Available online at www.sciencedirect.

com
Computers and Chemical Engineering 32 (2008) 15071522
DantzigWolfe decomposition and plant-wide MPC coordination
Ruoyu Cheng
a
, J. Fraser Forbes
a,
, W.San Yip
b
a
Department of Chemical and Materials Engineering, University of Alberta, Edmonton, AB T6G 2G6, Canada
b
Suncor Energy Inc., Fort McMurray, AB T9H 3E3, Canada
Received 16 December 2005; received in revised form 11 July 2007; accepted 12 July 2007
Available online 28 July 2007
Abstract
Due to the enormous success of model predictive control (MPC) in industrial practice, the efforts to extend its application from unit-wide to
plant-wide control are becoming more widespread. In general, industrial practice has tended toward a decentralized MPCarchitecture. Most existing
MPC systems work independently of other MPC systems installed within the plant and pursue a unit/local optimal operation. Thus, a margin for
plant-wide performance improvement may be available beyond what decentralized MPC can offer. Coordinating decentralized, autonomous MPC
has been identied as a practical approach to improving plant-wide performance. In this work, we propose a framework for designing a coordination
system for decentralized MPC which requires only minor modication to the current MPC layer. This work studies the feasibility of applying
DantzigWolfe decomposition to provide an on-line solution for coordinating decentralized MPC. The proposed coordinated, decentralized MPC
system retains the reliability and maintainability of current distributed MPC schemes. An empirical study of the computational complexity is used
to illustrate the efciency of coordination and provide some guidelines for the application of the proposed coordination strategy. Finally, two case
studies are performed to show the ease of implementation of the coordinated, decentralized MPC scheme and the resultant improvement in the
plant-wide performance of the decentralized control system.
2007 Elsevier Ltd. All rights reserved.
Keywords: Decentralized MPC; Coordination; DantzigWolfe decomposition; Complexity analysis
1. Introduction
Model predictive control (MPC) strategies have gained great
success in a wide range of industrial applications. The MPC
framework can be divided into a steady-state calculation and
a control calculation (or dynamic optimization) (Kadam et al.,
2002; Qin & Badgwell, 2003). The goal of the steady-state cal-
culation is to determine the desired targets for output, input,
and state variables at a higher frequency than those computed
from local economic optimizers. The target calculation pro-
vides optimal achievable set-points that are passed to control
calculation.
With considerable development effort in recent years, there
has been a trend to extend MPCto large-scale applications, such
as plant-wide control. Two commonly used strategies for plant-
wide MPCcontrol and optimization are centralized schemes and
decentralized schemes. A fully centralized or monolithic MPC

Corresponding author. Tel.: +1 780 492 0873; fax: +1 780 492 2881.
E-mail address: fraser.forbes@ualberta.ca (J. Fraser Forbes).
for an entire large plant is often undesirable and difcult, if not
impossible, to implement (Havlena &Lu, 2005; Lu, 2003). Such
a scheme can exhibit poor fault-tolerance, require an expensive
centralized computational platform, and can be difcult to tune
andmaintain. Alternatively, inmanychemical plants, large-scale
control problems are solved by a group of MPC subsystems via
decentralized schemes, in which each MPCcontroller takes care
of a specied operating unit. This decentralized MPC scheme
yields the desired operability, exibility and reliability, but may
not provide an appropriate level of performance. In this paper,
reliability refers to the possibility that some control subsystems
or portions thereof are able to function when other subsystems
fail.
Most MPC implementations, when considered in a plant-
wide context, have a decentralized structure, with individual
controllers working in an autonomous manner without coordina-
tion. Such a decentralized scheme can only provide the optimum
operation for each unit with respect to its own objective function.
Thus, this approach may lead to a signicant deviation from the
plant-wide optimum. Lu (2003) claims that the estimated latent
global benet for a typical renery is 210 times more than what
0098-1354/$ see front matter 2007 Elsevier Ltd. All rights reserved.
doi:10.1016/j.compchemeng.2007.07.003
1508 R. Cheng et al. / Computers and Chemical Engineering 32 (2008) 15071522
MPC by itself can capture. The key to exploiting the potential
of decentralized control systems, yet still retaining its structure
and advantages, is cooperation.
This potential benet has garnered increasing interest of
manyresearchers. (Camponogara, Jia, Krogh, &Talukdar, 2002)
proposed a distributed MPC scheme, where local control agents
broadcast their states and decision results to every other agent
under some pre-speciedrules andthis procedure continues until
noagent needs todoso. Recently, coordination-basedMPCalgo-
rithms were discussed in Venkat, Rawlings, and Wright (2004),
in which augmented states are used to model interactions, to
improve plant-wide performance via the coordination of decen-
tralized MPC dynamic calculation. One common characteristic
of the above two schemes is that the coordination of decen-
tralized or distributed MPC controllers is completed without a
coordinator, and thus controllers stand at equal status within
their negotiation. Alternatively, in Lu (2003), a cross-functional
integration scheme was developed, in which a coordination col-
lar performed a centralized target calculation for decentralized
MPC. This idea matches the wide-spread belief among industrial
practitioners (Scheiber, 2004) that the trend toward decentraliza-
tion will continue until the control systemconsists of seamlessly
collabrating autonomous and intelligent nodes with a supervi-
sory coordinator overseeing the whole process. There are two
key factors that determine the desirability of the coordinated
MPCsystem: computational efciency of the coordination strat-
egy to ensure a real-time solution; and required information ow
load throughout the plant communication network.
This work discusses a framework for the coordination of
steady-state MPC target calculation level, which aims to pro-
vide a timely response to local or plant-wide disturbances and
setpoint changes. The proposed approach exploits the exist-
ing plant computing, communication, and information systems
with minimal modication, to provide signicant performance
improvement.
In industrial practice, a variety of optimization methods are
applied to solve MPCtarget calculation problems, among which
linear programming (LP) and quadratic programming (QP) are
most commonly used (Qin &Badgwell, 2003). Many MPCtech-
nology products use a linear programto do the local steady-state
optimization (e.g., the Connoisseur controller offered by Inven-
sys, Inc.). Inthis paper, we discuss the plant-wide coordinationof
LP-based MPCtarget calculation by using the ideas drawn from
decentralized economic planning. The DantzigWolfe decom-
position principle is studied and applied to the development of
coordination scheme. An empirical study of complexity illus-
trates the feasibility of the proposed strategy for industrial
applications.
2. DantzigWolfe decomposition
The Dantzig and Wolfe (1960) decomposition principle is
depicted in Fig. 1.
In this decomposition approach, a large-scale linear program-
ming problem can be separated into independent subproblems,
which are coordinated by a master problem (MP). The solu-
tion to the original large-scale problem can be shown to be
Fig. 1. Mechanism of DW decomposition.
equivalent to solving the subproblems and the MP through a
nite number of iterations Dantzig and Thapa (2002). Within
each iteration, the MP handles the linking constraints that con-
nect the subproblems, using information [f
i
, u
i
] supplied by the
subproblems (note that f
i
is the objective function value and
u
i
is the solution of the ith subproblem). Then, the MP sends
its solution [,
i
] as price multipliers to all the subproblems
for updating their objective functions. Subsequently, the sub-
problems with updated objective functions are re-solved. The
iterative procedure continues until convergence to the solution
of the large-scale problem.
The DantzigWolfe decomposition hinges on the theorem of
convex combination and column generation techniques (Dantzig
& Thapa, 2002; Lasdon, 2002). The theorem of convex combi-
nation, or DW transformation, states that an arbitrary point x
in a convex polyhedral set X = {x|Ax = b, x 0} can be writ-
ten as a convex combination of the extreme points of X plus a
nonnegative linear combination of the extreme rays (normalized
homogeneous solutions) of X, or
x =
L

