fjx+1 sin q j
- + - +
L j cos f j m j w LL j cos f j m j w LL j cos f j m j +1w 2 LL j cos f j
2 2

Fjy+1 cos q j
-
m j +1w 2 LL j cos f j

e sin n L¢j é 2 3cos 2 q j ù m u

f ¢¢j = 2 f ¢ - 2 f ¢j - êê ( q ¢j + 1 ) + ú sin f j cos f j + j-1 j-1 (cos f j sin f j-1
k Lj k ú mj L j
êë úû
u j +1
- cos q j sin f j cos q j-1 cos f j-1 - sin q j sin f j sin q j-1 cos f j-1 ) + (cos f j sin f j +1
Lj
(17)
Fjz cos f j
- cos q j sin f j cos q j +1 cos f j +1 - sin q j sin f j sin q j +1 cos f j +1 ) +
m j w 2 LL j
Fjz+1 cos f j Fjy+1 sin q j sin f j Fjy sin q j sin f j Fjx cos q j sin f j Fjx+1 cos q j sin f j
- 2
+ 2
- 2
- 2
+
m j +1w LL j m j +1w LL j m j w LL j m j w LL j m j +1w 2 LL j
e sin n é 2 1 ù mj
L¢¢j = 2 L¢j + L j ê ( q ¢j + 1 ) cos 2 f j + f ¢j 2 - ( 1 - 3 cos 2 q j cos 2 f j ) ú - u j - u
k êë k úû m j +1 j
Fjz sin f j Fjz+1 sin f j Fjy sin q j cos f j Fjy+1 sin q j cos f j
+ - + -
m jw 2L m j + 1w 2 L m jw 2L m j + 1w 2 L
Fjx cos q j cos f j Fjx+1 cos q j cos f j m j-1
+ - + u (cos q j cos f j cos q j-1 cos f j-1 (18)
m jw L 2 2
m j + 1w L m j j-1
+ sin f j sin f j-1 + sin q j cos f j sin q j-1 cos f j-1 ) + u j +1 (cos q j cos f j cos q j +1 cos f j +1
+ sin f j sin f j +1 + sin q j cos f j sin q j +1 cos f j +1 )
Equations (16) through (18) utilize the orbit true anomaly n as independent variable, and L
is a scaling length representing the length of each tether element when fully deployed. The
applicable boundary conditions are

m0 = 0, u0 = 0, mn+1 = ¥, un+1 = 0 (19)

The equations (16) through (18) define the dynamics of the tethered satellite system using
spherical coordinates. These are not as general as Cartesian coordinates due to the
singularity introduced when f j = - p2 , p2 . This represents very large out of plane librational
motion or very large out of plane lateral motion. Although this is a limitation of the model,
such situations need to be avoided for most practical missions.
230 Model Predictive Control

2.1.1 Variable Length Case

The tether is modeled as a collection of lumped masses connected by inelastic links, which
makes dealing with the case of a variable length tether more difficult than if the tether was
modeled as a single link. In particular, it is necessary to have a state vector of variable
dimension and to add and subtract elements from the model at appropriate times. When
the tether is treated as elastic, great care needs to be exercised to ensure that the introduction
of new elements does not create unnecessary cable oscillations. This can happen if the
position of the new mass results in the incorrect tension in the new element. However, for
an inelastic tether, the introduction of a new mass occurs such that it is placed along the
same line as the existing element. Thus, the new initial conditions for the incoming element
are that it has the same angles and angle rates as the existing element (closest to the
deployer). Alternative formulations based on the variation principle of Hamilton-
Ostrogradksi and which transform the deployed length to a fixed interval by means of a
new spatial coordinate have also been used (Wiedermann et al., 1999). However, this was
not considered in this work.
If the critical length for introduction of a new element is defined as L*  1 + k * , then the
new element is initialized with a length of k * in nondimensional units, and the same length
rate as the previous nth element. During retrieval, elements must be removed. Here, the
nth element to be removed and the (n-1)th element need to be used to update the initial
conditions for the new n* th element. In this work, the position and velocity of the (n-1)th
mass is used to initialize the n* th element. Thus, let

xn-1 = Ln cos qn cos fn + Ln-1 cos qn-1 cos fn-1

y n-1 = Ln sin qn cos fn + Ln-1 sin qn-1 cos fn-1 (20)
zn-1 = Ln sin fn + Ln-1 sin fn-1
From which
Ln* = xn2-1 + yn2-1 + zn2-1 , qn* = atan2( yn-1 , xn-1 ), fn* = sin-1 ( zn-1 / Ln* ) (21)

where atan2 represents the four quadrant inverse tangent where the usual arctangent is
y
defined by tan-1 ( xn-1 ) . Similarly, the relative velocity of the (n-1)th mass in the rotating
n-1

frame is given by

xn¢ -1 = L¢n cos qn cos fn - Lnqn¢ sin qn cos fn - Lnfn¢ cos qn sin fn
+L¢n-1 cos qn-1 cos fn-1 - Ln-1qn¢ -1 sin qn-1 cos fn-1 - Ln-1fn¢ -1 cos qn-1 sin fn-1
y n¢ -1 = L¢n sin qn cos fn + Lnqn¢ cos qn cos fn - Lnfn¢ sin qn sin fn (22)
+L¢n-1 sin qn-1 cos fn-1 + Ln-1qn¢ -1 cos qn-1 cos fn-1 - Ln-1fn¢ -1 sin qn-1 sin fn-1
zn¢ -1 = L¢n sin fn + Lnfn¢ cos fn + L¢n-1 sin fn-1 + Ln-1fn¢ -1 cos fn-1
From which
L¢n* = xn¢ -1 cos fn* cos qn* + yn¢ -1 cos fn* sin qn* + zn¢ -1 sin fn*
qn¢ * = ( yn¢ -1 cos qn* - xn¢ -1 sin qn* ) / ( Ln* cos fn* ) (23)
fn¢ * = ( zn¢ -1 cos fn* - xn¢ -1 sin fn* cos qn* - yn¢ -1 sin fn* sin qn* ) / Ln*
Predictive Control of Tethered Satellite Systems 231

It should be noted that these updates keep the position and velocity of the (n-1)th mass the
same across the update. The reason for this is that the positions and velocities of all
subsequent masses depend on the position/velocity of the nth mass. Hence, if this is
changed, then the position and velocity of all masses representing the tether change
instantaneously. The accuracy of the updates depend on the transition parameter k ** , which
is used to monitor the length of the nth segment. An element is removed when Ln < k ** .
Because the tether is inelastic, altering the length of the new nth element does not keep the
total tether length or mass constant unless the nth and (n-1)th elements are tangential.
Therefore, by choosing k ** small enough, the errors in the approximation can be made
small.
For control purposes, it is assumed that the rate of change of reel-rate is controlled. Thus,
L¢¢n is specified or determined through a control law. This means that the nth element is
allowed to vary in length, but all other segments remained fixed in length. The problem is
to then solve for the unknown tension constraints that enforce constant total length of the
remaining segments, as well as the acceleration of the nth segment. Once these are known,
they are back-substituted into Equations (16) and (17), as well as Equation (18) for the nth
element. The equations formed by the set (18) are linear in the tensions u j , and can thus be
solved using standard techniques. This assumes that the segment lengths, length rates, and
length accelerations are specified. In this work, LAPACK is utilized in solving the
simultaneous equations.

2.1.2 Fixed Length Case

To simulate the case of a fixed length tether, Equations (18) are set to zero for j = 1,..., n ,
allowing the unknown tensions u j , j = 1,..., n to be determined. The resulting tensions are
substituted back into the librational dynamics to determine the evolution of the system
dynamics.

2.2 Control Model

The predominant modeling assumption that is used in the literature insofar as control of
tethered satellite systems is concerned is that the system can be modeled with three degrees
of freedom (Williams, 2008). In other words, when dealing with the librational motion of the
system, it is sufficient to model it using spherical coordinates representing the dynamics of
the subsatellite. This effectively treats the tether as a straight body, which can either be
modeled as an inelastic or elastic rod. Early work has neglected the tether mass since its
contribution to the librational motion can be considered relatively small (Fujii & Anazawa,
1994). This is due to the fact that the tether is axisymmetric. When large changes in length
are considered, the effect of tether mass becomes more important. Moreover, it is essential to
include the effects of tether mass when designing tension control laws because there is a
nonlinear relationship between tension and tether mass. However, when performing
preliminary analyses, it is sufficient to ignore such effects and compensate for these later in
the design.
Although the assumption of treating the tether as a straight rod is often a good one, it can
create some problems in practice. For example, all tether string vibrations are neglected,
232 Model Predictive Control

which play a very important role in electrodynamic systems or systems subjected to long-
term perturbations. Furthermore, large changes in deployment velocity can induce
significant distortions to the tether shape, which ultimately affects the accuracy of the
deployment control laws. Earlier work focused much attention on the dynamics of tethers
during length changes, particularly retrieval (Misra & Modi, 1986). In the earlier work,
assumed modes was typically the method of choice (Misra & Modi, 1982). However, where
optimal control methods are employed, high frequency dynamics can be difficult to handle
even with modern methods. For this reason, most optimal deployment/retrieval schemes
consider the tether as inelastic.

2.1 Straight, Inelastic Tether Model

In this model, the tether is assumed to be straight and inextensible, uniform in mass, the end
masses are assumed to be point masses, and the tether is deployed from one end mass only.
The generalized coordinates are selected as the tether in-plane libration angle, q, the out-of-
plane tether libration angle, f, and the tether length, l.
The radius vector to the center of mass may be written in inertial coordinates as

R = R cos n i + R sin n j (24)

From which the kinetic energy due to translation of the center of mass is derived as

Tt = 21 m( R 2 + R 2n 2 ) (25)

where m = m1 + mt + m2 is the total system mass, m1 = m10 - mt is the mass of the mother
satellite, mt is the tether mass, m2 is the subsatellite mass, and m10 is the mass of the mother
satellite prior to deployment of the tether.
The rotational kinetic energy is determined via

Tr = 21 wT [ I ]w (26)

where w is the inertial angular velocity of the tether in the tether body frame

w = ( n sin f + q sin f ) i - ( f ) j + ( n cos f + q cos f )k (27)

Thus we have that

2
Tr = 21 m*l 2 [f 2 + ( n + q ) cos2 f ] (28)

(
and m* = m1 +
mt
2 )( m2 + m2 ) / m - mt / 6
t
is the system reduced mass. The kinetic energy
due to deployment is obtained as
Predictive Control of Tethered Satellite Systems 233

1 m1 ( m2 + mt ) 2
Te = l (29)
2 m

which accounts for the fact that the tether is modeled as stationary inside the deployer and
is accelerated to the deployment velocity after exiting the deployer. This introduces a
thrust-like term into the equations of motion, which affects the value of the tether tension.
The system gravitational potential energy is (assuming a second order gravity-gradient
expansion)
mm mm*l 2
R
V =-+ ( 1 - 3 cos2 q cos2 f ) (30)
2 R3
The Lagrangian may be formed as
2
L = 1 m( R 2 + R 2n 2 ) + 1 m* l 2 [f 2 + ( n + q ) cos 2 f ]
2 2
(31)
m ( m + mt ) 2 mm mm* l 2
+ 21 1 2
m
l +
R
- ( 1 - 3 cos2 q cos2 f )
2 R3

Under the assumption of a Keplerian reference orbit for the center of mass, the
nondimensional equations of motion can be written as

é m1 ( m2 + 2t ) L¢ ùú 3
m
e sin n Qq
q ¢¢ = 2 ( q ¢ + 1 ) êê + f ¢ tan f - * ú - k sin q cos q + * 2 2 2 (32)
ê k mm L ú m L Lr n cos 2 f
ë û
m1 ( m2 + ) L¢ f ¢ - é
mt
2 e sin n 2 2 3 ù Qf
f ¢¢ = f¢ - 2 ê ( q ¢ + 1 ) + cos 2 q ú sin f cos f + * 2 2 2 (33)
k mm* L êë k úû m L Lr n
( 2 m1 - m ) 2t L¢ 2 æç m2 + 2t ö÷÷
m m
2 e sin n 2
L¢¢ = L¢ - + çç ÷ L[f ¢ 2 + ( q ¢ + 1 ) cos 2 f
k m1 ( m2 + mt ) L çè m2 + mt ÷÷ø
(34)
1 T
+ ( 3 cos 2 q cos 2 f - 1 )] -
k m1n 2 Lr ( m2 + mt ) / m

where L = l / Lr is the nondimensional tether length, Lr is a reference tether length, T is the

tether tension, and ()¢ = d() / dn . The generalized forces Qq and Qf are due to distributed
forces along the tether, which are typically assumed to be negligible.

3. Sensor models
The full dynamic state of the tether is not directly measurable. Furthermore, the presence of
measurement noise means that some kind of filtering is usually necessary before directly
using measurements from the sensors in the feedback controller. The following
measurements are assumed to be available: 1) Tension force at the deployer, 2) Deployment
rate, 3) GPS position of the subsatellite. Models of each of these are developed in the
subsections below.
234 Model Predictive Control

3.1 Tension Model

The tension force measured at the deployer differs from the force predicted by the control
model due to the presence of tether oscillations and sensor noise. The magnitude and
direction of the force in the tether is obtained from the multibody tether model. The tension
force in the orbital frame is given by

Tx = unmnw 2 L cos qn cos fn + wTx

Ty = unmnw 2 L sin qn cos fn + wTy (35)
2
Tz = unmnw L sin fn + wTz
where the w terms are zero mean, Gaussian measurement noise with covariance RT.

3.2 Reel-Rate Model

In general, the length of the deployed tether can be measured quite accurately. In this
chapter, the reel-rate is measured at the deployer according to

L = L¢nwLn + wL (36)

where wL is a zero mean, Gaussian measurement noise with covariance RL .

3.3 GPS Model

GPS measurements of the two end bodies significantly improve the estimation performance
of the system. The position of the mother satellite is required to form the origin of the orbital
coordinate system (in case of non-Keplerian motion), and the position of the subsatellite
allows observations of the subsatellite range and relative position (libration state). Only
position information is used in the estimator. The processed relative position is modeled in
the sensor model, as opposed to modeling the satellite constellation and pseudoranges. The
processed position error is modeled as a random walk process

wx wy w
dx = , d y = , d z = z (37)
tGPS tGPS tGPS

where wx,y,z are zero mean white noise processes with covariance RGPS, and tGPS is a time
constant. This model takes into account that the GPS measurement errors are in fact time-
correlated.

4. State Estimation
In order to estimate the full tether state, it is necessary to combine all of the measurements
obtained from the sensors described in Section 3. The most optimal way to combine the
measurements is by applying a Kalman filter. Various forms of the Kalman filter are
available for nonlinear state estimation problems. The two most commonly used filter
implementations are the Extended Kalman Filter (EKF) and the Unscented Kalman Filter
(UKF). The UKF is more robust to filter divergence because it captures the propagation of
Predictive Control of Tethered Satellite Systems 235

uncertainty in the filter states to a higher order than the EKF, which only captures the
propagation to first order. The biggest drawback of the UKF is that it is significantly more
expensive than the EKF. Consider a state vector of dimension nx. The EKF only requires the
propagation of the mean state estimate through the nonlinear model, and three matrix
multiplications of the size of the state vector (nx × nx). The UKF requires the propagation of
2nx + 1 state vectors through the nonlinear model, and the sum of vector outer products to
obtain the state covariance matrix. The added expense can be prohibitive for embedded
real-time systems with small sampling times (i.e., on the order of milliseconds). For the
tethered satellite problem, the timescales of the dynamics are long compared to the available
execution time. Hence, higher-order nonlinear filters can be used to increase performance of
the estimation without loss of real-time capability.
Recently, an alternative to the UKF was introduced that employs a spherical-radial-cubature
rule for numerically integrating the moment integrals needed for nonlinear estimation. The
filter has been called the Cubature Kalman Filter (CKF). This filter is used in this chapter to
perform the nonlinear state estimation.

4.1 Cubature Kalman Filter

In this section, the CKF main steps are summarized. The justification for the methodology is
omitted and may be found in (Guess & Haykin, 2009).
The CKF assumes a discrete time process model of the form

x k +1 = f ( x k , u k , v k , t k ) (38)
y k = h( x k , uk , wk , tk ) (39)

ny
where x k Î  nx is the system state vector, uk Î  nu is the system control input, y k Î  is
nv
the system measurement vector, vk Î  is the vector of process noise, assumed to be
white Gaussian with zero mean and covariance Qk Î  nv ´nv , wk Î  nw is a vector of
measurement noise, assumed to be white Gaussian with zero mean and covariance
Rk Î  nw ´nw . For the results in this paper, the continuous system is converted to a discrete
system by means of a fourth-order Runge-Kutta method.
In the following, the process and measurement noise is implicitly augmented with the state
vector as follows
é xk ù
ê ú
x k = êê vk úú
a
(40)
êw ú
ëê k ûú

The first step in the filtering process is to compute the set of cubature points as follows

 k-1 = ëé xˆ ka-1 + I na ´na Pk , xˆ ka-1 - I na ´na Pk ûù (41)

where xˆ a is the mean estimate of the augmented state vector, and Pk is the covariance
matrix. The cubature points are then propagated through the nonlinear dynamics as follows
236 Model Predictive Control

 k*|k-1 = f ( k-1 , uk , tk ) (42)

The predicted mean for the state estimate is calculated from

2n
1 a

xˆ -
k = å *
2na i =0 i , k|k-1
(43)

The covariance matrix is predicted by

2n
1 a

Pk- = å *  *T
2 na i=0 i , k|k-1 i , k|k-1
- xˆ -
kxˆ-
k
T
(44)

When a measurement is available, the augmented sigma points are propagated through the
measurement equations

k|k-1 = h( k|k-1 , uk , tk ) (45)

The mean predicted observation is obtained by

2n
1 a
2 na å
yˆ -
k = i , k|k-1 (46)
i=0

The innovation covariance is calculated using

2n
1 a

Pkyy = å  T - yˆ -
kyˆ- T
(47)
2na i =0 i , k|k-1 i , k|k-1 k

The cross-correlation matrix is determined from

2n
1 a

Pkxy = å  T - xˆ -
kyˆ- T
(48)
2na i =0 i , k|k-1 i , k|k-1 k

The gain for the Kalman update equations is computed from

k = Pkxy ( Pkyy )-1 (49)

The state estimate is updated with a measurement of the system y k using

k + k ( y k - y
xˆ k = xˆ - ˆ-k ) (50)

and the covariance is updated using

Predictive Control of Tethered Satellite Systems 237

Pk+ = Pk- - k Pkyy kT (51)

It is often necessary to provide numerical remedies for covariance matrices that do not
maintain positive definiteness. Such measures are not discussed here.

5. Optimal Trajectory Generation

Most of the model predictive control strategies that have been suggested in the literature are
based on low-order discretizations of the system dynamics, such as Euler integration.
Dunbar et al. (2002) applied receding horizon control to the Caltech Ducted Fan based on a
B-spline parameterization of the trajectories. In recent years, pseudospectral methods, and
in particular the Legendre pseudospectral (PS) method (Elnagar, 1995; Ross & Fahroo, 2003),
have been used for real-time generation of optimal trajectories for many systems. The
traditional PS approach discretizes the dynamics via differentiation operators applied to
expansions of the states in terms of Lagrange polynomial bases. Another approach is to
discretize the dynamics via Gauss-Lobatto quadratures. The approach has been more fully
described by Williams (2006). The latter approach is used here.

