Вы находитесь на странице: 1из 13

Journal of Process Control 21 (2011) 787799

Contents lists available at ScienceDirect

Journal of Process Control


journal homepage: www.elsevier.com/locate/jprocont

A hierarchical distributed model predictive control approach to irrigation canals:


A risk mitigation perspective
Ascensin Zafra-Cabeza a, , J.M. Maestre a , Miguel A. Ridao a , Eduardo F. Camacho a , Laura Snchez b
a
b

Department of Automatic and System Engineering, University of Seville Escuela, Superior de Ingenieros, Camino de los Descubrimientos, s/n, 41092 Seville, Spain
INOCSA S.L.Ingeniera, Spain

a r t i c l e

i n f o

Article history:
Received 23 July 2010
Received in revised form
30 November 2010
Accepted 23 December 2010
Available online 12 February 2011
Keywords:
Irrigation canals
Risk management
Control
Model predictive control
Water treatment

a b s t r a c t
This paper presents a hierarchical distributed model predictive control approach applied to irrigation
canal planning from the point of view of risk mitigation. Two levels in optimization are presented. At
the lower level, a distributed model predictive controller optimizes the operation by manipulating ows
and gate openings in order to follow the water level set-points. The higher level implements a risk
management strategy based on the execution of mitigation actions if risk occurrences are expected. Risk
factors such as unexpected changes in demand, failures in operation or maintenance costs are considered
in the optimization. Decision variables are mitigation actions which reduce risk impacts that may affect
the system. This work shows how model predictive control can be used as a decision tool which takes
into account different types of risks affecting the operation of irrigation canals.
2011 Elsevier Ltd. All rights reserved.

1. Introduction
Water canal system management involves operating pumps,
valves and gates to satisfy user demands and minimize costs. The
main objective of irrigation canals (IC) is to supply water to farmers
according to a specic schedule. A set of constraints imposed by the
physical system and by management policy has to be considered.
IC operation may be affected by many important factors. Operations can be interrupted for several reason such as scheduled
maintenance, response to warnings, subsystem failure or accidents.
Furthermore, variability in demand, changing weather and raw
water conditions are parameters that should be considered for the
optimal operation and maintenance of the plant. Several studies
have been carried out to take into consideration the inuence of
uncertain factors in the performance of water systems such as [21],
[26] or [30] where climate change, drought or demand uctuations
have been analysed.
Uncertain factors not previously considered in the planning
may affect IC operation. These uncertainties originate from various
causes: political (changes in politics can change water strategy),
operational (water level reference varies from that forecast, adjacent land becomes water logged, etc.), nancial, maintenance tasks

Corresponding author. Tel.: +34 954487360; fax: +34 954487340.


E-mail addresses: asun@esi.us.es (A. Zafra-Cabeza), pepemaestre@cartuja.us.es
(J.M. Maestre), ridao@esi.us.es (M.A. Ridao), eduardo@esi.us.es (E.F. Camacho).
0959-1524/$ see front matter 2011 Elsevier Ltd. All rights reserved.
doi:10.1016/j.jprocont.2010.12.012

(failure in reach or devices due to wear and tear, seepage loss, sensor theft) or ecological. Most of the factors mentioned are sources
of risks that can affect IC performance and should, therefore, be
taken into account. Quantifying these risks and incorporating them
into mathematical planning and operation models may result in
improved water system policies.
Risk management (RM) in plant operation is a discipline that
is giving rise to great interest amongst researchers and industrial
sectors [4,27]. The objective of RM in engineering systems is to
establish risk-based policies to obtain better trade offs in safety
and productivity. This technique begins at the conceptual design
phase and continues through out the execution and commissioning of the system. In recent years, risk has been intensively studied
in water resource engineering, and many signicant achievements
have been made. In Ref. [31] a risk analysis is developed where
parameters such as reliability, resilience and vulnerability are evaluated. In Ref. [2] an identication of different risks in water systems
is provided and in Ref. [3] a guideline is presented for mitigating
risks. An optimization of the operation of a desalination plant is
carried out by risk mitigation in Ref. [28]. It can be concluded that
risk in water systems may decrease if risk mitigation actions are
adopted beforehand.
From the point of view of IC control, many contributions can be
found in the literature. The introduction of automatic control has
been increasingly promoted when technical and socio-economic
circumstances make it possible [6,19]. There are applications ranging from classical approaches such as PI controllers [17] to model