i=1

i
u
i
+
M

j=1

j
(1)
L

i=1

i
= 1,
i
,
j
0 (2)
where u
i
and
i
are the nite set of all extreme points and
the nite set of all normalized extreme homogeneous solutions
respectively. If the feasible region is bounded, we can reformu-
late the problem by using the extreme points only.
Although any large-scale linear program problem can be
decomposed and solved by DantzigWolfe decomposition
(Chvatal, 1983), the approach is particularly powerful for
structured linear programs. Consider a block-wise linear pro-
gramming problem that has been converted to Simplex standard
form
min z
1
=
p

i=1
c
T
i
x
i
subject to
p

i=1
A
i
x
i
= b
0
(3)
B
i
x
i
= b
i
, x
i
0, i = 1, 2, . . . , p (4)
R. Cheng et al. / Computers and Chemical Engineering 32 (2008) 15071522 1509
where (3) represents the linking constraints associated with p
subproblems, and the constraints in (4) are the local constraints
of independent subproblems. Via the theorem of convex combi-
nation, the master problem (MP) can be formulated as follows
using the linking constraints in (3) and the convex combination
of extreme points from (4), assuming that the feasible regions
of subproblems are bounded
1
min z
2
=
p

i=1
N(i)

j=1
f
ij

ij
subject to
p

i=1
N(i)

j=1
p
ij

ij
= b
0
(5)
N(i)

j=1

ij
= 1,
ij
0, i = 1, 2, . . . , p (6)
where N(i) represents the number of extreme points of the fea-
sible region in the it Ith LP subproblem, and
x
i
=
N(i)

j=1

ij
u
j
i
(7)
f
ij
= c
T
i
u
j
i
(8)
p
ij
= A
i
u
j
i
(9)
with u
j
i
being the jth extreme point of ith subproblem.
The resulting master problem has fewer rows in the coef-
cient matrix than the original problem; however, the number of
columns in the MP is larger due to an increase in the number of
variables associated with the extreme points of all subproblems.
2.1. Multi-column generation algorithms
For a large-scale problem, it can be a formidable task to obtain
all the extreme points and formulate a full master problem. If
the MP is solved via the Simplex method, only the basic set
is needed and it has the same number of basic variables as the
number of rows. Thus, we do not need to explicitly know all
the extreme points of subproblems. This leads to solving an
equivalent problem, the restrictedmaster problem(RMP), which
can be dynamically constructed at a xed size by incorporating
column generation techniques (Dantzig &Thapa, 2002; Gilmore
& Gomory, 1961).
Assume that we have a starting basic feasible solution to the
RMPandit has a unique optimum. The optimal solutionprovides
us with the Simplex multipliers [,
1
,
2
, . . . ,
p
] for the basis
in the current RMP, with associated with (5) and
i
with the
ith constraint in (6), respectively. Then, the subproblems are
modied and solved to nd the priced-out column associated
1
Unbounded cases are discussed in Lasdon (2002) and Dantzig and Thapa
(2002).
with
ij
:
f
ij
= (c
T
i
A
i
)u
j
i

i
(10)
and the ith subproblem is
min z
0
i
= (c
T
i
A
i
)x
i
subject to B
i
x
i
= b
i
, x
i
0
(11)
Here we notice that only minor modication is made to each
optimizing subproblem, i.e., an augmented termcontaining sen-
sitivityinformationis introducedintoeachsubproblemobjective
function. An optimal solution is reached when the following
condition is satised:
min
i,j
f
ij
= min
i
(z
0
i

i
) 0, i = 1, . . . , p (12)
Acomplete proof of the optimality and nite convergence can
be found in Dantzig and Thapa (2002). When condition (12) is
not satised, a column generation strategy is used to determine
the column or columns that will enter the basis of the RMP.
The coordination of subproblems can be regarded as a direc-
tional evaluation procedure for the feasible extreme points of the
subproblems, in which the RMP evaluates and selects subprob-
lem solutions under the guidance of some rules. The selected
or priced-out extreme points will be used to generate columns
needed for updating the RMP. The column generation tech-
niques are the rules that show good performance in directing
the evaluation of subproblem feasible solutions. With column
generation techniques, instead of an exhaustive traversal of all
extreme points of the subproblems, usually only a small subset
of the extreme points are required to be evaluated during the
coordination procedure.
Several column generation techniques can be found in the
literature. In the single-column generation scheme (Lasdon,
2002), the minimum objective function value of problem (12) is
assumed to come from subproblem s (1 s p), i.e., the solu-
tion x
s
() solves subproblem s. Then, the column to enter the
basis is generated by

A
s
x
s
()
i
s

(13)
where i
s
is a n-component vector with a 1 in position s and
zeros elsewhere. The generated column is associated with the
most favorable subproblem (i.e., that with the most negative
reduced cost). Thus, with single-column generation algorithm,
the updated RMP can be expressed as
min z
3
=
p

i=1

J
basis
f
ij

ij
+f

subject to
p

i=1

J
basis
p
ij

ij
+p

= b
0
(14)

J
basis

ij
+

= 1, i = 1, 2, . . . , p (15)

ij
0,

0 (16)
1510 R. Cheng et al. / Computers and Chemical Engineering 32 (2008) 15071522
where the
ij
is the current basic variables and

is the variable
entering the basis. The terms associated with

can be derived
from (7) to (9) and (13), which contribute to the generated col-
umn. J
basis
is an index set whose elements coorespond to the
indices of subproblem feasible solutions, which are now in the
RMP basis. Particularly, in the above formulation, the number
of variables in the basis is |J
basis
| = m
0
+p,
2
which shows the
dimension of the basic set in the RMP is (m
0
+p), where m
0
is
the number of linking constraints in (3).
It should be noted that any other subproblem with a nega-
tive reduced cost has the potential to generate a column to enter
the basis of the master problem (Lasdon, 2002). To take advan-
tage of other favorable subproblems, multiple columns could
be considered for generating a new RMP. Several variants of
multi-column generation techniques are discussed in Dantzig
and Thapa (2002) and Lasdon (2002). In this work, we study the
multi-column generation scheme suggested in Lasdon (2002).
Thus, to incorporate all potential favorable proposals, a new
column is generated in the RMP for each subsystemby applying
(13):
min z
4
=
p

i=1

J
basis
f
ij

ij
+
p

i=1
f

i
subject to
p

i=1

J
basis
p
ij

ij
+
p

i=1
p

i
= b
0
(17)

J
basis

ij
+

i
= 1, i = 1, 2, . . . , p (18)

