Download as pdf or txt
Download as pdf or txt
You are on page 1of 6

Proceedings

The of
of the
International
Proceedings 20th
20th World
Federation
the Worldof Congress
Automatic Control
Congress
The
The International
Proceedings
Toulouse, France,
International Federation
of the 20th9-14,
July World
Federation of Automatic
of Congress
2017
Automatic Control
Control
Toulouse,
The France,Federation
International July 9-14,
9-14, 2017
2017 Available online at www.sciencedirect.com
of Automatic Control
Toulouse, France, July
Toulouse, France, July 9-14, 2017
ScienceDirect
IFAC PapersOnLine 50-1 (2017) 518–523
A
A generalized
generalized approach
approach to
to Economic
Economic Model
Model
A generalized
APredictive
generalized approach
approach to
to Economic
Economic Model
Model
Predictive Control
Control with
with terminal
terminal penalty
penalty
Predictive Control
Predictive Control with terminal penalty
with terminal penalty
functions
functions
functions
functions
∗ ∗
Zihang Dong ∗ David Angeli ∗
Zihang
Zihang Dong Dong ∗ DavidDavid Angeli
Angeli ∗
∗ Zihang Dong ∗ David Angeli ∗
∗ Department of Electrical and Electronic Engineering,
∗ Department of Electrical and Electronic Engineering,
∗ Department of Electrical
Imperial College andLondon,
Electronic UK.Engineering,
(e-mail: Department of Electrical
Imperial
Imperial
zihang.dong14@imperial.ac.uk, College
College andLondon,
Electronic
London, UK.Engineering,
UK.
d.angeli@imperial.ac.uk)
(e-mail: Imperial
zihang.dong14@imperial.ac.uk, College London, UK.
d.angeli@imperial.ac.uk)
(e-mail: zihang.dong14@imperial.ac.uk, d.angeli@imperial.ac.uk)
(e-mail: zihang.dong14@imperial.ac.uk, d.angeli@imperial.ac.uk)
Abstract: In this paper, we first introduce upper and a lower bounds of the best asymptotic
Abstract:
Abstract:
average In
In this
performancethis paper,
paper, we
we first
for nonlinear first introduce
introduce upper
upper and
control systems and
based aa lower
on thebounds
lower bounds
concepts of
of the
the best
best asymptotic
of dissipativity asymptotic and
Abstract:
average
average storage
control In
performance
performancethis paper,
for we
for nonlinear
functions. first
nonlinear
This allows introduce
control
control
to extend upper
systems
systems and
based a lower
based on the
the formulation on the bounds
concepts
andconcepts
analysisof of the
of best asymptotic
dissipativity
of dissipativity
Economic Modeland
and
average
control performance
storage
control storage
Predictive Control for
functions.
functions.
to more nonlinear
This
This allows control
to
allows optimal
general systems
extend the
to extendoperation based
formulation
the formulation on
regimes,andtheand concepts
analysis
analysis
such as periodic ofof dissipativity
Economic
of Economic solutions. Model
Modeland A
control
Predictive
Predictive storage
Control
Control functions.
to
to more
more This allows
general
general to extend
optimal
optimal the formulation
operation
operation regimes,
regimes,
performance and stability analysis is carried out within this generalized framework. Finally two and analysis
such
such as
as of
periodic
periodic Economicsolutions.
solutions. Model A
A
Predictive
performance
performance
examples Control
are andand to more
stability
stability
proposed general
analysis
andanalysis
discussed is optimal
carried
is carried
to showout operation
out within
thewithin regimes,
this
meritsthis such
generalized
generalized
of the as periodic
framework.
proposedframework. solutions.
approach. Finally two Finally twoA
performance
examples are and stability
proposed and analysis
discussed is carried
to show out
thewithin
merits
examples are proposed and discussed to show the merits of the proposed approach. thisof generalized
the proposed framework.
approach. Finally two
© 2017, IFAC
examples are (International
proposed andFederation discussedof to Automatic
show the Control)
meritsHosting
of the by Elsevier approach.
proposed Ltd. All rights reserved.
Keywords: nonlinear predictive control, optimal control, asymptotic stabilization, performance
Keywords:
Keywords:Lyapunov
analysis, nonlinear
nonlinearmethods predictive control, optimal control, asymptotic
predictive control, optimal control, asymptotic stabilization, performance stabilization, performance
Keywords:
analysis, nonlinear
Lyapunov
analysis, Lyapunov methods predictive
methods control, optimal control, asymptotic stabilization, performance
analysis, Lyapunov methods
1. INTRODUCTION Teel (2013). Based on the generalized terminal equality
1. INTRODUCTION
1. INTRODUCTION Teel (2013).
Teel (2013).several Basedupdate
Based on the
on therules generalized
generalized terminal terminal
terminal equality
equality
constraint, for the self-tuning
1. INTRODUCTION Teel (2013).
constraint, Based
several on
update the rulesgeneralized
for the terminal
self-tuning equality
terminal
Because of its ability of handling nonlinearity and system constraint, weight are
several update
weight are illustrated
constraint, several
illustrated update in
rules for
in Müller
rules
Müller
et al.
for
et
the(2013).
the
al.
self-tuning
self-tuning
(2013).
terminal
Furthermore,
terminal
Furthermore,
Because
Because
constraints, of its
of its ability
ability
model of handling
of handling
predictive nonlinearity
nonlinearity
control (MPC) isand and system
system weight
becoming in Müller areetillustrated
al. (2014),inthe Müller et al. (2013).
closed-loop asymptotic Furthermore,
average
Because of its ability of weight
in Müller
Müller areet
etillustrated
al. (2014),in Müller
the et al. (2013).
closed-loop if theFurthermore,
asymptotic average
constraints,
constraints,
increasingly model
model
popular in handling
predictive
predictive nonlinearity
control
control
industrial (MPC) is
(MPC)
applications isand system
becoming
becoming
and pro- in performance
in Müller
performance et
al.
al.
(2014),
bounds
(2014),
bounds
the
can
the
can
be
be
closed-loop
improved
closed-loop
improved
asymptotic
asymptotic
