N MPC 56 Passivity

Download as pdf or txt
Download as pdf or txt
You are on page 1of 12

Nonlinear Model Predictive Control: A

Passivity-Based Approach

Tobias Raff, Christian Ebenbauer, and Frank Allgöwer

Institute for Systems Theory and Automatic Control, University of Stuttgart,


Pfaffenwaldring 9, 70550 Stuttgart, Germany
{raff,ce,allgower}@ist.uni-stuttgart.de

Summary. This paper presents a novel approach for nonlinear model predictive con-
trol based on the concept of passivity. The proposed nonlinear model predictive control
scheme is inspired by the relationship between optimal control and passivity as well as
by the relationship between optimal control and model predictive control. In particular,
a passivity-based state constraint is used to obtain a nonlinear model predictive control
scheme with guaranteed closed loop stability. Since passivity and stability are closely
related, the proposed approach can be seen as an alternative to control Lyapunov func-
tion based approaches. To demonstrate its applicability, the passivity-based nonlinear
model predictive control scheme is applied to control a quadruple tank system.

1 Introduction

In this paper, nonlinear model predictive control is used to control unconstrained


nonlinear systems. Due to the fact that this control strategy does not naturally
guarantee closed loop stability [1], many different approaches have been devel-
oped to circumvent this problem [11]. Most of these nonlinear model predictive
control schemes achieve closed loop stability by using the concept of Lyapunov
stability, e.g., in the model predictive control setup a control Lyapunov function
is included as a state constraint [16], or as a terminal cost [6], or as a terminal
cost in conjunction with a terminal region [3] to achieve stability.
A nonlinear model predictive control scheme based on the relationships be-
tween optimal control, nonlinear model predictive control, and control Lyapunov
function was developed in [16]. In this paper, a nonlinear model predictive con-
trol scheme is developed based on the concept of passivity in this spirit, i.e., the
proposed nonlinear model predictive control scheme is inspired by the relation-
ship between optimal control and passivity [8, 12] as well as by the relationship
between optimal control and nonlinear model predictive control [16]. The rela-
tionship between optimal control and passivity is that an input affine nonlinear
system can be optimal if and only if it satisfies a passivity property with respect
to the optimal feedback [8, 12]. Optimal control and nonlinear model predictive
control are linked by the fact that model predictive control is a computational
tractable approximation of the optimal control problem by repeatedly solving
on-line a finite horizon optimal control problem. Based on these relationships it

R. Findeisen et al. (Eds.): Assessment and Future Directions, LNCIS 358, pp. 151–162, 2007.
springerlink.com 
c Springer-Verlag Berlin Heidelberg 2007
152 T. Raff, C. Ebenbauer, and F. Allgöwer

is shown that passivity-based concepts and nonlinear model predictive control


can be merged into a nonlinear model predictive control scheme which main-
tains the advantages of the individual concepts, i.e., closed loop stability due
passivity and good performance due to one-line optimization in nonlinear model
predictive control.
The remainder of the paper is organized as follows: In Section 2, the necessary
background of passivity is given. Furthermore, the connection between passivity
and optimal control is briefly reviewed which is the underlying motivation of
the passivity-based nonlinear model predictive control scheme. In Section 3,
nonlinear model predictive control is introduced and a short overview of different
model predictive control schemes with guaranteed stability is given. In Section 4,
the main result is presented and discussed, namely, a passivity-based nonlinear
model predictive control scheme. To demonstrate its applicability, the proposed
nonlinear model predictive control scheme is applied to control a quadruple tank
system in Section 5. Finally, conclusions are drawn in Section 6.