ij
0,

i
0 (19)
The above problem has n more variables than constraints,
rather than one more as in the single-column generation case.
If we use the size of the coefcient matrix in Simplex standard
form to represent the size of the problem, the RMP with multi-
column generation has a size of (m
0
+p) (m
0
+p +p) while
the RMP with single-column generation has a size (m
0
+p)
(m
0
+p +1). One wouldexpect a greater decrease inz
4
through
every iteration, and thus, a reduction in the number of iterations.
It was discussed in Lasdon (2002) and veried by the computa-
tional experiments (Cheng, Forbes, Yip, &Cresta, 2005) that the
advantage of having more columns in the RMP outweighs the
disadvantage of increased RMP size. Since the DantzigWolfe
decomposition algorithm with multi-column generation shows
a higher computational efciency, we adopt this approach in this
paper.
3. Plant-wide MPC coordination
Generally, because the process plant is built by connecting
individual operating units, the plant model has block-wise struc-
ture and the steady-state coefcient matrix is usually sparse.
With appropriate structural transformation, we can obtain a
2
Here the notation || is used to denote the number of elements or the length
of a data set as is common practice in computing science.
Fig. 2. Coordinated MPC target calculation.
block-angular model with a small number of block off-diagonal
elements. The off-diagonal elements indicate the interactions
among operating units. Most of the existing decentralized (e.g.,
unit-wide) MPC controllers only consider the block-diagonal
elements of the plant-wide process model. Therefore, most pro-
cess plants with decentralized MPC are potential candidates for
strategies to coordinate independent MPC.
Illustrated in Fig. 2 is a structural diagram of a typical plant-
wide decentralized MPCstructure. Note that each MPCcontains
a steady-state target calculation and a dynamic control calcu-
lation (Ying & Joseph, 1999). A coordinator can be designed
by considering different kinds of interactions among operat-
ing units. Such interactions can be formulated as the linking
constraints. Recall that the two key factors to efcient coor-
dination are the computational efciency of the coordination
strategy and the required information ow throughout the plant
communication network. In the previous section we dealt with
the coordination, and in this section we are going to introduce
two linking constraints construction approaches based on dif-
ferent interaction modeling methods, which ensure reasonable
data trafc through the plant communication network.
3.1. Approach 1: interstream consistency
For an individual MPC subsystem, which contains a target
calculation and a control calculation, we can formulate an LP
problem, for the steady-state target calculation at time k as
min z = c
T
x(k)
subject to A
eq
x(k) = b
eq
(k), Lx(k) b(k)
(20)
where x(k) = [u
s
(k), y
s
(k)] is a vector of steady-state values
(i.e., targets or setpoints) for the input and output variables for
the subsystem. The equality constraints in (20) are taken from
the linear dynamic model:
Y(s) = G(s)U(s) +G
d
(s)D(s) +E(s) (21)
which yields the steady-state model
y = Ku +K
d
d +e (22)
where d are the disturbances and e is the unmeasured noise.
The inequality constraints result fromphysical limitations on the
R. Cheng et al. / Computers and Chemical Engineering 32 (2008) 15071522 1511
Fig. 3. Demonstration of interstream consistency.
inputs and outputs, such as actuator limits or production quality
requirements. Other MPC target calculation formulations are
possible, suchas includingmodel bias andsoft output constraints
(Kassmann, Badgwell, & Hawkins, 2000; Lestage, Pomerleau,
& Hodouin, 2002). These can easily be incorporated into the
proposed method, but are omitted to simplify discussion.
The solution from the decentralized MPC target calculation
may not be optimal with respect to the entire plant operation
because of plant/model mismatch, which results from ignoring
the interactions among operating units. In this case, the inter-
actions between two operating units can always be modeled by
equating the appropriate output variables fromthe upstreamunit
and the input variables to the downstream unit. Shown in Fig. 3,
the hexagon labeled A represents the process interstream con-
necting individual operating units.
In formulating the subproblems, those streams connecting
different operating units are torn and consistency relationships
can be used to model the interactions between different units.
Assume that we have p separate operation units, each of which
is controlled by one MPC subsystem. By introducing interpro-
cess stream consistency as the linking constraints (3), we can
formulate an LP problem that includes constraints (3) and (4)
by incorporating those decentralized target calculation prob-
lems. This problem has a block-wise structure which can be
solved efciently by DantzigWolfe decomposition strategy.
At each execution of the plant-wide MPC target calculation,
a coordinated LP problem will be updated and solved. Then
all the calculated steady-state targets both for inputs and out-
puts, including interprocess streamvariables, are passed to MPC
control calculation.
3.2. Approach 2: off-diagonal element abstraction
Quite often, advanced control strategies are designed and
implemented at different times for different operating units. For
example, in a copper ore concentrator plant, model-based and
model-assisted APC controllers were separately installed in the
crushing, grinding and otation sections (Herbst & Pate, 1996).
In this case, certain controlled variables (CVs) and manipulated
variables (MVs) have been specied and grouped in a unit-wide
sense. As we discussed before, these unit-based implementation
of APC strategies can only use the block-diagonal information
of the plant model. Ignoring the off-diagonal information may
lead to signicant loss in the control performance.
Giventhe coordinationmechanismof DantzigWolfe decom-
position, if we can abstract off-diagonal elements from the
overall plant model for the construction of the linking con-
straints, the coordination only requires minor modications to
the objective functionandconstraints of eachdecentralizedMPC
problem. For the ith decentralized MPC system, the LP-based
target calculation has the form:
min c
T
u
i
u
i
+c
T
y
i
y
i
subject to y
i
= K
ii
u
i
+e
i
, u
i
lb
u
i
u
i
ub
, y
i
lb
y
i
y
i
lb
(23)
where u
i
and y
i
are the deviation variables for the MVs and
CVs, respectively. The vector e
i
represents the unmeasured dis-
turbances, u
i
lb
and y
i
lb
are the lower bounds, while u
i
ub
and y
i
ub
are the upper bounds for the decision variables.
If we have a full gain matrix for the entire plant with N
operating units
K =

K
11
K
12
. . . K
1N
K
21
K
22
. . . K
2N
. . . . . . . . . . . .
K
N1
K
N2
. . . K
NN

(24)
The matrix K can be ordered such that a unit-based imple-
mentation of MPC as in problem (23) uses the block-diagonal
information K
ii
of the plant model in its calculations. In such
a case, the off-diagonal blocks may be treated as disturbances.
This way of dealing with the off-diagonal information can result
in undesirable closed-loop behavior, when the interactions are
signicant. Assume the plant-wide model is
Y(k) = KU(k) +E(k) (25)
where Y(k) and U(k) are the CVs and MVs of N local operating
units concatenated into vectors; E(k) is the concatenation of
local disturbance variables (DVs). This is equivalent to
y
i
(k) = K
ii
u
i
(k) +e
i
m
(k) +e
i
u
(k) (26)
e
i
m
(k)
N

j=1
K
ij
u
j
(k) = 0 j = i (27)
where the auxiliary variable e
i
m
, which is an abstraction of the
off-diagonal elements, represents the inuence of the inputs of
other operating units on the local system; e
i
u
stands for unmea-
sured noise. In the proposed coordination scheme, a coordinator
will be developed to handle the constraints (27) and drive the
auxiliary variable e
i
m
to the values corresponding to the plant-
wide optimum operations. In this case, in each unit-based MPC
target calculation, the auxiliary vector e
i
m
is treated as a decision
variable, since Eq. (27) are not included in each unit-based MPC
calculation. Thus, the decision variables in each individual MPC
is augmented to [u
i
, y
i
, e
i
m
].
1512 R. Cheng et al. / Computers and Chemical Engineering 32 (2008) 15071522
Furthermore, the objective function of each MPC system
gets dynamically updated based on the sensitivity information
[,
i
], as in (10), fromthe solution of the master problemby the
coordinator. After each execution of the optimization, only the
original CVs or MVs will be implemented. The coordination
scheme is particularly efcient for coordinating decentralized
control systems based on plant-wide models with sparse off-
diagonal matrices.
Remark 1. Using the above two approaches, we are able to
gather interaction information that plays an important role in the
coordination of decentralized MPC. Compared with the locally
processed information within a decentralized control system, the
information ow trafc conveyed by the communication net-
work is not heavy. In this scheme, only the optimal solution
(one feasible point and the objective function value) and the
constraints sensitivity information are passed between the coor-
dinator and each subsystem. For example, within the solution a
plant-wide MPC problem containing 10 linking constraints and
100 decision variables from 5 subsystems, the information ow
required to be transmitted through the communication network
within one communication cycle is less than 1000 bytes.
3
In
addition, for the off-diagonal element abstraction method, it may
require additional efforts for the identication of off-diagonal
elements (gains).
4. Complexity analysis
The computational efciency of a coordination strategy
is a key factor in determining the viability of using coor-
dinated decentralized optimization approaches in industrial
applications. In the following sections, we evaluate the computa-
tional efciency of the DantzigWolfe decomposition algorithm
through an empirical complexity study.
Without loss of generality, we consider a large-scale block-
angular LP problem with p subproblems in its standard form
max