if the
average
generalized
average
generalized
constraints,
control,model
increasingly
increasingly
cess popular
popular predictive
see examplesin in control
in industrial
industrial (MPC)
applications
Qin applications
and is becoming
Badgwell and
and pro- performance
pro-
(2003). terminal equality bounds is relaxed can bebyimproved regional if the generalized
constraint.
increasingly performance
terminal equalitybounds is can
relaxed be by improved
regional if the
constraint. generalized
cess
This control popular
cess control,
control, see examples
see
paradigm in normally
examples industrial
in Qin
in applications
Qinrelegates
and Badgwell
and Badgwell
economic and pro-
(2003).
and terminal
(2003). However,
equality is relaxed by regional constraint.
terminal equalityoptimalisregimes relaxed of by operation
regional constraint. may have com-
cess
This
This control,
control
control see examples
paradigm
paradigm in
normally
normally Qin and
relegates
relegates
profitability issues to the design of optimal set-points Badgwell
economic
economic (2003).
and
and However, optimal regimes of operation
operation may have havesteady- com-
This control
profitability paradigm
issues to the normally
design ofrelegates
optimal economic
set-points and However,
plex nature, optimal
periodic regimes operation of can outperform may com-
profitability
suitable pointwise issues to in the
timedesign of optimal
constraints. set-points
Real-time and plex
control However,
plex nature,
nature, optimal
periodic
periodic regimes operation
operation of operation
can
can outperform may
outperform have
of operationssteady- com-
steady-
profitability
suitable
suitable
is, instead, issuesconcerned
pointwise
pointwise
only to
in the
in timedesign
time of the
optimal
constraints.
constraints.
with set-points
Real-time
Real-time
resulting and state
control
control
tracking plex
state
and even
nature,
and even
more general
periodic
more operation
general
regimes
regimescan outperform
of operations
could
steady-
could
suitable
is, instead,
is,
probleminstead,pointwise
whichonly in time constraints.
onlyis concerned
concerned
translated with
with
as an thethe Real-timeproblem
resulting
resulting
optimization control state
tracking
tracking
sometimes
state
sometimes
and even
and even
arise.more
arise.
To deal
more
To dealdeal
general
general
withregimes
withregimes
such instances,
such instances,
of operations
of operations
instances,
this could
this
work
could
worka
is, instead,
problem whichonly is concerned
translated with
as an the resulting
optimization tracking
problem sometimes
will remove arise.
terminalTo equalitywith such
constraints and this employ work
problem
over a finite which timeis horizon
translated andaswithan optimization
an objective function problem will sometimes
will remove
remove
suitable arise.
notion terminal
terminalTo deal
of “control equality
equalitywith
storage such
constraints
constraints instances,
function”and and
(CSF) this
employ
employ aswork theaa
problem
over aa is
over
which which
finite
finite time
(fortime
theis sake
translated
horizon
horizon andaswith
and
of stability) an optimization
with an objective
an
chosenobjective problem
function
to be function will remove
suitable
positive suitable notion terminal
of “control equality storage constraints
function” and
(CSF) employ as the a
over a is
which finite
is (fortime
the horizon
sake ofsomeand with
stability) an objective
chosen be function
to interest.
positive terminal notionpenaltyoffunction.
“controlThe storagenotefunction”
is organized (CSF) as the
as follows.
which
definite (for
with the sake
respect toof stability) chosen
equilibrium to
of be positive suitable
terminal
In terminal notion
penalty
penalty of “control
function.
function. Thestorage
Theare note function”
is
notedescribed organized
is organized (CSF) as as
follows.
as follows. the
which
definiteis
definite
recent (for however,
with
with
years, the saketo
respect
respect ofsome
toan stability)
some chosen
equilibrium
equilibrium
alternative approach,to interest.
of
of beeconomic
positive
interest. In Notation
In terminal
Notation3penalty
and problem
and function.
problem
setup
setupTheare note
are is organized
described
in Section
as
in Section follows.
Section
2.
2.
definite
recent with
years, respect
however, toan some equilibrium
alternative approach,of interest.
economic In Notation
Section and
provides probleman estimatesetup todescribed
some upper in and lower 2.
recent
MPC (EMPC), years, however,has looked an alternative
into the issue approach, economic
of directly Notation
Section
ad- Section 3 and
provides
3 provides probleman setup
estimate
an estimateaverage areto described
some
to someperformance. upper in Section
and
upper and lower lower 2.
recent
MPC (EMPC),
MPC
dressing years,
(EMPC), however,
economic has an alternative
hasoptimization
looked
looked into in
into thereal
the approach,
issue
issue of
time, and economic
of directly
directly ad- bound
ad-
to this Section
bound
for
3
for
system asymptotic
provides
system an estimate
asymptotic to
average some upper
performance. and
The
lower
The
MPCadopts (EMPC), bound
extension for ofsystem
EMPCasymptoticformulation average
and the performance.
closed-loop The sta-
dressing
dressing
end, cost has
economic
economic looked into
optimization
optimization
functionals which theare
in
in issue
real
real of
time,
time,
not directly
and to
and
required ad-
totothis
this bound
extension
be extension
bility arefordiscussed
ofsystem
of EMPCasymptotic
EMPC informulation
formulation
Section 4.average
and examples
and
Two the
theperformance.
closed-loop
closed-loopindicating The
sta-
sta-
dressing
end, adopts
end,
positive adopts economic
cost with
cost
definite optimization
functionals
functionals
respectwhich whichin
to thearereal
are time,
not
equilibrium and
not required
required to
point.tothis
to be
be extension
bility of EMPC
are discussed
discussed informulation
Section 4. Twoandsolution
