You are on page 1of 11

51st IEEE Conference on Decision and Control

December 10-13, 2012. Maui, Hawaii, USA

Fundamentals of Economic Model Predictive Control


James B. Rawlings, David Angeli and Cuyler N. Bates

Abstract— The goal of most current advanced control systems The hierarchical partitioning of information between these
is to guide a process to a target setpoint rapidly and reliably. layers has significant effects whenever the process deviates
Model predictive control has become a popular technology in from its setpoint. The objective function used by the con-
many applications because it can handle large, multivariable
systems subject to hard constraints on states and inputs. The troller is usually shaped to achieve fast asymptotic tracking
optimal steady-state setpoint is usually provided by some other to setpoint changes and low output variance in the face of
information management system that determines, among all disturbances, and is usually unrelated to the economic costs
steady states, which is the most profitable. For an increasing of operating the plant. For some industries, the vast majority
number of applications, however, this hierarchical separation of a plant’s operating time is spent near a constant setpoint. In
of information and purpose is no longer optimal or desirable.
A recently proposed alternative to the hierarchical decompo-
these cases, the transient costs stemming from economically
sition is to take the economic objective directly as the objective suboptimal controller action can be ignored [5]. Even in
function of the control system. In this approach, known as processes for which the best operating regime fails to be an
economic MPC, the controller optimizes directly in real time equilibrium, this hierarchical approach may still be applied;
the economic performance of the process, rather than tracking see, for example, optimal periodic control [6]–[8]. In this
to a setpoint. The purpose of this tutorial is to explain how
to design these kinds of control systems and what kinds of
case, an economically optimal periodic trajectory is com-
closed-loop properties one can achieve with them. We cover puted on a slow time scale, and then a controller is designed
the following issues: asymptotic average performance; closed- to asymptotically guide the system to that trajectory.
loop stability and convergence, strong duality and dissipativity;
designing terminal costs, terminal regions, and terminal pe- II. M OTIVATION FOR ECONOMIC MPC
riodic constraints. Several examples are included to illustrate
While the classical approach has shown great versatility
these results.
and has seen widespread application, there are an increasing
number of problems for which dynamic economic perfor-
I. I NTRODUCTION mance is crucial and the hierarchical separation of economic
The goals of advanced, large-scale feedback control sys- analysis and control is either inefficient or inappropriate.
tems are typically economic ones. Issues like profitability, re- The general reason for this proliferation is that the time
turn on capital, margins, efficiency of operation, cost cutting, scales on which continuous chemical processes react to
lean operation, etc. dominate the performance discussion economic demands has shrunk as industries have matured
at this level of performance assessment. In the process and products have diversified; flexibility and efficiency are
industries especially, but also in many other applications, the requirements the control system must now address [9]. At the
current paradigm for achieving overall economic objectives same time, the power of software and hardware available
is to partition the information management, decision making, for optimization has grown to the point where nonlinear
and control system into two layers [1]–[3]. The first layer, problems with thousands of variables can now be solved in
often referred to as real-time optimization (RTO), performs a seconds and used for real time process control [10], [11].
steady-state economic optimization of the plant’s variables, Traditionally, several methods have been used to improve
updated on a timescale of hours or days. The RTO sends the the economic performance of control systems without di-
results of its optimization as a setpoint to the second layer, rectly providing economic information to the controller. The
usually referred to as the advanced process control system. It two earliest methods for increasing the economic perfor-
is the task of this layer to guide the plant’s transient state to mance of control are minimum variance control and eco-
the setpoint and, once arrived, to reject dynamic disturbances nomic target selection. Minimum variance control involves
that enter the system. This process control layer is often minimizing the variance around the setpoint so that the
implemented with model predictive control (MPC) because controller can safely operate closer to the plant’s constraints,
of MPC’s flexibility, performance, robustness, and its ability increasing profit [12]. The target selector, used in conjunction
to directly handle hard constraints on both inputs and states with MPC, is required because the RTO and MPC layers use
[4]. different models with different disturbance information; the
setpoint given by the RTO layer need not always be feasible
J. B. Rawlings and C. N. Bates are with Dept. of Chemical and to the controller. By linearizing the steady-state economic
Biological Engineering, Univ. of Wisconsin, Madison, WI 53706, USA model of the RTO around the desired operating point, a
(rawlings@engr.wisc.edu, cnbates@wisc.edu) linear program can be solved to find the economically best
D. Angeli is with Dept. Electrical and Electronic Engineering, Imperial
College London, U.K. (d.angeli@imperial.ac.uk) and with the feasible target for the controller based off of an infeasible
Dip. Sistemi e Informatica, University of Florence, Italy. RTO setpoint [13]. These approaches, however, both keep
978-1-4673-2066-5/12/$31.00
978-1-4673-2064-1/12/$31.00 ©2012
©2012 IEEE
IEEE 3851
the same fundamental separation of information between the analysis on periodic processes have been done by [32], [33]
steady-state optimization and control layers. and computational studies of distributed economic model
There are several proposals to improve the effective use of predictive controllers have been published in [34].
dynamic and economic information throughout the hierarchy.
One is to move dynamic information into the RTO layer. III. M ATHEMATICAL BACKGROUND AND ANALYSIS OF
Termed dynamic real time optimization, a dynamic economic AVERAGE ASYMPTOTIC PERFORMANCE
problem is solved on a slow time scale to provide the We begin by comparing the forms, properties and con-
underlying regulatory layer with a target trajectory instead of sequences of the tracking and economic objective functions
a target steady state [14], [15]. The alternative approach is to as they are used in the model predictive control framework.
move economic information into the control layer; when the The tracking objective function, used in what we refer to as
control problem is posed as an optimization problem, such standard MPC, is designed to ensure asymptotic stability of
as in MPC, this approach involves modifying the traditional the desired steady state. This is accomplished by choosing
tracking objective function. One method is to include the the stage cost to be minimal at the best feasible pair of steady
economic model as a terminal cost, that is, the economic state and input, denoted (xs , us ).
model is evaluated at the predicted final state and input and Assumption 1: Standard MPC cost function
a multiple of the result is added to the tracking objective [16].
This has been shown to be effective, but an RTO layer is still 0 = `(xs , us ) ≤ `(x, u) for all admissible (x, u) (1)
required to guide the system to a new operating point, and the where ` : X × U → R is the stage cost, X is the set of
controller still has no knowledge of its economic effects prior admissible states and U is the set of admissible inputs.
to the end of the horizon. Another method is to intentionally This stage cost need not bear any functional relationship to
give the optimizer an unreachable setpoint with no terminal the operating cost of the plant and, in fact, is usually chosen
constraint [17], [18]. Of interest here, however, is the method to be quadratic. The weights to the quadratic objective can
of economic model predictive control, which replaces the be used to prioritize target tracking on some states more than
tracking stage cost function directly with the economic stage others or change the size of the control action for a given
cost function used in the RTO layer (or elsewhere) [19], [20]. tracking error.
Thus, the controller directly and dynamically optimizes the In economic MPC, the operating cost of the plant is used
economic operating cost of the process, doing so without directly as the stage cost in the MPC objective function.
reference to any steady state. Ostensibly, this renders the Since the cost of operating the plant need not be zero at the
RTO layer irrelevant, although in practice, including steady optimal steady-state operating conditions, (1) cannot hold
state information from the controller’s model may still be in general. More importantly, it may happen that `(x, u) <
desirable. `(xs , us ) for some feasible pair (x, u) that is not a steady
The idea of optimizing dynamic economic performance state. As a simple example, consider a CSTR with a first
directly is not new. Infinite horizon control problems with un- order reaction; the reactant inflow is the input and the
bounded costs were first considered in the field of economics product outflow is the output. If a linear cost is assigned
in the 1920s [21]. For discussion on the existence of optimal to the reactant and a linear profit is assigned to the product,
control trajectories and their convergence to steady state given any steady state (xs , us ) we can immediately find a
under various conditions the reader is referred to [22]. For more profitable transient state, namely (xs , 0). Clearly, this
nonlinear systems with constraints and disturbances it is usu- situation is not some rare pathology.
ally infeasible to implement infinite horizon control in real To begin analysis, we consider a discrete-time finite-
time. Control involving economic optimization over a finite dimensional nonlinear system of the form
horizon has, until recently, been primarily used in the design
and operation of batch systems; see, for an example, [23]. x+ = f (x, u) (2)
Theoretical analysis of economic model predictive control
with state x ∈ X ⊆ Rn , input u ∈ U ⊂ Rm and state-transition
for continuous processes began with stability proofs based on
map f : X × U → X. The objective function is given as the
convexity for linear systems and convex objectives [18], [19].
sum of stage costs:
Average asymptotic cost guarantees and average constraints
were then demonstrated in [24] and [20]. Lyapunov-based ∑ `(x(k), u(k)) (3)
stability proofs for nonconvex systems were provided in [25] k
and expanded in [26] and [27]. The design of economic subject to the model (2) as well as pointwise constraints:
model predictive controllers using Lyapunov techniques is
presented in [28] and an analysis of performance in the (x(k), u(k)) ∈ Z k ∈ I≥0 (4)
absence of any terminal constraints or penalties is presented for some compact, time-invariant set Z ⊆ X × U. Here, we
in [29]. Since its inception, economic MPC has been applied avoid considering the infinite horizon problem [22] and
to a small but growing variety of continuous processes. With define the finite horizon equivalent:
linear models and economics, it has been applied to building
N−1
heating and cooling and power management on a smart grid VN (x, u) =
[30], [31]. For nonlinear systems, computational studies and
∑ `(x(k), u(k)) (5)
k=0