5.1 Discretization approach

Instead of presenting a general approach to solving optimal control problems, the Gauss-
Lobatto approach presented in this section is restricted to the form of the problem solved
here. The goal is to find the state and control history { x(t ), u(t ) } to minimize the cost
function
tf
 =  éë x(t f ) ùû + ò  éë x(t * ), u(t * ), t * ùû dt * (52)
t0

subject to the nonlinear state equations

x (t ) = f [ x(t ), u(t ), t ] (53)

the initial and terminal constraints

y0 [ x(t0 ) ] = 0 (54)
y f éë x(t f ) ùû = 0 (55)

the mixed state-control path constraints

g L £ g [ x(t ), u(t ), t ] £ gU (56)

and the box constraints

x L £ x(t ) £ xU , uL £ u(t ) £ uU (57)

238 Model Predictive Control

where x Î  nx are the state variables, u Î  nu are the control inputs, t Î  is the time,
 :  nx ´    is the Mayer component of cost function, i.e., the terminal, non-integral
cost in Eq. (52),  :  nx ´  nu ´    is the Bolza component of the cost function, i.e., the
integral cost in Eq. (52), y0 Î  nx ´    n0 are the initial point conditions,
nf ng
nx
yf Î  ´    are the final point conditions, and g L Î  nx ´  nu ´    and
ng
gU Î  nx ´  nu ´    are the lower and upper bounds on the path constraints.
The basic idea behind the Gauss-Lobatto quadrature discretization is to approximate the
vector field by an N th degree Lagrange interpolating polynomial

f ( t ) » fN (t ) (58)

expanded using values of the vector field at the set of Legendre-Gauss-Lobatto (LGL) points.
The LGL points are defined on the interval t Î [-1,1] and correspond to the zeros of the
derivative of the N th degree Legendre polynomial, LN (t ) , as well as the end points –1 and
1. The computation time is related to the time domain by the transformation

(t f - t0 ) (t f + t0 )
t= t+ (59)
2 2
The Lagrange interpolating polynomials are written as
N
fN (t ) = å fkfk ( t ) (60)
k =0

where t = t(t ) because of the shift in the computational domain. The Lagrange
polynomials may be expressed in terms of the Legendre polynomials as

( t 2 - 1 )LN¢ (t )
fk (t ) = , k = 0,..., N (61)
( t - t k ) N ( N + 1)LN ( t k )

Approximations to the state equations are obtained by integrating Eq. (60),

( t f - t0 ) 1 N
x k = x0 +
2 ò-1 å
j =0
f j (t ) f (t j ) dt , k = 1,..., N (62)

Eq. (62) can be re-written in the form of Gauss-Lobatto quadrature approximations as

( t f - t0 ) N
x k = x0 +
2 å k-1, j f (t j ), k = 1,..., N (63)
j=0

where the entries of the N ´ ( N + 1 ) integration matrix  are derived by Williams (2006).
The cost function is approximated via a full Gauss-Lobatto quadrature as
Predictive Control of Tethered Satellite Systems 239

(t f - t0 ) N
 N = [ x N ] +
2 å  éë x j , u j , t j ùû w j (64)
j =0

Thus the discrete states and controls at the LGL points ( x0 ,..., x N , u0 ,..., uN ) are the
optimization parameters, which means that the path constraints and box constraints are
easily enforced. The continuous problem has been converted into a large-scale parameter
optimization problem. The resulting nonlinear programming problem is solved using
SNOPT in this work. In all cases analytic Jacobians of the cost and discretized equations of
motion are provided to SNOPT.
Alternatives to utilization of nonlinear optimization strategies have also been suggested. An
example of an alternative is the use of iterative linear approximations, where the solution is
linearized around the best guess of the optimal trajectory. This approach is discussed in
more detail for the pseudospectral method in (Williams, 2004).

5.2 Optimal Control Strategy

Using the notation presented above, the basic notion of the real-time optimal control
strategy is summarized in Fig. 2. For a given mission objective, a suitable cost function and
final conditions would usually be known a priori. This is input into the two-point boundary
value problem (TPBVP) solver, which generates the open-loop optimal trajectories
x * (t ), u * (t ) . The optimal control input is then used in the real-system, denoted by the
“Control Actuators” block, producing the observation vector y(t k ) . This is fed into the CKF
to produce a state estimate, which is then fed back to update the optimal trajectory by letting
t0 = t , and using t f - t as the time to go.
Imposing hard terminal boundary conditions can make the optimization problem infeasible
as t f - t  0 . In many applications of nonlinear optimal control, a receding horizon
strategy is used, whereby the constraints are always imposed at the end of a finite horizon
T = t f - t , where T is a constant, rather than at a fixed time. This can provide advantages
with respect to robustness of the controller. This strategy, as well as some additional
strategies, are discussed below.

Cost function,
control
constraints, initial Discrete Optimal x * (t ), u * (t ) Control y(t k )
and final Control Problem: Actuators
conditions TPBVP

x(t k )

Cubature
Kalman Filter

Fig. 2. Real-Time Optimal Control Strategy.

240 Model Predictive Control

5.3 Issues in Real-Time Optimal Control

Although the architecture for solving the optimal control problem presented in the previous
section is capable of rapidly generating optimal trajectories, there are several important
issues that need to be taken into consideration before implementing the method. Some of
these have already been discussed briefly, but because of their importance they will be
reiterated in the following subsections.

5.3.1 Initial Guess

One issue that governs the success of the NLP finding a solution rapidly is the initial guess
that is provided. Although convergence of SNOPT can be achieved from random guesses
(Ross & Gong, 2008), the ability to converge from a bad guess is not really of significant
benefit. The main issue is the speed with which a feasible solution is generated as a function
of the initial guess. It is conceivable for many scenarios that good initial guesses are
available. For example, for tethered satellite systems, deployment and retrieval would
probably occur from fixed initial and terminal points. Therefore, one would expect that this
solution would be readily available. In fact, in this work, it is assumed that these “reference”
trajectories have already been determined. Hence, each re-optimization would take place
with the initial guess provided from the previous solution, and the first optimization would
take place using the stored reference solution. In most circumstances then, the largest
disturbance or perturbation would occur at the initial time, where the initial state may be
some “distance” from the stored solution. Nevertheless, the stored solution is still a “good”
guess for optimizing the trajectory. This essentially means that the study of the
computational performance should be focused on the initial sample, which would
conceivably take much longer than the remaining samples.

5.3.2 Issues in Updating the Control

For many systems, the delay in computing the new control sequences is not negligible.
Therefore, it is preferable to develop methods that adequately deal with the computational
delay for the general case. The simplest way of updating the control input is illustrated in
Fig. 3. The method uses only the latest information and does not explicitly account for the
time delay. At the time t = ti , a sample of the system states is taken x(ti ) . This information
is used to generate a new optimal trajectory xi (t ), ui (t ) . However, the computation time
required to calculate the trajectory is given by Dti = ti +1 - ti . During the delay, the
previous optimal control input ui-1 (t ) is applied. As soon as the new optimal control is
available it is applied (at t = ti +1 ). However, the new control contains a portion of time that
has already expired. This means that there is likely to be a discontinuity in the control at the
new sample time t = ti +1 . The new control is applied until the new optimal trajectory,
corresponding to the states sampled at x(ti +1 ) , is computed. At this point, the process
repeats until t = t f . Note that although the updates occur in discrete time, the actual
control input is applied at the actuator by interpolation of the reference controls.
Predictive Control of Tethered Satellite Systems 241

Actual state/control
x( t )
Optimal state/control xi (t ), ui (t )
Optimal state/control xi +1 (t ), ui +1 (t )
Optimal state/control xi + 2 (t ), ui + 2 (t )

u(t )

ti ti +1 ti + 2 ti + 3 t
Fig. 3. Updating the Optimal Control using Only Latest Information.

Due to sensor noise and measurement errors, the state sampled at the new sample time
x(ti +1 ) is unlikely to correspond to the optimal trajectory that is computed from xi (ti +1 ) .
Therefore, in this approach, it is possible that the time delay could cause instability in the
algorithm because the states are never matching exactly at the time the new control is
implemented. To reduce the effect of this problem, it is possible to employ model prediction
to estimate the states. In this second approach, the sample time is not determined by the
time required to compute the trajectory, but is some prescribed value. The sampling time
must be sufficient to allow the prediction of the states and to solve the resulting optimal
control problem, tsol . Hence, Dti > tsol . The basic concept is illustrated in Fig. 4. At time
t = ti , a system state measurement is made x(ti ) . This measurement, together with the
previously determined optimal control and the system model, allows the system state to be
predicted at the new sample time t = ti +1 ,

ti + 1
xˆ (ti +1 ) » x(ti ) + ò x ( ui (t ) ) dt (65)
ti

The new optimal control is then computed from the state xˆ (ti +1 ) . When the system reaches
t = ti +1 , the new control signal is applied, ui +1 (t ) . At the same time, a new measurement is
taken and the process is repeated. This process is designed to reduce instabilities in the
system and to make the computations more accurate. However, it still does not prevent
discontinuities in the control, which for a tethered satellite system could cause elastic
vibrations of the tether. One way to produce a smooth control signal is to constrain the
initial value of the control in the new computation so that
242 Model Predictive Control

ui +1 (ti +1 ) = ui (ti +1 ) (66)

That is, the initial value of the new control is equal to the previously computed control at
time t = ti +1 . It should be noted that the use of prediction assumes coarse measurement
updates from sensors. Higher update rates would allow the Kalman filter to be run up until
the control sampling time, achieving the same effect as the state prediction (except that the
prediction has been corrected for errors). Hence, Fig. 4 shows the procedure with the
predicted state replaced by the estimated state.

5.3.3 Implementing Terminal Constraints

In standard model predictive control, the future horizon over which the optimal control
problem is solved is usually fixed in length. Thus, the implementation of terminal
constraints does not pose a theoretical problem because the aim is usually for stability,
rather than hitting a target. However, there are many situations where the final time may be
fixed by mission requirements, and hence as t f - t  0 the optimal control problem
becomes more and more ill-posed. This is particularly true if there is a large disturbance
near the final time, or if there is some uncertainty in the model. Therefore, it may be
preferable to switch from hard constraints to soft constraints at some prespecified time
t = tcrit , or if the optimization problem does not converge after ncrit successive attempts. It
is important to note that if the optimization fails, the previously converged control is used
until a new control becomes available. Therefore, after ncrit failures, soft terminal
constraints are used under the assumption that the fixed terminal conditions can not be
achieved within the control limits. The soft terminal constraints are defined by


 = 21 éë x(t f ) - x f ùû S f éë x(t f ) - x f ùû (67)

The worst case scenario is for fixed time missions. However, where stability is the main
issue, receding horizon strategies with fixed horizon length can be used. Alternatively, the
time to go can be used up until t = tcrit , at which point the controller is switched from a
fixed terminal time to one with a fixed horizon length defined by T = t f - tcrit . In this
framework, the parameters tcrit and ncrit are design parameters for the system.
It should also be noted that system requirements would typically necessitate an inner-loop
controller be used to track the commands generated by the outer loop (optimal trajectory
generator). An inner-loop is required for systems that have associated uncertainty in
modeling, control actuation, or time delays. In this chapter, the control is applied
completely open-loop between control updates using a time-based lookup table. The loop is
closed only at coarse sampling times.
Predictive Control of Tethered Satellite Systems 243

Actual
Optimal control ui (t )
x( t ) Optimal control ui +1 (t )
Optimal control ui + 2 (t )
Predicted state
Predicted state
Predicted state

u(t )

ti ti +1 ti + 2 ti + 3 t
Fig. 4. Updating the Optimal Control with Prediction and Initial Control Constraint.

5.4 Rigid Model In-Loop Tests

To explore the possibilities of real-time control for tethered satellite systems, a simple, but
representative test problem is utilized. Deployment and retrieval are two benchmark
problems that provide good insight into the capability of a real-time controller. Williams
(2008) demonstrated that deployment and retrieval to and from a set of common boundary
conditions leads to an exact symmetry in the processes. That is, for every optimal
deployment trajectory to and from a set of boundary conditions, there exists a retrieval
trajectory that is mirrored about the local vertical. However, it is also known that retrieval is
unstable, in that small perturbations near the beginning of retrieval are amplified, whereas
small perturbations near the beginning of deployment tend to remain bounded. Therefore,
to test the effectiveness of a real-time optimal controller, the retrieval phase is an ideal test
case.
The benchmark problem is defined in terms of the nondimensional parameters as: Minimize
the cost
tf 2
J = ò ( L¢¢ ) dt (68)
t0

subject to the boundary conditions

[ q , q ¢ , L , L¢ ]t=t0 = [ 0,0,1,0 ] , [ q , q ¢ , L , L¢ ]t=t f = [ 0,0,0.1,0 ] (69)

and the tension control inequality

244 Model Predictive Control

0.01 £ u £ 4 (70)

which is designed to prevent the tether from becoming slack, and to prevent the tether from
severing. The control input for this test case is defined as u = T /[m1n 2 Lr ( m2 + mt ) / m] .

5.4.1 Preliminary Study on Computation Time

To gauge the effectiveness of performing computations of the optimal control in real-time,
the problem of tether retrieval was solved using cold-starts with random perturbations to
the initial conditions. Since the computation of the control is most critical at the initial time
(because the initial state may be very far from the reference state), a numerical study of the
performance of the solution algorithm was run for 1000 computations. In terms of actual
implementation, if the sampling time is short enough, subsequent convergence is almost
always quicker than the initial computation.
The retrieval problem is posed in nondimensional units, with a nondimensional time of 6
rad. For a tether system in low Earth orbit at an altitude of 500 km, the total maneuver time
is roughly 5450 sec. The update time with a good guess of the trajectory averages 0.09 sec in
MATLAB 2009a on a Core 2 Processor running Windows XP. Clearly, this easily allows
real-time computation of the trajectory with over 50000 samples. However, as noted, the
critical time is the first update when the trajectory may be far from the reference or when a
good initial guess may not be available. A study of 1000 computations with different initial
conditions, but with the same infeasible guess for the trajectory was performed. The initial
conditions were distributed randomly in the ranges dq(0) £ 0.2 rad, dq ¢(0) £ 0.1 , and
dL(0) £ 0.02 . Fig. 5 shows a summary of the results from these computations. The level of
discretization was set to be N = 30 for this study. The mean computation time was
determined to be 0.164 sec.
0.35

0.3
Clock time (sec)

0.25

0.2

0.15

0.1
0 200 400 600 800 1000
Sample number
Fig. 5. Summary of Results from Study of Computation of Optimal Trajectories.

The minimum time was 0.102 sec and the maximum time was 0.290 sec. Even in the worst
possible case, it would still be possible to implement a sampled-data feedback controller
(using MATLAB) with roughly 18000 samples. It should also be noted that convergence
was achieved in every case. The CPU time as calculated in Windows represents the worst
Predictive Control of Tethered Satellite Systems 245

case that could be achieved using a dedicated embedded system. The Windows scheduler
can schedule the control process in- and out- at different times. The resolution of the
scheduler can be seen in the discrete banding of the mean CPU time in Fig. 5, rather than
completely random times.

5.4.2 Closed-Loop Control

To examine the actual performance of the controller for dealing with disturbances, the
control model is used with external perturbations included via the Qq and QL terms in the
equations of motion. For simplicity, the perturbations are generated randomly such that
Qq £ 0.05 and QL £ 0.05 . This corresponds to disturbances on the subsatellite on the
order of several Newtons, whose actual values depend on the system geometry. The
number of major iterations was limited to 50.
The terminal weighted matrix is selected as S f = diag[100,100,100,100] , and the controller
is switched at 4 rad from hard terminal constraints to soft constraints. Numerical results are
shown in Fig. 6. Fig. 6a and 6b shows that the terminal constraints are met reasonably
accurately, despite not being enforced with hard constraints. The mean CPU time for the
whole trajectory is 0.159 sec, the standard deviation is 0.0744 sec, the minimum time is 0.04
sec, and the maximum time is 1.442 sec. Prior to the change in controller, the mean CPU
time is 0.1265 sec, whereas after the change the mean CPU time increases to 0.223 sec.
Therefore, the smooth control input in the terminal phases of the trajectory comes at the
expense of a 76% increase in mean computation time. This is still well within the sampling
time of the controller.

6. Closed-Loop Control in Simulation Environment

The results presented in the previous section utilized tension as the control input. Tension
has been widely used as the control input in the literature, but it has several drawbacks. It
introduces long-term errors in the trajectories because of inaccuracies in the system
properties, errors in the gravity model, and tether oscillations. A better choice is to control
the reel speed or rate of change of reel speed. In the high fidelity simulation environment,
the control is implemented as the rate of change of nondimensional reel rate.
40 1

30 0.8

20 0.6
 (rad)

10 0.4

0 0.2

-10 0
0 2 4 6 8 10 0 2 4 6 8 10
 t (rad)  t (rad)

0.5 0.1

0 -0.1
'
'

-0.2

-0.5 -0.3
0 2 4 6 8 10 0 2 4 6 8 10

a)
 t (rad)
b)  t (rad)
246 Model Predictive Control

3.5 0.8

3 0.7

0.6
2.5
Control Tension, u

0.5

CPU Time (sec)

2
0.4
1.5
0.3

1
0.2

0.5 0.1

0 0
c)
0 2 4 6 8 10 0 1000 2000 3000 4000 5000 6000 7000 8000
d)  t (rad)
Sample Number

Fig. 6. Real-Time Computation of Retrieval Trajectory with 1 sec Sampling Time, Receding
Horizon after wt = 4 rad and Model Prediction of States with Continuous Control Enforced,
a) Libration Dynamics, b) Length Dynamics, c) Control Tension, d) Computation Time.

6.1 Simulation Environment

The simulation environment used for testing the closed-loop control behavior is built in
Simulink™, which is itself based on the MATLAB environment. Simulink provides a
graphical approach for modeling and control of complex systems. It has the distinct
advantage of being able to provide generated C-code targeting real-time operation directly
from the underlying model. This feature requires additional supporting tools available from
Mathworks. In the context of the current chapter, a Simulink model is used to simulate four
distinct elements of the system. Fig. 7 illustrates the interconnections of the four system
elements. These are: 1) Variable-Step, Multibody Propagation (bead tether model), 2) Sensor
models, 3) Tether state estimation, and 4) Pseudospectral predictive control. One of the
complicating factors in simulating the predictive control system is that a high-fidelity,
variable step integration algorithm is needed to propagate the multibody dynamic
equations.

Time
SimulationTime
ObservationTime

SampleTime

MPC_Time Truth_Observations Truth_Observations SensorMeasurements SensorMeasurements

1
SampleTime SensorModels
z
Pseudospectral MPC_Time
Predictive
Control StateEstimate
MPC_Control
MPC_Control

Variable-Step, Multibody Propagation Tether State Estimation

1
StateEstimate
z

solveTime

Fig. 7. Simulink simulation model for closed-loop model predictive control.

Predictive Control of Tethered Satellite Systems 247

Although Simulink supports variable-step integration algorithms, it does not easily allow
for the combination of variable-step integration and discrete sampling updates of the system
being propagated. For example, the multibody model requires regular checks on the length
of the deploying segment for the introduction or removal of an element from the model. To
overcome this, a custom S-function block is used which employs the LSODA variable-step
integration library. The LSODA library is coded in Fortran, but was ported to C via f2c.
The sensor models block implements the tension and GPS models for the system. The tether
state estimation block implements the Kalman filter for estimating the tether state in a
discrete-time manner. Finally, the pseudospectral predictive control block implements the
predictive controller.