Two the closed-loop
examples indicating sta-
end,
positiveadopts cost with
definite functionals
with respectwhich theare
to the not required
equilibrium point. bility
the
to be bility are
convergence to in
the Section
best 4.
periodic examples are indicating
included
positive definite respect to equilibrium point. the are discussed
convergence to in
the Section
best 4.
periodicTwo examples
solution are indicating
included
In this definite
positive respect, with various
respect tools in literature
to the equilibriumhave point. been the the
convergence
in Section
convergence
in Section
Section
5. Section
5. Section
to the
to the
Section
best periodic
6 concludes
best periodic
6 concludes
concludes
thissolution
paper. are included
thissolution
paper. are included
InIn this
this respect,
respect, various
various tools
tools in
in literature
literature have
have been
been in
proposed and studied in the economic optimization setup. in Section 5. Section 6 concludes this paper. 5. 6 this paper.
In
proposedthis
proposed
In analogyand respect,
and various
studied in
studiedetinal.
to Mayne thetools in
economic
the(2000),
economic literature
where optimization
optimization have
three ingredients been
setup.
setup.
Inproposed
In analogy
are analogy and
elaborated to studied
to Mayne
Mayne etinal.
et
in stabilizing the(2000),
al. economic
(2000), where
MPC,where optimization
three of
three
consisting setup.
ingredients
ingredients
terminal
In
are
are analogy
elaborated
cost,elaborated to Mayne
in et
in stabilizing
terminal constraint al.
stabilizing (2000),
MPC,
and MPC, where three
consisting
consistingsimilar
local controller, ingredients
of terminal
of terminal
tools 2. PRELIMINARIES AND SETUP
are
cost,elaborated
cost, terminal
terminal in stabilizing
constraint
constraint and MPC,
and local MPC
local consisting
controller,
controller, of terminal
similar
similar tools
tools 2. PRELIMINARIES
2. PRELIMINARIES AND AND SETUP SETUP
have been proposed for Economic and have allowed
cost,
have
have terminal
been constraint
proposed
been stability
proposedand for and
Economic
for performancelocal
Economic MPC controller,
MPC and
and of similar
have
have tools
allowed 2. PRELIMINARIES AND SETUP
feasibility, analysis theallowed
closed- 2.1 Notation
have been
feasibility,
feasibility,
loop system. proposed
stability
stability for
and Economic
performance
and performance
In Rawlings MPCanalysisand
analysisAngeli
et al. (2012), have
of the allowed
closed-
of the closed-
et al. 2.1 2.1 Notation
Notation
feasibility,
loop system.
loop
(2012) andstability
system. In
Amrit and
In Rawlings
Rawlings
et al.performance
et al.
et
(2011), analysisAngeli
al. asymptotic
(2012),
(2012), ofstability
Angelithe closed-
et al.
et al.
of 2.1 Notation
loop
(2012)system.
(2012)
EMPC and
and
withAmrit In Rawlings
Amrit
terminal et al.
et al. et al. asymptotic
(2011),
(2011),
constraints (2012),
asymptotic
or terminal Angeli
stability
stability
costs et has
al.
of The Euclidean norm of x is |x|. Let symbols R and I denote
of
(2012)
EMPC
EMPC and
with Amrit
with by terminal
terminal et al. (2011),
constraints asymptotic
or terminal stability
costs of the
has The sets
The Euclidean
Euclidean
of realnormnorm of x
of x is
isand |x|. integers,
|x|. Let symbols
Let symbols R and
R and II denote
denote
I[a,b]
been proved using aconstraints
rotated stage or cost
terminal
in an costs
auxiliaryhas numbers respectively.
EMPC
been with
proved terminal
by using a constraints
rotated stageor terminal
cost in an costs
auxiliaryhas The
the Euclidean
sets of real norm
numbers of x isand |x|. Let symbols
integers, R and
respectively. I II[a,b]
denote
been proved by
optimization using a provided
problem, rotated stage that cost in an auxiliary
a condition called the sets the
denotes of real numbers
integers {a, aand + 1, integers,
· · · , b} respectively.
and I≥0 denotes [a,b]
been proved by using a provided
rotatedMoreover,
stage
that cost called the sets the of real numbers aand integers, and II≥0
respectively. I
optimization
optimization
strict problem,
problem,
dissipativity provided
is satisfied. that ininthese
an auxiliary
aa condition
condition called
papers, denotes
denotes the
the non-negative integers
integers
integers. {a,
{a, a A+ 1, ·· ·· ·· ,, b}
+continuous
1, and
b}function denotes
[0, [a,b]
≥0α:denotes ∞)
optimization
strict dissipativity
strict
concepts problem,
dissipativity
on EMPC is areprovided
is satisfied.
satisfied. thatto aperiodic
Moreover,
Moreover,
extended condition
in these
in theseterminalcalled denotes
papers,
papers, the
the
→ [0, theis integers
non-negative
non-negative
∞) integers.
integers.
of class {a,
K, ifaAA+ 1, · · · , b}
itcontinuous
continuous
is zero and Iand
atfunction
function
zero α:denotes
≥0α: [0, ∞)
[0,
strictly ∞)
strict
concepts
concepts
constraint dissipativity
onand
on EMPC
EMPC is satisfied.
average are constraints.
are Moreover,
extended
extended to in these
toInperiodic
periodic
order obtain the
papers,
terminal
toterminal →
→ non-negative
[0,
[0, ∞)
∞) is
increasing. isA of integers.
class
ofcontinuous
class K,
K, ififAit itcontinuous
function is
is zero
zero ρ : at Rnnfunction
at zero
→ Rand
zero and [0, ∞)
isα:positive
strictly
strictly
aconcepts
constraint onand
and EMPCaverage areaconstraints.
extended
constraints. toIn
Inperiodic
order state →
toterminal
obtain [0, ∞)
increasing. isA
A of class to
continuous K,some
if it point
function is zero ρ x:: eat
ρ R
R∈ →nn R
nzero
→ Rifand
is strictly
positive
constraint
larger feasibility average
set, new generalized order
terminalto obtain increasing.
definite with continuous
respect function n R is
ρ(x positive
e) = 0
constraint
a larger
aconstraint and
larger feasibility
feasibilityaverage
set,
set, a
a constraints.
new
new generalized
generalizedIn order
terminal
terminalto obtain
state
state increasing.
definite
definite with
with A continuous
respect
respect to
to function
some
some point
point ρ :
x
x R∈
∈ R