2 Passivity
The concept of passivity is often used in the analysis and synthesis of nonlinear
systems [10, 17]. In the following, the necessary background, adjusted for the
paper, is given. Consider the nonlinear system given by

ẋ = f (x) + g(x)u
(1)
y = h(x),

where x ∈ Rn is the state, u ∈ Rm is the input and y ∈ Rm the output. It is


assumed that all data is locally Lipschitz and that x = 0, u = 0 is an equilibrium
point. The system (1) is said to be passive if there exists a positive semidefinite
storage function S such that the inequality

/t1
S(x(t1 )) − S(x(t0 )) ≤ uT (t)y(t)dt (2)
t0

is satisfied for all t0 ≤ t1 when (u(t), x(t), y(t)) satisfy the system dynamics (1).
The definition of passivity is motivated by the following consideration. In the
context of electrical network theory S can be considered as the energy stored in
the network, u as the port voltage, and y as the port current. Passivity of such
a network means that it cannot supply more energy to its environment than
energy was supplied to the network. If S is differentiable as a function of time
then inequality (2) can be written as

Ṡ(x(t)) ≤ uT (t)y(t), (3)

which is often a more useful notion for analyzing passive systems than inequality
(2). A further characterization of passive systems is also possible in terms of
Nonlinear Model Predictive Control: A Passivity-Based Approach 153

relative degree and minimum phase property. In case the system (1) has a well-
defined normal form, it must be weakly minimum phase and must have a vector
relative degree of one. Of importance in the concept of passivity is also the
relationship between passivity and stabilization, which is summarized next.

2.1 Passivity and Stabilization


The relationship between passivity and stabilization can be established by us-
ing the storage function S as a control Lyapunov function. However, the concept
of passivity requires only a positive semidefinite storage function S [17, 18]. To
achieve stability, one has to assume in addition that the system (1) is zero-state
detectable [10, 17], i.e., the solution of ẋ = f (x) satisfies limt→∞ x(t) = 0 for
y(t) = 0 for t ≥ 0. Under these assumptions, the system (1) can be stabilized
with the simple feedback u = −y [17] due to

Ṡ(x(t)) ≤ uT (t)y(t) ≤ −y T (t)y(t) ≤ 0. (4)

Furthermore, if S is radially unbounded, and all solutions of (1) are bounded,


then the system (1) is globally stabilized by the feedback u = −y. At first glance,
it seems that the concept of passivity is restrictive for stabilization purposes.
However, in case of optimal control, the concept of passivity appears in a natural
way.

2.2 Passivity and Optimality


The infinite horizon optimal control problem is defined by
/∞


V (x0 ) = min q(x(τ )) + uT (τ )u(τ ) dτ


u(·) (5)
0
s.t. ẋ = f (x) + g(x)u, x(0) = x0 ,

where q is a positive semidefinite function and V  is the value function, i.e., the
minimal cost of (5). The optimal feedback u which stabilizes the system (1)
and minimizes the performance index (5) is given by

1 ∂V  T
u = −k  (x) = − g T (x) , (6)
2 ∂x
under the assumption that V  is the positive semidefinite and continuously dif-
ferentiable solution of the Hamilton-Jacobi-Bellman equation

∂V  1 ∂V  ∂V  T
f (x) − g(x)g T (x) + q(x) = 0. (7)
∂x 4 ∂x ∂x
Note, that it is in general very difficult to solve the equation (7) and therefore the
infinite horizon optimal control problem (5). The relationship between optimal
control and passivity can be established [8, 12, 17] by using the value function
154 T. Raff, C. Ebenbauer, and F. Allgöwer

as a storage function and the optimal feedback as an output of the system (1).
Then the feedback (6) stabilizes the system (1) and minimizes the performance
index (5) if and only if the system

ẋ = f (x) + g(x)u (8)


y = k  (x)

is zero-state detectable and output feedback passive, i.e., Ṡ(x(t)) ≤ uT (t)y(t) +


1 T 1 
2 y (t)y(t) with S = 2 V [17]. This equivalence becomes clear considering that
the system (8) is stabilized by the feedback u = −y.

3 Nonlinear Model Predictive Control


Since stabilization of nonlinear systems (1) subject to the infinite horizon optimal
problem (5) is difficult due to solving equation (7), computationally tractable
approximations of (5) have been developed. One approximation approach is non-
linear model predictive control. The basic idea of nonlinear model predictive
control is to determine the control input by repeatedly solving on-line the finite
horizon optimal control problem

/
t+T

min q(x(τ )) + uT (τ )u(τ ) dτ


u(·) (9)
t
s.t. ẋ = f (x) + g(x)u.