6.2 Example: Closed-Loop Control with State Estimator

One of the future applications of tethered satellite systems is for capture and rendezvous of
a satellite in a coplanar orbit. In such an application, timing is critical for mission success. A
similar application where timing is not as critical is the deorbit of a payload, similar to the
idea of the YES-2 mission. In these examples, the control objective is similar in that it
requires the generation of a large in-plane swinging motion. As an example, the control
objective of rendezvous with a target satellite is used. The rendezvous conditions have been
derived in detail by Williams (2006) for the general case of circular and elliptical orbits as a
function of tether length.
The objective in this section is to deploy the tether from a length of 1 km to a length of 20 km
to achieve a nondimensional in-plane libration rate of -1.5. For a target satellite in a circular
orbit, the reel-rate at capture must be zero. The cost function that aids in minimizing tether
oscillations is given in Eq. (68). The tether mass density is 1 kg/km, the subsatellite mass is
200 kg, and the orbit radius is 500 km. The tension measurement noise is 0.5 N, the reel-rate
noise is 0.05 m/s, and the GPS error terms noise are RGPS = 25 m2, GPS = 0.01
(nondimensional). Solutions are obtained using N = 30, with a fixed sampling time of 0.01
rad  9 sec. The final time is set at 12 rad in nondimensional units.

80
Truth
20 60 Estimate

40
15
20
 (deg)
x (km)

10 0

-20
5
-40
0
-60
-10 -5 0 5 10 15 0 2 4 6 8 10 12
a) y (km) b) Nondimensional Time
248 Model Predictive Control

1.4 1.5

1.2 1

1 0.5

0.8 0
Truth

'


Estimate
0.6 -0.5
Truth
0.4 -1 Estimate

0.2 -1.5

0 -2
0 2 4 6 8 10 12 0 2 4 6 8 10 12
c) Nondimensional Time d) Nondimensional Time

Measured Tension (N)

8 40

6 20
Truth
Reel-Rate (m/s)

4 Estimate
0
0 2 4 6 8 10 12
2
0.4
CPU Time (sec)

0
0.2
-2

0
-4 0 2 4 6 8 10 12
e)
0 2 4 6 8
Nondimensional Time
10 12 f) Nondimensional Time

Fig. 8. Closed-loop optimal control of tethered satellite system, a) Tether tip trajectory, b) In-
plane libration angle, c) Nondimensional tether length, d) Nondimensional libration rate, e)
Reel-rate, f) Measured tension and computation time.

Fig. 8 shows the results of a closed-loop simulation in Simulink using the multibody tether
model in combination with the CKF. The results show that the tether is initially over-
deployed by about 20%, then reeled back-in to generate the swing velocity required for
capture. The final conditions are met to within a fraction of a percent in all state variables
despite the measurement errors and uncertainties. The peak reel-rate is approximately 7
m/s, and the variation in reel-rate is smooth throughout the entire maneuver. The average
CPU time is 0.23 sec, peaking to 0.31 sec.

7. Conclusion
Modern computing technology allows the real-time generation of optimal trajectories for
tethered satellites. An architecture that implements a closed-loop controller with a nonlinear
state estimator using a subset of available measurements has been demonstrated for
accurately deploying a tether for a rendezvous application. This strategy allows the
controller to adapt to large disturbances by recalculating the entire trajectory to satisfy the
mission requirements, rather than trying to force the system back to a reference trajectory
computer offline.
Predictive Control of Tethered Satellite Systems 249

8. References
Barkow, B.; Steindl, A.; Troger, H. & Wiedermann, G. (2003). Various methods of controlling
the deployment of a tethered satellite. Journal of Vibration and Control, Vol. 9, 187-
208.
Blanksby, C. & Trivailo, P. (2000). Assessment of actuation methods for manipulating tip
position of long tethers. Space Technology, Vol. 20, No. 1, 31-39.
Colombo, G.; Gaposchkin, E. M.; Grossi, M. D. & Weiffenbach G. C. (1975). The ‘skyhook’: a
shuttle-borne tool for low-orbital-altitude research. Meccanica, March, 3-20.
Dunbar, W. B.; Milam, M. B.; Franz, R. & Murray, R. M. (2002). Model predictive control of a
thrust-vectored flight control experiment. 15th IFAC World Congress on Automatic
Control, Barcelona, Spain.
Elnagar, G.; Kazemi, M. A. & Razzaghi, M. (1995). The pseudospectral legendre method for
discretizing optimal control problems. IEEE Transactions on Automatic Control, Vol.
40, No. 10, 1793-1796.
Fujii, H. & Ishijima, S. (1989). Mission function control for deployment and retrieval of a
subsatellite. Journal of Guidance, Control, and Dynamics, Vol. 12, No. 2, 243-247.
Fujii, H. A. & Anazawa, S. (1994). Deployment/retrieval control of tethered subsatellite
through an optimal path. Journal of Guidance, Control, and Dynamics, Vol. 17, No. 6,
1292-1298.
Fujii, H.; Uchiyama, K. & Kokubun, K. (1991). Mission function control of tethered
subsatellite deployment/retrieval: In-plane and out-of-plane motion. Journal of
Guidance, Control, and Dynamics, Vol. 14, No. 2, 471-473.
Gill, P. E.; Murray, W. & Saunders, M. A. (2002). SNOPT: An SQP algorithm for large-scale
constrained optimization. SIAM Journal on Optimization, Vol. 12, No. 4, 979-1006.
Arasaratnam, I. & Haykin, S. (2009). Cubature kalman filters. IEEE Transactions on Automatic
Control, Vol. 54, 1254-1269.
Kim, E. & Vadali, S. R. (1995). Modeling issues related to retrieval of flexible tethered
satellite systems. Journal of Guidance, Control, and Dynamics, Vol. 18, 1169-1176.
Kruijff, M.; van der Heide, E. & Ockels, W. (2009). Data analysis of a tethered spacemail
experiment. Journal of Spacecraft and Rockets, Vol. 46, No. 6, 1272-1287.
Lakso, J. & Coverstone, V. L. (2000). Optimal tether deployment/retrieval trajectories using
direct collocation. AIAA/AAS Astrodynamics Specialist Conference, 14-17 Aug. 2000,
AIAA Paper 2000-4349.
Lorenzini, E. C.; Bortolami, S. B.; Rupp, C. C. & Angrilli, F. (1996). Control and flight
performance of tethered satellite small expendable deployment system-II. Journal of
Guidance, Control, and Dynamics, Vol. 19, No. 5, 1148-1156.
Misra, A. K. & Modi, V. J. (1982). Deployment and retrieval of shuttle supported tethered
satellites. Journal of Guidance, Control, and Dynamics, Vol. 5, No. 3, 278-285.
Misra, A. K. & Modi, V. J. (1986). A Survey on the dynamics and control of tethered satellite
systems. Advances in the Astronautical Sciences, Vol. 62, 667-719.
Nordley, G. D. & Forward, R. L. (2001). Mars-earth rapid interplanetary tether transport
system. I – Initial feasibility analysis. Journal of Propulsion and Power, Vol. 17, No. 3,
499-507.
Ross, I. M. & Fahroo, F. (2003). Legendre pseudospectral approximations of optimal control
problems. Lecture Notes in Control and Information Sciences, Vol. 295, 327-342.
250 Model Predictive Control

Ross, I. M. & Gong, Q. (2008). Guess-free trajectory optimization. AIAA/AAS Astrodynamics

Specialist Conference and Exhibit, August, Honolulu.
Rupp, C. C. (1975). A tether tension control law for tethered subsatellites deployed along the
local vertical. NASA TM X-64963, Marshall Space Flight Center, Alabama.
Vadali, S. R. & Kim, E. S. (1991). Feedback control of tethered satellites using lyapunov
stability theory. Journal of Guidance, Control, and Dynamics, Vol. 14, No. 4, 729-735.
Wiedermann, G.; Schagerl, M.; Steindl, A. & Troger, H. (1999). Simulation of deployment
and retrieval processes in a tethered satellite system mission. Paper presented at the
International Astronautical Congress, Amsterdam, The Netherlands, Oct.
Williams, P. & Blanksby, C. (2008). Optimal prolonged spacecraft rendezvous using tethers.
International Review of Aerospace Engineering, Vol. 1, No. 1, 93-103.
Williams, P. (2004). Application of pseudospectral methods for receding horizon control.
Journal of Guidance, Control, and Dynamics, Vol. 27, No. 2, 310-314.
Williams, P. (2004). Guidance and control of tethered satellite systems using pseudospectral
methods. AAS/AIAA Spaceflight Mechanics Meeting, Wailea, Hawaii, Feb., Paper
AAS 04-169.
Williams, P. (2005). Receding horizon control using gauss-lobatto quadrature
approximations. AAS/AIAA Astrodynamics Specialist Conference, Aug. 7-11, Embassy
Suites Hotel, Lake Tahoe Resort, Paper AAS 05-349.
Williams, P. (2006). Dynamics and control of spinning tethers for rendezvous in elliptic
orbits. Journal of Vibration and Control, Vol. 12, No. 7, 737-771.
Williams, P. (2006). A gauss-lobatto quadrature approach for solving optimal control
problems. ANZIAM Journal (E), Vol. 47, July, C101-C115.
Williams, P. (2008). Optimal deployment/retrieval of tethered satellites. Journal of Spacecraft
and Rockets, Vol. 45, No. 2, 324-343.
Williams, P.; Blanksby, C. & Trivailo, P. (2004). Tethered planetary capture maneuvers.
Journal of Spacecraft and Rockets, Vol. 41, No. 4, 603-613.
Williams, P.; Hyslop, A.; Stelzer, M. & Kruijff, M. (2008) YES2 optimal trajectories in
presence of eccentricity and aerodynamic drag. Acta Astronautica, to be published.
Xu, D. M.; Misra, A. K. & Modi, V. J. (1981). Three dimensional control of the shuttle
supported tethered satellite systems during deployment and retrieval. Proceedings
of the 3rd VPISU/AIAA Symposium on Dynamics and Control of Large Flexible Spacecraft,
Blacksburg, VA, 453-469.
Xu, D. M.; Misra, A. K. & Modi, V. J. (1986). Thruster-augmented active control of a tethered
subsatellite system during its retrieval. Journal of Guidance, Control, and Dynamics,
Vol. 9, 663-672.
MPC in urban traffic management 251

11
x

MPC in urban traffic management

Tamás Tettamanti1, István Varga1,2 and Tamás Péni2
1Budapest University of Technology and Economics
2HAS Computer and Automation Research Institute
Hungary

1. Introduction
More and more people are concerned about the negative phenomenon resulted by the
negative effects of the growing traffic motorization. Traffic congestion is the primary direct
impact which became everyday occurrence in the last decade. As world trade is
continuously increasing, it is obvious that congestions represent also a growing problem.
The capacity of the traffic networks saturates during rush hours. At the same time, the
traditional traffic management is getting less effective in sustaining a manageable traffic
flow. Therefore, external impacts appear causing new costs for the societies. As a possible
solution the predictive control based strategy can be applied. The chapter investigates the
applicability of MPC strategy specialized in urban traffic management in order to relieve
traffic congestion, to reduce travel time and to improve homogeneous traffic flow. Over the
theory the realization of the control method is also presented. Firstly we give a historical
summary of adaptive traffic control. The brief results of MPC and its related methods in
urban traffic control are presented. Then we introduce the modeling possibilities of urban
traffic as the appropriate model means an important aspect of the control design. The use of
MPC requires a state space theory approach. Therefore the so called Store-and-forward
model is chosen which can be directly translated to state space. We analyze the model in
details showing the real meaning of system matrices. The constraints of urban traffic system
is also discussed which heavily influence modeling and control. The next section presents
the simulation environment which is used to demonstrate the developed control methods.
Thereinafter we present the main results of MPC in traffic application. The idea to apply
MPC in urban traffic network is induced by the fact that the distance is relatively short
between several intersections with traffic lights. Hence it is advisable to coordinate the
operation of the intersection controller devices. Several intersections are near to each other
in smaller or bigger networks, primarily in cities, the coordination is especially emphasized.
The development of new control strategies is a real demand of nowadays. One of the
possible solutions is the practical application of MPC. The aim of the control is to increase
capacity. To test and validate our control strategy we apply it to a real-word transportation
network where the actual system is not efficient enough to manage the traffic in rush hours.
The simulation results show the effectiveness of the control design. After the presentation of
the practical urban traffic MPC the distributive solution of MPC has to be discussed. As the
computational demand depends on the size of the network an efficient calculation method is
252 Model Predictive Control

sought to solve online the MPC problem. The classical scheme for adaptive road traffic
management structure is usually based on control center which processes and computes all
signal control for the network. Another method for the control system architecture is the
decentralized and distributed control scheme. This approach has numerous economical and
technological advantages. Distributed traffic control is developed using iterative solution.
The so-called Jacobi iteration algorithm is an efficient method to solve constrained and
nonlinear programming problem which the original problem can be transformed for. An
additional feature of the developed strategy is the ability to manage priority. If a preferred
vehicle arrives to any junction of the network it will be automatically indicated. Its stage will
be handled with priority getting maximum green time as possible in every cycle until the
vehicle will not leave the intersection. It means practically that the cost function is
dynamically modified by the system weights depending of presence of any preferred
vehicles. Finally we would like to introduce the robust MPC problem in traffic management
as our future work. The robustness of the traffic management means that even with the
presence of some disturbances the system is able to find optimal control solution. We
discuss the modification of the traffic model introduced in third section since the chosen
method requires a special model structure.

2. Brief historical summary of adaptive road traffic control

In case the distance is relatively short between several intersections with traffic lights it is
advisable to co-ordinate the operation of the intersection controller devices. The
coordination may include public transport devices and pedestrian traffic besides vehicles.
Where several intersections are near to each other in smaller or bigger networks, primarily
in cities, the coordination is especially emphasized.
In the 1970's a new control strategy appears in the road traffic management. Beside the
already extant fixed-time and traffic-actuated strategies the traffic-adaptive control is
invented. A traffic control system that continuously optimizes the signal plan according to
the actual traffic load is called an adaptive traffic control system. The essence of the
functioning is that the changes to the active signal plan parameters are automatically
implemented in response to the current traffic demand as measured by a vehicle detection
system. Such system can be used as local or network-wide control.
The appearance of the adaptivity induces new developments of traffic control methods. The
first adaptive systems like SCOOT (Hunt et al., 1982) or SCATS (Lowrie, 1982) are based on
heuristic optimization algorithms. In the 1980's new optimization methods are introduced
based on rolling horizon optimization using dynamic programming. Some examples are
OPAC (Gartner, 1983), PRODYN (Farges et al. 1983), and RHODES (Sen & Head, 1997).
In the middle of the 1990's the first control method is introduced which adopts results of the
modern control theory. The TUC system (Diakaki et al., 1999) applies a multivariable
regulator approach to calculate in real time the network splits, while cycle time and offsets
are calculated by other parallel algorithms. The basic methodology employed for split
control by TUC is the formulation of the urban traffic control problem as a linear-quadratic
(LQ) optimal control problem. The advantage of LQ control is the simplicity of the required
real-time calculations which is an important aspect in network-wide signal control.
However the algorithm has a main disadvantage. LQ control is not able to manage
constraints on the control input (its importance is discussed in the next section). Therefore a
MPC in urban traffic management 253

posteriori application is needed to force the constraints which may lead to suboptimal
solution.
In the early 2000's the first results are published in the subject of MPC based traffic control.
However these publications (e.g. Bellemans et al., 2002; Hegyi et al., 2003) are related to
ramp metering and variable speed limit control of the freeway traffic management. MPC
based urban traffic control approach is published by Tettamanti et al. (2008). The paper
consists theory, realization and a real-word example. The main result is the possibility to
overcome the disadvantage of the LQ problem mentioned above as the MPC method can
take the constraints into consideration. These results constitute the basis of the chapter. The
paper of Aboudolas et al. (2009) is published investigating large-scale traffic control problem
and introducing the open-loop quadratic-programming control (QPC) as a possible method
for optimal traffic management. The paper concludes that for the application of the QPC
methodology in real time, the corresponding algorithms may be embedded in a rolling-
horizon (model-predictive) scheme which constitutes the part of future works.
In 2010 as a development result of Tettamanti et al. (2008) the paper of Tettamanti & Varga
(2010) is published which introduces a distributed realization of an MPC based traffic
control system. The publication's results will be also enlightened in detail in the chapter.

3. Urban traffic modeling

Modeling and control are coherent notions in control theory as the model highly determines
the applicable methods for control. In the previous chapter various control approaches were
presented. All of them use an appropriate traffic modeling technique for functioning.
Apparently, the modern control theory based traffic management strategies apply the state
space approach. The state space modeling is derived from the so called Store-and-forward
model (Gazis & Potts, 1963) which introduces a model simplification that enables the
mathematical description of the traffic flow process. This modeling technique opens the way
to the application of a number of highly efficient optimization methods such as LQ control,
MPC, or robust LMI based control. Before to begin to investigate the MPC based traffic
control the properties of the model have to be discussed in detail.

3.1 From Store-and-forward traffic modeling to state space representation

The following derivation of the state space model reflects the results of the paper of Diakaki
et al. (1999).