788

A. Zafra-Cabeza et al. / Journal of Process Control 21 (2011) 787799

predictive control (MPC) applications (see Refs. [5,11,18]). MPC is


an optimal control strategy based on the explicit use of a model
to predict the process output at future time instants [7]. MPC
approaches have been widely applied in industry and also in water
systems. However, MPC is a technique with strong computational
requirements which hinder its application to large-scale systems
such as transportation systems including trafc, water or power
networks [20]. In these systems, the computational requirements
or the impossibility of obtaining a centralized model are major
problems that MPC cannot overcome. For this reason, most large
scale and networked control systems are based on a decentralized
control architecture where the system is divided into several subsystems, each controlled by a different control agent which may
or may not share information with the rest. Each of the agents
implements an MPC based on a reduced model of the system and
on partial state information which, in general, results in an optimization problem with a lower computational burden. In the case
where the agents communicate in order to obtain a cooperative
solution, we speak of distributed MPC (DMPC), otherwise the term
decentralized MPC is used.
Different decentralized and distributed MPC schemes can be
found in the literature; see Ref. [23] for an extensive review of
the most relevant results on the topic. The most basic schemes are
based on a decentralized architecture in which there is no communication between the agents. For example, in Ref. [1] the centralized
problem is decentralized considering a one step horizon, which
guarantees small deviations in the values of the variables the agent
shares. Another example can be found in Ref. [15], where a decentralized control architecture for nonlinear systems with continuous
and asynchronous measurements is presented. We can properly
speak of DMPC schemes when communication comes into play.
The most basic DMPC algorithms use communication to broadcast
information amongst the agents. In general, this information consists of bounds on the state or input trajectories, in such a way
that neighboring agents can take these bounds into account in their
calculations. For example, in Ref. [14] a DMPC scheme for linear systems coupled by the state is considered. In this scheme, the agents
exchange predictions about their state at the end of each sample step. The procedure of information broadcasting can be done
in an iterative fashion; for example, in Ref. [25] this procedure
is presented as communication-based control. Finally, the most
sophisticated DMPC algorithms provide a means of agent collaboration, that is, the agents exchange information in order to obtain
a consensus on the values of the shared variables. For example,
a gradient-based distributed dynamic optimization method based
on the exchange of sensitivities is proposed in Ref. [24]. In Ref. [8],
basic collaboration algorithms are provided with an extensive list
of conditions to ensure convergence and stability. The use of dual
decomposition for the distribution of the centralized problem [22]
is another technique which can be classied in this group. Dual
decomposition has been applied to the control of irrigation canals in
Ref. [18]. In this paper we will use a variation of another cooperative
scheme [16] which provides a good trade off between performance
and communicational burden in such a way that agents can reach
a cooperative solution with a low number of communications.
This paper presents a hierarchical DMPC (HDMPC) to optimize
the IC operation, the benets and the costs associated to choosing a
set of risk mitigation actions. At the higher level, a risk-based strategy is implemented which is based on mitigation actions execution
to reduce risk impacts. A centralized MPC is used in the optimization problem. At the lower level, a DMPC drives the plant to the
correct values. The main point of innovation this work is considering mitigation actions to reduce exposure of identied risks in the
operating process.
In this paper, the term risk is dened as an event that may take
place when operating IC and which may cause unexpected con-