However, this control strategy does not naturally guarantee closed loop stability
[1]. To overcome this problem, several nonlinear model predictive control schemes
have been developed which achieve closed loop stability [3, 6, 11, 14, 15, 16]. In
the following, some approaches are summarized with respect to their stability
conditions. For a detailed and more rigorous treatment of nonlinear model pre-
dictive control, see for example [11] and the references quoted therein. In [11] it
was shown that most nonlinear model predictive control schemes with guaran-
teed stability can be summarized in the setup

/
t+T

min ϕ(x(t + T )) + q(x(τ )) + uT (τ )u(τ ) dτ


u(·)
t (10)
s.t. ẋ = f (x) + g(x)u,
x(t + T ) ∈ W,

where ϕ is a terminal cost and W a terminal6region. In [6, 14] it was shown that

with the setup W = Rn and ϕ(x(t + T )) = t+T (q(x(τ )) + k T (x(τ ))k(x(τ )))dτ
closed loop stability can be achieved, where u = −k(x) is a locally stabilizing
feedback of the system (1) and x is the closed loop state trajectory with the
feedback u = −k(x). In this approach, ϕ is the cost of stabilizing the system (1)
Nonlinear Model Predictive Control: A Passivity-Based Approach 155

with the feedback u = −k(x) over the time interval [t, ∞). Another possibility
is to choose W = Rn and ϕ(x(t + T )) = V (x(t + T )), where V is a control Lya-
punov function of the nonlinear system (1) [6]. In this approach, the cost-to-go
is approximated by the control Lyapunov function V . It is further possible to
achieve stability by using a terminal region W and no terminal cost, i.e., ϕ = 0.
Using this idea, it is possible to achieve stability by imposing the terminal state
constraint x(t + T ) = 0, i.e., W = {0} [9]. Since this approach is computation-
ally demanding, a relaxed setup was developed in [15]. In [15] it was shown that
closed loop stability is achieved by steering the final system state x(t + T ) in
the terminal region W and by stabilizing the system with a locally stabilizing
feedback u = −k(x) inside the terminal region W . Other approaches with guar-
anteed stability use both a terminal cost ϕ and a terminal region W [3]. For
example, the approach developed in [3] uses W as a level set of the terminal cost
ϕ(x(t + T )) = 12 xT (t + T )P x(t + T ), where 12 xT P x is a local control Lyapunov
function of the linearized system of (1). Finally, another approach, using a form
different from (10) but which is in the same line as the approach presented in
Section 4, is given by

/
t+T

min q(x(τ )) + uT (τ )u(τ ) dτ


u(·)
t
s.t. ẋ = f (x) + g(x)u, (11)
∂V
[f (x) + g(x)u] < 0
∂x
x(t + T ) ∈ W,

where V is a control Lyapunov function for the system (1) [16]. In this approach,
closed loop stability is achieved by requiring that the derivate of the control
Lyapunov function V is negative along the state trajectory of the closed loop
system. Furthermore, the terminal region W is used a performance constraint to
recover the optimal controller in case the level curves of the control Lyapunov
function V correspond to the level curves of the optimal value function V  .
In summary, all nonlinear model predictive schemes summarized above share
one common property, namely, the stability is achieved by using a control Lya-
punov function in their setups. Of course, as shown in [16], there is a strong
relationship between optimal control, nonlinear model predictive control, and
control Lyapunov functions. However, as reviewed in Section 2, there is also a
strong relationship between optimal control and passivity. Based on this second
relationship, a nonlinear model predictive control scheme is developed in the
next section.

4 Passivity-Based Nonlinear Model Predictive Control


In this section, the passivity-based nonlinear model predictive control scheme is
introduced. Inspired by the idea of combining control Lyapunov functions and
156 T. Raff, C. Ebenbauer, and F. Allgöwer

nonlinear model predictive control based on their relationship to optimal control


[16], passivity is merged with nonlinear model predictive control in the same
spirit. Suppose that the system (1) is passive with a continuously differentiable
storage function S and zero-state detectable. Then the passivity-based nonlinear
model predictive control scheme for the system (1) is given by

/
t+T

min q(x(τ )) + uT (τ )u(τ ) dτ


u(·)
t
s.t. ẋ = f (x) + g(x)u (12)
y = h(x)
uT (t)y(t) ≤ −y T (t)y(t).