3852
where u = [u(0), u(1), . . . , u(N −1)] and x(0) = x. The set I≥0 Having set the mathematical foundation, we now turn
in (4) denotes the non-negative integers, while I0:N refers to attention to one of the most striking features of economic
the set {0, 1, . . . , N}. This finite horizon optimization problem MPC: without considering stability, we demonstrate that
is solved in a rolling horizon fashion, applying the first the (nominal) average asymptotic performance of economic
element of the optimal control sequence to the plant, waiting MPC is at least as good as that of the best admissible steady
for the next state measurement or estimate and then solving state. The theorem and its proof, as first presented in [24],
the problem again by using the new initial state. Formally, are as follows:
at each discrete time k, the following nonlinear program is Theorem 1: Let x(0) ∈ XN . Then there exists at least one
solved: admissible control sequence that steers the state to xs at time
N without leaving XN and the closed-loop system (2) and
min VN (x, u)
u
 + (10) has an asymptotic average performance that is at least
 x = f (x, u) as good as the best admissible steady state.
subject to (x(k), u(k)) ∈ Z k ∈ I0:N−1 (6) Proof: Let x ∈ XN . If u is a feasible control sequence

x(N) = xs , x(0) = x at time k for initial state x and u(0) is applied to the plant,
then {u(1), . . . , u(N − 1), us } is a feasible control sequence at
where we have added the terminal constraint x(N) = xs .
time k + 1, from initial state f (x, u(0)) because the problem
We define the admissible set ZN as the set of (x, u) pairs
formulation (6) includes a terminal constraint. By induction,
satisfying the constraints:
feasibility of (6) follows for all non-negative times, for all
ZN = {(x, u) | ∃ x(1), . . . , x(N) : x+ = f (x, u), x ∈ XN .
As is usually derived in stability proofs of MPC with a
(x(k), u(k)) ∈ Z, ∀k ∈ I0:N−1 , x(N) = xs , x(0) = x} (7)
terminal constraint (see, for example [35, pp. 115-116,145-
Next, we define the set of admissible states XN as the 146]), the following inequality holds along solutions of the
projection of ZN onto X: closed-loop system:

XN = {x ∈ X | ∃ u such that (x, u) ∈ ZN } (8) VN0 (x+ ) −VN0 (x) ≤ `(xs , us ) − `(x, u) (11)

Finally, an input sequence u = {u(0), u(1), . . . u(N − 1)} is where VN0 (x)denotes the optimal cost of (6). Taking asymp-
called feasible for initial state x if (x, u) ∈ ZN . With these totic averages of both sides of (11), we compute:
definitions, we formally define the optimal steady state as ∑Tk=0 VN0 (x(k + 1)) −VN0 (x(k))
the pair (xs , us ) that satisfies lim inf
n o T →+∞ T +1
T
`(xs , us ) = min `(x, u) | (x, u) ∈ Z, x = f (x, u) (9) ∑ `(x s s ) − `(x(k), u(k))
, u
x,u ≤ lim inf k=0
T →+∞ T +1
Strictly speaking (9) could be fulfilled by more than one ∑Tk=0 `(x(k), u(k))
state-control pair, but, for the sake of simplicity, we assume = `(xs , us ) − lim sup (12)
T →+∞ T +1
it to be unique. The optimization problem (6) defines an
implicit feedback law κN : XN → U which we denote Assuming without loss of generality that `(x, u) is nonnega-
tive:
u = κN (x) = u0 (0; x) x ∈ XN (10)
∑Tk=0 VN0 (x(k + 1)) −VN0 (x(k))
lim inf
where u0 (x) is the optimal solution of (6) for initial state x, T →+∞ T +1
and u0 (k; x) denotes this optimal solution at time k ∈ I0:N−1 VN0 (x(T + 1)) −VN0 (x(0))
in the horizon. Since we can handle the case of multivalued = lim inf (13)
T →+∞ T +1
optima by arbitrarily assigning a constant selection map,
VN0 (x(0))
we assume without loss of generality that u0 (x) is uniquely ≥ lim inf − =0
T →+∞ T +1
defined. Note that since we have not yet invoked Assumption
1, the control framework defined by (6) and (10) applies to We complete the proof by combining (12) and (13):
both standard and economic MPC. Finally, to guarantee the ∑Tk=0 `(x(k), u(k))
existence of a solution to the optimal control problem (6) we lim sup ≤ `(xs , us )
T →+∞ T +1
require the following assumption [26].
Assumption 2 (Model, cost, and admissible set): It should be noted that this result applies only to the asymp-
totic average cost; transient averages up to some arbitrary
time may take any value.
1) The model f (·) and stage cost `(·) are continuous. The
admissible set XN contains xs in its interior. IV. S TABILITY AND CONVERGENCE ANALYSIS
2) There exists γ : Rn → R≥0 of class K∞ such that for
each x ∈ XN there exists a feasible u, with We now turn attention to the second major consequence of
the economic objective function: while asymptotic economic
|u − [us , . . . , us ]T | ≤ γ(|x − xs |) performance is guaranteed, stability is not. A common ap-