M qz N
hz

sz rz

Fig. 1. The Store-and-forward traffic model

The two basic parts of an urban road traffic network are intersection and link. The
combination of these elements constitutes the traffic network with link z  Z and junction
254 Model Predictive Control

j  J which are defined geometrically exactly. Each signalized junction j has its own sets
of incoming I j and outgoing O j links. Figure 1. shows the coherence (link z ) of two
neighboring intersections ( M , N ) in the transportation network where z  OM and z  I N .
The dynamic of link z is described by the conservation equation:

x z k  1 = x z k   T q z k   hz k   rz k   s z k  (1)

where x z k  measures the number of vehicles within link z , practically the length of queue,
at time kT . q z k  and hz k  are the inflow and outflow, rz k  and s z k  are the demand
and the exit flow during the sample period kT , k  1T  . T is the control interval and
k = 0,1.. is the discrete time index. For simplicity we assume henceforth that the cycle times
are equal for each junction j  J , namely Tc , j = Tc . Moreover T is also equal to Tc . rz k 
and s z k  represent typically the fluctuation between a parking lot and link z or the effects
of any non-controlled intersection between M and N . These disturbing flows can be
considered as known perturbations if they can be well measured or estimated. In case of
unknown disturbances robust control system is needed.
Equation (1) is linear scalar equation for the portrayal of vehicles movement of a given link.
But if we wish to define a whole traffic network each link has to be described by its
conservation equation and what is more the equations needs to be interconnected. At this
point we can change for state space representation which means the appearance of the state
and control input vectors together with the coefficient system matrices. The general discrete
LTI state space representation is the following:

xk  1 = Axk   Bu k   Ed k 
y k  = Cxk  (2)

Using Equation (2), it is possible to describe the dynamics of an arbitrary urban traffic
network (see Fig. 2 as an example).
pz

M N
qz hz

sz rz

Fig. 2. Dynamics in the urban traffic network

MPC in urban traffic management 255

The physical meaning of matrices and vectors is elementary to understand the model. The
state equation form can be achieved using all conservation equations, arranging them in one
linear matrix equality. In our case the state matrix A is practically considered as an identity
matrix. The elements of the state vector x(k ) represent the number of vehicles of each
controlled link. The second term of the state equation is the product of input matrix B and
control input u . Vector u contains the green times of all stages. Their numerical values are
the results of a corresponding controller at each cycle. Naturally the number of states is
equal to the number of controlled links in the network. The product Bu k  is arising from
the part T q z k   hz k  of Equation (1) which means the difference of the inflow and the
outflow of a link during the control interval. q z k  and hz k  are directly related to control
input (green time), saturation flow ( S ) and turning rate ( t ) in a signalized network. To
understand the construction of B the parameters S and t have to be discussed. Saturation
flow represents the outflow capacity of link z  Z during its green time. A standard value
for saturation flow is S  0.5 veh / sec which is considered constant in practice. Turning rate
represents the distribution of turnings of vehicles from link z  O j to links w I N . These
parameters are defined by the geometry and the rights of way in the traffic network and
 
assumed to be known and constant or time varying. Then matrix B  bij can be constructed
by the appropriate allocation of the combinations of saturation flow and turning rates. The
diagonal values of B are negative S z as the product S z u z k  represents the outflow from
link z . At the same time the inflow to the link z has to be also characterized. Therefore the
products S z t w, z are placed in matrix B such that bij  S z t w, z when i  j . The parameters
t w, z ( w  I M ) are the turning rates towards link z from the links that enter junction M .
Hence the inflow is resulted from the appropriate matrix-vector multiplication for all z .
In state space representation the third term Ed k  of Equation (2) represents an additive
disturbance where E = I . d k  is composed of two type of data. On the one hand it is
coming from the part T rz k   s z k  of Equation (1) where rz k  and s z k  are considered as
measured disturbances. They reflect the difference of the demand and the exit flows of a link
during the control interval. On the other hand there is demand p z k  at the boundary of the
traffic network (Figure 2.) which also has to be taken into consideration in the model. The
traffic p z k  intending to enter is a measurable value. Therefore it is simply added to the
appropriate row of d k  .
To end the state space description of the urban traffic the measurement equation has to be
mentioned. As each output inside of the network is a measured state (number of vehicles of
the link z  Z ) the output equation is simplified to y k  = xk  with C = I . Note that as the
exit links of the network are not controlled they do not have to be confused with the
outputs y k  .
Finally, as three of the system matrices are identity matrix (discussed above) the general
discrete LTI state space representation for urban traffic simplifies to the following form:

xk  1 = xk   Bu k   d k 
y k  = xk  (3)
256 Model Predictive Control

3.2 Constraints of urban traffic control

As store-and-forward modeling technique tries to express the real dynamics and states of
the urban traffic there are several constraints which have to be taken into account. The most
essential constraints of the urban network are determined by the geometry. It is evident that
the maximum number of vehicles is defined by the length of link between two junctions.
Naturally the vehicles are considered as passenger car unit (PCU) resulting from
appropriate transformation (Webster & Cobbe, 1966). Thus the states are subject to the
constraints:
0  x z k   x z , max (4)

If we consider a network the use of the states constraints can contribute to avoid the
oversaturation in the controlled traffic area. In a control scheme beside the state constraints
one can define output limitations too. However in our case the states constraints are
identically to the output constraints as C = I .
The control input is the next variable restricted by some constraints. The first constraint on
u is the interval of seconds of green time:

u z , min  u z k   u z , max (5)

Depending on the system setting u z ,min (for lack of vehicles on link z ) can be zero. It means
permanent red signal for the stage in the next control interval. The second control input
constraint is represented by the linear combination of green times at junction j  J . The sum
of the green times has to be lower as T j , max :

Oj

u k   T
z =1
z j , max , j  1, 2, ..., J , (6)

where O j is the number of stages at junction j , T j , max = T  L j ( L j is the fixed lost time
resulted from the geometry of junction j ), and J is the number of controlled intersections.

4. Simulation environment
In the previous sections traffic modeling was introduced which can be used in control
design. Moreover the simulation environment has to be discussed similarly as all the
methods presented in this chapter were simulated and tested. For simulation we used traffic
simulator (VISSIM, 2010), numerical computing software (MATLAB, 2010) and C++
programming language.
VISSIM is a microscopic traffic simulation software for analyzing traffic operations. It is able
to simulate network consisting of several intersections and allow the use of external control
algorithm in the control processes. These properties make it suitable to use this software by
reason of the several junctions and the control algorithms written in MATLAB. VISSIM uses
a so-called psycho-physical driver behavior model based on the car-following model of
(Wiedemann, 1974). The model describes all the cars found in the system. The vehicles are
defined by both physical and psychical parameters (origin, destination, speed, driver
MPC in urban traffic management 257

behavior, vehicle type, etc.). The VISSIM simulation is based on an iteration process of
acceleration and deceleration.
The communication does not work directly between MATLAB and VISSIM as the
simulation can only be accessed via Component Object Model (COM) interface (Roca, 2005).
To control the communication a C++ application has to be created. The created C++
program manages the simulation process and controls the data transfer between the
software (Figure 3.).

COM INTERFACE

MATLAB C++ VISSIM

PROGRAM
CONTROL TRAFFIC
SIMULATION
LOGIC AND NETWORK
COMMUNICATION
CONTROL

Fig. 3. The simulation process of the system model

5. MPC based urban traffic management

The aim of our research was to elaborate a control process related to network consisting of
several junctions which perform the control of all the traffic lights in its sphere of action in a
coordinated way depending on the traffic. The controller must be able to dynamically make
the traffic signal set of the intersections. From the point of view of realization, this means
that before every period a new traffic sign must be generated regarding all the traffic lights,
in harmony with the present traffic. To solve the above, MPC technology was chosen since it
is able to take all the constraints into consideration in course of the control input setting. To
show the efficiency of MPC the control design was tested simulating a real-word traffic
network.
5.1 The MPC cost function
The control objective is the minimization and balancing of the numbers of vehicles within
the streets of the controlled network. This control objective is approached through the
appropriate manipulation of the green splits at urban signalized junctions, assuming given
cycle times and offsets. By employing the predictive control model, the dynamic
determination (per cycle) of the traffic light’s period is possible either with the consideration
of the natural constraints existing in the system introduced in Section 3.2.
The state space equation for MPC design can be given as follows:

 xk  1 | k    xk   d k    B 0  0   u k | k  
      
 xk  2 | k     xk   2d k     B B  0   u k  1 | k   (7)
             
      
 x

k  N | k   xk   Nd k   B B  B  u k  N  1 | k 
     
~
x ( k 1) c(k ) ~ g (k )
B

where x , d , B and u are elements of Equation (3) already discussed. ~ x is a hyper vector
of the state vectors, representing the number of vehicles standing at each controlled link of
258 Model Predictive Control

the intersections. c is a hyper vector combination of the previous state vector and d . The
disturbance d is considered measured and constant during the horizons of kth step. Hence
~
it is multiplied by the value of the current horizon. B is a lower triangular hyper matrix
including the matrix B . g is a hyper vector of the control input vectors (green times),
k  1,2... a is the discrete time index, and N is the length of the MPC horizon.
The MPC algorithm needs the current values of the states at each control interval which
means the exact knowledge of the numbers of vehicle. However the states can not be
directly measured only estimated using appropriate measurement system (e.g. loop
detectors) and estimation algorithm. A possible realization for state estimation was
published in paper of Vigos et al. (2007) which is based on the well-known Kalman Filter
algorithm (Welch & Bishop, 1995). The estimation error is neglected in the paper.
The elements of B are the combinations of turning rates and saturation flow as discussed in
Section 3.1. Saturation flow is not measurable hence a standard value is determined
( S  0.5 veh / sec ). Usually the values of turning rates are also considered constant.
Nevertheless, in practice the turnings vary around the nominal rates. Thus a continuous
estimation may be applied to ameliorate the MPC algorithm. A possible way to estimate
turning rates is to use a finite back stepped state observer, e.g. Moving Horizon Estimation
(MHE) method (Kulcsár et al., 2005).
Several choices of the objective function in the optimization literature have been reported. In
this chapter we consider the following quadratic cost function characterized by the
weighted system states and control inputs:

J k  
2

1 ~T
x k Q ~ 
x k   g T k R g k   min (8)

where Q  0 and R  0 are scalar weighting matrices. Q and R have appropriately chosen
tuning parameters in their diagonals. The weightings reflect that the control input variation
is lightly punished compared to the state variation. The selection of the appropriate
weightings is important, because this could influence (especially the end-point weight) the
stability of the closed loop (Kwon & Pearson, 1978). To solve this minimization problem
several mathematical software can be applied which provide built-in function for quadratic
constrained optimization. The solution of optimization problem (8) leads to the
minimization of the vehicle queues waiting for crossing intersections. The control input
green time is defined corresponding to the states of intersection branches representing a
fully adaptive traffic management.
Different stability proofs exist for receding horizon control algorithms. Maciejowski (2002),
Rawlings & Muske (1993) or Mayne et al. (2000) offer different methodological approaches.
However the urban traffic is a special case. It is ensured that the system will not turn
instable because of the hard physical constraints coming from the network geometry.
Accordingly, there is a natural saturation in the system. The states can never grow
boundlessly. The instability can appear only if there is an oversaturation in the network. To
solve this problem we intend to apply the results of the invariant set theory (Blanchini &
Miani, 2007) in the future. It is also has to be noted that if we choose a traffic area to control
we do not deal with the traffic outside of the boundary of the network. Obviously the sphere
of control action is also an important question in traffic management.
MPC in urban traffic management 259

5.2 Test network for simulation

To test MPC technology in urban traffic management we choose a real-world test area
situated in the 10th district of Budapest. The test network includes seven neighboring
intersections (Figure 4.).

2 1

Fig. 4. Schematic representation of the test network consisting of seven junctions

The dimension of the system is 36 which means that we intend to control 36 links. This area
is suitable for testing our new control system since the included road stretches have a heavy
traffic volume in rush hours. The current traffic management system is offline. The seven
junctions are controlled individually. Three of them use fix time signal plan. In the other
four intersections detectors help the controllers. They can slightly modify their fix programs.
The current control is effective but only in case of normal traffic flow. If the volume of
vehicles increases extremely, the system cannot manage the situation and traffic becomes
congested before the stop lines. The biggest problem is that the controllers work locally and
independently. Our new control design, however, takes the seven junctions into
consideration as a real network.
As the MPC cost function (8) represents a quadratic optimization problem the control input
was calculated using the built-in quadprog function of MATLAB.

5.3 Simulation results

To prove the applicability of the MPC based control design it was compared with the
current control system of the test network, which is a partly adaptive control strategy.
The same input traffic volumes were set for both simulations. We used volume data for
which the traffic lights were originally designed. The simulation provided similar results for
both strategies as we expected. This means the current system is correctly designed, and
manages non-extreme traffic flow with good results.
To test the effectiveness of the two systems in case of heavier traffic we generated more
intensive traffic flow during the simulation. The original input volumes were increased by
10% in the network. This simulation showed different results to the previous case. The
current system could manage the traffic less efficiently compared with the MPC based
control system. The simulation time was 1 hour long. The results are presented in Table 1.
All important traffic parameters changed in a right way. The new system can provide a very
effective control in the test network.
260 Model Predictive Control

OLD MPC based

Parameter Variation
STRATEGY strategy
Total travel time per vehicle [sec] 114 96 ↓ 16%
Average speed [km/h] 20.6 24.9 ↑ 21%
Average delay time per vehicle [sec] 68 56 ↓ 18%
Average number of stops per vehicles 3.8 3.1 ↓ 18%
Table 1. Average simulation results of the test network

At the same time these simulations were run in a reduced environment. We diminished the
number of junctions in the test network from seven to four. Namely the traffic lights at
junctions 4., 5., 6. (see Figure 4.) work totally offline. The capacities of these locations
increased apparently. So only the junctions 1., 2., 3., and 4. were kept in order to focus on the
comparison of the two adaptive strategies.

MPC based
Parameter Old strategy Variation
strategy
Total travel time per vehicle [sec] 105 96 ↓ 9%
Average speed [km/h] 20.5 23.5 ↑ 15%
Average delay time per vehicle [sec] 64 52 ↓ 19%
Average number of stops per vehicles 1.2 1.2 0%
Table 2. Average simulation results of the test network with design input volumes

MPC based
Parameter Old strategy Variation
strategy
Total travel time per vehicle [sec] 110 96 ↓ 13%
Average speed [km/h] 18.4 23.6 ↑ 28%
Average delay time per vehicle [sec] 71 52 ↓ 27%
Average number of stops per vehicles 1.5 1.2 ↓ 20%
Table 3. Average simulation results of the test network with 10% augmentation of the design
input volumes

Alike above, the behavior of the reduced network was analyzed with normal and heavier
input traffic volumes. The results ameliorated in both cases (see Table 2. and 3.). The
simulation time was 2 hours long.
The aim of the MPC based control is the minimization of the number of vehicles waiting at
the stop line. The current system cannot adapt to the increased volume. The average queue
length grew strongly during the simulations. However, the MPC strategy is able to manage
heavier traffic situations real-time. Figure 5. represents the effectiveness of our system. It
shows the variation of average queue lengths in the network.
MPC in urban traffic management 261

Fig. 5. The variation of average queue lengths in the two different control cases

6. Distributed traffic management system based on MPC

The classical scheme for adaptive road traffic management structure is based on control
center which processes and computes all signal control for the network. Another method for
the control system architecture is the decentralized and distributed control scheme. This
approach has numerous economical and technological advantages.
In this section we present a distributed control system scheme for urban road traffic
management. The control algorithm is based on MPC involving Jacobi iteration algorithm to
solve constrained and nonlinear programming problem. The distributed control design was
also simulated and tested.
6.1 The MPC cost function
We refer to the results of Section 5.1. Substituting ~ x k  and g k  in Equation (8) one arrives
to:
1
2
 ~ ~
 ~ 1
2
1
J k   g T qB T B  rI g  qcT Bg  qcT c  g T g   T g  
2
(9)

where q and r are constants coming from the diagonal of the scalar matrices Q and R .
As  is a constant term, finally one has the objective function to minimize:

1 T
J k   g k g k    T k g k   min (10)
2

where  is constant matrix as it contains the combination of constant turning rates,

saturation rates and fixed tuning parameters. At the same time  contains varying values
coming from the current dynamics of the traffic area.

6.2 Multivariable nonlinear programming to solve MPC problem

The solution of the MPC cost function (10) represents a multivariable nonlinear problem
subject to linear constraints. It formulates a standard quadratic optimization problem
(Bertsekas & Tsitsiklis, 1997):
262 Model Predictive Control

1 T
J k   g g   T g  min
2
s. t. Fg  h  0 (11)

where matrix inequality Fg  h incorporates the constraints (4), (5) and (6) already
discussed in Section 3.1.
If  is a positive semi definite matrix, (11) gives a convex optimization problem (Boyd &
Vanderberghe, 2004). Otherwise one has to use the singular value decomposition method to
 which results a convex problem. This means a linear transformation to the original
problem (11).
Using the duality theory (Bertsekas & Tsitsiklis, 1997) the primal problem can be formulated
into Lagrange dual standard form. The basic idea in Lagrangian duality is to take the
constraints into account by augmenting the objective function with a weighted sum of the
constraint functions. We define the Lagrangian associated with the problem as:

Lg ,    J k   T Fg  h  (12)

We refer to i as the Lagrange multiplier associated with the ith inequality constraint of
(11). The dual function is defined as the minimum value of the Lagrangian function. This
can be easily calculated by setting gradient of Lagrangian to zero (Boyd & Vanderberghe,
2004). This yields an optimal green time vector (16) which minimizes the primal problem.
Hence one arrives to the dual of the quadratic programming problem:

1 T
J DUAL k    P  wT   min
2
s. t.   0 (13)

where P and w are coming from the original problem:

P  F 1F T (14)
1
w  F   h (15)

It is shown that if  provides optimal solution for the J DUAL k  problem then


g    1   F T   (16)

gives also an optimal solution for the primal problem (Rockafellar, 1970).
The dual problem has a simple constraint set compared with the primal problem’s
constraints. Hence expression (13) represents a standard minimization problem over
nonnegative orthant.
A very efficient method, the Jacobi iteration was found to solve the optimization problem.
Since  is a positive semi definite matrix the jth diagonal element of P , given by
MPC in urban traffic management 263

p jj  f jT  1 f j (17)

is positive. This means that for every j the dual cost function is strictly convex along the jth
coordinate. Therefore the strict convexity is satisfied and it is possible to use the nonlinear
Jacobi algorithm. Because the dual objective function is also quadratic the iteration can be
written explicitly. Taking into account the form of the first partial derivative of the dual cost

n
w j   p jk k (18)
k 1

the method is given by:

   n 
 j t  1  max 0,  j t   wj  p jk k t  , j  1,  , n (19)
 p jj  k 1 

Where   0 is the stepsize parameter which should be chosen sufficiently small and some
experimentation may be needed to obtain the appropriate range for  .
The importance of this method, over its efficiency, is the ability to satisfy the positivity since
equation (19) excludes negative solution for  . Thus, during the MPC control process at
each (kth) step the optimal green times can be directly calculated from equation (16) after
solving the problem (11).

6.3 Realization of MPC based distributed traffic management system

The economical and technological innovation of the above described control method is
represented by the state-of-the-art control design and the optional decentralized realization
at the same time.
Generally, the architectures of traffic control systems can be central, decentralized, or mixed.
The central management architecture is a frequent strategy based on a central processor
which controls all signal controllers in the transportation network. Decentralized and mixed
control systems are not so common applications yet. However they have many advantages
and represent a new way in traffic control technology. Decentralized management systems
carry a higher performance since they can distribute their computations between the traffic
controllers. As well as they represent a higher operation safety because of their structural
redundancy. Some of these distributed realizations are for example SCATS (Wolshon &
Taylor, 1999) or Utopia (UTOPIA, 2010).
The distributed technology can be used in any road traffic network which is equipped with
adequate signal controllers and detectors, as well as communication between controllers is
also required.
Since the solution of the Jacobi algorithm (19) is an iteration process the computers can
distribute their calculations during the operation cycle. Therefore it is suitable for the
distributed realization of the MPC problem. Considering a large traffic network the
following practical system realization can be applied. Firstly we define the nodes
represented by the red cubes on Figure 6. The nodes are the head traffic controllers which
participate in the resolution procedure. Every node covers a few intersections (traffic
264 Model Predictive Control

controllers) which do not participate in the computation. The distributed control network is
represented by Figure 6.

x, 
1

x,  x, 

4 3
x, 

Fig. 6. Distributed MPC control in urban traffic network

The distributed computation is executed during the operation cycle as follows:

1. Communication: At the end of the kth cycle all traffic controllers send their
measurements (number of vehicles) to their node.
2. Communication: The head controllers share the measurement data with the other
nodes.
3. Calculation (which is not the third step practically since it can be started parallel
with step 1.): Node 1. starts iteration procedure. After some predefined iteration
steps it forwards their computational results to the next node and so on. The
transmitted data is the currently calculated vector s where s  1,2,...,  ( 
represents the final iteration step number specified previously).
4. Communication: When the last node finishes the computation (which means
practically that s   ) it shares the optimal result (  ) with the other nodes.
5. Calculation: The head controllers calculate their final calculation. Using Equation
(16) the nodes do not need to execute the whole multiplication. But only the
specified part of g  which contains the optimal green times of their traffic
controllers.
6. Communication: Finally the nodes pass the optimal green times to every traffic
controllers for the next (k+1) cycle.
If one wishes to control small traffic network with a few intersections the distributed
solution is not certainly required. The calculation of one Jacobi iteration step means simple
multiplications and additions of scalars. In case of few states (number of the controlled
links) a single controller’s performance is sufficient to compute all signal sets of the network.
Thus the system is working with redundancy which can be very useful at the same time.
The controllers can continuously check their operation comparing their computation results.
On the other hand, if one of the signal controllers fails in the network the system can go on
with safe functioning.
However with the growth of the number of the system states the computational demand
increases quadratically. Therefore larger network requires the distributed solution of the
MPC in urban traffic management 265

MPC control. Certainly the solution method is also largely depends on the performance of
the actual signal controllers and the communication system.