The passivity-based state constraint in the last line in the setup (12) is moti-
vated by the fact that in case the system (1) is passive and zero-state detectable,
it can be stabilized with the feedback u = −y. Hence, the passivity-based state
constraint is a stability constraint which guarantees closed loop stability. Fur-
thermore, if the storage function S is radially unbounded, and all solutions of
the system are bounded, then the closed loop system is globally asymptotically
stable. In contrast to many other nonlinear model predictive control schemes [11]
which achieve stability by enforcing a decrease of the value function along the
solution trajectory, the stability of the proposed nonlinear model predictive con-
trol scheme is achieved by using directly a state constraint. Hence, one obtains
the following stability theorem of the passivity-based nonlinear model predictive
control scheme (12):
Theorem 1. The passivity-based nonlinear model predictive control scheme (12)
locally asymptotically stabilizes the system (1) if it is passive with a continously
differentiable storage function S and zero-state detectable.

Proof. The proof of Theorem 1 is divided into two parts. In the first part it is
shown that the nonlinear model predictive control scheme (12) is always feasible.
In the second part it is then shown that the scheme (12) asymptotically stabilizes
the system (1).
Feasibility: Feasibility is guaranteed due to the known stabilizing feeback u =
−y. Stability: Let S be the storage function of the passive system (1). With the
differentiable storage function S and the state constraint in the model predictive
control scheme (12), one obtains

Ṡ(x(t)) ≤ uT (t)y(t) ≤ −y T (t)y(t).

Using the fact that the system (1) is zero-state detectable, the same arguments
presented in Theorem 2.28 of [17] can be used in order to show asymptotic sta-
bility of the origin x = 0. Hence, the passivity-based nonlinear model predictive
control scheme (12) asymptotically stabilizes system (1) if it is passive and zero-
state detectable.
Nonlinear Model Predictive Control: A Passivity-Based Approach 157

At the first glance, it seems that the nonlinear model predictive scheme (12)
is very restrictive since it is only applicable for passive systems. However, for
stabilization purposes, no real physical output y is needed. Instead it is enough
to know one fictitious output η = σ(x) in order to stabilize the system. Once
such a fictitious output η is known, the fictitious passive system
ẋ = f (x) + g(x)u
(13)
η = σ(x)
can be stabilized with the passivity-based scheme (12). Note that a fictitious
output η always exists, as long as a control Lyapunov function exists. Since then,
by definition, Lg V (x) is a fictitious output. Unfortunately, there is no efficient
way to construct a passive output. However, it is often possible to find a fictitious
passive output since passivity is physically motivated concept. Furthermore, if
the linearized system ẋ = Ax+Bu of (13) is stabilizable, a passive output for the
linearized system is given by σ(x) = B T P x, where P is the solution of the Riccati
equation AT P +P A+Q−P BB T P = 0. This may help to find a fictitious passive
output for local stabilization purposes. In the following, another property of the
nonlinear model predictive control scheme (12) is discussed. Namely, the scheme
(12) recovers the passivity-based feeback u = −y as the prediction horizon T
goes to zero. This property is summarized in the next theorem.
Theorem 2. The passivity-based nonlinear model predictive control scheme re-
covers the passivity-based feedback u = −y for T → 0.
Proof. To show this property, the same arguments are used as in [16]. By rewrit-
6 t+T

ing the the performance index to T1 t q(x(τ )) + uT (τ )u(τ ) dτ the optimiza-


tion problem is not changed. For T → 0 one obtains q(x(t)) + uT (t)u(t). Since
the system state x(t) at time instant t is fixed, one finally obtains

min uT (t)u(t)
u(t)

s.t. ẋ = f (x) + g(x)u (14)


y = h(x)
uT (t)y(t) ≤ −y T (t)y(t).

Based on the concept of the pointwise min-norm controller [5], the state con-
straint will be always active in order to minimize the control input u and the
resulting feedback is therefore the passivity-based feedback u = −y.

If the value function V  of the infinite horizon optimal control problem (5) is
known and the output y is set to be equal to y = k  (x), then the optimal
performance is recovered for T → 0, which is summarized below:
Corollary 1. The optimal performance of the infinite horizon optimal control
problem (5) is recovered by the passivity-based model predictive control scheme
 T
(12), if y = 12 g T (x) ∂V
∂x and T → 0, where V  is the value function of the
infinite horizon optimal control problem (5).
158 T. Raff, C. Ebenbauer, and F. Allgöwer

Proof. Proof follows from Theorem 2.