i
c
T
i
x
i
subject to

i
A
i
x
i
= b
0
(28)
B
i
x
i
= b
i
(29)
x
i
0 i = 1, 2, . . . , p (30)
where vectors x
i
(n
i
1), b
i
(m
i
1), b
0
(m
0
1), c
i
(n
i
1),
and matrices A
i
(m
0
n
i
), B
i
(m
i
n
i
) are specic to subprob-
lem i.
Computational complexity is the study of determining the
cost of solving a numerical computing problemusing a scientic
algorithm. In this work, the cost of the proposed DantzigWolfe
decomposition approach can be interpreted as the required arith-
metic or other computational operations.
4
Although the cost of
3
For platforms that use double (64-bit) precision oating-point numbers to
represent a real number according to IEEE Standard 754 oating point.
4
In this paper, the denition of an arithmetic operation, such as an addition
or a multiplication, is applied to two real numbers.
Fig. 4. Worst-case behavior illustration.
an algorithm can be measured in several ways, the worst-case
behavior and average-case behavior are two typical measures
(Nash & Sofer, 1996).
4.1. Worst-case behavior
Unlike a complete enumeration approach, the process of
coordination can be viewed as a directional evaluation of sub-
problem extreme points in the RMP. When the multi-column
generation technique is incorporated into a coordinated opti-
mization scheme, shown in Fig. 4, the RMP (coordinator) will
evaluate a new set of p extreme points submitted by each sub-
problem at every iteration. Note that the new set of p extreme
points in black is not generated by an arbitrary combination
but under the direction of coordinator (RMP). In Fig. 4, the
(m
0
+p) extreme point set in gray is associated with the RMP
optimal basis at the previous coordination iteration. The p point
set is a new set of extreme points submitted by subproblems at
current iteration, and N
i
is the number of extreme points of the
Ith subproblem feasible region.
The worst-case behavior analysis depends on the LP tech-
niques that are used for solving the RMP and subproblems. The
worst-case behavior of Simplex methods is non-polynomial;
whereas, the interior point methods are polynomial time
algorithms. If Simplex methods are used and we take its worst-
case performance, the DantzigWolfe decomposition algorithm
cannot be a polynomial time algorithm. Even though each sub-
problem can generate an (optimal) extreme point in polynomial
time, using interior point methods (IPM), it may take exponen-
tial time to get all the extreme points for a subproblem. Note
that each subproblem has m
i
constraints and n
i
decision vari-
ables. Thus, the number of extreme points can be as many as
the combination
niCmi
. Therefore, in the worst-case, whatever
methods are used to solve either RMP or subproblems, the
coordination process might theoretically evaluate every com-
bination of the subproblem extreme points. Then, worst-case
behavior analysis for DantzigWolfe decomposition implicates
that it may not be a polynomial time algorithm.
Usually, only when the worst-case behavior well reects the
average-case behavior of an algorithm, is it used to provide an
upper bound of the cost of solving a problem (Nash & Sofer,
1996). For example, the worst-case performance of interior point
methods (IPM) gives a rather tight upper bound; however, a
worst-case analysis does not reect the observed performance of
the Simplex method (Chvatal, 1983; Nash &Sofer, 1996). Since
R. Cheng et al. / Computers and Chemical Engineering 32 (2008) 15071522 1513
average-case behavior is more relevant for our work, average-
case performance will be emphasized here.
4.2. Average-case behavior
Considering the coordination mechanism discussed in pre-
vious sections, the overall complexity for the decomposition
strategy can be expressed as
T = {T(RMP) +
p

i=1
T(SP
i
)} CCN (31)
where T represents the number of arithmetic operations for
solving the LP problem, and the communication cycle number
(CCN) is used to distinguish the number of Simplex iterations
fromthat of coordination iterations (the times to solve the RMP).
Thus, the required arithmetic operations are attributed to two
parts: the operations for solving the subproblems and the RMP
in a single communication cycle, as well as the number of com-
munication cycles.
If we dene the rst part as the non-coordination complexity:
T(NonCo) = T(RMP) +
p

i=1
T(SP
i
) (32)
the overall complexity can be expressed as
T = T(NonCo) CCN (33)
In the Simplex method, for an LP problem with m con-
straints and n variables in standard form, the cost of solving
one iteration is O( m n) for Gaussian elimination plus O( m
3
)
arithmetic operations for periodic re-factorization of the basis
matrix,
5
thus the arithmetic operation needed in one Simplex
iteration is O( m
3
+ m n) (Nash & Sofer, 1996). Here, we con-
sider the average-case performance of Simplex method
6
and
take the average number of Simplex iterations as O( m+ n) as
in Andrei (2004). Therefore, the average behavior bound to be
used is O( m
4
+ n m
2
+ m
3
n + n
2
m) for the Simplex method,
which shows polynomial time complexity. For the LP prob-
lem described in (2830), we can derive the non-coordination
complexity from
T(RMP) O(M
4
0
+N
0
M
2
0
+M
3
0
N
0
+N
2
0
M
0
) where
M
0
= m
0
+p, N
0
= m
0
+2p (34)
and
T(SP
i
) O(m
4
i
+n
i
m
2
i
+m
3
i
n
i
+n
2
i
m
i
), i = 1, 2, . . . , p
(35)
Since the above numbers of arithmetic operations are all
polynomials in corresponding m and n, the complexity of
5
The O() notation for a given function g(n) is given as O(g(n)) = {f(n) :
a
+
and n
+
0
such that 0 f(n) ag(n) for all n n
+
0
}.
6
The observed scaling behavior of Simplex method is between m and 3 m
(Nash & Sofer, 1996).
the non-coordination computation in (32) can be expressed
as a polynomial with respect to the number of constraints
and decision variables. In other words, the non-coordination
arithmetic operations T(NonCo) can be computed very ef-
ciently if we consider the average-case behavior of Simplex
method.
The other part of the complexity analysis deals with the com-
municationcycles. Toour knowledge, there is nosimilar analysis
in the literature and therefore resort to a study of the average
behavior of CCN via a comprehensive empirical study.
4.3. Empirical study of complexity
For analysis of the computational complexity of the decom-
position algorithm, since the algorithms are implemented on a
sequential machine, we may express the overall complexity in
terms of computational time
t =
CCN

{t(RMP) +
p

i=1
t(SP
i
)} (36)
where t represents the total computational time to solve a prob-
lem. Through the comparison between the computational time
of each subproblem and the total computational time, we can
identify and focus on the bottleneck subproblem, which may
be the largest in dimension or hardest to solve.
Similarly, if we denote the non-coordination computational
time as
t(NonCo) = t(RMP) +
p

i=1
t(SP
i
) (37)
and assume a distributed/parallel computing environment, for
example, one CPUfor each subproblem, then we have the equiv-
alent computational time (parallel computing)
t
eqv
(NonCo) = t(RMP) +max
p
i=1
{t(SP
i
)} (38)
Such a distributed computing environment coincides with
what is encountered in current plant-wide decentralized MPC
applications. In this case, it may be desirable to balance the
computational load on each computing node because the non-
coordination computational complexity (t
eqv
) relies heavily on
the largest subproblem. We will reemphasize this point in next
section. Then, the computational time within a decentralized
computing environment can be expressed as
t
eqv
=
CCN

t
eqv
(NonCo) (39)
Next the focus is on the analysis of CCN, where the rela-
tionship between CCN and the characteristic parameters of the
LP problems such as m
0
, p, and |I
i
| = (m
i
n
i
) is to be deter-
mined. Note that, each parameter may have physical meaning in
real systems. For instance, in the plant-wide MPC coordination,
m
0
may reect the density of interactions among operating units;
p could be the number of decentralized MPC controllers or dis-
tributed industrial computers; while |I
i
| can represent the size
1514 R. Cheng et al. / Computers and Chemical Engineering 32 (2008) 15071522
of the control problems handled by an MPC subsystem. More-
over, the inuence of the relative subproblemratio (RSR), which
is dened as RSR = {max|I
i
|/|I
j
|, i, j = 1, 2, . . . , p}, will also
be studied. The RSR gives some idea on the computational load
balance throughout the distributed computing network.
In the following Monte Carlo simulations, besides the
CCN, the computational efciency and scaling behavior of
DantzigWolfe decomposition will also be investigated by com-
paring the performance between the decomposition algorithm
and the centralized LP solvers on the platform of MatLab