3853
proach to prove the stability of standard MPC is to use the Regardless of λ (·), we have
optimal cost VN0 (x) as a Lyapunov function for the closed-
min L(x, u) ≤ min L(x, u)
loop system. Together, the inequalities (1) and (11) imply (x,u)∈Z (x,u)∈Z,x= f (x,u)
VN0 (x+ ) ≤ VN0 (x), that is, the optimal cost is monotonically = min `(x, u) = `(xs , us )
decreasing along solutions of the closed loop system. Under (x,u)∈Z,x= f (x,u)
mild additional conditions, one can use Lyapunov arguments Combining this with (17),
to prove the asymptotic stability of standard MPC.
If Assumption 1 does not hold, these stability arguments max min `(x, u) + λ̄ 0 (x − f (x, u))
λ̄ ∈Rn (x,u)∈Z
fail. The optimal cost function VN0 (x) need not be monotoni-
cally decreasing even if the system is stable. More fundamen- = min `(x, u)
(x,u)∈Z,x= f (x,u)
tally, for general nonlinear systems and cost functionals, it
we can see that condition (17) represents the absence of
is not even guaranteed that xs is an equilibrium point of the
a duality gap. Of note is the fact that strong duality is
closed loop system (2)–(10). Since there exists (x, u) such
a sufficient condition for dissipativity, and dissipativity is
that `(x, u) < `(xs , us ), it may be the case that the optimal
itself a sufficient condition for optimal operation at steady
trajectory from xs at time 0 to xs at time N is different than
state, in the sense that any feasible trajectory has an average
x(k) = xs for all k ∈ I0:N . These problems may be ignored
asymptotic cost no better than the best steady state [26].
for the case of linear systems with convex constraints and
a convex economic objective function, as shown in [18].
However, to prove the stability of economic MPC without A. Terminal Constraint
such harsh restrictions, we must make one crucial additional
assumption: dissipativity [26]. We now state and prove stability for the formulation (6),
Definition 1: The system (2) is dissipative with respect which includes a terminal constraint.
to the supply rate s : X × U → R if there exists a function Theorem 2: Suppose Assumption 2 holds and consider the
λ : X → R such that: closed-loop system defined by (2) and (10), where (xs , us ) is
the best feasible steady-state pair as defined in (9). If the
λ ( f (x, u)) − λ (x) ≤ s(x, u) (14) system (2) is strictly dissipative with respect to the supply
for all (x, u) ∈ Z ⊆ X×U. If in addition ρ : X → R≥0 positive rate:
definite1 exists such that: s(x, u) = `(x, u) − `(xs , us ) (19)

λ ( f (x, u)) − λ (x) ≤ −ρ(x) + s(x, u) (15) then xs is an asymptotically stable equilibrium point of the
closed-loop system with region of attraction XN .
then the system is said to be strictly dissipative. Proof: We begin the proof by defining the auxiliary
Before proceeding to the resulting stability proof, it is optimization problem
worthwhile to investigate the meaning of dissipativity. First,
let us consider the supply rate s(x, u) = `(x, u) − `(xs , us ). min ṼN (x, u)
u
The assumption of dissipativity is then equivalent to the  +
 x = f (x, u)
following: there exists λ : X → R such that subject to (x(k), u(k)) ∈ Z k ∈ I0:N−1 (20)
x(N) = xs , x(0) = x

min `(x, u) + λ (x) − λ ( f (x, u)) ≥ `(xs , us ) (16)
(x,u)∈Z
where the auxiliary objective function is defined as
With this supply rate, dissipativity depends on both the
N−1
system to be considered, f (·), and the stage cost function, ṼN (x, u) = (21)
`(·). If we were to choose `(·) to be positive definite around
∑ L(x(k), u(k))
k=0
(xs , us ) as is the case in standard MPC, note that equation
and the rotated stage cost is defined by (18). Since the only
(16) would be trivially satisfied for any system f (·) by
difference between (6) and (20) is the objective function,
choosing λ (x) = 0.
their feasible sets, XN , coincide. We next claim that the op-
Suppose now we restrict ourselves to linear Lyapunov-like timal trajectory u0 (x) of the rotated formulation is identical
functions λ (x) = λ̄ 0 x. Then, condition (16) becomes to the original formulation. We compute:
min `(x, u) + λ̄ 0 (x − f (x, u)) ≥ `(xs , us ) (17) N−1
(x,u)∈Z
ṼN (x, u) = ∑ `(x(k), u(k)) + λ (x(k)) − λ ( f (x(k), u(k)))
This equation is known as strong duality in the context of k=0
infinite horizon optimal control [22]. To see why, we define N−1

the rotated stage-cost = ∑ `(x(k), u(k)) + λ (x(k)) − λ (x(k + 1))


k=0
L(x, u) = `(x, u) + λ (x) − λ ( f (x, u)) (18) N−1
= λ (x(0)) − λ (x(N)) + ∑ `(x(k), u(k))
k=0
1A function is positive definite with respect to some point xs ∈ X if it is
continuous, ρ(xs ) = 0 and ρ(x) > 0 for all x 6= xs . = λ (x(0)) − λ (xs ) +VN (x, u)