Hence, for T → 0 the passivity-based feedback u = −y is obtained and for any


horizon T > 0 one can expect a better closed loop performance with the approach
(12) than with the passivity-based feedback u = −y due to on-line optimization
in nonlinear model predictive control. Summarizing, the passivity-based nonlin-
ear model predictive control combines passivity with model predictive control
while maintaining the advantages of the individual concepts, i.e., stability due
to passivity and good performance due to one-line optimization in nonlinear
model predictive control. Furthermore, the passivity-based state constraint was
incorporated in the model predictive control setup (12) in such a way that the-
oretically interesting properties are obtained, e.g., stability due to passivity and
recovery of the passivity-based feedback u = −y for T → 0. Finally, the underly-
ing idea of of the approach (12) are the relationships of passivity and nonlinear
model predictive control to optimal control.

4.1 Extension of the Passivity-Based Nonlinear Model Predictive


Control Scheme
In Corollary 1 it was shown that the passivity-based nonlinear model predictive
control scheme (12) can recover the optimal feedback u = −k  (x) for T → 0 in
case the optimal feeback is available. However, nothing can be said about the
performance for a prediction horizon T > 0. The reason for this fact is that
the proposed approach (12) does not take into account the cost-to-go. Many
nonlinear model predictive control schemes [6, 11, 14, 16] approximate the cost-
to-go in order to improve the performance. As seen in Section 3, there exists
many approaches which incorporate an approximation of the cost-to-go in the
nonlinear model predictive control setup. These approaches can be also used
in order to improve the performance of the passivity-based nonlinear model
predictive controller for any prediction horizon T > 0. One possibility is to
introduce a terminal cost ϕ. By introducing a terminal cost in (12), the passivity-
based nonlinear model predictive control scheme becomes
/
t+T

min ϕ(x(t + T )) + q(x(τ )) + uT (τ )u(τ ) dτ


u(·)
t
s.t. ẋ = f (x) + g(x)u (15)
y = h(x)
uT (t)y(t) ≤ −y T (t)y(t).
One possibility is to choose the terminal cost as the storage function of the system
(1), i.e., ϕ(x(t+T )) = S(x(t+T )). In case the optimal value function V  is known
and used as a storage function of the system (1), the optimal performance can be
recovered for any prediction horizon T > 0. Note, in case the value function V 
is not available, stability is achieved due to the passivity-based state constraint,
since the storage function of the system (1) is in general not positive definite and
Nonlinear Model Predictive Control: A Passivity-Based Approach 159

can therefore not be used as a control Lyapunov


6∞ function. Another possibility
based on [6, 14] is to choose ϕ(x(t + T )) = t+T (q(x(τ )) + hT (x(τ ))h(x(τ )))dτ ,
where u = −h(x) is the stabilizing feedback of the system (1) and x is the
corresponding closed loop system trajectory.

5 Example

In this section, the passivity-based nonlinear model predictive scheme is applied


to control a quadruple tank system. The dynamical model of the quadruple tank
systems [7] is given by

ẋ = f (x) + g(x)u,

with
⎡ √ a3 √ ⎤
−A
a1
2gx1 + A 2gx3
√ 1 1

⎢− a2 2gx2 + a4 2gx4 ⎥
⎢ ⎥
f (x) = ⎢ A2 a3 √
A2
⎥,
⎣ −A 2gx 3 ⎦
a4 √
3

−A 4
2gx 4

⎡ γ1 ⎤
A1 0
⎢ γ2 ⎥
⎢ 0 A2 ⎥
g(x) = ⎢


(1−γ2 ) ⎥ ,
⎣ 0 A3 ⎦
(1−γ1 )
A1 0

and x = [x1 , x2 , x3 , x4 ]T , u = [u1 , u2 ]T . The variables xi , i = 1, ..., 4 represent


the water levels of the tanks, ui , i = 1, 2 the control inputs, Ai , i = 1, ..., 4 the
cross-sections of the tanks, ai , i = 1, ..., 4 the cross-sections of the outlet holes,
γi = 0.4, i = 1, 2 positive constants, and g = 981 cm s2 the gravitational constant.
The parameter values of Ai and ai are taken from the laboratory experiment
[4] and are given in Table 1. In order to control the quadruple tank system via
the passivity-based nonlinear model predictive control, one has to search for a
fictitious passive output. By using the storage function S(x) = 12 x23 + 12 x24 , one
obtains for x3 , x4 ≥ 0 [the levels of the tanks cannot be negative].