.
In both solvers, ILOG

CPLEX 9.0 is used to solve all LP

problems. In our study, we focus on the average behavior of
different optimization strategies in solving the problems with
the following assumptions:
(1) No cycling: many techniques can be applied to efciently
deal with cycling.
(2) No degeneracy: when degeneracy occurs in practice, it can
be well handled using techniques such as perturbation tech-
niques in Simplex method.
(3) The studied problems have bounded feasible regions and
optimal solutions.
Due to the random nature of Monte Carlo simulation, we
would like to acknowledge that there is possibility that an
average-case problem may not be represented in our test set.
The scheme of test problem generation is introduced in
Appendix A. It randomly generates a set of LP problems with
a block-angular structure. We start from a reference problem
model, whose problem size and structure should be a good ref-
erence for the comparison experiments, i.e., we can observe
algorithm performance changes when we change the problem
with respect to the reference model. With some preliminary
tests, we choose the following set of parameters as the reference
problem model:
{p = 17, m
0
= 30, m
i
= 40, n
i
= 30 . . . i = 1, 2, . . . , p}
(40)
Note that the reference problemhas subproblems of identical
size, i.e., the problem is thus a well-balanced decomposable
problem with RSR = 1. We also assume that each subproblem
has been allocated to a distributed CPU. Therefore, the equiva-
lent computational time t
eqv
for the decomposition algorithm is
estimated by summing up the time for solving the master prob-
lem and the most difcult subproblem, assuming a distributed
computational environment.
Scenario 1. We x p and |I
i
|, change m
0
(see Appendix B).
In this case, we can study the performance of decomposition
and coordination with respect to the dimension of linking
constraints in Eq. (28).
Fig. 5 shows that the CCN increases almost linearly
with the dimension of linking constraints increases. In
Fig. 6, when the number of linking constraints is small, the
decomposition algorithm gives comparable performance
Fig. 5. Coordination complexity.
to the centralized LP solver. When the number of linking
constraints increases, the computational performance gets
worse. This shows the computational complexity of the
decomposition algorithm has strong dependence on the
dimension of linking constraints.
Scenario 2. For xed it p and m
0
, we change subproblem size
|I
i
| by simultaneously changing m
i
and n
i
(see Appendix B).
In this case, we study the algorithm performance with respect
to subproblem sizes.
Fig. 7 shows a rather surprising result. Intuitively, one may
think that the CCN would increase when subproblem size
increases, because the overall problem gets bigger. By anal-
ogy to the coordination of plants in a large company, larger
plant proposals are usually less sensitive to coordination and
thus do not tend to change dramatically. In such a case, the
central planning board may end up with less coordination iter-
ations. Since the solution of a larger subproblem is more time
consuming, Fig. 8 shows an increase in the computational
time of DantzigWolfe decomposition algorithm, but its per-
Fig. 6. Computational performance.
R. Cheng et al. / Computers and Chemical Engineering 32 (2008) 15071522 1515
Fig. 7. Coordination complexity.
formance is much better than the centralized LP solver. In
addition, as the linking constraints contribute much less than
the subproblems to the overall problem size, the changes of
overall problem size reect the changes of subproblem size.
Thus, the overall problem coefcient number is used in the
gures.
Scenario 3. We keep m
0
, m
i
and n
i
constant, and change
the number of subproblems p (see Appendix B). In this case,
we investigate the performance of the coordination algorithm
whenmore andmore subproblems are integratedintothe coor-
dination system, assuming a rather well-balanced subproblem
computational load.
Fig. 9 shows that the number of subproblems p slightly
inuences the coordination complexity. When the number of
subproblems increases, there is a minor increase inCCN. Sim-
ilarly, in Fig. 10, the number of subproblems slightly inuence
the computational performance when distributed computing
environment is considered. In other words, the incorporation
of a similar-size subsystem does not deteriorate the compu-
Fig. 8. Computational performance.
Fig. 9. Coordination complexity.
tational performance too much, which also indicates a good
scaling behavior of the DantzigWolfe decomposition algo-
rithm.
Scenario 4. We x m
0
,

p
i=1
m
i
and

p
i=1
n
i
, i.e., we x the
overall problem size, we can study the inuence of relative
subproblem ratio (RSR). In this case, we change p by com-
bining subproblems into groups (see Appendix B) according
to different partition patterns of the original LP problem. For
example, when RSR = 4, the original 17 subproblems in the
reference problemcan be combined into a set of subproblems
which have a problem size ratio {4 : 4 : 4 : 4 : 1} (pattern 1)
or {4 : 1 : : 1} (pattern 2). The above patterns reect two
typical situations in which we have a smaller subsystem or a
larger subsystem compared with others.
In Fig. 11, the CCNof both cases is monotonically increasing
as the RSR increases; while in Fig. 12, the computational time
also increases as the RSR increases. Note that an identical CPU
is assumed to be allocated to each subproblem and thus the
Fig. 10. Computational performance.
1516 R. Cheng et al. / Computers and Chemical Engineering 32 (2008) 15071522
Fig. 11. Coordination complexity.
number of CPUs, p
0
, corresponds to the number of resulting
subproblems. We can see that even for the same LP problem,
different decomposition patterns lead to different computational
performance. Moreover, if we dene
D
subm
=
max
p
0
i=1
t(Sub
i
)
p
0

i=1
t(Sub
i
)
(41)
as a quantity that represents the computational time dominance
of the largest subproblem, and
p
0

i=1
t(Sub
i
) = t(NonCo) t(RMP) (42)
can be obtained from our simulation.
Shown in Fig. 13, for the two decomposition patterns, the
computational time taken by the largest subproblem dominates
the computational time for solving subproblems when RSR
increases. This also addresses the reason we should balance the
computational load on each computing node in our experiment
design.
Fig. 12. Computational performance.
Fig. 13. Computational dominance of Max subproblem.
Remark 2. Although the complexity study has not yielded
an accurate mathematical description for the complexity of the
proposed approach, it does provide new insight and reveal some
inherent features of the complexity of DantzigWolfe decom-
position algorithm. Moreover, this study also provides some
guidelines for coordination system design with DantzigWolfe
decomposition principle. Advantages over a centralized LP
solver are gained by using the decomposition algorithm to solve
an LP problem, which has the following properties:
a block-angular structure;
a large-scale;
a small number of linking constraints;
relatively high-dimensional subproblems;
well-balanced computational load for each computing node;
a large number of subproblems, if distributed computing
power is available.
In many industrial applications, the plant-wide MPC target
calculation problemcan be formulated as an LP problemhaving
most of the above properties, and as a result, can be efciently
solved by DantzigWolfe decomposition.
5. Illustrative case studies
In this section, we illustrate the implementation of the pro-
posed coordination scheme through two case studies: the rst is
used to investigate the interstream consistency approach and
the second investigates the off-diagonal element abstraction
approach.
5.1. Case study 1
Let us consider a system shown in Fig. 14 (Cheng, Forbes,
& Yip, 2004). The normalized gains for the system are given in
(43) through (45). An identity matrix is chosen for K
d
in (22)
assuming that the disturbances inuence the outputs directly.
The locations where the disturbances entering the plant are
R. Cheng et al. / Computers and Chemical Engineering 32 (2008) 15071522 1517
Fig. 14. Interacting MIMO operating units.
shown as dashed lines in Fig. 14
K
A
= G
A
(0) =