6.4 Simulation of MPC based distributed control algorithm

To verify the designed control system scheme a closed loop simulation environment was
created (Section 4.). The traffic network used for the simulation is equivalent with the one
applied in Section 5.2.
As discussed before the appropriate setting for the stepsize parameter  requires some
practical experimentation. This value strongly influences the performance of the calculation.
Convergence can be shown when   n 1 . However this value may lead to an unnecessarily
slow rate of convergence for some problems.
The values in Table 4. shows the variation of the number of steps to achieve convergence. In
case of our test network the smallest value with convergence was   n 0.0525 .

 Number of steps
n 1 6000

n 0.5 1000

n 0.1 200

n 0.0525 150
Table 4. The variation of the number of steps to achieve convergence

We also compared the computation times of the applied methods. Using the quadprog
function of MATLAB the computation time was about 20 seconds. Conversely the Jacobi
algorithm required less than 1 second on average which means 20 times faster calculation. It
has to be noted that the Jacobi algorithm was not tested in a distributed way. However even
with some communication time the Jacobi iteration is more efficient. On the one hand in our
test network the number of states was quite few. The distributed solution is not needed. On
the other hand the distributed realization is highly dependent on the current system
configuration (measurement accuracy, communication speed, etc.).

6.5 Vehicle priority management in MPC based urban control

The design of an adaptive traffic control system comes with the desirable demand to
incorporate vehicle priority management as well. Therefore an additional feature of the
designed system is the ability to manage priority.
The scope of the priority management has to be specified as some special vehicle classes
(e.g. emergency vehicles) have top-level priority. Therefore they do not need any help from
traffic lights to cross the intersections anytime. Our control deals vehicles which are favored
compared to the others but not by all means. Vehicles of the public transport are typically of
this sort. However one may differentiate the levels of importance even between public
vehicles (e.g. an overland bus compared to a local bus).
To operate such system these vehicles have to be able to communicate with the traffic
controllers. If a preferred vehicle arrives to any junction of the network it may be
automatically indicated by the traffic controller through radio frequency. Its stage can be
handled with priority getting maximum green time as possible in every cycle until the
266 Model Predictive Control

vehicle will not leave the intersection. It means practically that the cost function is
dynamically modified by the system weights depending of the presence of any preferred
vehicles. Accordingly for the sake of immediate reaction the given junction falls out of the
scope of the coordinated traffic control until the vehicle will not leave the intersection.
However it can be considered as disturbance.
We refer to the original MPC cost function (8) where Q is a diagonal weighting matrix:

q1 
 
q
Q 2  (20)
  
 
 qn 

Each diagonal element tunes a state (queue length of controlled links). If there is no preferred
vehicle in the scope of control: q1  q2    qn . By online modifying the weight qi (according to
the preferred vehicle's direction) one can assure priority. The measure of the modification of qi
depends on the current level of priority. In practice, the appropriate choice of the weights is an
empirical process as it strongly depends on the junction's properties.

7. Future work: Robust MPC in urban traffic management

As future work we introduce the problem of robustness in urban traffic management. In
Section 3.1 all disturbances in the state space model were considered as known (measured)
values and all possible uncertainties were neglected. These assumptions were taken by
practical reasons. However for more precise traffic modeling these factors can be involved in
the control scheme determining upper and lower bounds of the uncertainties. This implies
the use of a suitable robust control method as well.
The simplest approach to represent disturbances in the system is the bounded unknown
external additive disturbance. In this case an additive term appears in the LTI state space
model. This approach can deal with state disturbances. As a part of the Ph.D thesis of
Löfberg (2003) a Minimax MPC is presented which can be eligible for traffic systems too.
Another possibility to model the uncertainties is the polytopic paradigm. The system
matrices Ak  and Bk  of an LTV state space description can be defined by a prespecified
polytopic set:
 = CoA1 B1 , A2 B2 ,  , AL BL  (21)

where Co devotes to the convex hull and L is the number of the vertices. Matrix Ak  can
be used to express uncertainties of the states. In practice it means for example parking places
along the road or non-controlled junctions in the network which result unmeasured state
variation. Matrix Bk  can be used to represent uncertainties of the saturation flow rates
which are also non-measurable parameters. For polytopic system Kothare et al. (1996)
provide an efficient Minimax MPC solution which can be potentially applied in urban traffic
management as well.
There is another factor which can be taken into consideration in robust traffic control. In
Section 5.1 the demands ( d ) intending to enter the network were assumed constant and
MPC in urban traffic management 267

measured disturbances. In effect they vary continuously. Therefore for fully exact solution
varying demands should be considered in the MPC cost function.

8. Conclusion
This chapter introduced the aspects of MPC applied in urban traffic management. As the
urban traffic is a complex system having special attributes the appropriate traffic model had
to be discussed in details as well. At the same time MPC technology is suitable to control
such complex system optimally and real-time. The main control aim was the optimal and
coordinated control which can be satisfied. The applicability was demonstrated by several
simulations. Furthermore a distributed technology was presented which can be very useful
in practice particularly in large traffic network. As an additional feature of MPC based
system we showed that an optional vehicle priority management can be easily implemented
in the control design. Finally we introduced the possibility of the robust control in urban
traffic which is a planned research scope in the future.

9. Acknowledgement
This work is connected to the scientific program of the "Development of quality-oriented
and harmonized R+D+I strategy and functional model at BME" project. This project is
supported by the New Hungary Development Plan (Project ID: TÁMOP-4.2.1/B-
09/1/KMR-2010-0002) and by the Hungarian Scientific Research Fund (OTKA) through
grant No. CNK 78168 and by the János Bólyai Research Scholarship of the Hungarian
Acadamy of Sciences which are gratefully acknowledged.

10. References
Aboudolas, K.; Papageorgiou, M. & Kosmatopoulos, E. (2009). Store-and-forward based
methods for the signal control problem in large-scale congested urban road networks,
Transportation Research Part C: Emerging Technologies, 17:163_174,
doi:10.1016/j.trc.2008.10.002
Bellemans, T.; De Schutter, B. & De Moor, B. (2002). Model predictive control with repeated
model fitting for ramp metering. Singapore, Proceedings of the Fifth IEEE Intelligent
Transportation Systems Conference, doi:10.1109/ITSC.2002.1041221
Bertsekas, D. P. & Tsitsiklis, J. N. (1997). Parallel and distributed computation: Numerical
methods. ISBN 1-886529-01-9, 731 pages
Blanchini, F. & Miani, S. (2007). Set-Theoretic Methods in Control. ISBN:0817632557,
Birkhäuser, Boston
Boyd, S. & Vanderberghe, L. (2004). Convex optimization. Cambridge University Press, ISBN
0 521 83378 7
Diakaki, C.; Papageorgiou, M. & McLean, T. (1999). Application and evaluation of the integrated
traffic-responsive urban corridor control strategy In-TUC in Glasgow. In CD-ROM of the
78th Annual Meeting of the Transportation Research Board, number 990310,
Washington, D.C., USA
Farges, J. L.; Henry, J. J. & Tufal, J. (1983). The PRODYN real-time traffic algorithm. In 4th
IFAC Symposium on Transportation Systems, pages 307_312
268 Model Predictive Control

Gartner, N. H. (1983). OPAC: A demand-responsive strategy for traffic signal control.

Transportation Research Record, 906:75_84
Gazis, D. C. & Potts, R. B. (1963). The oversaturated intersection. In: Proceedings of the Second
International Symposium on Traffic Theory, London, UK, pp. 221–237.
Hegyi, A.; De Schutter, B. & Hellendoorn, J. (2003). MPC-based optimal coordination of variable
speed limits to suppress shock waves in freeway traffic. Denver, USA, American Control
Conference.
Hunt, P. B.; Robertson, D. L. & Bretherton, R. D. (1982). The SCOOT on-line traffic signal
optimisation technique. Traffic Engineering & Control, 23:190_192
Kwon, H. K. & Pearson, A. E. (1978). On Feedback Stabilization of Time-Varying Discrete Linear
System. IEEE Trans. Autom. Control, 23_3_, pp. 479–481
Maciejowski, J. M. (2002): Predictive Control with Constraints. Harlow, England: Prentice Hall
MATLAB, (2010). The MathWorks Inc., available on http://www.mathworks.com
Rawlings, J. B. & Muske, K. R. (1993). The Stability of Constrained Receding Horizon Control.
IEEE Trans. Autom. Control, 38_10_, pp. 1512– 1516., doi:10.1109/9.241565
Mayne, D. Q.; Rawlings, J. B.; Rao, C. V. & Scokaert, P. O. (2000). Constrained Model Predictive
Control: Stability and Optimality. Automatica, 36_6_, pp. 789–814.,
doi:10.1016/S0005-1098(99)00214-9
Kothare, M. V.; Balakrishnan, V. & Morari, M. (1996). Robust Constrained Model Predictive
Control using Linear Matrix Inequalities. Automatica, vol. 32, no. 10, pp. 1361--79.,
doi:10.1016/0005-1098(96)00063-5
Kulcsár, B.; Varga, I. & Bokor, J. (2005). Constrained Split Rate Estimation by Moving Horizon.
16th IFAC World Congress Prague, Czech Republic
Lowrie, P. R. (1982). The Sydney co-ordinated adaptive traffic system – principles, methodology
logrithms. In: International Conference on Road Traffic Signaling, London, pp. 67–70
Löfberg, J. (2003). Minimax approaches to robust model predictive control. Ph.D thesis, Linköping
University, Sweden
Roca, V. (2005). User Manual for the VISSIM COM interface, PTV AG.
Rockafellar, R. T. (1970). Convex analysis, Princeton University Press
Sen, S. & Head, L. (1997). Controlled optimization of phases at an intersection. Transportation
Science, 31:5_17, doi: 10.1287/trsc.31.1.5
Tettamanti, T. & Varga, I. (2010). Distributed traffic control system based on model predictive
control. Periodica Polytechnica ser. Civil. Eng., 54
Tettamanti, T.; Varga, I.; Kulcsár, B. & Bokor, J. (2008). Model predictive control in urban traffic
network management. 16th IEEE Mediterranean Conference on Control and
AutomationIn CD ISBN:978-1-4244-2505-1 pp. 1538-1543, Ajaccio, Corsica, France
UTOPIA, (2010). Swarco AG., available on http://www.swarco.com
Vigos, G.; Papageorgiou, M. & Wang, Y. (2007). Real-time estimation of vehicle-count within
signalized links. Transportation Research Part C, doi: 10.1016/j.trc.2007.06.002
VISSIM, (2010). PTV AG., available on http://www.ptvag.com
Webster, F. V. & Cobbe, B. M. (1966). Traffic Signals. Technical Paper 56. HMSO, London
Welch, G. & Bishop, G. (1995). An Introduction to the Kalman Filter. University of North
Carolina at Chapel Hill, TR95-041
Wiedemann, R. (1974). Simulation des Straßenverkehrsflusses, Schriftenreihe des Instituts für
Verkehrswesen der Universität Karlsruhe, Heft 8
Off-line model predictive control of dc-dc converter 269

0
12

Off-line model predictive

control of dc-dc converter
Tadanao Zanma and Nobuhiro Asano
Mie University
Japan

1. Introduction
Control systems with switching modes in which different dynamics are assigned are called
hybrid dynamical systems and are being actively researched (1–6). The continuous behavior
in the hybrid dynamical system is expressed generally by differential or difference equations,
while the discrete behavior is described by logics or state machines such as automata. If a sys-
tem can be regarded as a hybrid dynamical system, both continuous and discrete properties
can be dealt with concurrently. Therefore, a hybrid dynamical system has the ability to repre-
sent many systems as a single model without dividing into separate continuous and discrete
systems.
Power electronic circuits can also be regarded as hybrid dynamical systems as they share both
continuous and discontinuous behaviors(7–14). The continuous behavior of current or voltage
in such a system is subject to passive elements such as resistance, capacitance and inductance,
whereas the discontinuous element of switching devices such as MOSFETs and IGBTs yields
an on-off signal that is essentially discrete.
A conventional method currently being used for the control of dc-dc converters is PWM (Pulse
Width Modulation) with triangular wave. The average output voltage is controlled by PWM,
which determines on-off switching timing by employing relatively high carrier frequency.
However, the reference may vary in the half period of triangular wave carrier if the carrier
frequency is lowered to decrease switching loss for saving energy. Then, the average voltage
can no longer approximate the voltage reference. One possible reason is that the control fre-
quency is determined by the carrier frequency only. Another reason may be that the PWM
method focuses only on the average output characteristic and excludes switching property.
Therefore, a novel method is desired for dc-dc converters by considering switching property
explicitly as hybrid dynamical systems.
For synthesis of the hybrid dynamical system, various approaches have been proposed.
Specifically, modeling and synthesis based on mixed logical dynamical (MLD) systems
has much potential since the formulation is similar to the linear discrete time state space
representation(19). The solution of the design is obtained by solving an optimization prob-
lem with the help of model predictive control (MPC)(16; 17). It derives the optimal input to
minimize an estimation of a given cost function by predicting controlled variables for an MLD
system. Specifically, the problem is reduced to a mixed-integer linear or quadratic program-
ming (MILP or MIQP) problem. The method is expected to achieve better control performance
than that achieved by conventional methods when applied to the output control of a power
270 Model Predictive Control

Fig. 1. Topology of the step-down dc-dc converter.

converter. However, it is difficult to solve the optimal problem online because of the com-
putation burden caused by the control period of power converters being considerably short
compared to that in mechanical or process control systems.
This paper proposes a control method using the MPC for the output control problem of the
dc-dc converter. The considered system is described as an MLD system form. In our work(14),
one control period is divided into N submodels. Thus, additional auxiliary variables are
needed. In addition, the state variable among the submodels is handled as an averaged one.
The method in this paper, however. requires no averaging. The explicit switching law is given
as a direct gate signal for the switching devices. Moreover, it is emphasized that a quadratic
cost function which was not adopted in a previous work(14) is addressed in this paper so that
not only the tracking error but also the switching losses can be considered. The proposed
control method achieves quick tracking to the reference in transient state, while keeping the
switching frequency as small as possible in steady state. To verify the effectiveness of the
proposed method, numerical simulations and experimental results are illustrated.
This paper is organized as follows. In Section 2, a step-down dc-dc converter and MLD system
are introduced. Next, the optimization problem for the control is described. Following several
simulation results, Section 3 proposes a modified control method taking into account the com-
putation delay. Then, experimental results are shown in Section 4. Finally, Section 5 concludes
this paper. In the Appendix, formulation of constraints and transformation to mp-MIQP are
explained.

2. Preliminaries
In this section, a step-down dc-dc converter is considered as an example of power elec-
tronic circuits. After the formulation, an MLD system(19) and multi-parametric MIQP (mp-
MIQP)(18) are reviewed.

2.1 Step-down dc-dc converter

The circuit topology of the step-down dc-dc converter is shown in Fig. 1. The dc-dc con-
verter controls the load voltage vo with on-off switches S1 and S2 . The resistance ro expresses
the load. The equivalent series resistance of the capacitor and the internal resistance of the
inductor are denoted by rc and rl , respectively, while xl and xc represent inductance and ca-
pacitance of the low-pass filtering stage, respectively. Switches S1 and S2 cannot be conducted
simultaneously. Together with diode D, switch S2 provides a path for the inductor current il
Off-line model predictive control of dc-dc converter 271

regardless of whether it is positive or negative. The continuous time state-space representa-

tion of the dc-dc converter shown in Fig. 1 is given by

ẋ (t) = Ac x (t) + Bc u(t), (1)

y(t) = Cc x (t), (2)
 
where x (t) = il (t) vo (t) . Denoted by il (t) and vo (t) are the inductor current
and
 output voltage, respectively.  The matrices  Ac , Bc and Cc are given by Ac =
rl 1 1  
− xl − xl xl
ro 1 rc rl ro 1 rc , Bc = rc ro and Cc = 0 1 , respectively.
r c +r o ( x c − x l ) − r c +r o ( x c r o + x l ) x l (r c +r o )
Eqs. (1) and (2) are sampled by Ts . Hereafter, the discrete time is described anew as t. Thus,
the considered system is recast as follows.

x (t + 1) = Ax (t) + Bu(t), (3)

y(t) = Cx (t), (4)
 Ts
where A = e Ac Ts , B = e Ac τ dτBc and C = Cc . Note that the value of input is limited to
0
either 0 or vs , which can be rewritten as follows.

(∀t) u(t) ∈ {0, vs }. (5)

2.2 Representation by MLD system(19)

A mixed logical dynamical (MLD) system is described by a linear dynamical equation with
linear mixed-integer inequalities so that discrete properties included in the process can be
introduced into the system using logical variables. One advantage is that the logical formula
can be described with linear inequalities and model predictive control can be applied.
The model of the dc-dc converter is rewritten to the MLD system representation. The auxiliary
δ of 0-1 variable is introduced as a new input variable to describe the discrete variable. The
variable is associated as follows.

[ δ ( t ) = 1] → [ z ( t ) = v s ], (6)
[ δ ( t ) = 0] → [ z ( t ) = 0], (7)

where z(t) is,

0 ≤ z(t) ≤ vs . (8)

Eqs. (6) and (7) indicate that z(t) = vs if δ(t) = 1, whereas z(t) = 0, otherwise. By replacing
Eqs. (6) and (7) with their equivalent linear inequalities,

E1 δ(t) + E2 z(t) ≤ E3 u(t) + E4 x (t) + E5 , (9)

where,
 
E1 = 0 vs −vs 0 , (10)
 
E2 = 1 −1 1 −1 , (11)
E3 = E4 = O, (12)
 
E5 = vs 0 0 0 . (13)
272 Model Predictive Control

is obtained. Inequality (9) reflects that z(t) = vs if δ(t) = 1 whereas z(t) = 0 if δ(t) = 0.
Namely, δ(t) can be considered as the state of the switch: δ(t) = 1 if the switch is on, δ(t) = 0
otherwise. Note that z(t) in inequality (8) is an apparent continuous auxiliary variable.
As a result, Eqs. (3), (4) and (5) can be transformed into an MLD system consisting of one
standard linear discrete time state space representation and linear inequalities associated with
the constraints on the system,

x (t + 1) = Ax (t) + Bz(t), (14)

y(t) = Cx (t), (15)
subject to Eq. (9). (16)

2.3 Multi-parametric MIQP(18)

Multi-parametric MIQP (mp-MIQP) is a type of MIQP(18) parameterized by multiple param-
eters. The mp-MIQP parameterized by state x of the system is described as follows.

min ν Hν + 2x  Fν + x  Yx + 2C f ν + 2Cx x, (17)

ν
subject to Gν ≤ W + Ex, (18)

where ν is
 
ν = ∆ Ξ , (19)
 
∆ = δ0 ... δ Np −1 , (20)
 
Ξ = z0 ... z Np −1 . (21)

In Eqs. (20) and (21), the predictive horizon in MPC is denoted by Np .

If solved, the optimal solution of mp-MIQP is given as the piece-wise affine state feedback
form. Namely, the explicit control law parameterized by the state x is obtained as follows.

ν = Ki x + h i if x ∈ Xi , (22)

where Xi (i = 1, 2, . . .) are regions partitioned in the state space, and Ki and hi are the cor-
responding constant matrices and vectors, respectively. As Eq. (22) is available off-line, the
optimal input is determined online according to the state measured at each sampling.

3. Numerical simulation and revision of control method

In this section, the effectiveness of the method proposed in the previous section and the Ap-
pendix is shown by applying it to the output control of the dc-dc converter shown in Fig. 1.
The control objective is to achieve quick tracking to the reference in transient state with mini-
mal switching in steady state. For the purpose, mp-MIQP is exploited.