Table 1. Parameter values of the system

Ai ai
i=1 50.3 cm2 0.2 cm2
i=2 50.3 cm2 0.2 cm2
i=3 28.3 cm2 0.1 cm2
i=4 28.3 cm2 0.1 cm2
160 T. Raff, C. Ebenbauer, and F. Allgöwer

16
15

14

12
1
x

2
10

x
10

6 5
0 50 100 150 200 250 0 50 100 150 200 250
Time Time

Fig. 1. Closed loop trajectory x1 : Fig. 2. Closed loop trajectory x2 :


Passivity-based MPC (gray) and MPC Passivity-based MPC (gray) and MPC
without guaranteed stability (black) without guaranteed stability (black)

35
25

30
20
25

20 15
x4
3
x

15
10
10

5
5

0
0 50 100 150 200 250 0
Time 0 50 100 150 200 250
Time

Fig. 3. Closed loop trajectory x3 : Fig. 4. Closed loop trajectory x4 :


Passivity-based MPC (gray) and MPC Passivity-based MPC (gray) and MPC
without guaranteed stability (black) without guaranteed stability (black)

300 300

250 250

200 200
2
1

150 150
u
u

100 100

50 50

0 0
0 50 100 150 200 250 0 50 100 150 200 250
Time Time

Fig. 5. Control input u1 : Passivity-based Fig. 6. Control input u2 : Passivity-based


MPC (gray) and and MPC without guar- MPC (gray) and and MPC without guar-
anteed stability (black) anteed stability (black)
Nonlinear Model Predictive Control: A Passivity-Based Approach 161

' ( ' (
a3  (1 − γ2 ) a4  (1 − γ1 )
Ṡ(x) ≤ − 2gx3 + u2 x3 + − 2gx4 + u1 x4 ,
A3 A3 A4 A1
≤ x3 u2 + x4 u1 .
Consequentely, y = [x3 , x4 ]T is a passive output for the quadruple tank system
with the storage function S(x) = 12 x23 + 12 x24 . Since the quadruple tank system is
zero state detectable with respect to the output y = [x3 , x4 ]T , the passivity-based
nonlinear model predictive control (12) can be used to asymptotically stabilize
the quadruple tank system. In the following, the control task is to stabilize the the
quadruple tank system at the equilibrium point xs = [14 cm 14 cm 15.3 cm
20.3 cm]T . The steady state control input us at the equilibrium point xs is us =
s 38.2 s ] . The performance index (5) was chosen as 1000([x1 − x1s ] +
[43.2 ml ml T 2

[x2 − x2s ] ) + (u1 − u1s ) + (u2 − u2s ) . Furthermore, the nonlinear model pre-
2 2 2

dictive control schemes were implemented with a prediction horizon T = 60 s


and a sampling time δ = 1 s in the nonlinear model predictive control toolbox
[13]. Figure 1 to Figure 6 show the dynamic behavior with the initial condition
x0 = [5 cm 5.6 cm 4 cm 4.5 cm]T of the passivity-based nonlinear model pre-
dictive controller (12) and the model predictive controller (9), i.e., a model pre-
dictive controller without guaranteed closed loop stability. As it can be seen from
the figures, the closed loop system is unstable with the nonlinear model predictive
controller (9). This instability is not a new fact in model predictive control [1] and
has motivated the development of model predictive control schemes with guaran-
teed closed loop stability. In contrast to the model predictive controller (9), the
passivity-based nonlinear model predictive controller (12) asymptotically stabi-
lizes the quadruple system. Hence, this examples illustrates nicely that the pro-
posed approach achieves closed loop stability while improving the performance.