sequences. After identication of the risks and the setting of the


mitigation actions to reduce the risks have been established, the
risk model to be integrated in the operation is provided. The decision problem is formulated as an MPC that computes the original
control variables of the system and the risk mitigation actions over
a nite control receding horizon. An HDMPC approach to consider
the topology of the system and to decentralize control depending
on the geographical region, is described.
The mitigation actions to be carried out can be discrete or continuous, thus, the resulting optimization problem is formulated as a
mixed integer quadratic problem (MIQP) which belongs to the class
of NP-complete problems [12]. The objective function considered in
this paper is to minimize a multicriteria weighted function where
operating costs, demand satisfaction, mitigation actions and control efforts are involved. The outcome of the problem will strongly
depend on the weights of the different terms.
The paper is organized as follows. First, a description of irrigation canals modelling is shown. Section 2 describes the risk model
used. The optimization problem for planning is described in Section
3, where the distributed model predictive controller and risk mitigation approach are joined in the objective function of the problem.
In order to illustrate the benets of the method, a simulation model
of a real IC and a risk structure are used with different congurations in Section 4. Finally, some concluding remarks are provided
in Section 5.
2. Irrigation canal modelling for control
The system considered is an open-canal used for water distribution (for irrigation and drinking water supply), made up of several
reaches connected by gates for regulation purposes. The dynamics of water owing along open irrigation canals can be obtained
by applying the Saint Venant equations [9,10], which are nonlinear partial differential equations. Because these equations are very
complex to be used directly for control, they are often linearized
around a set point. First-order systems plus a delay are normally
used to model the canal dynamics (see Ref. [9]).
A typical irrigation canal may be divided into several reaches
separated by gates; the controlled variables are the downstream
water levels, hi (t) R+ (m) and the manipulated variables are the
check point to the gates, ui (t) R+ (m).
Each canal reach has an inow from an upstream canal reach,
Qin,i R+ (m3 /s), and an outow to a downstream canal reach,
Qo,i R+ (m3 /s). Furthermore, other ows are considered as perturbation variables:
qin,i R+ (m3 /s), ows due to rainfall, failure in upstream gate.
qo,i R+ (m3 /s), known offtake outows by farmers, considered
as measurable perturbations.
The discrete model considered using the dened variables is:
Ai (hi (k + 1) hi (k)) = Td (Qin,i (k td ) + qin,i (k) Qo,i (k) qo,i (k))
(1)
where Td (s) is the length of the sampling time, Ai the surface of
the reach and td the delay of the input Qin (the level is measured
downstream).
The discharge through a submerged ow gate is usually determined [9]:
Qo (t) = Cd L

2gu(t)

hup (t) hdn (t),

(2)

where Cd is the gate discharge coefcient, L is the gate width, u(t)


the gate opening and hup (t), hdn (t) the upstream and downstream
water levels, respectively.

A. Zafra-Cabeza et al. / Journal of Process Control 21 (2011) 787799

External risk information

Cost of mitigation
Higher level

MPC
Mitigation actions

Internal risks (plant data)

Lower level
DMPC
Flow head and gate openings

Internal risks (plant data)

PLANT

Fig. 1. Hierarchical scheme of controllers.

3. HDMPC and planning of IC


As previously mentioned, two control levels are implemented in
this approach. At the higher level, risk mitigation is introduced to
determine the mitigation actions to be executed (see Fig. 1). With
this strategy, cost is optimized and parameters at the lower level
may be modied as a consequence of the mitigation action applied.
For example, level references may be changed as result of a risk
mitigation action. On the other hand, the lower controller drives
the plant to the corrected values by opening and closing gates and
valves. A distributed control system is implemented at this level.
Note that both controllers will work with different sampling times
due the nature of the process.
As can be observed, the proposed controller has a hierarchical
and distributed scheme. MPC has been selected at both level to
carry out the optimization problems.
MPC is an optimal control strategy based on the explicit use
of a dynamic model to predict the process output at future time
instants [7] over a prediction horizon (N). The set of future control
signals is calculated by optimizing a criterion or objective function.
The predicted outputs depend on the past input and output values
known up to instant t and on the future control signals. Only the
control signal calculated for instant t is sent to the process whilst the
next control signals are ignored. Benets of MPC include the relative
ease of implementation, the ready extension to the multivariable
case, time delays, and the natural addition of constraints in the
optimization.
3.1. Higher level: risk management and MPC
This section describes the risk model to be integrated in the
optimization problem. This model takes into consideration external
risks such as changing weather or nancial data and internal risks.
Internal risks comes from IC operation. Examples of internal risks
may be failure in the gates or seepage losses.
In this work, the term risk is dened as an event that could take
place and have an impact on some of the units U ={U1 , . . ., Un } that
make up the IC system. These units are either the departments of
the plant (maintenance, operation, nance, etc.) or the different
stages that make up the plant (water pretreatment, distribution,
storage, etc.). In order to dene the risk strategy, several elements
have to be previously identied:
Operation policies, objectives and priorities.
Risks that may cause impacts on the system, denoted by set R.
Strategic plan to reduce exposures of risks by mitigation actions,
denoted by set A.
Consider the set of parameters Z ={Z1 , . . ., Zq } that risks can
change, q being the number of parameters; examples of these

789

parameters are time delays, demand or economic costs. We dene


R ={R1 , . . ., Rm } as the set of identied risks for the plant. Each risk Rr
is characterized by a probability of occurrence at each time instant
Pr (t) and some initial impacts IIrc , with c = 1, . . ., q on the different
parameters of the plant. Note that a unit can be inuenced by any
risk and each risk may have an impact on any parameter. Therefore,
risk impacts can change the values of parameter set Z when they
occur and no mitigation action is carried out.
Once risk identication has taken place, the next step to be
undertaken is the design of a strategic mitigation plan. This makes
the reduction of the impact of the identied risks possible. In this
way, each risk can be associated to a set of actions that could mitigate it risk. We assume the mitigation action set to be A ={A1 , . . .,
Ap } with p representing the number of mitigation actions. Each
mitigation action is described by a set of three elements:
Aa = {uMa , Fa , Ga }

a = 1, . . . , p.