0.4 0.6 0.1

0.5 0.4 0.1

(43)
K
B
= G
B
(0) =

0.3 0.4 0.3

0.1 0.2 0.1

(44)
K
C
= G
C
(0) =

0.7 0.3
0.6 0.5

(45)
Each operating unit has its own objective, which is a subset
of information used by plant-wide optimizers, and the prot
function cost coefcients are:
c
T
A
= [ 1 1 1 3 3] (46)
c
T
B
= [ 1 1 1 3 3] (47)
c
T
C
= [ 1 2 5 5] (48)
So for each operating unit, by tearing the interprocess stream,
a linear program for the k th target calculation is:
min c
T
j
x
j
(k)
subject to K
j
x
j
(k) = b
j
eq
(k)
(49)
L
j
x
j
(k) b
j
(k), j = A, B, C (50)
where L
j
stands for the coefcient matrix associated with all the
inequality constraints when it is in standard form. The R.H.S. of
the equality constraints b
j
eq
(k) represent the updated model bias
at each target calculation execution. The R.H.S. of the inequality
constraints b
j
(k) contain the lower bounds lb and upper bounds
ub of the variables in the operating units. The bounds on the
variables in this case study are shown in Eqs. (51) and (52).
lb = [ 0.3 0.3 0.15 0.45 0.4 0.45 0.4 0.3 0.45 0.15 0.45 0.3 0.45 0.5] (51)
ub = [ 0.5 0.5 0.25 0.55 0.5 0.55 0.5 0.5 0.55 0.25 0.55 0.5 0.55 0.6] (52)
Three different MPC strategies, centralized MPC, decen-
tralized MPC and coordinated, decentralized MPC, are
implemented to evaluate their abilities to track the changing
optimum in steady-state target-calculations. For the central-
ized MPC target calculation, a direct LP problem is formulated
treatingall the inputs andoutputs, includinginterprocess interac-
tions, as decision variables. For the decentralized MPC scheme,
separate LP problems are formulated by passing the upstream
decisions to downstream units as disturbances. Finally, the
coordinated MPC target calculation incorporates the linking
constraints in modeling the interactions and solves the RMP
and independent subproblems iteratively.
In our case study, unknown disturbances are generated by
ltering random series of uniformly distributed variates in order
to restrict these disturbances within the interval 0.05. These
unknown disturbances are directly imposed on the outputs when
the optimized targets are implemented in our simulation.
d(t) =
1
1 +c
1
q
1
+c
2
q
2
e(t) (53)
Byusingthe autoregressive models inequation(53) as simpli-
ed disturbance models, we predict one-step ahead disturbances
based on the past information of estimated disturbances. The
estimated disturbances used to update the disturbance model
in (53) are calculated by comparing the measured outputs and
model predictions at every control execution. At the current con-
trol calculation, the parameters, c
1
and c
2
, in the disturbance
model are estimated using the estimated disturbances in the past
10 control execution periods. The one-step ahead disturbances
are predicted using the estimated c
1
and c
2
. Then the process
models are updatedusingthe predicteddisturbances. The steady-
state targets are then calculated by using the updated process
models.
The following accumulated prot function is dened for per-
formance comparison:
P =

Z(k) T
s

k=i
V(k) T
s
(54)
where Z(k) is the actual prot per unit time from the k th target
calculation; and T
s
is the sampling period between two target
calculations. As per the denition of the cost coefcients in Eqs.
(46)(48), theprot Z(k) has the same value as the objective
function of the LP problemdescribed in (49) and (50). V(k) rep-
resents the penalty for constraint violations when we implement
the calculated targets:
V(i) = w
T
y
v
(i) (55)
where w is a specied penalty vector, which is used in all three
cases; and the violation of output constraints y
v
(i) is dened as:
y
v
(i) =

y
act
(i) y
max
, if y
act
(i) y
max
y
act
(i) y
min
, if y
act
(i) y
min
(56)
where y
act
(i) is the actual output vector when the calculated
targets are implemented in the process; while y
max
and y
min
are
subsets of the upper and lower bounds given in Eqs. (51) and
(52).
A benchmark is dened for comparing the performance of
different MPCtarget calculation strategies. The benchmark used
for comparison is dened as the maximumprot achieved when
the plant is operated at the true optimum, which is calculated
1518 R. Cheng et al. / Computers and Chemical Engineering 32 (2008) 15071522
Fig. 15. Calculated targets by different approaches.
using the perfect process model and exact knowledge of distur-
bances. Although this maximum prot is not achievable, it is a
useful basis for performance comparison.
Fig. 15 shows the prot achieved as a function of control
execution using different MPC steady-state target calcula-
tion strategies. The coordinated target calculation gives the
same achievable optimum as the centralized MPC scheme
does (because the same interactions are considered in both
approaches), while the decentralized scheme yields a subop-
timum as interactions are ignored in the calculation.
Table 1 compares the performance of different MPC strate-
gies for a simulation of 150 target calculation executions. From
Table 1, we can see that the centralized and the coordinated,
decentralized target calculation give the same best achievable
prot and achievable ratio to the true optimum, while the fully
decentralized target calculation only captures around 94.48%of
the maximum prot.
Table 1
Performance comparison
Centralized Decentralized Coordinated True. opt.
Prot 5800.9 5481.7 5800.9 5801.7
Achiev. ratio 99.98 94.48 99.98 100
Prob. dim. 46 42 15 15 3 7 7 +15
15 3
NA
Table 1 also reports the problem sizes for different steady-
state target calculation strategies. The problem size is dened
as the size of the coefcient matrix in the LP standard form
used in the Simplex method. Therefore, slack and excess vari-
ables are added to convert the inequality constraints to equality
constraints, and the columns of the coefcient matrices corre-
spond to the process variables as well as the slack and excess
variables. We can see that the centralized scheme has the largest
problemsize, which will growsignicantly when the dimension
of separate problems and the number of operation units in the
owsheet increase. The problemsize for the decentralized MPC
is reported as the dimension of the largest subproblem multi-
plied by the number of units. For the coordinated scheme, the
problemsize is expressed as the addition of two components, the
dimension of RMP (the coordinator) and the problemdimension
in the decentralized scheme (the coordinated parts).
This case study shows that the interstreamconsistency can be
used for interaction modeling, and when such interactions are
handled by the coordinator, the resulting coordinated, decentral-
ized control system does produce signicant improvement on
the plant-wide performance. As such, it provides an approach to
plant-wide control that does not require centralized computing
environment.
5.2. Case study 2
This section discusses a potential application of the proposed
coordinated, decentralized control scheme in upgrading a decen-
tralized control system for an energy services system (i.e., a
Fig. 16. A utility plant in Oil Sand industry.
R. Cheng et al. / Computers and Chemical Engineering 32 (2008) 15071522 1519
utility plant) in an Oil Sands plant site. While this scheme is
being evaluated at the scoping study stage, the coordinated,
decentralized optimization system will be run in parallel with
the existing control system and its solution will be used as
a reference to support operator decision making for set-point
adjustment. In this case study, an overall steady-state model of
the energysystemis identied, andthenthe off-diagonal element
abstraction method is applied to constructing linking constraints
as is discussed in Section 3.2.
As is shown in Fig. 16, the utility plant includes several types
of major equipment, such as gas turbines and waste heat recov-
ery steam generators (e.g., G1 and H1), boilers (e.g., B1B5),
and steam turbines (e.g., T1 and T2), etc. The utility plant must
satisfy the power, steam and hot process water demand (e.g.,
D1D3 for process use, HE1 for heat exchanger use to produce
HPW) for the Oil Sands operations. Different levels of steam,
such as high pressure, mediumpressure, and lowpressure steam,
are required by the processes. It should be noted that, the lower
pressure steam can also be supplied by letting down high pres-
sure steam. As is shown in Fig. 16, L1 and L2 are de-superheat
valves while L3 is for steam venting and thus, its use should be
well controlled. S1 represents the low pressure steam partially
returned from the processes that use high pressure steam.
For the purpose of illustration, a simpliedsimulationis given
with the parameters and values modied to respect condential-
ity.
Due to the different levels of steam demand and the geo-
graphical distributions of facilities, the current control systems
include decentralized rule-based expert systems for each major
piece of equipment and distributed control systems (DCS) for
regulatory control. Presently the DCS systems receive set-points
fromdecentralized rule-based expert systems, which are respon-
sible for adjusting the set-points (e.g., boiler steam production
rates, gas turbines power generation, and let-down steam ow
rates) based on fuel cost and actual steam and power demand
information. The decision from the expert system is based on a
complex rule base, which contains a large amount of empirical
information. For example, the power generation of gas turbines
and steam turbines is related to the ratio between the electricity
and natural gas price; while the boilers are assigned priorities
for adding incremental load to accommodate steam demand
changes.
The decentralized rule-based expert system can provide sat-
isfactory performance; however, its solutions are usually not
optimum. To improve energy efciency and reduce energy cost,
an optimization-based approach is desired. In this case, we
investigate replacingthe existingdecentralizedrule-basedexpert
systems with a coordinated, decentralized model-based opti-
mization system (e.g., coordinated, decentralized MPC target
calculation) in order to optimally adjust the set-points to reduce
fuel cost and thus maximize the site-wide prot.
Based on the historical data collected from Interplant