3854
Since the objective functions of the original and rotated that
formulations differ by only a constant on feasible solutions
V f ( f (x, κ f (x))) ≤ V f (x) − `(x, κ f (x)) + `(xs , us ) (28)
and the constraints are the same, we conclude that the
solutions of (6) and (20) are equal. (x, κ f (x)) ∈ ZN,p ∀x ∈ X f
Next, using (16) and the definition of the rotated stage In this assumption, we implicitly also require the set X f to
cost (18), we note that be invariant under the control law u = κ f (x).
While Assumption 3 is identical to the assumption re-
L(xs , us ) = `(xs , us ) ≤ min L(x, u) (22) quired to prove the convergence of standard MPC with a
(x,u)∈Z
terminal cost and region, we note that in the economic case,
where we assume that L(xs , us ) = 0 without loss of generality. V f (x) need not be positive definite with respect to xs . We
Equation (22) is, in fact, the key point of the proof: we now state the companion to Theorem 2.
have used dissipativity to show that the augmented problem Theorem 3: Suppose Assumptions 2 and 3 hold and con-
(20) satisfies Assumption 1, and we may now use the sider the closed-loop system defined by (2) and (27), where
same stability arguments as for standard MPC. Specifically, (xs , us ) is the best feasible steady-state pair as defined in
the cost-to-go ṼN0 (x) serves as a Lyapunov function of the (9). If the system (2) is strictly dissipative with respect
closed-loop system. This optimal cost is continuous due to to the supply rate (19) then xs is an asymptotically stable
Assumption 2, positive definite with respect to xs and, due equilibrium point of the closed-loop system with region of
to strict dissipativity, strictly decreasing: attraction XN .
ṼN0 (x+ ) ≤ ṼN0 (x) + L(xs , us ) − L(x, u) Proof: We begin the proof by defining the equivalent
(23)
≤ ṼN0 (x) − ρ(x) auxiliary optimization problem
We conclude that xs is an asymptotically stable equilibrium min ṼN,p (x, u)
u
of the closed-loop system with region of attraction XN .  +
 x = f (x, u)
subject to (x(k), u(k)) ∈ Z k ∈ I0:N−1 (29)
x(N) ∈ X f , x(0) = x

B. Terminal Cost and Terminal Region
where the new auxiliary objective function is defined as
We now consider the stability of a new system, defined N−1
by the following optimization problem: ṼN,p (x, u) = ∑ L(x(k), u(k)) + Ṽ f (x(N))
k=0
min VN,p (x, u) We also define the rotated terminal penalty:
u
 +
 x = f (x, u) Ṽ f (x) = V f (x) + λ (x) − λ (xs )
subject to (x(k), u(k)) ∈ Z k ∈ I0:N−1 (24)

x(N) ∈ X f , x(0) = x Once again, the only difference between (24) and (29) is
the objective function, and so their feasible sets, XN,p ,
Instead of a terminal equality constraint, we only require
coincide. The optimal trajectory u0 (x) is also identical for
that x(N) be a member of a compact terminal region X f
each formulation. We compute:
which contains xs in its interior. The admissible set of (x, u)
N−1
becomes
ṼN,p (x, u) = ∑ `(x(k), u(k)) + λ (x(k))
+ k=0
ZN,p = {(x, u) | ∃ x(1), . . . , x(N) : x = f (x, u),
− λ ( f (x(k), u(k))) + Ṽ f (x(N))
(x(k), u(k)) ∈ Z, ∀k ∈ I0:N−1 , x(N) ∈ X f , x(0) = x} (25)
N−1
while the feasible set XN,p remains the projection of ZN,p = ∑ `(x(k), u(k)) + λ (x(k))
k=0
onto X. In addition, we augment the objective function:
− λ (x(k + 1)) +V f (x(N)) + λ (x(N)) − λ (xs )
N−1
N−1
VN,p (x, u) = ∑ `(x(k), u(k)) +V f (x(N)) (26)
k=0
= λ (x(0)) − λ (xs ) + ∑ `(x(k), u(k)) +V f (x(N))
k=0
where V f : X f → R is a penalty on the terminal state. As = λ (x(0)) − λ (xs ) +VN,p (x, u)
before, the optimization problem (24) defines an implicit
control law Since the objective functions of each formulation differ by
only a constant and the constraints are the same, we conclude
u = κN,p (x) = u0 (0; x) x ∈ XN,p (27) that the solutions of (24) and (29) are identical.
In addition to Assumption 2, we require one additional We now note the following: the pair (Ṽ f , L) has the
assumption on the terminal region and terminal penalty. property
Assumption 3 (Terminal penalty stability assumption): Ṽ f ( f (x, κ f (x))) ≤ Ṽ f (x) − L(x, κ f (x)) + L(xs , us ) ∀x ∈ X f
There exists a terminal region control law κ f : X f → U such (30)

3855
if and only if the pair (V f , `) satisfies Assumption 3. This Theorem 4 says that we can recover stability by adding a
can be shown by adding λ ( f (x, κ f (x))) + λ (x) to both sides sufficiently large positive definite penalty to the stage cost.
of (3) and rearranging: In theory, we may search over all continuous functions λ (·)
and all positive definite functions α(·) to find the one that
Ṽ f ( f (x, κ f (x))) − Ṽ f (x) ≤ −(`(x, κ f (x)) + `(xs , us )
maximizes the economic performance of the controller while
− λ ( f (x, κ f (x))) − λ (x)) satisfying (33). The key benefit of this tuning method is
= −L(x, κ f (x)) + L(xs , us ) that the tracking term α(·) restores the flexibility that was
lost by treating the controller objective function as fixed
Having demonstrated that the pair (Ṽ f , L) satisfies As- entirely by the economic model. By changing the weight
sumption 3 and noting that dissipativity once again guaran- of the tracking term, the emphasis the controller places on
tees that (22) holds, one can use the same proof techniques stability and economic profit changes. For weights which
used for terminal regions in standard MPC (see, for instance produce a closed-loop stable system, decreasing the tracking
[27]) to show that weight increases economic performance but may, though not
0 (x+ ) ≤ Ṽ 0 (x) + L(x , u ) − L(x, u)
ṼN,p necessarily, also increase the time taken to approach steady
N,p s s
0 (x) − ρ(x) (31) state.
≤ ṼN,p
We conclude that xs is an asymptotically stable equilibrium V. E XTENSIONS
of the closed-loop system with region of attraction XN .
We now discuss two extensions of the basic theory of
C. Enforcing Convergence
economic MPC that are relevant to processes not operated
The stability of economic MPC relies on the assumption at steady state. The first is an application to systems with
of dissipativity. Dissipativity itself is a property of both the precomputed optimal periodic solutions. As was previously
process and the objective function. For model predictive the case for systems with steady state targets, asymptotic
control, the process is fixed but the objective function is average cost is guaranteed to be at least as good as the
available to tune the performance of the controller. For optimal periodic solution using a terminal constraint formu-
economic model predictive control, the objective function lation, while asymptotic convergence to that optimal periodic
is fixed as well because it comes from an economic model solution requires an extra assumption. The second extension
of the plant. It is therefore prudent to ask how an economic to the theory is the inclusion of average constraints. For
model predictive controller should be tuned. In particular, stable closed-loop operation, average constraints do not need
how can dissipativity, and therefore stability, be achieved for to be considered by the control layer, as all average values
a process which is unstable when operated with controller are equal to their steady states and can be handled by steady
(10). One tuning method to achieve dissipativity is to modify state optimization. The motivation for considering average
the stage cost of the objective function by adding a function, constraints here is that constraints on state and input averages
α : X × U → R≥0 , that is positive definite around the desired of unsteady closed-loop processes cannot be dealt with in
equilibrium: such a manner, and should therefore be provided directly to
¯ u) = `(x, u) + α(x, u)
`(x, (32) the controller (see, for an example, [19]).
To satisfy strict dissipativity, Definition 1, we require the
existence of a continuous function λ : X → R such that for A. Periodic Terminal Constraint
all (x, u) ∈ Z Here, we consider a process for which an optimal periodic
solution of period Q has been computed offline. We denote
α(x, u) ≥ λ (x) − λ ( f (x, u)) + ρ(x) − `(x, u) + `(xs , us ) (33)
this optimal periodic solution as (x∗ (k), u∗ (k)), k ∈ I0:Q−1 .
In particular, we can define h(x, u) := λ̄ 0 (x− f (x, u))+ρ(x)− We modify the optimization problem of (6) to include a
`(x, u) + `(xs , us ) and compute periodic terminal constraint.
h̄(r) = max h(x, u) min VN,Q (x, u, q)
(x,u)∈Z u
 +
|(x,u)−(xs ,us )|≤r  x = f (x, u)
where we are guaranteed that the maximum exists because subject to (x(k), u(k)) ∈ Z k ∈ I0:N−1 (34)
x(N) = x∗ (q), x(0) = x