(3)

where the decision variable for mitigation action Aa is denoted by


uMa . Fa = {fca : R R} with c = 1, . . ., q being the set of functions
that determine the risk impact reduction as a function of uMa at
each time; thus, fca is the reduction of the initial impact affecting
parameter Zc when action Aa is applied. Actions that are chosen
to mitigate risks may have an associated cost of execution; this
characteristic is modelled by dening functions Ga = {gca : R R}
that describe the extra costs to be added if action Aa is also carried
out as a function of the corresponding decision variable uMa .
Fig. 2 shows an example of a risk-based structure (RBS) that
illustrates the relationship between risks and actions in a possible
strategic plan. It can be observed that a unit may be associated to
some specic risks (i.e. unit 1 is susceptible to risks R1 and R2 ); a
risk can be mitigated by different actions. In Fig. 2 for example, Rm is
mitigated by Ap1 and Ap . One action may mitigate different risks;
note in Fig. 2 how Ap1 mitigates R2 and Rm . Mitigating actions will
reduce the initial impact of a risk but, usually, the system will incur
additional costs as a result. Even if the impact is stochastic in nature
(i.e., only assessed if the risk actually occurs), costs associated to
mitigating actions will be incurred regardless.
In previous work [29], decisions about mitigation actions were
in the form of binary execute/do not execute decisions. The intensity of the action has often to be taken into account when deciding
how to execute it; this decision will depend on the nature of the
mitigation action control variable uMa , which can either be a continuous (uMa R) or integer (uMa R) variable.
The term denoted by RE and named Risk Exposure is dened for
each risk taking into account all the previous information about
risks. Hence, RErc (uM , t) means the exposure of risk Rr affecting

Irrigation Canal System

U2

U1

R1

A1

R2

...

Un

...

...

Rm

Ap-1

Ap

Fig. 2. Risk-based structure of an IC with relationships between units, risks and


mitigation actions.

790

A. Zafra-Cabeza et al. / Journal of Process Control 21 (2011) 787799

parameter Zc at instant t. It takes the form:


p

RErc (uM , t) = Pr (t)(IIrc


p

RAra fca (uMa )) +

a=1

RAra gca (uMa ),

(4)

a=1

where Pr (t) is the probability of risk Rr at instant t and IIrc denotes


the initial impact of risk Rr affecting parameter Zc ; both of these can
be time dependent. The sum of functions f means the reduction of
the initial impact by taking actions; RAra = 1 if risk Rr is mitigated by
action Aa , otherwise, RAra = 0. gca (uMa ) is the extra cost of mitigation
action Aa on the parameter Zc .
3.2. Optimization problem at the higher level
At this level, the decision variables are the mitigation actions
to be executed (uM ). These actions optimize the cost of the canal
operation and may change parameters of the lower controller,
if necessary. The index performance optimizes a multicriteria
weighted function where internal and external risks are involved.
This optimization problem is solved by MPC.
minJ = 1 Jint (uM , t) + 2 Jext (uM , t)

(5)

uM ,t

where Jint represents the optimization of the cost associated to


internal risks such as operational risks or maintenance risks, and
Jext represents the optimization of the cost associated to external
risks such as changes in reference levels due to rainfalls, nancial
opportunities or market issues. These terms are expressed as the
standard term in MPC:
Ji (uM , t) =

q
N 


(c)(Y ic (t + k|t) wc (t + k)) ,

k=1 c=1

with i = {int, ext}

(6)

where Y ic is the predicted output on parameter ZC at instant (t + j)


on the class of risk i. wc is the reference to follow for parameter Zc .
(c) is the weight for each parameter Zc . Note that wc corresponds
to the reference to follow in the MPC. In this study, usually this term
is 0, because we want to optimize the parameters. Y ic is calculated
as:
Y ic =

m


RErc (u, t + k)