, a set
of (linear) steady-state models of the equipment are identied
for the energy services system. Then a linear programming prob-
lem can be formulated for the overall energy services system.
Such an linear programming problem involves a large num-
ber of variables and thus is in large-scale; however, it presents
Fig. 17. Cost of energy services.
special structure and can be partitioned into several small sub-
problems with a linking constraint set. The actual problem
could be in large-scale, but for the purpose of illustration, the
simplied system in Fig. 16 involves 30 variables and 22 con-
straints. The process variables are described in Table C.1 in
Appendix C. Besides a few linking constraints that intercon-
nect multiple steam headers, the coefcient matrix is a sparse
block-diagonal matrix with very few off-diagonal elements.
Withthe off-diagonal element abstractionmethod, the linear pro-
gramming problem can be decomposed into three subproblems,
which correspond to three steam headers, and a set of common
constraints including four augmented linking constraints.
In this case study, based on the actual operations, the natu-
ral gas price is assumed to vary from 5 to 11 dollars per kscf
(a thousand standard cubic feet); power price varies from 20
to 160 dollars per MWh (a million watts hour). As steam let-
down causes energy loss, it is priced based on the natural gas
cost to make up for the heat loss during steam let-down, i.e.,
the let-down from HP to MP steam costs $0.01 NG price per
klbs (a thousand pounds), the let-down from MP to LP steam
costs $0.23NG price per klbs, and the LP steam venting costs
$1.33 NG price per klbs. Although coke is a byproduct of
the Upgrading processes, which is usually considered as free
fuel, it is appropriate to price coke because a ue gas desul-
phurization (FGD) facility has to be operated at some cost to
reduce SO
2
emission for environmental concerns. In the actual
operation, an hourly prole of the next days power pool price
is based on a forecast and the NG price will be held con-
stant for each 24 h period. In this case, the DantzigWolfe
decomposition is employed to solve the large-scale LP prob-
lem by taking advantage of the existing distributed computing
environment.
The cost reported in Fig. 17 is dened as the objective func-
tion value of the optimization problem plus a xed operational
cost.
7
The objective function in the case study is developed
7
The xed operational cost includes budgeted maintenance fees, labor and
administration fees, etc.
1520 R. Cheng et al. / Computers and Chemical Engineering 32 (2008) 15071522
Table 2
Performance comparison
Decentralized Centralized Coordinated
Oper. cost ($) 3,171,815 2,296,702 2,296,702
Saving ratio N/A 28% 28%
Comput. time (s) 0.08 0.1 0.7
for MPC target calculation with targets and incentives coming
from upper level economic optimization. So it is formulated
so as to minimize the overall energy cost, which includes the
cost of fuel (e.g., coke and natural gas) and some penalty
cost on steam let-down and venting, less the credit from the
sale of power. Therefore, a smaller value of this function is
desired. A simulation of 336 h of operation is performed,
where the rst 168 h reect summer operations and the sec-
ond 168 h reect winter operations. In summer operations, the
steam demand is usually lower than that in winter; and the
electricity and natural gas prices are usually lower in summer
time.
From Fig. 17, it is clear that the coordinated control scheme
and the centralized control scheme provide the same operating
cost. This occurs as the same interactions are considered in both
schemes. The decentralized (rule-based) control system yields
poorer performance (i.e., larger cost function value) because it
ignores the interactions.
Table 2 compares the performance of different optimization
schemes for the 336-h simulation. Seen from Table 2, the cen-
tralized and the coordinated, decentralized schemes yield the
same operation cost and saving ratio to the operational cost of
decentralized scheme. The table also reports the computational
time for each scheme. Although the coordinated scheme does
not show better computational efciency when solving a prob-
lem of this size, it easily meets the solution time requirements
for this application.
The key point drawn from this case study is that, when we
have satisfactorily accurate subsystem steady-state models, the
DantzigWolfe decomposition algorithm can be applied to the
coordination of decentralized model-based target calculation,
which provides the same solution as the centralized scheme if
the interactions are appropriately considered.
6. Summary and conclusions
Industrial practice has revealed the deciencies of existing
decentralized MPC systems in nding plant-wide optimal oper-
ations. Using a coordinator with decentralized controllers can
address these issues. This work introduces a novel approach
to coordinating decentralized MPC target calculation by tak-
ing advantage of the DantzigWolfe decomposition algorithms.
It also proposes several linking constraints construction meth-
ods for coordination system design, in which the constraints
associated with multiple units can be incorporated.
In this work, we proposed a framework of designing a coor-
dination systemfor decentralized MPCwith minor modication
to current decentralized control layer. Our work shows that
the proposed coordinated target calculation scheme substan-
tially improves the performance of the existing decentralized
control scheme, while it can utilize decentralized comput-
ing environment to ensure acceptable real-time calculation
speeds.
The computational complexity analysis presented in this
work, which is based on a comprehensive empirical study,
addresses several implementation issues for the potential indus-
trial applications of the proposed coordination scheme. It
reveals the structural features that an LP problem should have
for advantageous use of DantzigWolfe decomposition. The
computational complexity analysis veries the efciency and
applicability of the proposed coordination strategy, and also pro-
vides some guidelines for its application in industrial control
problems.
We are continuing the development of this framework and
methodologies. A key issue that should be investigated is
the determination of MPC subsystems scope and structure to
ensure highperformance withminimal computation, e.g., should
decomposition balance the computational load of each subprob-
lem.
Acknowledgments
This study is part of a project on decomposition and coor-
dination approaches to large-scale operations optimization. The
authors wouldlike tothankDr. Guohui Lin(Department of Com-
puting Science, University of Alberta) for his helpful discussions
and constructive comments on this paper. We also acknowledge
the nancial support from NSERC, Syncrude Canada Ltd., and
the Department of Chemical and Materials Engineering at the
University of Alberta.
Appendix A. Test problem generation
To generate a test problem set, we follow the block-wise
structure in (28) and (29). To generate one LP problem, we
take the following steps, assuming the optimization problem
is formulated with some scaling operations:
(1) Generate p sets of subproblem constraint B
i
b
i
: generate
a random vector x
i
with n
i
elements in [1, 10]generate a
random m
i
n
i
matrix B
i
with elements in [10
6
, 10
3
]
calculate b
o
i
= B
i
x
i
perturb b
i
= b
o
i
+b
o
i
, where is
a m
i
vector with randomly generated elements in [0, 0.5],
then we have generated subproblem constraints which have
feasible solutions.
(2) Generate m
0
linking constraints: combine X = [x
1
, . . . , x
p
]
of a dimension N generate a random m
0
N matrix A
with elements in [10
6
, 10
3
] calculate b
o
0
= AX per-
turb b
0
= b
o
0
+b
o
0
, where is a m
0
vector with randomly
generated elements in [0, 0.5].
(3) Generate a N vector c with random elements in [0, 10] (in
theory, we can generate an unrestricted c vector).
The generated LP problem should have feasible solutions.
Degeneracy and cycling is avoided by careful design of the
problem instance generation algorithm.
R. Cheng et al. / Computers and Chemical Engineering 32 (2008) 15071522 1521
Appendix B. Monte carlo simulations
Numerical experiments were designed for the following sce-
narios:
(1) An appropriate reference problem model must be speci-
ed. The reference problem size and structure should be
a good reference for the comparison experiments, i.e., we
can observe the algorithm performance changes when we
change the problem with respect to the reference model. In
the preliminary study, the reference model is chosen from
p = 17, m
0
= 30R
2
/10, m
i
= 40R
2
/10,
n
i
= 30R
2
/10, R = {1, 2, 3, 4, 5, 6, 7}
In this case, the overall problem size can be repre-
sented as the number of elements in the coefcient matrix
I = (m
0
+