R nR if is
ρ(x positive
e) = Rnn0
e of an point xe )∈=
if ρ(x 0
where the terminal state-input pair can be a free and ρ(x) > 0 for all x = xe . The distance e
a larger in
constraint
constraint
variable feasibility
where
where theset,
the
optimization aprocess
terminal
terminal newstate-input
generalized
state-input
is studiedpair pair
byterminal
can be
can
Fagianobe aastate
free to
free
and definite
and
and set with
a ρ(x)
ρ(x) Π> respect
>is00 denoted
for
for all
all xx to
=
= xsome
as x e .. The
|x| The:=point
distance
distance
min xe ∈|x R−
of
of if ρ(xx
aa z|.
point
point xe )∈
∈ =RR n0
e Π z∈Π
constraint
variable in where the
optimization terminalprocess state-input
is studied pair
by can
Fagianobe a free
and and
to a ρ(x)
set
variable in optimization process is studied by Fagiano and to a set Π is denoted as |x|Π := minz∈Π |x − z|. Π> is 0 for
denotedall x 
= as x|x|
e . ΠThe:= distance
min z∈Π |xof− a point
z|. x ∈ R n

variable in optimization process is studied by Fagiano and to a set Π is denoted as |x|Π := minz∈Π |x − z|.
Copyright © 2017 IFAC 520
2405-8963 ©
Copyright © 2017
2017, IFAC (International Federation of Automatic Control) Hosting by Elsevier Ltd. All rights reserved.
Copyright
Peer review 2017 IFAC
© under IFAC
responsibility of International Federation of
520
520 Control.
Automatic
Copyright © 2017 IFAC 520
10.1016/j.ifacol.2017.08.114
Proceedings of the 20th IFAC World Congress

Toulouse, France, July 9-14, 2017 Zihang Dong et al. / IFAC PapersOnLine 50-1 (2017) 518–523 519

2.2 Problem setup Recall the notion of dissipativity as given in Definition 4.1
in Angeli et al. (2012),
We consider finite dimensional discrete-time nonlinear Definition 4. A discrete time system is dissipative with
control systems described by difference equations respect to a supply rate s : Z → R if there is a continuous
x+ = f (x, u) (1) storage function λ : X → R such that:
with state x ∈ X ⊂ Rn , input u ∈ U ⊂ Rm , and a λ(f (x, u)) − λ(x) ≤ s(x, u) (7)
continuous state transition map f : X × U → X. Together for all (x, u) ∈ Z. If in addition a positive definite function
with system (1), let us consider a time-invariant, nonlinear, ρ : X → R≥0 exists such that:
nonconvex, but continuous stage cost given as
λ(f (x, u)) − λ(x) ≤ −ρ(x) + s(x, u), (8)
(x, u) : Z → R (2)
then the system is said to be strictly dissipative.
where Z is a compact set capturing the pointwise-in-time
state and input constraints which our system is subject to: Alternatively, given the role of dissipativity in providing
(x(k), u(k)) ∈ Z, ∀k ∈ I≥0 . (3) lower bounds to the best asymptotic performance, one may
Our goal is to enhance profitability by minimizing the consider the following quantity,
economic costs incurred in the long term system operation: Definition 5. The tightest lower bound of ∗av is defined
 as:
V (x, u) = (x(k), u(k)), x+ = f (x, u), x(0) = x.
 := sup{ c | ∃ λ(·) : X → R, continuous, such that
k
(4)
c (9)
To this end, we need to identify a viable subset of state λ(f (x, u)) ≤ λ(x) + (x, u) − c, ∀(x, u) ∈ Z}.
space and corresponding control actions. As well known, Next, along the lines of the well known tool of Control
the notion of control invariant set is crucial in this respect. Lyapunov Function (CLF) (see definition in Rawlings and
Definition 1. A control invariant set is any non-empty Mayne (2009)), we propose a similar concept referred to
closed set X ⊆ X, such that ∀x ∈ X, ∃u : f (x, u) ∈ as Control Storage Function (CSF).
X and (x, u) ∈ Z. The corresponding input which keeps Definition 6. A control storage function is a function Vf :
the system state inside X is denoted as U(x) := {u | (x, u) ∈ Xf → R that is continuous and such that for all x ∈ Xf
Z and f (x, u) ∈ X}. The set of  state and corresponding inf Vf (f (x, u)) − s(x, u) ≤ Vf (x), (10)
admissible input pairs is Z := x∈X [{x} × U(x)]. u∈Uf (x)

Remark 2. We consider the largest control invariant set where s : Zf → R is the supply rate.
X ⊆ X. This contains all control invariant sets in X and As a special case of CSF, a CLF, in which s(x, u) = 0,
any given intial condition x(0) ∈ / X generates trajectories is frequently used to approximate the tail of the infinite
which violate system constraints (3) at some point in time. horizon cost of tracking MPC. Our CSF is meant to be an
Therefore, constraints (3) can be strengthened as follows: appropriate choice of terminal cost in an economic setup.
(x(k), u(k)) ∈ Z, ∀k ∈ I≥0 , (5)
In order to estimate an upper bound for the best asymp-
and viability is still guaranteed for initial state x ∈ X. totic performance, the quantity below can be specified,
It will also be convenient to define an additional control Definition 7. The tightest upper bound of ∗av is defined
invariant set for later use as in the assumption below as:
 := inf { c | ∃ Vf : Xf → R, such that ∀x ∈ Xf ,
Assumption 1. There is a control invariant set Xf ⊆ X and c
a set of admissible control which keeps the state inside Xf (11)
inf Vf (f (x, u)) + (x, u) ≤ Vf (x) + c}.
u∈Uf (x)
as Uf (x) := {u ∈ U(x) | f (x, u) ∈ Xf }.
Remark 8. Notice that the above CSF inequality in Def-
The set of state and
corresponding admissible input pairs inition 6 follows the same form of the Hamilton-Jacobi-
is denoted as Zf := x∈Xf [{x} × Uf (x)]. Bellman (HJB) inequality, so any CSF can also be re-
garded as a solution of the HJB inequality.
3. DISSIPATIVITY AND CONTROL STORAGE
FUNCTIONS We are now ready to state the main result of this Section:
Theorem 9. Consider system (1) subject to constraints
In order to have a grasp of the system long-run optimal (5), then, the following inequality holds:
average performance, three quantities ∗av ,  and , which  ≤ ∗av (x), ∀x ∈ X. (12)
are explicitly defined below, will be addressed.
In addition, if Assumption 1 is fulfilled, we have the
Definition 3. Let x ∈ X be a given initial state, then the following upper bound for ∗av (x):
best average asymptotic cost is defined as:
T −1 ∗av (x) ≤ , ∀x ∈ Xf . (13)
∗ (x(t), u(t))
av (x) := inf lim inf t=0 . Proof. The inequality (12) is derived from the dissipativ-
u(·) T →+∞ T
ity with supply rate s(x, u) = (x, u) −  along any feasible
x(0) = x, x+ =f (x, u) solution, whereas (13) is obtained by applying the same
(x(t), u(t)) ∈ Z technique on CSF with supply rate s(x, u) =  − (x, u). 
(6)
Moreover, we denote by ∗av = inf ∗av (x). If a system is controllable within finite time to the best
x∈X optimal operation, every initial condition gives the same