(7)

r=1

Note that risks can appear modifying the estimated cost. Therefore, term RErc (u, t + k) models the effect of risk Rr on parameter
Zc (risk exposure); m denotes the total number of risks and N the
prediction horizon. In the case of internal risks, Y int,c , risks associated to maintenance and operation will be included; in the case of
Y ext,c , risks associated to demand, nancial and market issues will
be included.
The output of the problem will strongly depend on the weights
of the different terms. Additional terms can be added to the index
function in order to incorporate other operating objectives. For
example, the control effort can be added, if required. Thus, Eq. (5)
may be expanded:
minJ = 1 Jint (uM , t) + 2 Jext (uM , t) + 3 J3 (uM , t)
uM ,t

(8)

where
J3 =

N


uM (t + k 1)

(9)

k=1

The decision about run the higher controller off line or on line,
will depend on when was executed. If it is run before IC is operating, may be considered as a planning tool where all the data are

estimated. Otherwise, if the execution in on line, actual values of


variables are taken for optimization at time instant t.
Because decision variables of mitigation actions may be boolean,
the resulting optimization problem is a mixed integer quadratic
problem. This belongs to the class of NP-complete problems [12].
The complexity of the problem depends on the number of real
and integer variables (mitigation actions) and the number of constraints. The computation time required to solve the problem is
worst-case exponentially with the problem size. If the problem has
ni binary inputs, the complexity is 2ni (2ni QP problems to be solved).
The number of QP problems to be solved is nite and, therefore, the
algorithm nds a feasible solution (if there is one) at a nite time.
The sampling time should be sufciently broad to satisfy that.
3.3. Lower level: DMPC
In this paper we employ a distributed algorithm based on the
DMPC scheme presented in Ref. [16]. This algorithm provides a
reasonable trade-off between performance and the number of communications needed to reach a cooperative solution, which is a
convenient characteristic for the type of physical system considered. It assumes that for each subsystem, there is an agent that has
access to the model and the state of that subsystem. The agents do
not have any knowledge of the dynamics of any of their neighbors,
but can communicate freely amongst them in order to reach an
agreement. The proposed strategy is based on negotiation between
agents. At each sampling time, agents make proposals to improve
an initial feasible solution on behalf of their local cost function, state
and model, following a given protocol. The neighbors affected by
the changes of a proposal evaluate the proposal when they receive
it and answer with the cost increment (or decrement) implied by
the proposal. Next, the agent that made the proposal computes
a neighborhood cost as the sum of the corresponding neighbor
cost increments plus its own increment (or decrement). If this sum
results in a cost decrement, then the proposal is accepted and all
the neighbors are notied so that they can update their information.
The procedure we propose is valid if, and only if, a given agent
is not evaluating two different proposals at the same time. Such
simultaneous evaluation by an agent might lead to a global cost
increment due to possible crossed interactions between two proposals. For this reason, the various negotiation/communication
protocols that may be implemented must guarantee that each proposal is evaluated independently. In this paper, we propose to
implement a controller in which, a xed number of proposals made
sequentially by random agents are considered at each sampling
time. A simple way to do this is the following: before trying to
communicate, an agent listens to the communication channel to
check if there are other agents communicating. In case that the
channel is free, the agent places its proposal. Otherwise it waits a
small random time before retrying to make a proposal. Every time
a proposal is made, the agents update a counter. Once it exceeds a
given threshold no more proposals can be made during the current
sample time.
Note that this algorithm can be easily enhanced to admit the
parallel evaluation of proposals. As we said, several proposals can
be evaluated in parallel as long as they do not involve the same set
of agents; that is, at any given time an agent can only evaluate a
single proposal. However, note that the communication protocol
to implement the algorithm in parallel is beyond the scope of this
work.
3.3.1. Problem setting at lower level
Consider the following class of distributed linear systems in
which there are Mx subsystems coupled to their neighbors through
Mu inputs.