3.1 Simulation condition and state partition

The circuit and control parameters for simulation are listed in Tables 1 and 2, respectively.
Let us consider Eqs. (14) to (16) as the model for the dc-dc converter shown in Fig. 1. In
Eq. (45), H̃ and L are first set as zeros. Then, the setting of these matrices imply that focus
is only on tracking performance. The state partition obtained by off-line model predictive
control, (mp-MIQP) and its enlarged view are shown in Fig. 2. In each region of Fig. 2, the
optimal input sequence is assigned. The figure of state partition shown in Fig. 2 is generated
Off-line model predictive control of dc-dc converter 273

Table 1. Circuit parameters

source voltage vs 5.0 [V]
inductance xl 20 [µH]
internal resistance rl 25 [mΩ]
capacitance xc 2.2 [mF]
equivalent series resistance rc 60[mΩ]
load resistance ro 1[Ω]

Table 2. Control parameters

control period Ts 10 [µs]
predictive horizon Np 1, 3, 5
upper limit il,max 8.0 [A]
reference value vref 2.0 [V]

using of Multi-Parametric Toolbox(20). In Fig. 2, the number of state partitions is limited to

at most 2 Np . Each partition is specified by linear inequalities. In each partition, the solution
of mp-MIQP given by Eq. (22) is assigned. To investigate to which partition it belongs, the
 
state il vo at each sampling can be performed simply since the obtained state partition
is constructed by linear inequalities. Focus on the white region at the right bottom corner
 
in Fig. 2. Whenever the state il vo enters the region, switch S1 shown in Fig. 1 is forced
to turn off since the constraint about the inductor current given by Eq. (37) can no longer be
satisfied.

3.2 Consideration of delay for computation of state distinction

Figs. 3 and 4 show simulation results for Np = 3 and Np = 5, respectively. Note that the
method described in the Appendix is utilized for each of the calculations. Figs. 3 and 4, also
indicate that the output voltage is kept at the specified value 2.0 [V] in steady state, while the
inductor current does not exceed its limit of 8[A]. In the simulation, the computation time of
state distinction for optimal input is assumed to be negligible. Little difference exists between

5 2.6

4
2.4

3
2.2
vo

vo

2
2
1

1.8
0

-1 1.6
-5 0 5 10 -5 0 5 10
il il

Fig. 2. State partition for Np = 5 (left: whole, rigtht: closeup).

274 Model Predictive Control

2.5 12

2 10
8
1.5
6
vo [V]

il [A]
1
4
0.5
2
0 0
−0.5 −2
0 1 2 3 4 5 0 1 2 3 4 5
time [ms] time [ms]

Fig. 3. Simulation result in case computation delay is negligible for Np = 3 (left: vo , right: il ).

2.5 12

2 10
8
1.5
6
vo [V]

il [A]
1
4
0.5
2
0 0
−0.5 −2
0 1 2 3 4 5 0 1 2 3 4 5
time [ms] time [ms]

Fig. 4. Simulation result in case computation delay is negligible for Np = 5 (left: vo , right: il ).

the two outputs shown in Figs. 3 and 4. In other words, the performance is almost identical
for Np = 3 and Np = 5 as long as the computation time is minimal.
On the other hand, as described later in the next section, the computation time should be
considered. because of the effects of various factors such as DSP performance and the number
of state partitions. In preliminary experiments, 5 [µs] and 8 [µs] for Np = 3 and Np = 5,
respectively, are obtained as average computation delay. Using the values, we set the delay
for determination of the switching signal after measurement of the state in the simulation.
Figs. 5 and 6 illustrate the simulation results under the assumption that the computation delay
is not negligible, i.e., the delay is assumed to exist for the computation. From Figs. 5 and 6,
the switching intervals that exceed 20[¯s] can be seen. Thus, the ripple effect increases as the

2.5 12

2 10
8
1.5
6
vo [V]

il [A]

1
4
0.5
2
0 0
−0.5 −2
0 1 2 3 4 5 0 1 2 3 4 5
time [ms] time [ms]

Fig. 5. Simulation result in case computation delay is 5 [µs] for Np = 3 (left: vo , right: il ).
Off-line model predictive control of dc-dc converter 275

2.5 12

2 10
8
1.5
6
vo [V]

il [A]
1
4
0.5
2
0 0
−0.5 −2
0 1 2 3 4 5 0 1 2 3 4 5
time [ms] time [ms]

Fig. 6. Simulation result in case computation delay is 8 [µs] for Np = 5 (left: vo , right: il ).

2.5 12

2 10
8
1.5
6
vo [V]

il [A]
1
4
0.5
2
0 0
−0.5 −2
0 1 2 3 4 5 0 1 2 3 4 5
time [ms] time [ms]

Fig. 7. Simulation result with consideration of computation time for Np = 5 (left: vo , right: il ).

difference widens between the value of the measured state and that of the input which is
determined after the delay.

3.3 Modification of control method

In the method proposed(21) in the Appendix, input is applied after examination of the region
in which the state belongs. However, as mentioned above, the performance is not necessarily
satisfactory due to the computation delay even if the horizon is small. Therefore, the con-
trol method should be slightly modified in order to consider the computation delay so that
performance is not degraded. Specifically, instead of the first one, the second element of the
optimal input sequence is applied to the system at the beginning of the next control period.
In addition, the first element of the optimal input sequence has to be used as that given at the
last sampling. In other words, the first element is not solved but is set as that given at the last
period, i.e., in the modified control method, δ0 and z0 in Eqs. (20) and (21), respectively, are
given in advance as the constants of the last optimized input sequence, not solved as the opti-
mized variables. Note that the modified control method requires Np > 1 due to the structure.
Fig. 7 depicts the simulation result by the modified method above mentioned. Compared with
Fig. 6, the result shown in Fig. 7 is improved in the sense that the ripple is reduced in steady
state.

4. Experimental result
In this section, we show the effectiveness of the modified proposed method(21) through exper-
iments. In addition, the effectiveness for consideration of the switching loss is demonstrated.
276 Model Predictive Control

2.5 12

2 10
8
1.5
6
vo [V]

il [A]
1
4
0.5
2
0 0
−0.5 −2
0 1 2 3 4 5 0 1 2 3 4 5
time [ms] time [ms]

Fig. 8. Experimental result without consideration of computation delay (left: vo , right: il ).

2.5 12

2 10
8
1.5
6
vo [V]

il [A]
1
4
0.5
2
0 0
−0.5 −2
0 1 2 3 4 5 0 1 2 3 4 5
time [ms] time [ms]

Fig. 9. Experimental result with consideration of computation delay (left: vo , right: il ).

The experiments are carried out on a DSP (Texas Instruments TMS3200C/F2812, operating
frequency: 150 [MHz], AD-converter: 12 [bit], conversion time: 80 [ns]).

4.1 Comparison of proposed method(21) and its modified method

Fig. 8 shows the experimental result obtained without considering the computation delay for
state distinction for Np = 5. Similar to simulation results shown in Fig. 4, many switchings
are described with intervals exceeding 20 [µs] although the control period is 10 [µs]. The
reason for the results is that the state transits to another which is not the predictive one, due
to the computation delay. Therefore, the computation delay for state distinction should be
considered in the experiments. Fig. 9 shows the experimental result upon consideration of the
computation delay. Note that the results shown in Fig. 9 are obtained by the modified control
method mentioned in the previous section.
Compared with the results shown in Fig. 8, the ripple effect is reduced as shown in Fig. 9. This
reduction occurs because the computation delay is considered in the latter result. Thus, the
effectiveness of the modified control method in Subsection 3.3 is demonstrated.

4.2 Consideration of switching loss

The shorter the control period, the more the switching losses tend to increase, as do the num-
ber of switchings. In the proposed method, the switching loss can be considered by incorpo-
rating it into the cost function. This can be achieved by setting Q = qINp −1 where q = 10−3
in Eq. (42). The experimental result is shown in Fig. 10. From Fig. 10, the output voltage is
tracked to the voltage reference even though the term to reduce switching is added into the
cost function. Fig. 10 also shows that the inductor current does not severely exceed the limit
Off-line model predictive control of dc-dc converter 277

2.5 12

2 10
8
1.5
6
vo [V]

il [A]
1
4
0.5
2
0 0
−0.5 −2
0 1 2 3 4 5 0 1 2 3 4 5
time [ms] time [ms]

Fig. 10. Experimental result with consideration of computation delay and the switching loss
for Np = 5 (left: vo , right: il ).

6 6
switching signal u(t)

switching signal u(t)

4 4

2 2

0 0

−2 −2
2.0 2.2 2.4 2.6 2.8 3.0 2.0 2.2 2.4 2.6 2.8 3.0
time [ms] time [ms]

Fig. 11. Experimental result of switching signal without/with consideration of the switching
loss for Np = 5 (left: without, that in Fig. 9, right: with, that in Fig. 10).

of 8 [A]. Fig. 11 shows the switching signals for Figs. 9 and 10. From the right of Fig. 11, the
switching frequency is reduced by considering the switching loss in the cost function given
by Eq. (45). Thus, both tracking performance and switching loss can be considered simultane-
ously in the proposed method.

5. Conclusions
In this paper, a novel control method for the dc-dc converter has been proposed. The dc-
dc converter has been modeled as a mixed logical dynamical (MLD) system since it has the
ability to combine continuous and discrete properties. For the control, a model predictive
control (MPC) based method has been introduced. The optimization problem has been solved
as a multi-parametric off-line programming problem. The result has been obtained as the
state space partition which makes the implementation feasible. As a result, computation time
is shortened without deteriorating control performance. Finally, it has been demonstrated that
the output voltage has been tracked to the reference at the expense of tracking performance by
introducing the term to reduce the switching in the cost function. In some cases, other factors
such as resistance loss in rl shown in Fig. 1 may need to be considered, although the cost
function given by Eq. (28) considers only the tracking performance and switching loss. Note,
however, that the factors represented as linear and/or quadratic forms of the state variable
can be incorporated into the cost function.
Further research includes robustness analysis in implementation and investigation of perfor-
mance for different cost functions as mentioned above.
278 Model Predictive Control

Acknowledgment
We are grateful to the Okasan-Kato Foundation. We also thank Professor Manfred Morari,
Ph.D, Sébastien Mariéthoz, Ph.D, Andrea Beccuti, Ph.D, of ETH Zurich for valuable comments
and suggestions.
Here, the proposed method(15) is reviewed in brief.
MIQP derives the values that minimize an estimation of a given cost function under con-
straints given by inequalities and/or equalities concerning integer variables. The MIQP for
Eqs. (14) to (16) is given as follows.
min νt S1 νt + 2(S2 + x (t) S3 )νt , (23)
νt
subject to F1 νt ≤ F2 + F3 x (t), (24)
where νt is
 
νt = ∆t Ξt , (25)
 
∆ t = δ (0| t ) . . . δ ( Np − 1| t ) , (26)
 
Ξ t = z (0| t ) . . . z ( Np − 1| t ) . (27)
To derive the optimal input sequence for Eqs. (14) to (16), the following cost function is set.
Np
J ( x ( t ), ∆ t , Ξ t ) = ∑ y(k|t) − vref 22
k =1
+ ∆t H̃∆t + 2L∆t , (28)
where vref denotes the constant voltage reference. In Eq. (28), the first term is associated with
the tracking performance whereas the switching loss can be also considered in the second and
third terms. Eq. (28) is rewritten as the general MIQP form of Eqs. (23) in order to solve the
minimization problem. By Eqs. (14) and (15), y(k|t) which is the predictive output k steps
ahead of t is described as follows.
k −1
y(k |t) = C ( Ak x (t) + ∑ Ak− j−1 Bz( j))
j =0

= C ( Ak x (t) + Gk Ξk ), (29)
 k −1 
where Gk = A B Ak−2 B . . . B . By substituting Eq. (29) for Eq. (28), the minimization
problem for Eq. (28) is formalized as follows.
 Np Np
min ∑ Ξt Gk C  CGk Ξt − 2 ∑ vref
CGk Ξt
∆t , Ξt
k =1 k =1
Np 
+2 ∑ x (t) Ak C  CGk Ξt + ∆t H̃∆t + 2L∆t . (30)
k =1

Note that the irrelative terms for the minimization problem are omitted in Eq. (30). Connected
with Eq. (23), the optimization problem of Eq. (30) is transformed as
     
∆t ∆ ∆
min S1 t + 2 ( S2 + x ( t )  S3 ) t , (31)
∆t , Ξt Ξ t Ξt Ξt
Off-line model predictive control of dc-dc converter 279

where S1 , S2 and S3 are,

 
H̃ O
S1 =  Np  ∈ R2Np ×2Np , (32)
O ∑ Gk C  CGk
k =1
 
Np
S2 = L − ∑  CG
vref k ∈ R1×2Np , (33)
k =1
 
Np
S3 = O ∑ Ak C  CGk ∈ R2×2Np , (34)
k =1

respectively.
Let us rewrite the constraint as the general form like inequality (24). Recall that only two
discrete inputs are permitted in the considered system. The constraint represented by Eq. (9)
is also transformed as
 
∆
F̃1 t ≤ F̃2 + F̃3 x (t), (35)
Ξt

where F̃1 , F̃2 and F̃3 are, respectively,

 
E1 O E2 O
 .. ..  4N ×2Np
F̃1 =  . . ∈R p ,
O E1 O E2
    (36)
E5 E4 E4
 .   . ..  ∈ R4Np ×2 .
F̃2 =  ..  ∈ R4Np , F̃3 =  .. . 
E5 E4 E4

The constraints imposed on the inductor current limitation is are necessary to prevent damage
to the switching device from excessive current. More specifically, if the predictive inductor
current at t + 1, i.e., il (1|t), exceeds its limit, il,max , then the switch is forced to be off. Such an
additional condition can be described as

[il (1|t) > il,max ] → [δ(0) = 0]. (37)

Transformed into the inequality, Eq. (37) is described as

il (1|t) − il,max ≤ M(1 − δ(0)), (38)

 
where M is the admissible upper limit of il . Since x = il vo , replaced the first row of A
and the first element of B with A1 and b1 , respectively, il (1|t) is recast as,
 
il (1|t) = a11 a12 x (t) + b1 z(0), (39)
 
where a11 a12 is the first row of A. Consequently, using Eq. (39), inequality (38) can be
expressed as
 
Mδ(0) + b1 z(0) ≤ ( M + il,max ) − a11 a12 x (t). (40)
280 Model Predictive Control

Add Eq. (40) as a new constraint to the last row of Eq. (36), then Eq. (36) is modified as follows.
 
F̃1
F1 = ,
M 0 . . . 0 b1 0 . . . 0
    (41)
F̃2 F̃
F2 = , F3 = 3 .
M + il,max A1
The switching loss can also be considered in the second and third terms in Eq. (28). In Eq. (28),
for example, L = O and H̃ is set with Q  0 as follows.
H̃ = (Π1 − Π2 ) Q(Π1 − Π2 ), (42)
where Π1 and Π2 are, respectively,
 
0
. 
Π1 =  .. INp −1  ∈ R ( Np −1)× Np , (43)
0
 
0

Π 2 =  I Np −1 ..  ∈ R ( Np −1)× Np . (44)
.
0

Note that when H̃ and L are set above, the estimation of the cost function of Eq. (28) increases
in response to the number of switchings required. Therefore, the switching loss can be reduced
depending on Q in Eq. (42).
If the cost function is described, the optimal input sequence can be derived. However, it is
impractical to apply it to the considered dc-dc converter with a short control period since
the computation requires much solution time for every control period. Then, the method
above is transformed into mp-MIQP so that solving the optimization problem on-line is no
longer necessary. Eq. (28) is adopted as the cost function again for mp-MIQP. Then, Eq. (28) is
described as follows.
J ( x, ∆, Ξ)
Np Np
= ∑ Ξ Gk C  CGk Ξ + 2 ∑ x Ak C CGk Ξ
k =1 k =1
Np Np
+ ∑ x  Ak C  CAk x − 2 ∑ vref

CGk Ξ
k =1 k =1
Np
−2 ∑ vref

CAk x + ∆ H̃∆ + 2L∆, (45)
k =1
   
where ∆ = δ0 . . . δNp −1 and Ξ = z0 . . . z Np −1 . Associated with Eq. (17), the opti-
mization problem of Eq. (45) is transformed as follows.
     
∆ ∆ ∆
min H + 2x  F + x  Yx
∆,Ξ Ξ Ξ Ξ
 
∆
+ 2C f + 2Cx x, (46)
Ξ
Off-line model predictive control of dc-dc converter 281

where H̃ = S1 , F = S3 and C f = S2 , respectively. Note that there exists a clear difference

between notations of νt and ν. The former is utilized for MIQP while the latter is used for
mp-MIQP. The others are
Np
Y= ∑ Ak C CAk , (47)
k =1
Np
Cx = − ∑ vref

CAk . (48)
k =1

The constraints are given by

 
∆
F1 ≤ F2 + F3 x. (49)
Ξ

Transformed as above, the optimization problem is solved offline as mp-MIQP. Then, the re-
sult is employed for on-line control.

6. References
[1] Hybrid systems I, II, III, IV, V, Lecture Notes in Computer Science, 736, 999, 1066, 1273, 1567,
New York, Springer-Verlag, 1993 to 1998.
[2] “Special issue on hybrid control systems," IEEE Trans. Automatic Control, Vol. 43, No. 4,
1998.
[3] “Special issue on hybrid systems," Automatica, Vol. 35, No. 3, 1999.
[4] “Special issue on hybrid systems," Systems & Control Letters, Vol. 38, No. 3, 1999.
[5] “Special issue hybrid systems: Theory & applications", Proc. IEEE, Vol. 88, No. 7, 2000.
[6] T. Ushio, “Expectations for Hybrid Systems," Systems, Control and Information, Vol. 46,
No. 3, pp. 105–109, 2002.
[7] S. Almer, H. Fujioka, U. Jonsson, C. Y. Kao, D. Patino, P. Riedinger, T. Geyer, A. G. Beccuti,
G. Papafotiou, M. Morari, A. Wernrud and A. Rantzer, “Hybrid Control Techniques for
Switched-Mode DC-DC Converters Part I: The Step-Down Topology," Proc. ACC , pp. 5450–
5457, 2007.
[8] A. G. Beccuti, G. Papafotiou, M. Morari, S. Almer, H. Fujioka, U. Jonsson, C. Y. Kao,
A. Wernrud, A. Rantzer, M. Baja, H. Cormerais, and J. Buisson, “Hybrid Control Tech-
niques for Switched-Mode DC-DC Converters Part II: The Step-Up Topology," Proc. ACC,
pp. 5464–5471, 2007.
[9] A. G. Beccuti, G. Papafotiou, R. Frasca and M. Morari, “Explicit Hybrid Model Predictive
Control of the dc-dc Boost Converter," Proc. IEEE PESC, pp. 2503–2509, 2007.
[10] I. A. Fotiou, A. G. Beccuti and M. Morari, “An Optimal Control Application in Power
Electronics Using Algebraic Geometry," Proc. ECC, pages 475–482, July 2007.
[11] R. R. Negenborn, A. G. Beccuti, T. Demiray, S. Leirens, G. Damm, B. D. Schutter and
M. Morari, “Supervisory Hybrid Model Predictive Control for Voltage Stability of Power
Networks," Proc. ACC, pp. 5444–5449, 2007.
[12] A. G. Beccuti, G. Papafotiou and M. Morari, “Optimal control of the buck dc-dc converter
operating in both the continuous and discontinuous conduction regimes," Proc. IEEE CDC,
pp. 6205–6210, 2006.
282 Model Predictive Control