Z is compact and ρ(·), `(·) and α(·) are continuous. By
choosing α(x, u) = h̄(|(x, u) − (xs , us )|) we can guarantee that The optimization problem now varies with time through the
the system is strictly dissipative. By invoking Theorem 2, we index of the periodic trajectory reference, q = t mod Q ∈
have the following result. I0:Q−1 . We denote the optimal solution for initial state x and
Theorem 4: Consider the nonlinear closed-loop system target index q as u0 (x, q). The resulting control law is
defined by (6) and (10) and let (xs , us ) be the best feasible
steady-state pair as defined in (9). If the stage cost is chosen κN,Q (x, q) = u0 (0; x, q)
according to (32) with α(x, u) = h̄(|(x, u) − (xs , us )|), then xs
which is applied in a rolling horizon manner by selecting
is an asymptotically stable equilibrium point of the closed-
loop system with region of attraction XN .

u(x,t) = κN,Q x,t mod Q t ∈ I≥0 (35)

3856
We now state the main result of this section. (38) is a generalization of the usual notion of an average
Theorem 5: Consider the closed-loop system defined by ∑tk=0 v(k)
(2) and (35). The average asymptotic performance of the Av∗ [v] = lim
t→+∞ t +1
closed-loop system fulfills:
in the sense that when Av∗ [v] exists, the set Av[v] is a
Q−1
∑Tk=0 `(x(k), u(k)) ∑k=0 `(x∗ (k), u∗ (k)) singleton equal to Av∗ [v]. If Av∗ [v] does not exist (see, for
lim sup ≤ (36)
T →+∞ T +1 Q an example, [26]), then Av[v] is a nonempty, nonsingleton
set.
Proof: We begin by noting that recursive feasibility
of the optimization problem (34) can be argued in exactly We define the auxiliary variable y
the same way as it was in the proof of Theorem 2 and y = h(x, u)
omit the details for brevity. Denoting as VN0 (x, q) the cost
corresponding to the optimal input u0 (x, q)), we define the where h : Z → R p is arbitrary, but continuous. The average
following Lyapunov-like function constraint set is defined as Y ⊆ R p and we require it to
contain h(xs , us ) and be closed and convex. The purpose of
V (x,t) := VN0 (x,t mod Q) including average constraints in the economic MPC formu-
The equivalent inequality to (11), developed in the same lation is to ensure that
manner, is Av[y] ⊆ Y (39)
 −V(t, x(t)) 
V(t + 1, f (x(t), u(t)))
∗ ∗ (37) The optimization problem to be solved is
≤ ` x (τ(t)), u (τ(t)) − ` x(t), u(t)
min VN,A (x, u,t)
where we use the shorthand τ(k) := k mod Q. Proceeding as u
 +
in the proof of Theorem 1, we compute 
 x = f (x, u)
(x(k), u(k)) ∈ Z k ∈ I0:N−1

V (T + 1, x(T + 1)) −V (0, x(0)) subject to (40)
0 = lim inf x(N) = xs , x(0) = x
T →+∞ T +1 

 N−1
T ∑k=0 h(z(k), v(k)) ∈ Yt
∑ V (k + 1, f (x(k), u(k))) −V (k, x(k))
= lim inf k=0 Here, Yt is a time-varying constraint set defined recursively
T →+∞ T +1
∑Tk=0 `(x∗ (τ(k)), u∗ (τ(k))) − `(x(k), u(k)) by
≤ lim inf
T →+∞ T +1 Yi+1 = Yi ⊕ Y h(x(i), u(i)) for i ∈ I≥0 (41)
∑Tk=0 `(x∗ (τ(k)), u∗ (τ(k)))
= lim inf The binary set operators ⊕ and are defined by
T →+∞ T +1
T
∑ `(x(k), u(k)) V⊕W = {z = v + w | v ∈ V, w ∈ W}
− lim sup k=0 V W = {z | {z} ⊕ W ⊆ V}
T →+∞ T +1
Q−1
∑q=0 `(x∗ (q), u∗ (q)) ∑T `(x(k), u(k)) The recursion is initialized by
= − lim sup k=0
Q T →+∞ T +1 Y0 = NY ⊕ Y00 (42)
with Y00 ⊂ R p an arbitrary compact set containing the origin.
If stability, in the sense of asymptotic convergence to As before, the optimization (40) defines an implicit control
the optimal periodic solution, is desired, then an addi- law κN,A (x,t) from which the input is selected.
tional assumption must be made. This assumption is dis-
sipativity with respect to the time varying supply rate u = κN,A (x,t) = u0 (0; x,t) (43)
`(x, u) − `(x∗ (q), u∗ (q)). A rigorous statement and proof, We are now ready to state the main result of this section.
using slightly different arguments, is presented in [32].
Theorem 6: Provided x(0) is feasible, the closed-loop
system (2) and (43) has an average performance that is at
B. Average Constraints least as good as the best admissible steady state and satisfies
the average constraint (39) asymptotically.
For a given, bounded signal v : I≥0 → Rnv we define its
average, denoted Av[v], as the set of accumulation points of Proof: The arguments for feasibility and asymptotic
the averaged partial sums of the signal: average performance are again very similar to those for
t
Theorem 1 and are presented in [26]. The proof for the
∑ n v(k) asymptotic satisfaction of the average constraint is as follows.
Av[v] = v̄ ∈ Rnv | ∃tn → +∞ : lim k=0