A. Zafra-Cabeza et al. / Journal of Process Control 21 (2011) 787799

xi (t + 1) = Ai xi (t) +

Bij uj (t)

(10)

j ni

Rqi

where xi
with i = 1, . . ., Mx are the states of each subsystem,
and uj Rrj with j = 1, . . ., Mu are the different inputs. The set of
indices ni indicates the set of inputs uj which affect state xi and the
set of indices mj indicates the set of states xi affected by input uj .
Note that Eq. (10) has the same structure as (1). We consider the
following linear constraints in the states and the inputs
xi i ,
uj uj ,

i = 1, . . . , Mx
j = 1, . . . , Mu

(11)

where i and uj are closed polyhedra that contain the origin in their
interior dened by the following set of linear inequalities
xi i Hxi xi bxi , i = 1, . . . , Mx
uj uj Huj uj buj , j = 1, . . . , Mu

791

value for the decision control vector Ud (t) is set to the value of
the shifted input trajectories, that is, Ud (t) = Us (t).
Step 2: Agents try to submit their proposals randomly. To this
end, each agent asks the neighbors affected if they are free to
evaluate a proposal (each agent can only evaluate one proposal
at any given time). If all the agents acknowledge the petition, the
algorithm continues, if not, the agent waits a random time before
trying again. We will use the superscript p to refer to the agent
granted permission to make a proposal.
Step 3: In order to generate its proposal, agent p minimizes Jp
solving the following optimization problem:
p

{Uj (t)}

j nprop

arg

min

{Uj }

j nprop

Jp (xp , {Uj }j n )

s.t.
xp (k + 1) = Ap xp (k) +

(12)

Bpj uj (k)
(15)

j np

Note that, as these polyhedra contain the origin in their interior,


then bxi > 0 and buj > 0.
The above inequalities allow us to model typical constraints in
this type of systems. For example, the level of the canals has to
be kept between minimum and maximum levels. Similar intervals
apply for the manipulated variables, such as the opening of the
gates. Finally, note that the requirement that the polyhedra contain
the origin in their interior can be addressed with a simple change
of variable.

xp (0) = xi (t)
xp (k) p , k = 0, . . . N
uj (k) uj , k = 0, . . . N 1,
Uj = Ujd (t), j
/ nprop

j np

In this optimization problem, agent p optimizes over a set nprop


of inputs that affect its dynamics, that is, nprop np . Based on the
optimal solution of this optimization problem, agent p presents a
p
proposal dened by a set of input trajectories {Uj (t)}
where
j nprop

3.3.2. DMPC algorithm at lower level


The control objective of the proposed scheme is to minimize a
global performance index dened as the sum of each of the local
cost functions. The local cost function of agent i is based on the
predicted trajectories of its state and inputs and it is dened as
Ji (xi , {Uj }j n ) =

N1


Li (xi (k), {uj (k)}j n )

(13)

k=0

where Uj ={uj (k)} is the future trajectory of input j, N is the prediction horizon, Li ( ) with i Mx is the stage cost function dened
as
T

Li (xi , {uj }j n ) = (xi h i (t)) Qi (xi h i (t)) +

uTj Sij uj

(14)

j ni

with Qi > 0, Sij > 0. Note that term h i (t) stands for the agent i reference.
We use notation xi (k) to denote state i, k-steps in the future,
obtained from the initial state xi applying the input trajectories
dened by {Uj }j n . Note that each of the local cost functions only
i

depends on the trajectories of its state and the inputs that affect it.
At the end of the negotiation rounds, the agents decide a set of
input trajectories denoted as Ud . The rst input of these trajectories
is applied, the rest of the trajectories are not discarded, however,
they are used instead to generate the initial proposal for the next
sampling round which is given by the shifted future input trajectories Us of all the inputs. The last input of each of these trajectories
is obtained simply by repeating the penultimate value.
We next dene the proposed distributed MPC scheme:
Initialization: A feasible initial set of input trajectories Ud ( 1) has
to be computed. Even if the calculation of Ud ( 1) is beyond the
scope of this work, it can be computed either by a centralized
supervisor or in a distributed manner by agents. Anyhow, note
that nding a feasible solution has a lower computational burden
than nding the optimal solution of the optimization problem.
Step 1: Each agent p measures its current state xp (t). The agents
communicate in order to obtain Us (t) from Ud (t 1). The initial

Uj (t) stands for the value of the trajectory of input j of the proposal of agent p. From the centralized point of view, the proposal
at time step t of agent p is dened as
(16)
where operation
p
atives to {Uj (t)}

j np

stands for the update of the components relin Ud (t) and leaving the rest unmodied.