p
i
m
i
) N, or in standard LP formI = (m
0
+

p
i=1
m
i
) (

p
i=1
m
i
+m
0
+N).
(2) For xed p = 17, |I
i
| = 40 30, we change m
0
in the fol-
lowing way:
m
0
= 30 2
R3
, R = {1, 2, 3, 4, 5, 6, 7}
(3) For xed p = 17 and m
0
= 30, change subproblem size
(m
i
n
i
) by factors of 2 to the reference problem model,
by changing m
i
and n
i
m
i
= 40 2
R3
; n
i
= 30 2
R3
,
R = {1, 2, 3, 4, 5, 6, 7}
(4) We keep m
0
= 30, m
i
= 40 and n
i
= 30 constant and
change the number of subproblems p:
p = 8R +1, R = {1, 2, . . . , 15}
In this case, we assume a rather well-balanced sub-
problem load, i.e., m
1
, m
2
, . . . , m
p
is in similar order of
magnitude and the same to n
1
, n
2
, . . . , n
p
.
(5) By xing m
0
= 30,

p
i
m
i
and

p
i
n
i
, i.e., we x the overall
problemsize, we can study the inuence of relative subprob-
lem ratio (RSR). In this case, we change p by combining
subproblems into groups following the patterns below:
{1, 1, . . . , 1, 1}, {2, 2,..,2, 1}, {4, 4, 4, 4, 1}, {8, 8, 1},
{16, 1}
and
{1, 1, . . . , 1, 1}, {2, 1,..,1, 1}, {4, 1, . . . , 1}, {8, 1, . . . , 1},
{16, 1}
in the above cases, RSR changes from 1 to 16.
Appendix C. Steady-state model variables
See Table C.1.
Table C.1
Description of process variables
Equipment CVs Unit MVs Unit
G1 Gas turbine power
gen.
MW Fuel gas ow kscf/h
Waste heat prod. mmBtu/h
H1 HRSG HP steam
prod.
klbs/h Waste heat from G1 mmBtu/h
Duct ring gas ow kscf/h
B1 Boiler HP steam klbs/h Coke fuel klbs/h
Fuel gas ow klbs/h
B2 Boiler HP steam klbs/h Coke fuel klbs/h
Fuel gas ow klbs/h
B3 Boiler HP steam klbs/h Fuel gas ow klbs/h
B4 Boiler MP steam klbs/h fuel gas ow klbs/h
B5 Boiler LP steam klbs/h Fuel gas ow klbs/h
T1 MP steam prod. klbs/h HP steam ow klbs/h
Power generation MW
T2 LP steam klbs/h HP steam ow klbs/h
Power generation MW
L1 HP let-down
steam
klbs/h HP let-down valve %
L2 MP let-down
steam
klbs/h MP let-down valve %
L3 LP steam venting klbs/h LP venting valve %
References
Andrei, N. (2004). On the complexity of MINOS package for linear program-
ming. Studies in Informatics and Control, 13, 3546.
Camponogara, E., Jia, D., Krogh, B. H., & Talukdar, S. (2002). Distributed
model predictive control. IEEE Control Systems Magazine, 02721708/02,
4452.
Cheng, R., Forbes, J. F., & Yip, W. S. (2004). DantzigWolfe decomposition
and large-scale constrained MPC problems. In Proceedings of the DYCOPS
7.
Cheng, R., Forbes, J. F., Yip, W. S., & Cresta, J. V. (2005). Plant-wide MPC: A
cooperative decentralized approach. In Proceedings of the 2005 IEEE-IAS
APC.
Chvatal, V. (1983). Linear Programming. W. H. Freeman and Company.
Dantzig, G. B., & Thapa, M. N. (2002). Linear programming 2: Theory and
extensions. Springer Verlag.
Dantzig, G. B., &Wolfe, P. (1960). Decomposition principle for linear programs.
Operation Research, 8, 101111.
Gilmore, P. C., & Gomory, R. E. (1961). A linear programming approach to the
cutting stock problem. Operation Research, 9, 849859.
Havlena, V., &Lu, J. (2005). Adistributed automation framework for plant-wide
control, optimization, scheduling and planning. In Proceedings of the 16th
IFAC world congress 2005.
Herbst, J. A., & Pate, W. T. (1996). Overcoming the challenges of plantwide
control. In Emerging separation technologies for metals II (pp. 313).
Kadam, J. V., Schlegel, M., Marguardt, W., Tousain, R. L., Hesssem, D. H.
V., Berg, J. V. D., & Bosgra, O. H. (2002). A two-level strategy of inte-
grated dynamic optimization and control of industrial processes - a case
study. In Proceedings of the European symposiumon computer aided process
engineering, 12 (pp. 511516). Elsevier.
Kassmann, D. E., Badgwell, T. A., & Hawkins, R. B. (2000). Robust steady-
state target calculation for model predictive control. AIChE Journal, 46,
10071024.
1522 R. Cheng et al. / Computers and Chemical Engineering 32 (2008) 15071522
Lasdon, L. S. (2002). Optimization theory for large scale systems (2nd ed.).
Dover Publications Inc.
Lestage, R., Pomerleau, A., & Hodouin, D. (2002). Constrained real-time
optimization of a grinding circuit using steady-state linear programming
supervisory control. Power Technology, 124, 254263.
Lu, J. Z. (2003). Challenging control problems and emerging technologies in
enterprise optimisation. Control Engineering Practice, 11, 847858.
Nash, S. G., & Sofer, A. (1996). Linear and nonlinear programming (1st ed.).
McGraw-Hill.
Qin, S. J., & Badgwell, T. A. (2003). A survey of industrial model predictive
control technology. Control Engineering Practice, 11, 733764.
Scheiber, S. (2004). Decentralized control. Control Engineering,44
47.
Venkat, A. N., Rawlings, J. B., &Wright, S. J. (2004). Plant-wide optimal control
with decentralized mpc. In Proceedings of the DYCOPS 7.
Ying, C., & Joseph, B. (1999). Performance and stability analysis of LP-
MPC and QP-MPC cascade control systems. AIChE Journal, 45, 1521
1534.