521
Proceedings of the 20th IFAC World Congress
520
Toulouse, France, July 9-14, 2017 Zihang Dong et al. / IFAC PapersOnLine 50-1 (2017) 518–523

best asymptotic average performance, that is ∗av (x) = By applying liminf on both sides and
exploiting boundness
K−1
∗av , ∀x ∈ X. However, if this is not the case, the strict of solutions, we have lim supK→∞ t=0 K
(x(t),u(t))
≤ .
inequality ∗av (x) > ∗av may hold. This gap arising from Therefore, the closed-loop performance of EMPC problem
uncontrollable systems will be illustrated in Example 5.2. is no worse than .
4. PERFORMANCE AND STABILITY OF
GENERALIZED EMPC 4.3 Stability analysis

4.1 Economic MPC formulation This sub-section explores the asymptotic stability of the
closed-loop system under EMPC control actions.
Now, we consider the EMPC problem for a given finite
prediction horizon N ∈ I[1,+∞) . The open loop EMPC In this context, we only consider the no-gap case, viz:
problem can be formulated as Assumption 4. There is no gap between the tightest upper
N
 −1 and lower bounds of ∗av :
VN0 (x) = min (x(k), u(k)) + Vf (x(N ))  = ∗av = ∗av (x) = , ∀x ∈ X. (19)
u
k=0
(14) Moreover, there exists a storage function λ(·) and a CSF
s.t. x(k + 1) = f (x(k), u(k)), x(0) = x
Vf (·) for which inequalities (8) and (10) are achieved with
(x(k), u(k)) ∈ Z, ∀k ∈ I[0,N −1] supply rates (x, u) − ∗av and ∗av − (x, u), respectively.
x(N ) ∈ Xf
The rotated stage cost and terminal cost are defined as:
where u := {u(0), u(1), · · · , u(N − 1)} is the control
sequence, Vf is the terminal penalty function and a CSF L(x, u) := (x, u) + λ(x) − λ(f (x, u)) − , (20)
according to Definition 6. The feasible set XN ⊆ X is V f (x) := Vf (x) + λ(x), ∀x ∈ Xf , (21)
XN := {x ∈ X | ∃x(1), · · · , x(N ) : x+ = f (x, u), x(0) = x, then, the new objective function is
x(N ) ∈ Xf , (x(k), u(k)) ∈ Z, ∀k ∈ I[0,N −1] }. N
 −1

(15) V N (x, u) := L(x(k), u(k)) + V f (x(N )). (22)


A classic assumption found in the literature: k=0

Assumption 2. The economic optimization problem (14) It is noticed that the strict dissipation inequality (8) with
admits a feasible solution for initial state x ∈ XN . supply rate s(x, u) = (x, u) −  implies that the modified
stage cost is positive definite with respect to some closed
If the above assumption holds, the optimal feedback policy subset of state space X, denoted by ΠX . Thus it is bounded
from EMPC controller is κN (x) := u0 (0, x), where u0 (k, x) from below by a K function L(x, u) ≥ α(|x|ΠX ) ≥ 0.
denotes the k th element of the optimal control sequence u0
at the given initial state x. Next, we recall lemmas introduced in Amrit et al. (2011),
Lemma 10. For every function λ(·) and constant  ∈ R,
In order to analyse the stability of the closed-loop trajec-
the solutions of optimization problem (14) and the one by
tories induced from optimal solutions of EMPC problem
using (22) as the objective function are identical.
(14), we propose the following assumption for later use,
Assumption 3. The optimal control κN (x) is a continuous Lemma 11. If Assumption 4 holds, then for all x ∈ Xf ,
function of x in a neighborhood of the optimal periodic the CSF inequality in (10) with s(x, u) =  − (x, u) holds
trajecctory. if and only if the following inequality is satisfied
V f (f (x, uf (x))) − V f (x) ≤ −L(x, uf (x)). (23)
4.2 Performance analysis
Then, the main result on asymptotic stability is given as:
At any time instant t, let x(t) ∈ XN , then ũ :=
{u0 (1, x(t)), · · · , u0 (N − 1, x(t)), uf (x(t + N ))} is feasible Theorem 12. Let Assumption 1, 2 and 4 hold, the set ΠX
but not necessarily optimal for the optimization problem by using control policy κN (x), is asymptotically stable
at t + 1, where uf (·) is the terminal control policy: within a region of attraction XN .
uf (x) ∈ argmin Vf (f (x, u)) + (x, u). (16) Proof. The proofs of Lemma 10, 11 and Theorem 12 are
u∈Uf (x)
similar to Lemma 14, 9 and Theorem 15 in Amrit et al.
Therefore, the optimal cost-to-go functions fulfills VN0 (x(t+ (2011). 
1)) ≤ VN (x(t), ũ), which is
VN0 (x(t + 1)) ≤ VN0 (x(t)) + Vf (x(t + N + 1)) 4.4 Periodic optimal regimes of operation
+ (x(t + N ), uf (x(t + N )))
− (x(t), u0 (0, x(t))) − Vf (x(t + N )) Denote the optimal period-T solution and associated con-
trol of the optimal control problem (4) on Z as Π :=
≤VN0 (x(t)) − (x(t), u0 (0, x(t))) + . {(x∗i , u∗i ), i ∈ I[0,T −1] } for some T ∈ I[1,+∞) . For con-
(17) venience, we denote the projection of Π on X as ΠX , and
Then, for any time K, we have, we simply consider i ∈ I[0,T −1] and use i + j to respresent
K−1
 K−1
 (i + j) mod T, ∀j ∈ I.
VN0 (x(t + 1)) − VN0 (x(t)) ≤ ( − (x(t), u0 (0, x(t)))).
t=0 t=0 The system cost over a time period T achieves its minimum
(18) at Π, viz.