= v̄ (38)
n→+∞ tn + 1 Combining (41) and (42) produces the following result for
Since v is bounded, the average of its partial sums is also Yt :
t−1
bounded. Bounded signals must have accumulation points Yt = Y00 ⊕ (t + N)Y ∑ y(k)
and therefore Av[v] is guaranteed to be nonempty. Note that k=0

3857
By (40), we then have sublevel set L≤0 = {(x, u) : `(x, u) ≤ 0}. Since this periodic
N−1 t−1 solution outperforms the best steady state, we can conclude
∑ h(x(k), u(k)) + ∑ y(k) ∈ Y00 ⊕ (t + N)Y from the converse of Theorem 2 that dissipativity does not
k=0 k=0 hold for this system. While there is no algebraic test to
Since Z is compact and h(·) is continuous, h(x(t), u(t)) is directly prove that this system is not dissipative for small
bounded independently of t. Therefore, α we can directly prove that it is dissipative for large α.
First, however, we test the system for strong duality. That
∑tk=0 y(k) Y00 ⊕ (t + 1 + N)Y is, we seek a scalar λ such that
lim ∈ lim
t→+∞ t +1 t→+∞ t +1
=Y λ x+ − λ x ≤ `(x, u) (45)

This completes the proof. Since the left-hand side of (45) is a linear function of (x, u)
As an alternative to adding a tracking term to the objective and `(x, u) has a saddle point at (0, 0), we conclude that this
function of economic MPC to guarantee stability, an average system does not satisfy strong duality.
constraint of the form Next, we seek to establish dissipativity by evaluating a
specific candidate storage function, λ (x) = kx2 . By (16),
y = h(x, u) = |x − xs |2 Av[y] ∈ {0} dissipativity holds if there exists a k and ε > 0 so that:
can be added to the problem. This procedure is described k(x+ )2 − kx2 ≤ −εx2 + `(x, u)
more thoroughly in [36], but generally produces a similar
tradeoff between the speed of convergence and transient for all (x, u) ∈ R2 . Since (x − u)4 ≥ 0, this can be converted
profit. into the quadratic equation
k(αx + (1 − α)u)2 − kx2 ≤ −εx2 + (x + u/3)(2u − x) (46)
VI. E XAMPLES
To conclude this work, we present two examples to Rearranging, (46) is satisfied only if the matrix
illustrate the concept of dissipativity and its relation to the k(1 − α 2 ) − 1 56 − kα(1 − α)
 
tradeoff between stability and profitability, both of which are Q= 5 2 2
6 − kα(1 − α) 3 − k(1 − α)
adapted from [26].
is positive definite. Requiring det(Q) > 0 yields the condition
A. Dissipativity and strong duality
−k2 (1 − α)2 + k(1 − α)(4α/3 + 5/3) − 49/36 > 0
The purpose of this example is to illustrate how dissipa-
tivity and strong duality can be computed, and demonstrate which is equivalent to choosing k between
that dissipativity is a relaxation of strong duality. We begin √
4α + 5 ± 16α 2 + 40α − 24
by considering the following linear single-input single-output
6(1 − α)
system:
x+ = αx + (1 − α)u (44) provided α > 1/2. We conclude that this system is dissipative
for all α ∈ (1/2, 1). In fact, this dissipativity condition
where α ∈ [0, 1) is an arbitrary scalar parameter. We also happens to be a tight condition because α = 1/2 is exactly
define the non-convex stage cost: the critical value of α for which the zero average period-2
`(x, u) = (x + u/3)(2u − x) + (x − u)4 solutions exit the sublevel set L≤0 ; see Fig. 1 for a qualitative
picture.
The system (44) has one unique steady state for each input
u regardless of α, namely x = u. At steady state, the cost B. Consecutive-competitive reactions
function therefore becomes
For the final example, we consider the control of a
4
`(x, u)|x=u = u2 nonlinear isothermal CSTR with consecutive-competitive re-
3 actions [37]. Specifically, we consider two such irreversible
so that (xs , us ) = (0, 0) is the best steady state and `(xs , us ) = reactions:
0. It is not the case, however, that `(xs , us ) ≤ `(x, u) for
all x and u, violating Assumption 1. In fact, the stage P0 + B −→ P1
cost has negative global minima located at (x, u) = P1 + B −→ P2
√ two √
±(21 6/64, 7 6/192). Significantly, (0, 0) is a saddle-point The following process model comes from the dimensionless
of `(x, u). mass balances around the reactor:
By selecting the input sequence +u0 , −u0 , +u0 , −u0 , . . .
which produces from (44) the corresponding state sequence ẋ1 = u1 − x1 − σ1 x1 x2
1+α
− 1−α u0 , 1+α 1+α
1−α u0 , − 1−α u0 , . . ., we can immediately verify the ẋ2 = u2 − x2 − σ1 x1 x2 − σ2 x2 x3
existence of zero-average period-2 solutions of (44). Choos-
ẋ3 = −x3 + σ1 x1 x2 − σ2 x2 x3
ing α to be sufficiently small and tuning the input amplitude
u0 , we can create period-2 solutions that belong to the ẋ4 = −x4 + σ2 x2 x3

3858
60
9
u
40
6
u1 u2
3 20
α
00 00
u = x/2 2 4 6 8 10 2 4 6 8 10
(a) (b)

1.6
(b) 4
1.2 3
x x10.8 x2
(a) 2
0.4 1
00 2 4 6 8 10 00 2 4 6 8 10
(c) (d)

Eco
1.2
R=0
u = −3x 0.8
0.4 Str. dual
x3 x4
0.4 0.2

0 2 4 6 8 10 0 2 4 6 8 10
Fig. 1: Qualitative picture of L0 : global minima (a),(b); α- Time (t) Time (t)
(e) (f)
dependent period-2 solution (black dots). Adapted from [26].
Fig. 2: Open-loop input (a), (b) and state (c), (d), (e), (f)
profiles with different cost functions. Adapted from [26].
where the states x1 , x2 , x3 and x4 are the concentrations of
10
P0 , B, P1 , and P2 respectively, and the inputs u1 and u2 are
inflow rates of P0 and B. The parameters σ1 and σ2 have
values 1 and 0.4, respectively. The time average value of u1 u1 5

is constrained to lie between 0 and 1.