6 Conclusions
In this paper a nonlinear model predictive control scheme based on the concept
of passivity was developed. It was shown that by using a specific passivity-
based state constraint closed loop stability is guaranteed. The basic idea of
the passivity-based nonlinear model predictive control scheme is to unify op-
timal control, passivity and nonlinear model predictive control based on their
relationships. Since passivity and stability are closely related, the proposed ap-
proach can be seen as an alternative to the control Lyapunov function based
nonlinear model predictive control scheme [16]. Finally, the passivity-based non-
linear model predictive control scheme was applied to control a quadruple tank
system in order to demonstrate its applicability.

References
[1] R. R. Bitmead, M. Gevers, I. R. Petersen, and R. J. Kaye, “Monotonicity and
Stabilizability Properties of Solutions of the Riccati Difference Equation: Proposi-
tions, Lemmas, Theorems, Fallacious Conjectures and Counterexamples”, Systems
and Control Letters , pages 309-315, (1985).
162 T. Raff, C. Ebenbauer, and F. Allgöwer

[2] C.I. Byrnes and A. Isidori and J.C. Willems, “Passivity, Feedback Equivalence,
and the Global Stabilization of Minimum Phase Nonlinear Systems”, IEEE Trans-
actions on Automatic Control, pages 1228-1240, (1991).
[3] C. Chen and F. Allgöwer, “A Quasi-Infinite Horizon Nonlinear Model Predictive
Control Scheme with Guaranteed Stability”, Automatica, pages 1205-1217, (1998).
[4] C. Ebenbauer, “Control of a Quadruple Tank System”, Institute for Systems The-
ory in Engineering Laboratory Lecture Notes, (2004).
[5] R.A. Freeman P.V. Kokotovic, “Robust Nonlinear Control Design: State-Space
and Lyapunov Techniques”, Birkhäuser, (1996).
[6] A. Jadbabaie, J. Yu, and J. Hauser, “ Stabilizing receding horizon control of
nonlinear systems: a control Lyapunov function approach”, In Proceedings of the
American Control Conference, pages 1535 - 1539, (1999).
[7] K.H. Johansson , “The Quadruple-Tank Process: A Multivariable Laboratory Pro-
cess with Adjustable Zero”, IEEE Transactions on Control Systems Technology,
pages 456-465 (2000).
[8] R.E. Kalman, “When is a Linear Control System Optimal”, Transactions of the
ASME, Journal of Basic Engineering, pages 1-10, (1964).
[9] S. Keerthi and E. Gilbert, “Optimal Infinite-Horizon Feeback Laws for a General
Class of Constrained Discrete-Time Systems: Stability and Moving-Horizon Ap-
proximations ”, Journal of Optimiztaion Theory and Applications, pages 265-293
(1988).
[10] H.K. Khalil, “Nonlinear Systems”, Prentice Hall, (1996).
[11] D.Q. Mayne and J.B. Rawlings and C.V. Rao and P.O.M. Scokaert, “Constrained
Model Predictive Control: Stability and Optimality ”, Automatica, pages 789-814,
(2000).
[12] P. Moylan, “Implications of Passivity for a Class of Nonlinear Systems”, IEEE
Transactions on Automatic Control, pages 373-381, (1974).
[13] Z.K. Nagy, “Nonlinear model predictive control toolbox”, (2005).
[14] G. De Nicolao, L. Magni, and R. Scattolni, “Stabilizing Receding-Horizon Control
of Nonlinear Time-Varying Systems”, IEEE Transactions on Automatic Control,
pages 1031-1037,(1998).
[15] H. Michalska and D. Q. Mayne, “Robust Receding Horizon Control of Constrained
Nonlinear Systems ”, IEEE Transactions on Automatic Control, pages 1623 - 1633,
(1993).
[16] J. A. Primbs, V. Nevistic, and J.C. Doyle, “A receding horizon generalization of
pointwise min-norm controllers”, IEEE Transactions on Automatic Control, pages
898 - 909, (2000).
[17] R. Sepulchre M. Jankovic P. Kokotovic, “Constructive Nonlinear Control”,
Springer, (1997).
[18] J.C. Willems, “Dissipative Dynamical Systems. Part I: General Theory”, Archive
for Rational Mechanics and Analysis, pages 321-351, (1972).

You might also like