[13] T. Geyer, G. Papafotiou, M. Morari, “On the Optimal Control of Switch-Mode DC-DC
Converters," Hybrid Systems: Computation and Control, Vol. 2993, pp. 342–356, Lecture Notes
in Computer Science, 2004.
[14] G. Papafotiou, T. Geyer, M. Morari, “Hybrid Modelling and Optimal Control of Switch-
mode DC-DC Converters," IEEE Workshop on Computers in Power Electronics (COMPEL),
pp. 148–155, 2004.
[15] K. Asano, K. Tsuda, A. Bemporad, M. Morari, “Predictive Control for Hybrid Systems
and Its Application to Process Control," Systems, Control and Information, Vol. 46, No. 3,
pp. 110–119, 2002.
[16] M. Ohshima, M. Ogawa, “Model Predictive Control –I– Basic Principle: history & present
status," Systems, Control and Information, Vol. 46, No. 5, pp. 286–293, 2002.
[17] M. Fujita, M. Ohshima, “Model Predictive Control –VI– Model Predictive Control for
Hybrid Systems," Systems, Control and Information, Vol. 47, No. 3, pp. 146–152, 2003.
[18] F. Borrelli, M. Baotic, A. Bemporad, M. Morari, “An efficient algorithm for computing
the state feedback optimal control law for discrete time hybrid systems," In Proc. ACC,
pp. 4717–4722, 2003.
[19] A. Bemporad, M. Morari, “Control of systems integrating logic, dynamics, and con-
straints," Automatica, Vol. 35, No. 3, pp. 407–427, 1999.
[20] M. Kvasnica, P. Grieder, M. Boatić and F. J. Christophersen, “Multi-Parametric Toolbox
(MPT)," Institut für Automatik, 2005.
[21] N. Asano, T. Zanma and M. Ishida, “Optimal Control of DC-DC Converter using Mixed
Logical Dynamical System Model," IEEJ Trans. IA, Vol. 127, No. 3, pp. 339–346, 2007.
Nonlinear Predictive Control of Semi-Active Landing Gear 283

13
x

Nonlinear Predictive Control of

Semi-Active Landing Gear
Dongsu Wu, Hongbin Gu and Hui Liu
Nanjing University of Aeronautics and Astronautics
China

1. Introduction
When airplane touches down and taxis on uneven runways with high speed, there is heavy
ground impact and huge vertical load to the airframe. To improve safety and make
passengers comfortable during landing, an effective landing gear capable of absorbing
impact energy as much as possible is indispensable for modern airplane. Besides the basic
function of reducing impact loads, the landing gear must also allow sufficient
maneuverability during ground operation, which leads to conflicting requirements in terms
of the suspension system (Krüger, 2000). Traditional landing gear consists of tires and
passive shock absorbers, which can only be optimized before leaving factory to ensure the
landing gear having a fairly good performance in particular design operational conditions,
typically hard landings. However, due to its fixed structure, the passive shock absorber
cannot always work well on various ground conditions and operational conditions. A heavy
landing or a coarse runway may lead to significant deterioration of its performance, which is
harmful to the fatigue life of the landing gear and of the airframe.
Active control and semi-active control are widely used approach in the field of construction
vibration control and vehicle suspension control. Compared with passive control, active and
semi-active control has excellent tunable ability due to their flexible structure. Active control
needs an external hydraulic source to supply energy for the system. The main drawback of
active control approach is that its structure is very complex and the external energy may
lead to instability of the system. The semi-active approach (Fig.1) modifies the damping
characteristics by changing the size of the orifice area and does not introduce any external
energy. Studies by Karnopp (Karnopp, 1983) for automotive applications also suggest that
the efficiency of semi-active dampers is only marginally lower than of a fully active system,
provided that a suitable control concept is used. In consideration of its simple structure and
high reliability, semi-active control approach could be a better choice for landing gear
systems.
The main component of semi-active landing gear system is a tunable oleo-pneumatic shock
absorber, which contains multidisciplinary and highly nonlinear dynamics. It is not an easy
task to design an effective controller for such complex system. Krüger (Krüger, 2000) focuses
his studies on optimization of taxiing performance of a semi-active landing gear. SIMPACK
software is used to run simulation with a complete aircraft FEA model. Ghiringhelli builds a
284 Model Predictive Control

complete aircraft landing simulation model in ADAMS software (Ghiringhelli et al., 2004). A
semi-active PID control method is used to control the orifice area. His also studies sensitivity
of the complete aircraft model to the variation of control parameters and compares the
results obtained in the simulated drop tests between passive and semi-active approach
(Ghiringhelli, 2000). Maemori et al. (Maemori et al., 2003) proposes an optimization method
for a semi-active landing gear to handle variations in the maximum vertical acceleration of
an aircraft during landing caused by the variation of the aircraft mass, which is always due
to the variations in the number of passengers, and the amounts of cargo and fuel. Wang et
al. (Wang et al., 1999) considers both taxiing and landing conditions. He uses a fuzzy
controller to optimize the performance of the semi-active landing gear. But he does not
consider the dynamics of the actuator. Mikulowski et al. (Mikulowski et al., 2008) discuss
the application of piezo-actuators and magneto-rheological damper in the adaptive landing
gear design. And there are some other researchers applying ER (Lou et al., 1993) or MR
(Batterbee et al., 2007) technology in semi-active land gear system. All of the semi-active
controllers designed above do not consider the actuator saturations (limited control
amplitude and rate), which may lead to significant, undesirable deterioration in the closed-
loop performance and even closed-loop instability.

Fig. 1. Structure of Semi-Active Controlled Shock Absorber

Model predictive control refers to a class of control algorithms in which a dynamic model is
used to predict and optimize control performance. The predictions are obtained from a
dynamic model and the optimization problem is solved subject to constraints on input and
output variables. So MPC is especially suited for constrained, digital control problems.
Initially MPC has been widely used in the industrial processes with linear models, but
recently some researchers have tried to apply MPC to other fields like automotive (Mehra et
al., 1997) and aerospace (Hyochoong et al., 2004), and the nonlinear model is used instead of
linear one due to the increasingly high demands on better control performance and rapidly
developed powerful computing systems (Michael et al., 1998). To the semi-active landing
gear control problem, the nonlinear model predictive control is a good choice considering its
effectiveness to constrained control problems and continuously optimized performance. The
goal of this paper is to introduce the design and the analysis of a nonlinear hierarchical
Nonlinear Predictive Control of Semi-Active Landing Gear 285

control strategy, for semi-active landing gear systems in civil and military aircrafts, based on
predictive control strategies.

2. Dynamic Model of Semi-Active Landing Gear

The structure mass of landing gear is divided into sprung mass and non-sprung mass.
Sprung mass defined in the figure includes the airframe, the cylinder etc. Non-sprung mass
includes the piston rod, wheel etc. The tire is modelled as a simple spring and the tunable
damping is realized by a variable size orifice which is controlled by a high-speed solenoid
valve.
The governing dynamic equations of semi-active landing gear can be presented as the
following:
ms zs  ms g  F (1)

mu zu  mu g  F  P (2)

Where mu is the unsprung mass, ms the sprung mass, zu the displacement of unsprung
part, z s the displacement of sprung part, P the vertical force on the tire, F the semi-active
damper shock strut force.

Fig. 2. System Model of Semi-Active Landing Gear

2.1 Shock Struct Force Model

Considering basics of the shock strut operation, a damping effect is produced by squeezing
the compressed oil through the tunable orifice. In the pneumatic chamber, the enclosed air is
compressed by the movement of the piston, which provides an air cushion spring. There is
also friction produced between sliding parts. All these forces comprise the shock strut force
(Yadav et al., 1991):
Oleo damping force:
Ao 3 sgn( z s  zu ) (3)
Foil  ( z s  zu ) 2
2 2
2( Ad C d ) 1  ( Ad / Ao )
286 Model Predictive Control

Air spring force:

V0
Fair  Pi Aa [ ]n  P0 Aa (4)
V0  Aa ( zs  zu )
Friction force:
F f  K m Fair (5)
Total axial force in the shock strut:
F  Fair  F f  Foil (6)

where ρ is the oil density, Pi is the initial pneumatic pressure of air chamber, Pi is the
atmospheric pressure, A0 is the effective oil action area, Aa is the effective air action area,
Ad is the tunable oil orifice area, C d is the tunable oil orifice flow coefficient, V0 is the
initial volume of air chamber, K m is the coefficient of kinetic friction.

2.2 Tire Force Model

The vertical force P on the tire is due to polytropic compression of air inside the tire. In
order to simplify the mathematical model, the tire is treated as a linear spring here:

P  K t zu  Ct zu (7)

where K t is the stiffness coefficient of tire, and Ct the damp coefficient of tire.

2.3 Model of High Speed Solenoid Valve

Traditional solenoid valve (Fig.3) are simple in construction, rugged, relatively cheap to
produce and have higher power-mass ratio, but they are not usually used for continuous
and proportional control due to its high nonlinearity. Recently, some attempts are made in
this kind of application using nonlinear control methods. According to our previous studies
(Liu H. et al, 2008), we model a high speed solenoid valve by considering its mechanical,
magnetic and electrical dynamics.

Fig. 3. High-speed Solenoid Valve’s Structure

Nonlinear Predictive Control of Semi-Active Landing Gear 287

(a) The mechanical dynamics of solenoid valve can be expressed as below:

mv xv  Cs xv  ( K s  K f ) xv  f 0  f  Fv (8)

where mv is the total mass of movable parts including armature, actuator pin, etc., C s
viscous damping coefficient, K s spring stiffness, K f static flow coefficient, f 0 preloading
force of spring, f Coulomb friction, and xv movable part displacement and is proportional
to oil orifice area Ad .
Ad  K v xv (9)

where K v is the proportionality coefficient.

(b) The magnetic dynamics of solenoid valve can be summarized as following:
The magnetomotive force is
 m  Ni  ΦRm (10)

where N is the coil turns, i current, Rm total magnetic reluctance, and  total magnetic
flux.
The electromagnetic force that acts on the armature of valve can be given by
2
Φair
Fv  (11)
20 r 2 2

where μ0 is the air permeability, r the radius of armature, λ the leakage coefficient of the
main air gap, and  air magnetic flux passing through the working air gap.

Ro RL
Φair  Φ (12)
Ro  RL

Ro and RL are corresponding to the magnetic reluctance of two part of magnetic flux paths.
Due to the fact that λ, Ro , RL , Rm are related to xv , and according to Eq. (10-12), the
magnetic equations of solenoid valve can be simplified as:

Fv  B ( xv )i 2 (13)

where B(xv) is a function of xv and represents nonlinear magnetic dynamics of valve. Fv

depends on i 2 , the square of electrical current. i is the control input for solenoid valve.
(b) Solenoid valve is also characterized by the electric equation:
288 Model Predictive Control

dL ( x v , i ) di
V  Ri  i  L ( xv , i ) (14)
dt dt

From the above equation, we can see that an inner loop to control current can be introduced
to improve current input accuracy. According to (Malaguti et al., 2002), mechanical
dynamics of solenoid valve is slow respect to electric one, so we obtain the simple electric
equation.
di
V  Ri  L( xv 0 ) (15)
dt

The inductance is supposed constant in the operating position and independent on current.
And specific values of valve’s parameters can be found in (Liu H. et al., 2008).

2.4 Full State Mode

Assigning the states as x1  z s  zu , x2  zs  zu , x3  zu , x4  zu , x5  xv , x6  xv ,
x7  i , and combining all the equations we obtain the full state model.

x1  x2 ; (16)

1 m  mu
x2  (Kt x3  Ct x4 )  s F; (17)
mu ms mu

x3  x4 ; (18)

1
x 4  g  [ F  K t x3  Ct x4 ]; (19)
mu

x5  x6 ; (20)

1
x6  [B(x5 )x72  Cs x6  (Ks  K f )x5  f  f0 ]; (21)
mv

1
x7  (V  Rx7 ); (22)
L
where F, Fair and Foil can be expressed as following,

F  (1  K m ) Fair  Foil (23)

Nonlinear Predictive Control of Semi-Active Landing Gear 289

V0
Fair  Pi Aa ( ) n  P0 Aa (24)
V0  Aa x1

Ao 3 sgn( x2 ) 2 (25)

Foil  x2
2 2
2( K v x 5 C d ) 1  ( K v x5 / Ao )

3. Design Objective of Oleopneumatic Shock Absorber in Landing Gear

System
The tasks of aircraft landing gears are complex and lead to a number of sometimes
contradictory requirements. At touchdown, the landing gear has to perform its task of
absorbing the aircraft vertical energy via the shock absorber and the horizontal energy by
the brakes. At taxiing, the landing gear has to carry the aircraft over taxiways and runways
of varying quality. The requirements for absorption of a hard touch-down and for
comfortable rolling lead to a design conflict.

3.1 Touchdown Phase

At touchdown phase, the design objective of oleo-pneumatic shock absorber is aimed at
reducing the maximum vertical load level introduced at the fuselage attachment and
producing a possibly “balanced” set of landing structural loads at touchdown. To get
optimal structural load, the impact energy should be equally distributed with respect to the
shock absorber stroke. So the optimal structural load during touchdown is a constant value:
z cc

Fsao 
 0
Fsa ( z )dz
(26)
z0
Fsao can be estimated by the total energy to be absorbed at touchdown, including kinetic
energy and potential energy in vertical direction, and the expected stroke of shock absorber
which is generally 90%-95%of the maximal stroke (the work done by drag and lift are
omitted).
It is hard for a conventional passive landing gear system to achieve this optimal target load.
Semi-active landing gear system has a better performance due to its flexible structure, and is
possible to reach the ideal effect if a suitable control method is used. Actually, stroke z1 is
needed to travel before structural load reaches Fsao , and this part of the gear compression
cannot overly reduced (Ghiringhelli et al., 2004). If z1 is too short, the gear stiffness will be
large and thus the longitudinal spin-up loads will increase sharply. That will lead to the
reduction of unitary efficiency. So a reasonable choice is to use passive control till the
structural load reaches Fsao , and then change to semi-active control afterwards. That results
Fsas , a sub-optimal structural load solution. By using this scheme, the unitary efficiency of a
landing gear system can be achieved though the efficiency of the shock absorber is
decreased.
290 Model Predictive Control

Fig. 4. Shock Absorber Efficiency Comparison

3.2 Taxiing Phase

At taxiing phase, the design objective of oleo-pneumatic shock absorber is aimed at filtering
the unevenness of runway surface and providing a comfortable ground ride. It is expected
that an aircraft rapidly returns to its original equilibrium state and have minimum vertical
displacement when influenced by a runway excitation such as bump or cave. So the
maximum vertical displacement of airframe over a test runway is an important criterion for
shock absorber design. Another design criterion is root-mean-square (RMS) of airframe
vertical acceleration by reason that ground induced vibrations become more and more of a
problem as structures of modern aircraft become increasingly flexible. That will lead to
shorten the fatigue life of the landing gear and of the airframe.
The RMS of acceleration is defined as follows

t
1 e
C RMS  
t e  t 0 t0
( zs  zsr ) 2 dt (27)

with airframe’s vertical acceleration zs and reference value for acceleration evaluation zsr .

3.3 Transition From Touchdown to Taxiing

The damping required to successfully encounter oscillations has to considerably larger for
taxiing than for touchdown because the oleo stroke velocity at taxiing is significantly smaller
than at touchdown. So there exists a transition between touchdown phase and taxiing phase.
For passive landing gear, a standard solution is the use of a double-stage air spring or a taxi
valve. At low stroke velocities (taxiing), high damping factor is achieved, while at high
stroke velocity (touchdown), the valve reduces its damping factor. For our semi-active
landing gear, the transition from touchdown phase to taxiing phase is also monitored by
Nonlinear Predictive Control of Semi-Active Landing Gear 291

measurement of stroke velocity and the damping factor of the shock absorber is changed by
variable-sized oil orifice.

3.4 Dual Mode Controller

Due to totally different design goal of landing gear during aircraft touchdown phase and
taxiing phase, the semi-active controller should be able to switch from one mode to another.
Thus a dual mode predictive controller will be proposed in the following sections. Fig.5
shows the structure of the dual mode controller.

Fig. 5. Controller Switching Between Touchdown Phase and Taxiing Phase

4. Semi-Active Predictive Controller Design for Touchdown Phase

It is noted that the hydraulic dynamics, pneumatic dynamics and fast valve dynamics make
controls design very difficult. In order to achieve the ideal objective, a proper semi-active
control method should be applied. Considering the highly nonlinear behaviour of landing
gear, the classical linear control theory will be useless. The advances of nonlinear control
theory make it possible to transform certain types of nonlinear systems to linear system
(Slotine et al., 1991).

4.1 Inverse Dynamics Controller

The semi-active landing gear dynamic model (eq.16-25) can be simplified as a following
SISO nonlinear system:
x  f ( x )  g ( x )u (28)

y  c( x) (29)

Where, u is the system input which stands for actuator’s driving voltage V , y is the
system output which stands for the shock absorber force F .
292 Model Predictive Control

To deal with strong nonlinearities, generally an input-output linearization can be adopted

during the system synthesis process. The basic approach of input-output linearization is
simply differentiating the output function y repeatedly until the input u appears, and then
designing u to cancel the nonlinearity (Slotine et al., 1991). However, the nonlinearity
cancelling can not be carried out here because the relative degree of the semi-active landing
gear system is undefined,
Since the semi-active landing gear dynamic model consists of shock absorber’s model and
high-speed solenoid valve’s model, we propose a cascade nonlinear inverse dynamics
controller. First, an expected oil orifice area Ad for the shock absorber is directly computed
by inversion of nonlinear model if control valve’s limited magnitude and rate are omitted,

A03
Ad  (30)
2 x22Cd2 ( Fsao  Fair  K m Fair )

Then a nonlinear tracking controller for high-speed solenoid valve can be designed to follow
the expected movable parts position of solenoid valve. However, the practical actuator has
magnitude and rate limitations. The maximum adjustable open area of the valve is 7.4mm2
and switch frequency is 100Hz. So the optimal performance is not achievable.

Fig. 6. Shock Absorber Efficiency and Control Input Comparison w/o Input Constraints

From the above figures, we can see that the high-speed solenoid valve’s limited rate and
magnitude have negative effects on the shock absorber if those input constraints are not
considered during the controller synthesis process.

4.2 Nonlinear Predictive Controller

Model predictive control (MPC) is suitable for constrained, digital control problems. Initially
MPC has been widely used in the industrial processes with linear models, but recently some
researchers have tried to apply MPC to other fields like automotive and aerospace, and the
nonlinear model is used instead of linear one due to the increasingly high demands on
better control performance. However, optimization is a difficult task for nonlinear model
predictive control (NMPC) problem. Generally a standard nonlinear programming method
such as SQP is used. But it is the non-convex optimization method for constrained nonlinear
Nonlinear Predictive Control of Semi-Active Landing Gear 293

problem, thus global optimum can not be obtained. Furthermore, due to its high
computational requirement, SQP method is not suitable for online optimization.
To the semi-active landing gear control problem, a nonlinear output-tracking predictive
control approach (Lu, 1998) is adopted here considering its effectiveness to constrained
control problems and real-time performance. The basic principle of this control approach is
to get a nonlinear feedback control law by solving an approximate receding-horizon control
problem via a multi-step predictive control formulation.
The nonlinear state equation and output equation are defined by eq. (28-29). And the
following receding-horizon problem can be set up for providing the output-tracking control:

1 t T T
min J [ x (t ), t , u ]  min
u u 2 t
[e ( )Qe( )  u T ( ) Ru ( )]d (31)

subject to the state equations (28) and

e(t  T )  0 (32)

where e(t )  y (t )  yd (t ) .
Then we shall approximate the above receding-horizon control problem by the following
multi-step-ahead predictive control formulation. Define h  T / N , with N is control
number during the prediction horizon. The output y (t  kh) is approximated by the first-
order Taylor series expansion

y (t  kh)  y (t )  C [ x (t )]{ x (t  kh)  x (t )}, 1  k  N (33)

where C  c ( x ) / x . The desired output yd (t  kh) is predicted similarly by recursive

first-order Taylor series expansions

y d (t  h)  y d (t )  hy d (t )

y d (t  2 h)  y d (t  h)  hy d (t  h)  y d (t )  hy d (t )  h[ y d (t )  hyd (t )]

where another first-order expansion y d (t  h)  y d (t )  hyd (t ) , then we have

k 1
y d (t  kh)  y d (t )  h[ (1  hp ) i py d (t )] (34)
i 0

where p  d / dt is the differentiation operator. Combining the predictions of y (t  kh)

and yd (t  kh) , we obtain the prediction of the tracking error
294 Model Predictive Control

k 1
e(t  kh)  y (t  kh)  y d (t  kh)  h{ [C ( I  hF ) i ] f
i 0
(35)
k 1
  [C ( I  hF ) gu (t  (k  1  i )h)  (1  hp ) py d (t )]}
i i

i 0

where F ( x )  f ( x) / x . Approximating the cost function by the trapezoidal rule, it can be

written as a quadratic function

1
J  v T H ( x)v  r T ( x)v  q(e, x, yd ) (36)
2

where v  col{u (t ), u (t  h),..., u[t  ( N  1) h]} .