Av[u1 ] ⊆ [0, 1] 0
0 2 4 6 8 10
Time (t)
The economic objective of this system is to maximize the R = 0.002 R = 0.004 R = 0.006 R = 0.008 R = 0.01

production of intermediate P1 , represented by the stage Fig. 3: Stabilization via variation in input penalty. Adapted
cost: `(x, u) = −x3 . For a single second-order reaction, it from [26].
is known that periodic operation outperforms steady-state
operation of a CSTR, so we might expect unsteady closed
loop
 behavior here [37]. The optimal0 steady  state is x0s = From the closed-loop trajectories in Figure 3 we can see
0.3874 1.5811 0.3752 0.2373 , us = 1 2.4310 . that the effect of increasing R is to increasingly dampen the
For implementation of the MPC controller, we discretize oscillations that the economic controller is taking. The loss in
the system with a sample time Ts = 0.1, use a horizon of profit demonstrates that this process is indeed suboptimally
100, add an upper bound of 10 to u1 (t) and a terminal operated at steady state.
state constraint. Not only is the economic model predictive While increasing the R penalty to 0.01 appears to make
controller unstable in the closed loop, but its trajectories also the system dissipative, it does not make it strongly dual.
appear chaotic. To make it so, we increase the Q penalty until the steady-
As recommended by Theorem 4, a tracking term is added state problem is strongly dual. Strong duality is achieved
to the objective function to explore the tradeoff between for Q = 0.36I4 and R = 0.002I2 and checked numerically by
stability and profit. The modified stage cost is given by solving the dual problem, presented in [25]. Figure 2 shows
that this does indeed stabilize the optimal steady state.
`(x, u) = −x3 + (1/2) |x − xs |2Q + |u − us |2R

(47)
Next, we perturb the system away from steady state and
We begin by keeping Q = 0 and increasing the R penalty in compare the performance of an economic MPC controller,
steps. Average economic performance, defined through (38) the strongly dual controller, a tracking controller with the
by Av[−`(x(k), u(k))], is presented in Table I. Note that the same R and Q penalties as the strongly dual controller,
negative sign arises because Table I presents profit, not cost. and a hybrid controller with the same Q as the strongly

3859
R avg profit 2
14
0 0.4472 1.5 12
10
0.002 0.4397
u1 1 u2 8
0.004 0.4233 6
0.006 0.4076 0.5 4
2
0.008 0.3938 00 00
2 4 6 8 10 12 14 2 4 6 8 10 12 14
0.01 0.3752
(a) (b)
TABLE I: Average economic profit for open-loop profiles 2 6
with varying R penalty. Adapted from [26]. 1.8
5
1.6
1.4 4
1.2
x1 1 x2 3
Case avg profit trans profit 0.8
0.6 2
0.4 1
Economic 0.4648 ∞ 0.2
00 2 4 6 8 10 12 14 00 2 4 6 8 10 12 14
R=0 0.3916 2.6201
Strongly dual 0.3848 1.6034 (c) (d)
Tracking 0.3812 1.0587 1.6 0.8
1.4 0.7 Eco
R=0
TABLE II: Average profit and transient profit for open-loop 1.2 0.6 Str. dual
1 0.5 track
profiles with varying R penalty. Adapted from [26]. x3 0.8 x4 0.4
0.6 0.3
0.4 0.2
0.2 0.1
00 2 4 6 8 10 12 14 00 2 4 6 8 10 12 14
dual controller, but R = 0. Figure 4 shows the open-loop Time (t) Time (t)
input and state profiles for these four cases. To evaluate the (e) (f)
performance of each controller, we consider both the average Fig. 4: Open-loop input (a), (b) and state (c), (d), (e), (f)
profit and transient profit. Transient profit is defined as total profiles with different cost functions and a random initial
profit made relative to steady state state. Adapted from [26].

− ∑ `(x(k), u(k)) − `(xs , us ) (48)
k=0 analysis tools for stability and convergence, and we have
Transient profit is therefore a measure to quantify how eco- developed some basic understanding of average performance.
nomically efficient each controller is at returning to steady Many interesting questions remain, and we feel the field is
state. For the case of an economic controller that does not ripe for rapid progress. For example, one interesting question
converge to steady state, this measure is infinite. is how best to design the control system when an economic
Simulation results are presented in Table II. All stable performance goal is augmented with other more traditional
controllers operating with no disturbances have the same goals such as closed-loop stability and convergence to the
average profit. The three stable controllers presented here optimal steady state. Another interesting open question is
appear to have different average profits only because of the how to quantify achieved closed-loop performance given that
finite simulation length. The average economic profit differs the MPC controller optimizes only an open-loop, predicted
from Table I because of the different initial state combined performance.
with the finite simulation length. The key observation is that One of the next important steps forward for the research
controllers which placed more emphasis on the economic community will be to develop illustrative case studies that
objective made up to 147% more transient profit. If this show when this new approach provides enough benefit to be
system were to be perturbed from its steady state frequently, disruptive to the widely implemented, two-layer hierarchical
these improvements to economic efficiency could increase approach of steady-state optimization coupled with dynamic
the average profit of the system significantly. This example setpoint tracking.
therefore highlights the potential benefits of using economic
R EFERENCES
terms in the objective of model predictive controllers even if
[1] W. Findeisen, F. N. Bailey, M. Bryds, K. Malinowski, P. Tatjewski,
tracking terms must be added for stability. and A. Wozniak, Control and Coordination in Hierarchical Systems.
New York, N.Y.: John Wiley & Sons, 1980.
VII. C ONCLUSION [2] W. L. Luyben, B. Tyreus, and M. L. Luyben, Plantwide Process
Control. New York: McGraw-Hill, 1999.
The purpose of this tutorial was to illustrate methods [3] T. Backx, O. Bosgra, and W. Marquardt, “Integration of model
for designing and analyzing behavior of control systems predictive control and optimization of processes,” in International
that optimize an economic criterion rather than a traditional Symposium on Advanced Control of Chemical Processes (ADCHEM
2000), vol. 1, June 2000, pp. 249–260.
setpoint tracking criterion. We should emphasize that this [4] M. L. Darby and M. Nikolaou, “MPC: Current practice and chal-
research area is in its infancy. We have developed some lenges,” Contr. Eng. Prac., vol. 20, no. 4, pp. 328 – 342, 2012.