522
Proceedings of the 20th IFAC World Congress

Toulouse, France, July 9-14, 2017 Zihang Dong et al. / IFAC PapersOnLine 50-1 (2017) 518–523 521

T
 −1 T
 −1 Then, according to Theorem 12, the optimal periodic
(x∗i , u∗i ) ≤ (xi , ui ), ∀(xi , ui ) ∈ Z, xi+1 = f (xi , ui ). solution is orbitally asymptotically stable.
i=0 i=0 Remark 15. Exploring the link of our concept with respect
(24) to those extended notions of dissipativity for periodic
For every state x∗i ∈ ΠX , three important terminal in- systems defined in Grüne and Zanon (2014) and analysing
gredients, as discussed on optimal steady state operation the case in which Assumption 4 does not hold is an
in Amrit et al. (2011), are adopted. A terminal region, interesting question for future investigations.
containing x∗i , is denoted by Xif , with the corresponding
Lemma 16. Suppose a continuous autonomous system
terminal control policy and penalty functions denoted by
x+ = f (x) has a periodic solution. Then this periodic
uif (x) and Vfi (x), respectively.
solution is Lyapunov asymptotically stable if and only if
Let us first make an assumption on the relationship it is orbitally asymptotically stable.
between Xif and uif (x):
Proof. According to Proposition 4.6 in Bredon (2013),
Assumption 5. There exists a family of compact sets i
Xif , i ∈ I[0,T −1] , each containing x∗i , such that for all the continuous autonomous system implies f (Xf ) is con-
i
x ∈ Xif , the solutions of x(t + 1) = f (x(t), ui+t nected, Thus, solutions initiated within any Xf ⊆ Xf will
f (x(t)))
converge exponentially to ΠX . evolve in phase with the optimal periodic solution which
means asymptotically stability and orbitally asymptoti-
Accordingly, the feedback law uif (x) which drives any cally stability are equivalent. 
state x ∈ Xif to Xi+1 takes values in Uif (x) := {u ∈ Corollary 17. If Assumption 3 holds, the optimal periodic
f
i+1 solution is orbitally asymptotically stable and asymptoti-
U(x) | f (x, u) ∈ Xf }. Then, the ith set of feasible state-
 cally stable in the sense of Lyapunov.
input pairs is given as Zif := x∈Xi [{x} × Uif (x)], and we
  f

let Xf = i Xif , Uf = i Uif (x). 4.5 Construction of terminal penalty function


With regards to the terminal penalty function Vfi (x) We propose a constructive formula for computing T -CSF
in period-T optimal operation, the concept of CSF in ∞

Definition 6 is adapted to give us a T -CSF as follows: i
V f (x) = L(xi (k), ui+k i
f (x (k))), (27)
Definition 13. A T -CSF is a family of continuous func- k=0
tions Vfi : Xif → R, i ∈ I[0,T −1] and the following holds with x (k) the solution of xi (k +1) = f (xi (k), ui+k
i i
f (x (k)))
min Vfi+1 (f (x, u)) − s(x, u) ≤ Vfi (x), ∀x ∈ Xif , (25) initiated at x (0) = x if x ∈ Xf .
i i
u∈Uif (x)
 To analyze the convergence of (27), we consider the dis-
where s : i∈I[0,T −1] Zif → R is the supply rate.
sipation inequalities along the optimal periodic solution
Notice that Vfi (x) is the ith component of a T -CSF defined so that one can derive λ(x∗i+1 ) − λ(x∗i ) = (x∗i , u∗i ) − .
Equivalently, the rotated stage cost at the periodic solution
for all x ∈ Xif . We can adopt the following convention:
L(x∗i , u∗i ) = 0, ∀i ∈ I[0,T −1] . (28)
Assumption 6. Vfi (x) is a bounded function if x ∈ Xif ,
Assumption 7. There exists a δL > 0 such that the func-
/ Xif .
whereas Vfi (x) := +∞, ∀x ∈ tions (·) and λ(·), and hence the rotated stage cost L(·)
are Lipschitz continuous for all (x, u) fulfilling |(x, u)|Π <
Then, for any point x ∈ Xf a single terminal penalty δL . Then, the corresponding Lipschitz constants are de-
function can be defined as: noted by L , Lλ and LL . Moreover, for all i ∈ I[0,T −1] ,
Vf (x) := min Vfi (x). (26) uif (x) is Lipschitz continuous with respect to x.
i
Lemma 14. The minimum of a T -CSF as in (25) and (26)
is a suitable CSF. Since Assumption 5 implies that there exists constant real
number A and |a| < 1 such that |(x(k), u(k))|Π ≤ A · |a|k .
Proof. Denote the T -CSF by {Vf0 , Vf1 , · · · , VfT −1 }, and Together with Assumption 7 and (28), we conclude that
let x ∈ Xf be arbitrary. Then, for any i∗ ∈ argmin Vfi (x), |L(x(k), u(k))| ≤ LL · A · |a|k , and therefore the T -CSF in
L ·A
∗ ∗ ∗ ∗
i