The constraint (eq. (32)) is then expressed as e(t  Nh )  0 which leads to

M T ( x )v  d (e, x, y d ) (37)
where
M T  C[( I  hF ) N 1 g, , ( I  hF ) g, g ] (38)

1 N 1
d   e  [C ( I  hF ) i f  (1  hp) i pyd (t )] (39)
h i 0

Now the output-tracking receding-horizon optimal control problem is reduced to the

problem of minimizing J with respect to v subject to eq. (37), which is a quadratic
programming problem. The closed-form optimal solution for this problem is

v  [ H 1  H 1 M ( M T H 1 M ) 1 M T H 1 ]r  [ H 1 M ( M T H 1 M ) 1 ]d (40)

Then the closed-loop nonlinear predictive output-tracking control law is

u (t ; x , N )  v (1) (41)

Unlike the input-output feedback linearization control laws, the existence of the proposed
nonlinear predictive output-tracking control does not depend on the requirement that the
system have a relative degree. And more important, the actuator’s amplitude and rate
constraints can be taken into account during the controller synthesis process.

4.3 Numerical Simulation

Based on the analysis described in previous sections, the numerical simulation of the semi-
active landing gear system responses are derived using MATLAB environment. The
prototype of the simulation model is a semi-active landing gear comprehensive
Nonlinear Predictive Control of Semi-Active Landing Gear 295

experimental platform we built, which can be reconfigured to accomplish tasks such as drop
tests, taxi tests and shimmy tests. The sprung mass of this system is 405kg and the
unsprung mass is 15kg. The other parameters of the simulation model can be found in (Wu
et al, 2007). Fig.7 is the photo of the experiment system.

Fig. 7. Landing gear experiment platform

Three kinds of control methods including passive control, inverse dynamics semi-active
control and nonlinear predictive semi-active control are used in the computer simulation.
The fixed size of oil orifice for passive control is optimized manually under following
parameters: sinking speed is 2 m/s and aircraft sprung mass is 405 kg. In the process of
simulation, the sprung mass remains constant and the comparison is taken in terms of
different sinking speed: 1.5 m/s, 2 m/s and 2.5 m/s. For passive control, the orifice size is
fixed. From the Figs. 8-10 and Table 1, when system parameters such as sinking speed
change, the control performance of the passive control decreases greatly, for the fixed orifice
size in passive control is designed under standard condition.

Fig. 8. Efficiency Comparison under Normal Condition

296 Model Predictive Control

Conventional passive landing gear is especially optimized for heavy landing load condition,
so the passive landing gear behaves even worse under light landing load condition. The
performance of semi-active control is superior to that of passive one due to its tunable
orifice size and nonlinear predictive semi-active control method has the best performance of
all. Due to its continuous online compensation and consideration of actuator’s constraints,
nonlinear predictive semi-active control method can both increase the efficiency of shock
absorber and make the output smoother during the control interval, which can effectively
alleviate the fatigue damage of both airframe and landing gear.

Fig. 9. Efficiency Comparison under Light Landing Load Condition

Fig. 10. Efficiency Comparison under Heavy Landing Load Condition

Control Method Passive Semi-Active IDC Semi-Active Predictive

Efficiency/( 2.0m  s
1
) 0.8483 0.8788 0.9048

Efficiency/( 1.5m  s )
1 0.8449 0.8739 0.9036
1
Efficiency/( 2.5m  s ) 0.8419 0.8554 0.8813
Table 1. Comparison of shock absorber efficiency
Nonlinear Predictive Control of Semi-Active Landing Gear 297

4.4 Sensitivity Analysis

Sometimes system parameters such as sinking speed, sprung weight and attitude of aircraft
at touch down may be measured or estimated with errors, which will lead to bias of
estimation for optimal target load. But the controller should behave robust to withstand
certain measurement or estimation errors within reasonable scope so that the airframe will
not suffer from large vertical load at touch down.
Simulation of sensitivity analysis is conducted under the standard condition controller
design: sinking speed is 2 m/s and aircraft sprung mass is 405 kg, introducing 10% errors
for sinking speed and sprung mass individually. The actual sinking speed is measured by
avionic equipments and the aircraft sprung mass is estimated by considering the weights of
oil, cargo and passengers. The measurement and estimation errors will be less than the
assumed maximal one.
From the above Figs.11,12 simulation results, it can seen that the reasonable measuring error
of sinking speed has little effect on the performance of nonlinear predictive semi-active
controller, whilst estimating error of sprung mass has side effect to the control performance
and shock absorber efficiency decreases a little. To further improve the performance under
mass estimating error, it is possible to either simply introduce measurement of aircraft mass
or develop robust controller which is non-sensitive to estimating the error of aircraft sprung
mass.

Fig. 11. Sensitivity to sink speed measuring error

Fig. 12. Sensitivity to sprung mass estimating error

298 Model Predictive Control

5. Semi-Active Predictive Controller Design for Taxiing Phase

In this section, we will propose a nonlinear predictive controller incorporating radial basis
function network (RBF) and backstepping design methodology (Kristic et al., 1995) for semi-
active controlled landing gear during aircraft taxiing.

5.1 Hierarchical Controller Structure

A hierarchical control structure which contains three control loops is adopted here. The
outer loop determines the expected strut force of the semi-active shock absorber. At
touchdown phase and taxiing phase, the computation of the expected strut force will be
different due to different design objective. The middle loop is responsible for controlling of
solenoid valve’s mechanical and magnetic dynamics. The high speed solenoid valve
contains high nonlinearity and can not be regulated by traditional linear controller i.e. PID.
We develop a RBF network to approximate the nonlinear dynamics which can not be
precisely modelled and adopt backstepping, a constructive nonlinear control design method
to stabilize the whole nonlinear system. The inner loop is the current loop. It ensures stable
tracking of commanded current that middle loop outputs.

Fig. 13. Hierarchical Controller Structure

5.2 Background for RBF network

A RBF network is typically comprised of a layer of radial basis activation functions with an
associated Euclidean input mapping. The output is then taken as a linear activation function
with an inner product or weighted average input mapping.
In this paper, we use a weighted average mapping in the output node. The input-output
relationship in a RBF with x  [ x1 ,, xn ]T as an input is given by

 w exp( x  c /  )  θ ξ(x)
m 2 2
i 1 i i T
 (x, θ)  (42)
 exp( x  c /  )
m 2 2
i 1 i

where
Nonlinear Predictive Control of Semi-Active Landing Gear 299

θ  [ w1 ,, wn ]T (43)

2
exp( x  ci /  2 ) (44)
i 
i1 exp( x  ci /  2 )
m 2

The RBF network is a good approximator for general nonlinear function. For a nonlinear
function FN, we can express it using RBF network with the following form,

~
FN  θT ξ    θˆ T ξ  θ T ξ   (45)

where θ is the vector of tunable parameters under ideal approximation condition, θ̂ under
~
practical approximation condition, θ parameter approximation error, ε function
reconstruction error.

5.3 Outer Loop Design

The function of the outer control loop is to produce a target strut force for semi-active shock
absorber by using active control law. Then middle loop and inner loop controller will be
designed to approximate the optimal performance that active controller achieves.

(a) Skyhook Controller

At the taxiing phase, the landing gear system acts like the suspension of ground vehicle. So
we first adopt the most widely used active suspension control approach – the skyhook
controller. At this control scheme the actuator generates a control force which is
proportional to the sprung mass vertical velocity. The equation of skyhook controller can be
expressed as the following form:

Fsky  K sky ( x1d  x1 )  Csky ( x2  x4 ) (46)

In order to blend out low frequency components of the vertical velocity signal which results
from the aircraft taxiing on sloped runways or long bumps, we modify it by adding high
pass filter to the skyhook controller.

s
xs  x1 (47)
s  wk

where wk is roll off frequency of high pass filter. Thus we get the desired strut force.

Fd  K sky ( x1d  x1 )  C sky ( x2  x4 )  K HP xs (48)

where K HP is a constant scale factor.

300 Model Predictive Control

(b) Nonlinear Predictive Controller

Compare with traditional skyhook controller, model predictive controller is more suitable
for constrained nonlinear system like landing gear system or suspension system. Input and
state constraints can be incorporated into the performance index to achieve best
performance.
The system model of outer loop controller is eq. (16-19), which can be expressed as follows:

x a  f ( x a )  g ( x a ) Fd (49)

where xa  [ x1 , x2 , x3 , x4 ] , Fd is the control input and the output equation is y  x1 . Then

a similar receding-horizon problem can be set up for providing the output-tracking control:

1 t T T
min J [ x a (t ), t , Fd ]  min
Fd Fd 2 t
[ea ( )Qa ea ( )  FdT ( ) Ra Fd ( )]d (50)

subject to the state equations (49) and

ea (t  T )  0 (51)

where ea (t )  x1 (t )  x1d (t ) .
Following a similar synthesis process as in section 4.2, we can get a closed-loop nonlinear
predictive output-tracking control law to achieve approximate optimal active control
performance.

5.4 RBF-based Backstepping Design (Middle Loop)

In this section we propose a RBF-based backstepping method to complete the design of the
semi-active controller. Stability proofs are given.
First we define the force tracking error as e1  Fd  F . Differentiate and substitute from Eq.
(16-25),
d
e1  Fd  F  Fd  [(1  Km )Fair  Foil ]
dt
d  Ao 3
 Fd  ( )x2 2 x 5
dx5 2(K x C )2 1  (K x / A )2
v 5 d v 5 o

 Ao 3
 x2 x 2
(K v x5C d )2 1  (K v x5 / Ao )2
d V0
(1  K m ) [ Pi Aa ( )n  P0 Aa ]x 1
dx1 V0  Aa x1
 Fd  G1 ( x2 , x5 )x6  H 1 ( x1 , x2 , x3 , x 4 , x5 )
where G1 ( x2 , x5 ) , H ( x1 , x2 , x3 , x4 , x5 ) is the nonlinear functions related to the strut
dynamics.
Nonlinear Predictive Control of Semi-Active Landing Gear 301

(a) First Step

Select the desired solenoid valve movable part velocity as

x6 d  G11 ( H1  Fd  k1e1 ) (52)

where k1 is a design parameter. Then we get

e1  Fd  G1(x2 , x5 )x6d  G1(x2, x5 )e2  H1(x1, x2 , x3, x4, x5 )

 F  G e  (H  F  k e )  H  G e  k e
d 1 2 1 d 1 1 1 1 2 1 1

where e2  x6  x6 d .
Consider the following Lyapunov function candidate

1 2
V1  e1
2
Differentiate V1 , thus we get
V1  e1e1  e1Ge2  k1e12

(b) Second Step

x6 is not the true control input. We then choose u  x72 as virtual input.
Differentiate e2 , we get
Cs
e2  x6  x6d  G2u  x6  H 2  W
mv

where G2  B( x5 ) / mv , H 2   f / mv  x6d and W  [(K s  K f ) x5  f 0 ] / mv .

Consider the following Lyapunov function candidate

1 1 ~ ~ 1 ~ ~
V2  V1  e22  tr ( θ1T Γ11θ1 )  tr ( θ2T Γ 21θ2 )
2 2 2

where 1 and 2 are positive definite matrices. Differentiate V2

C ~ ~ ~ ~
V2  V1  e2 (G2u  s x6  H2  W )  tr(θ1T Γ11θ1)  tr(θ2T Γ21θ2 )
mv

Then we choose the control input:

302 Model Predictive Control

C
u  Gˆ 21 ( Hˆ 2  s x6 d  W  k 2e2  G1e1 ) (53)
mv

where k 2 is a design parameter, Gˆ 2  θˆ 1 ξ1 is the estimation of G 2 ( x 2 , x5 ) , Hˆ 2  θˆ 2T ξ 2 is the

T

estimation of H ( x1 , x2 , x3 , x4 , x5 ) . Thus we get

~ C ~ ~ ~ ~ ~
V2  V1  e2G2u  e21u  s e2  e2 H2  e2 2  e2k2e2  e2G1e1  tr(θ1T Γ11θ1)  tr(θ2T Γ21θ2 )
mv
~ ~ ~ ~ C
 V1  e2θ1ξ1u  e2θ2ξ2  θ1ξ1ue2  θ2ξ2e2  e2k2e2  s e2  e21u  e2 2  e2G1e1
mv
~  ~ 
 tr[θ1T (Γ11θˆ 1  ξ1ue2 )]  tr[θ2T (Γ21θˆ 2  ξ2e2 )]

Choose the tuning law as:

 
θˆ 1   Γ1ξ1ue2 , θˆ 2  Γ 2 ξ 2 e2 (54)

So we have

C
V2  e1G1e2  k1e12  k2e22  e2G1e  s e2  e21u  e221
mv
1u  2  Cs / mv (1u  2  Cs / mv )2
 k1e12  k2 (e2  )2  0
2k2 4k2

Therefore, the system is stable and the error will asymptotically converge to zero.

5.5 Inner Loop Design

The function of the inner loop is to precisely tracking of solenoid valve’s current. We apply a
simple proportional control to the electrical dynamics as follows

V  K c ( x7 d  x7 )  K c ( u  x7 ) (55)

where K c is the controller gain.

The above three control loops represent different time scales. The fastest is the inner loop
due to its electrical characteristics. The next is the middle loop. It is faster than the outer loop
because the controlled moving part’s inertial of the middle loop is much smaller than that of
the outer loop.
Nonlinear Predictive Control of Semi-Active Landing Gear 303

5.6 Numerical Simulation

After touchdown, the taxiing process will last relatively a long time before aircraft stops. To
simulate the road excitation of runway and taxiway, a random velocity excitation signal
w(t ) is introduced into Eq. (18).
x3  x4  w(t ) (56)

The simulation result is compared using airframe vertical displacement, which is one of the
most important criterion for taxiing condition. Due to lack of self-tuning capability, the
passive landing gear does not behave well and passes much of the road excitation to the
airframe. That will be harmful for the aircraft structure and meanwhile make passages
uncomfortable. The proposed semi-active landing gear effectively filters the unfriendly road
excitation as we wish.

Fig. 14. System Response Comparison under Random Input

From the simulation results of both aircraft touch-down and taxiing conditions, we can see
that the proposed semi-active controller gives the landing gear system extra flexibility to
deal with the unknown and uncertain external environment. It will make the modern
aircraft system being more intelligent and robust.

6. Conclusion
The application of model predictive control and constructive nonlinear control methodology
to semi-active landing gear system is studied in this paper. A unified shock absorber
mathematical model incorporates solenoid valve’s electromechanical and magnetic
dynamics is built to facilitate simulation and controller design. Then we propose a
hierarchical control structure to deal with the high nonlinearity. A dual mode model
predictive controller as an outer loop controller is developed to generate the ideal strut force
on both touchdown and taxiing phase. And a systematic adaptive backstepping design
method is used to stabilize the whole system and track the reference force in the middle and
inner loop. Simulation results show that the proposed control scheme is superior to the
traditional control methods.
304 Model Predictive Control

7. References
Batterbee, D.; Sims, N. & Stanway, R. (2007). Magnetorheological landing gear: 1. a design
methodology. Smart Materials and Structures, Vol. 16, pp. 2429-2440.
Ghiringhelli, L. G. (2000). Testing of semiactive landing gear control for a general aviation
aircraft. Journal of Aircraft, Vol.7, No.4, pp.606-616.
Ghiringhelli, L. G. & Gualdi, S. (2004). Evaluation of a landing gear semi-active control
system for complete aircraft landing. Aerotecnica Missili e Spazio, No. 6, pp. 21-31.
Hyochoong, B.& Choong-Seok, O. (2004). Predictive control for the attitude maneuver of a
flexible spacecraft. Aerospace Science and Technology, Vol. 8, No. 5, pp. 443-452.
Karnopp, D. (1983). Active damping in road vehicle suspension systems. Vehicle System
Dynamics, Vol.12, No. 6, pp.291-316.
Kristic, M.; Kanellakopoulos, I. & Kokotovic, P.V. (1995). Nonlinear and Adaptive Control
Design, Wiley-Interscience, ISBN: 978-0-471-12732-1, USA.
Krüger，W. (2000). Integrated design process for the development of semi-active landing
gears for transport aircraft, PhD thesis, University of Stuttgart.
Liu, H.; Gu, H. B. & Chen, D. W. (2008). Application of high-speed solenoid valve to the
semi-active control of landing gear. Chinese Journal of Aeronautics, Vol.21, No.3,
pp.232-240.
Liu, H. & Gu, H. B. (2008). Nonlinear model for a high-speed solenoid valve and its
simulation. Mechanical Science and Technology for Aerospace Engineering, Vol.27, No.7,
pp.866-870.
Lou, Z.; Ervin, R. & Winkler, C. (1993). An electrorheologically controlled semi-active
landing gear. SAE Paper931403.
Lu, P. (1998). Approximate nonlinear receding-horizon control laws in closed form.
International Journal of Control, Vol. 71, No.1, pp.19-34.
Maemori, K.; Tanigawa, N. & Koganei R (2003). Optimization of a semi-active shock
absorber for aircraft landing gear, Proceedings of ASME Design Engineering Technical
Conference, pp.597-603.
Malaguti, F. & Pregnolato, E. (2002). Proportional control of on/off solenoid operated
hydraulic valve by nonlinear robust controller, Proceedings of IEEE International
Symposium on Industrial Electronics, pp.415-419.
Mehra, R. K.; Amin, J. N. & Hedrick, K. J. (1997). Active suspension using preview
information and model predictive control, Proceedings of the 1997 IEEE International
Conference on Control Applications, pp. 860-865.
Michael, A. H. (1998). Nonlinear model predictive control: current status and future
direction. Computers and Chemical Engineering, Vol.23, pp.187-202.
Mikulowski, G. & LeLetty, R. (2008). Advanced landing gears for improved impact
absorption, Proceedings of the 11th International conf. on New Actuators, pp.175-194.
Slotine, J.E. & Li, W.P. (1991). Applied Nonlinear Control, Prentice Hall, ISBN: 0-13-040890-5,
USA.
Wang, X. M. & Carl U. (1999). Fuzzy control of aircraft semi-active landing gear system,
AIAA 37th Aerospace Sciences Meeting and Exhibit.
Wu, D. S.; Gu, H.B.; Liu, H. (2007). GA-based model predictive control of semi-active
landing gear. Chinese Journal of Aeronautics, Vol.20, No.1, pp.47-54.
Yadav, D. & Ramamoorthy, R. P. (1991). Nonlinear landing gear behavior at touchdown.
Journal of Dynamic Systems, Measurement and Control, Vol.113, No.12, pp.677-683.