3860
[5] M. L. Darby, M. Nikolaou, J. Jones, and D. Nicholson, “RTO: An [29] L. Grüne, “Optimal invariance via receding horizon control,” in 50th
overview and assessment of current practice,” J. Proc. Cont., vol. 21, IEEE Conference on Decision and Control, 2011 European Control
no. 6, pp. 874 – 884, 2011. Conference, 2011.
[6] P. L. Silveston, “Periodic operation of chemical reactors - a review of [30] J. Ma, J. Qin, T. Salsbury, and P. Xu, “Demand reduction in building
the experimental literature,” Sādhanā, vol. 10, no. 1–2, pp. 217–246, energy systems based on economic model predictive control,” Chem.
1987. Eng. Sci., vol. 67, no. 1, pp. 92 – 100, 2012.
[7] S. Bittanti, G. Fronza, and G. Guardabassi, “Periodic control: A [31] T. G. Hovgaard, K. Edlund, and J. B. Jørgensen, “Economic MPC
frequency domain approach,” IEEE Trans. Auto. Cont., vol. 18, no. 1, for Power Management in the SmartGrid,” in 21st European Sympo-
pp. 33–38, 1973. sium on Computer Aided Process Engineering, ser. Computer Aided
[8] E. G. Gilbert, “Optimal periodic control: A general theory of necessary Chemical Engineering, 2011, vol. 29, pp. 1839 – 1843.
conditions,” in 1976 IEEE Conference on Decision and Control [32] R. Huang, E. Harinath, and L. T. Biegler, “Lyapunov stability of
including the 15th Symposium on Adaptive Processes, vol. 15, 1976. economically oriented NMPC for cyclic processes,” J. Proc. Cont.,
[9] S. Engell, “Feedback control for optimal process operation,” J. Proc. vol. 21, pp. 501–509, 2011.
Cont., vol. 17, pp. 203–219, 2007. [33] R. Huang, L. T. Biegler, and E. Harinath, “Robust stability of econom-
[10] L. T. Biegler, “An overview of simultaneous strategies for dynamic ically oriented infinite horizon NMPC that include cyclic processes,”
optimization,” Chem. Engg. Proc. PI, vol. 46, no. 11, pp. 1043–1053, J. Proc. Cont., vol. 22, no. 1, pp. 51 – 59, 2012.
2007. [34] X. Chen, M. Heidarinejad, J. Liu, and P. D. Christofides, “Distributed
[11] F. D’Amato, A. Kumar, R. Lopez-Negrete, and L. T. Biegler, “Fast economic MPC: Application to a nonlinear chemical process network,”
nonlinear model predictive control optimization strategy and industrial J. Proc. Cont., vol. 22, no. 4, pp. 689 – 699, 2012.
process applications,” in FOCAPO, 2012. [35] J. B. Rawlings and D. Q. Mayne, Model Predictive Control: Theory
[12] C. Zhao, Y. Zhao, H. Su, and B. Huang, “Economic performance and Design. Madison, WI: Nob Hill Publishing, 2009, 576 pages,
assessment of advanced process control with LQG benchmarking,” J. ISBN 978-0-9759377-0-9.
Proc. Cont., vol. 19, no. 4, pp. 557 – 569, 2009. [36] D. Angeli, R. Amrit, and J. B. Rawlings, “Enforcing convergence
[13] C. M. Jing and B. Joseph, “Performance and stability analysis of LP- in nonlinear economic MPC,” in IEEE Conference on Decision and
MPC and QP-MPC cascade control systems,” AIChE J., vol. 45, pp. Control (CDC), Orlando, FL, 2011.
1521–1534, 1999. [37] C. K. Lee and J. E. Bailey, “Modification of Consecutive-Competitive
[14] J. V. Kadam, W. Marquardt, M. Schlegel, T. Backx, O. H. Bosgra, P. J. Reaction Selectivity by Periodic Operation,” Ind. Eng. Chem. Proc.
Brouwer, G. Dünnebier, D. van Hessem, A. Tiagounov, and S. de Wolf, Des. Dev., vol. 19, no. 1, pp. 160–166, 1980.
“Towards integrated dynamic real-time optimization and control of
industrial processes,” in Proceedings Foundations of Computer Aided
Process Operations (FOCAPO2003), 2003, pp. 593–596.
[15] J. V. Kadam and W. Marquardt, “Integration of Economical Opti-
mization and Control for Intentionally Transient Process Operation,”
Lecture Notes in Control and Information Sciences, vol. 358, pp. 419–
434, 2007.
[16] A. C. Zanin, M. Tvrzská de Gouvêa, and D. Odloak, “Integrating
real-time optimization into the model predictive controller of the FCC
system,” Control Eng. Practice, vol. 10, pp. 819–831, 2002.
[17] M. Abidi, S. B. Jabrallah, and J. P. Corriou, “Optimization of the
dynamic behavior of a solar distillation cell by model predictive
control,” Desalination, vol. 279, no. 13, pp. 315 – 324, 2011.
[18] J. B. Rawlings, D. Bonné, J. B. Jørgensen, A. N. Venkat, and S. B.
Jørgensen, “Unreachable setpoints in model predictive control,” IEEE
Trans. Auto. Cont., vol. 53, no. 9, pp. 2209–2215, October 2008.
[19] J. B. Rawlings and R. Amrit, “Optimizing process economic perfor-
mance using model predictive control,” in Nonlinear Model Predictive
Control, ser. Lecture Notes in Control and Information Sciences,
L. Magni, D. M. Raimondo, and F. Allgöwer, Eds., vol. 384. Berlin:
Springer, 2009, pp. 119–138.
[20] D. Angeli and J. B. Rawlings, “Receding horizon cost optimization
and control for nonlinear plants,” in 8th IFAC Symposium on Nonlinear
Control Systems (NOLCOS), Bologna, Italy, September 2010.
[21] F. P. Ramsey, “A mathematical theory of saving,” Econ. J., vol. 38,
no. 152, pp. 543–559, December 1928.
[22] D. A. Carlson, A. B. Haurie, and A. Leizarowitz, Infinite Horizon
Optimal Control, 2nd ed. Springer Verlag, 1991.
[23] W. Paengjuntuek, P. Kittisupakorn, and A. Arpornwichanop, “On-line
dynamic optimization integrated with generic model control of a batch
crystallizer,” J. Ind. Engr. Chem., vol. 14, no. 4, pp. 442 – 448, 2008.
[24] D. Angeli, R. Amrit, and J. B. Rawlings, “Receding horizon cost op-
timization for overly constrained nonlinear plants,” in Proceedings of
the Conference on Decision and Control, Shanghai, China, December
2009.
[25] M. Diehl, R. Amrit, and J. B. Rawlings, “A Lyapunov function for
economic optimizing model predictive control,” IEEE Trans. Auto.
Cont., vol. 56, no. 3, pp. 703–707, 2011.
[26] D. Angeli, R. Amrit, and J. B. Rawlings, “On average performance
and stability of economic model predictive control,” IEEE Trans. Auto.
Cont., 2011, accepted for publication.
[27] R. Amrit, J. B. Rawlings, and D. Angeli, “Economic optimization
using model predictive control with a terminal cost,” Annual Rev.
Control, vol. 35, pp. 178–186, 2011.
[28] M. Heidarinejad, J. Liu, and P. D. Christofides, “Economic model
predictive control of nonlinear process systems using Lyapunov tech-
niques,” AIChE J., vol. 58, no. 3, pp. 855–870, 2012.

3861

You might also like