(27) is upper bounded by L1−a .
∀x ∈ Xif , Vfi +1 (f (x, uif (x))) ≤ Vfi (x) + s(x, uif (x))
Notice that, provided all the series converge for arbitrary
holds. Together with (26), it fulfills
∗ ∗ ∗
i ∈ I[0,T −1] , it is straightforward to have
Vf (f (x, uif (x))) ≤ Vfi +1
(f (x, uif (x))) i i+1
∗ ∗ V f (x) = L(x, uif (x)) + V f (f (x, uif (x))). (29)
≤ Vfi (x) + s(x, uif (x))

Thus, the proposed T -CSF follows Lemma 11 and the ter-
= Vf (x) + s(x, uif (x)). i
minal cost function can be selected as V f (x) = min V f (x).
∗ i
Therefore, there exists an admissible control policy uif (x) ∈
∗ To see the storage function λ(·) is not needed in designing
Uif (x) ⊆ Uf (x) such that
the EMPC controller, we consider the terminal cost

inf Vf (f (x, u)) − s(x, u) ≤ Vf (x), Vf (x) = min Vfi (x), (30)
u∈Uif (x)⊆Uf (x) i
and Vf (x) defined in (26) is a CSF.  where

523
Proceedings of the 20th IFAC World Congress
522
Toulouse, France, July 9-14, 2017 Zihang Dong et al. / IFAC PapersOnLine 50-1 (2017) 518–523


 Correspondingly, constraints of this linearized system are
i
Vfi (x) =V f (x) − λ(x) = L(xi (k), ui+k i
f (x (k))) − λ(x) − 0.3407 ≤ δx0 ≤ 0.6393, −2.9499 ≤ δu0 ≤ 0,
k=0 (38)
− 0.8918 ≤ δx1 ≤ 0.0882, −2.9499 ≤ δu1 ≤ 0.
∞ T
  −1
((xi (mT + k), ui+k i Since there is no [K0 K1 ]T being able to stabilize system
= f (x (mT + k))) − ).
m=0 k=0 (37) within the whole region in (38), we consider maxi-
(31) mizing the region of attraction of the linearized system
by solving a bilinear matrix inequality problem based
4.6 SOSTOOLs for storage function on the method in Vassilaki et al. (2004). The resulting
feedback gain is [K0 K1 ]T = [−5.2 10.75]T which is able
The MATLAB toolbox SOSTOOLs proposed in Prajna to stabilize the linear system for 0 ≤ δx0 ≤ 0.5681 and
et al. (2002) can be used to determine a candidate for λ(·) −0.2606 ≤ δx1 ≤ 0.
in polynomial form. Therefore, in this paper, we consider Next, we determine a potentially smaller set where the
the optimization problem as follows: above control policy works for the nonlinear system (34).
min  (32) If the terminal state feedback control law is implemented
,c0 ,c1 ,··· ,cP
into (34), the dynamic of state deviation becomes
subject to
P δx+ 3 2
0 = −10.75 δx1 − 12.6 δx1 − 2.23 δx1 ,
 (39)
λ(x) = c i xi , (33.1) + 3 2
δx1 = 5.2 δx0 − 5.513 δx0 − 0.0016 δx0 .
i=0 Trajectories described by these two equations admit a
λ(x) − λ(f (x, u)) + (x, u) −  is sum of squares (≥ 0), small region 0 ≤ δx0 ≤ 0.22 and −0.21 ≤ δx1 ≤ 0 such
∀(x, u) ∈ Z. that any interior point can be attracted to the origin with-
(33.2) out constraints violation. Therefore, the terminal region is
In this formulation, x and u are symbolic variables, Xf = X0f ∪ X1f = [0.3507, 0.5707] ∪ [0.6918, 0.9018].
whereas , ci are decision variables. (33.1) shows the poly- Then, according to the equations in (30) and (31), it is
nomial storage function of order P and (33.2) comes from sufficient to construct the terminal cost function as
the dissipation inequality. ∞ T −1
 
Vf (x) = min ((xi (mT +k), ui+k i
f (x (mT +k)))−),
5. EXAMPLES i
m=0 k=0
(40)
5.1 No gap:  =  - optimal period-2 operation
where ui+k
f (x i
(mT +k)) = u ∗
i+k +Ki+k (x i
(mT +k)−x∗
i+k ).

Consider the nonlinear system, known as logistic map, By SOSTOOLs, a storage function fulfilling dissipativity
described by the following difference euqation approximately is λ(x) = 0.30471x3 − 0.81183x2 + 1.2215x.
x+ = ux(1 − x) (34) Finally, Fig.1 shows the closed loop state transition and
in which x is the ratio between current population and input at initial condition x = 0.5 and prediction horizon
the maximum possible population with value in [0, 1] and N = 6. It can be seen that the state trajectory converges
a parameter u in interval (0, 4]. To avoid the trivial case to the optimal periodic solution in several steps with the
x = 0, we onlyconsider the robust control invariant set control input remaining at its upper bound. It also shows
X := [, 1 − ], x∈X U(x) := [ 1−
1
, 4(1 − )] and Z = X × the asymptotic average cost converges to  = .

x∈X U(x), where  is a small positive value.
1
The following stage cost
(x, u) = −x4
State

(35) 0.5
x

results in an optimal periodic regime of operation which Practical


Optimal
0
can be expressed as 0 5 10 15 20 25 30 35 40

∗ (5 − 4) ∓ (1 − 4)(5 − 4)
x0,1 = , u∗0,1 = 4(1 − ). 3.9
Input

8(1 − )
u

3.8
(36)
3.7
Asymptotic average cost

Particularly, when  = 0.01, the optimal periodic operation 0 5 10 15 20 25 30 35 40

is x∗0 ≈ 0.3507 and x∗1 ≈ 0.9018 with optimal control -0.1


u∗0 = u∗1 = 3.96, so that the best asymptotic average -0.2
ℓ∗av

l *av

cost for all initial states x ∈ X is  = ∗av = ∗av (x) = -0.3


(x∗ )4 +(x∗ )4
− 0 2 1 = −0.3382. 0 5 10 15 20 25 30 35 40
Time
To construct the terminal cost function, a terminal state
feedback is needed uif (x) = u∗i + Ki (x − x∗i ), i ∈ I[0,1] , Fig. 1. Closed-loop behaviour of the logistic map system
where Ki is determined from the linearized system of (34)
along the optimal solution that is
5.2 With gap:  >  - optimal period-4 operation
δx+
0 = −3.1821δx1 + 0.0886δu1
(37)
δx+
1 = 1.1821δx0 + 0.2277δu0 Consider a bidimensional nonlinear system of equations:

524
Proceedings of the 20th IFAC World Congress

Toulouse, France, July 9-14, 2017 Zihang Dong et al. / IFAC PapersOnLine 50-1 (2017) 518–523 523

 
1 − u2 2u Furthermore, a candidate of storage function is λ(x) =
 2 2  (xT Qx)2 . The system initialized at x = [0.48 0.64]T and
x+ =  1 + u 1 + u2  x (41) with a prediction horizon N = 6 is shown in Fig.2 in
−2u 1 − u
1 + u2 1 + u2 which we see that closed-loop system catches the periodic
solution after several control moves. The system average
where x is the state and u is the input. This nonlinear performance for this cost objectives convergences to ∗av (x)
system gives a state trajectory which rotates on a circle = -0.4096 instead of  = -1.
with radius that is equal to the magnitude of the initial
state. Thus, only states on this circle are reachable.
6. CONCLUSION
Consider the state space X := {|x| ≤ 1} and input
constraint u ∈ U := (−∞, 0], a stage cost In conclusion, this paper discusses a generalized approach
  for estimation of system asymptotic average performance
1 0
(x, u) = −(xT Qx)2 + (u + 1)2 , Q = (42) from above and below by means of the CSF and dissipation
0 −1
inequalities. Such tools are adapted to formulate EMPC
results into the global optimal periodic solution x∗0 = control schemes and eventually analyze their performance
[1, 0]T , x∗1 = [0, 1]T , x∗2 = [−1, 0]T and x∗3 = [0, −1]T . and stability. In the case of ”no-gap”, if the economic
MPC controller generates a continuous input, the optimal
However, due to the structure of reachable sets, any initial
periodic operation is Lyapunov asymptotically stable by
condition with distance to the origin |x| = r, 0 < r < 1
using the CSF as the terminal penalty function.
moves within a control invariant set X := {x | |x| = r}.
Then, a sub-optimal periodic operation is x∗0 = [r, 0]T ,
REFERENCES
x∗1 = [0, r]T , x∗2 = [−r, 0]T and x∗3 = [0, −r]T with optimal
control u∗0 = u∗1 = u∗2 = u∗3 = −1. Thus, there is a gap Amrit, R., Rawlings, J., and Angeli, D. (2011). Economic
between ∗av and , which is  = −1 ≤ ∗av (x) = −r4 . optimization using model predictive control with a ter-
minal cost. Annual Reviews in Control, 35(2), 178–186.
Since the terminal control policy is expected to be a
Angeli, D., Amrit, R., and Rawlings, J. (2012). On average
continuous function of state x, we consider uf (x), x ∈
performance and stability of economic model predictive
Xf = X \ { (2k+1)π
4 }, k ∈ R forcing the system to follow the control. IEEE transactions on automatic control, 57(7),
dynamic in polar coordinate, θ+ = θ+ π2 −K sin(4θ), where 1615–1626.
θ is the angle of the state vector with respect to the positive Bredon, G. (2013). Topology and geometry, volume 139.
horizontal axis and K is the state feedback gain. Since the Springer Science & Business Media.
optimal solution evolves π2 radius counterclockwise every Fagiano, L. and Teel, A. (2013). Generalized terminal state
step, K sin(4θ) reduces the bias between the practical state constraint for model predictive control. Automatica,
to the optimal ones. To keep in phase with the rotation of 49(9), 2622–2631.
periodic solution, K should be selected from (0, 14 ]. Grüne, L. and Zanon, M. (2014). Periodic optimal control,
dissipativity and mpc. MTNS, 1804–1807.
Then, with uf (·) defined above and K = 14 , the terminal Mayne, D., Rawlings, J., Rao, C., and Scokaert, P. (2000).
cost function for x ∈ Xf is Constrained model predictive control: Stability and op-
∞ T
  −1 timality. Automatica, 36(6), 789–814.
Vf (x) = ((x(mT + k), uf (x(mT + k))) − ). (43) Müller, M., Angeli, D., and Allgöwer, F. (2013). Economic
m=0 k=0 model predictive control with self-tuning terminal cost.
European Journal of Control, 19(5), 408–416.
Notice that, this terminal cost function will only be finite Müller, M., Angeli, D., and Allgöwer, F. (2014). On the
for x in {x| |x| = 1} and Lyapunov asymptotic stability performance of economic model predictive control with
discussed in Section 4 applies for {x| |x| = 1}. self-tuning terminal cost. Journal of Process Control,
24(8), 1179–1186.
Prajna, S., Papachristodoulou, A., and Parrilo, P.
Practical state Optimal state

0
(2002). SOSTOOLSSum of Squares Optimization Tool-
box, Users Guide.
x

-2 x1

-4
x2
Qin, S. and Badgwell, T. (2003). A survey of industrial
0 5 10 15 20 25 30 35 40
model predictive control technology. Control engineer-
0
ing practice, 11(7), 733–764.
x

-2 x1
x2 Rawlings, J., Angeli, D., and Bates, C. (2012). Fundamen-
-4
0 5 10 15 20 25 30 35 40 tals of economic model predictive control. 2012 IEEE
-0.5
51st IEEE Conference on Decision and Control (CDC),
Asymptotic average cost Input

-1
3851–3861.
u

-1.5
0 5 10 15 20 25 30 35 40
Rawlings, J. and Mayne, D. (2009). Model predictive
1
ℓ∗av
control: Theory and design. Nob Hill Pub.
ℓ Vassilaki, M., Hennet, J., and Bitsoris, G. (2004). Feed-
av

0
l*

-1 back control of linear discrete-time systems under state


0 5 10 15 20 25 30 35 40 and control constraints. International Journal of con-
Time
trol, 47(6), 1727–1735.
Fig. 2. Closed-loop behaviour of the circle system

525

You might also like