Step 4: Each agent i affected by the proposal of agent p evaluates


the predicted cost corresponding to the proposed solution. To do
that, the agent calculates the difference between the cost of the
new proposal Up (t) and the cost of the current accepted proposal
Ud (t) as
p

Ji (t) = Ji (xi (t), {Uj (t)}

j ni

) Ji (xi (t), {Ujd (t)}

j ni

(17)

This difference Ji (t) is sent back to agent p. If the proposal does
not satisfy the constraints of the corresponding local optimization
problem, an innite cost increment is assigned. This implies that
unfeasible proposals will never be chosen.
Step 5: Once agent p receives the local cost increments from
each neighbor, it can evaluate the impact of their proposal Jp (t),
which is given by the following expression

J p (t) =
i


j nprop

Ji (t)

(18)

mj

This global cost increment is used to make a cooperative decision


on the future input trajectories. If Jp (t) is negative, the agent
will broadcast the update on the control actions involved in the
proposal and the joint decision vector Ud (t) will be updated to the
value of Up (t), that is Ud (t) = Up (t), otherwise, it is discarded.
Step 6: The algorithm returns to Step 1 until the maximum number of proposals has been made or the sampling time ends. We
denote the optimal cost corresponding to the chosen inputs as
J(t) =

Mx

i=1

Ji (xi (t), {Ujd (t)}

j ni

(19)

792

A. Zafra-Cabeza et al. / Journal of Process Control 21 (2011) 787799

Step 7: The rst input of each optimal sequence in Ud (t) is applied


and the procedure is repeated the next sampling time from Step
1.
4. Case study
In this section, results of the method are presented, taking into
account the following data:

which can manipulate the upstream gate. The control objective of


an agent is to regulate the water level of its subsystem.
Table 1 shows a description of the gates, the off-take gates, and
their locations. This information has been provided by the company
INOCSA.
4.2. Data of controllers
As was before mentioned, two controllers compose the system.

Irrigation canal example


Models and parameters of controllers
Risk identication
Mitigation plan: actions to be executed.
The above items are described in the following.

4.1. Irrigation canal example


The proposed algorithm will be tested with data from a real
system, a section of the postrasvase Tajo-Segura in the south-east of
Spain. The postrasvase Tajo-Segura is a set of canals which distribute
water coming from the river Tajo in the basin of the river Segura.
This water is mainly used for irrigation (78%), although 22% of it is
drinking water. The selected section is Y-shaped (see Fig. 3), a main
canal split into two with a gate placed at the entry of each of them.
Canal de la Pedrera, the total length of this canal is 6.680 km.
Canal de Cartagena; in our case-study only a part of this canal is
used (17.444 km).
The total length of the canals is approximately 24 km.
The most important elements in the canals are the main gates
which regulate the level of water along the canals and the off-take
gates, where farmers take water from the canals for irrigation. There
are 7 main gates and 17 off-take gates in the section selected, hence,
we have considered that the system is composed of seven subsystems at the lower level. Each subsystem begins at one of the main
gates and ends at the next one. There are, however, two exceptions.
The subsystems that begin at the gates labeled CCMICAR 08 and
CCMIPED 01 do not have a gate at the end because they are the last
canals considered. Each of the subsystems is controlled by an agent

1. Higher level.
Main target: to minimize cost due to internal and external
risks.
To minimize costs, the objective function described in Eq. (5)
has been used with values in = [1 1], that is, internal and
external risks will be optimized. To calculate the predicted
output, Eqs. (6) and (7) are used.
There is a 365 day study period (1 year) and a 1 day sampling
time.
Prediction horizon, N = 5 days.
Manipulated variables: mitigation actions, uM .
2. Lower level
Main target: to control water management in canals in order
to guarantee ow demanded by users. For this purpose, it is
necessary to maintain the level of the canal over the off-take
gate when ow is requested.
Controlled variables: upstream levels at the gates, hi .
Manipulated variables: ow at the head of the canal and the
position of the gates
Constraints:
(a) Maximum and minimum levels to guarantee that off-take
points are submerged.
(b) The total amount of water at the head over a determined
time period is limited.
(c) Maximum and minimum gates opening.
The study period is 1440 min (1 day) and the sampling time
1 min. This period corresponds to what happens during the
day 150 of the higher controller period.
The control horizon is set to Nc = 5 for all the agents. The prediction horizon for the agent i is equal to the control horizon
plus the delay i of the reach, that is, Np (i) = Nc + i .

Fig. 3. Scheme of the canal.

A. Zafra-Cabeza et al. / Journal of Process Control 21 (2011) 787799

793

Table 1
Data of Cartagena-La Pedrera irrigation canal.
Code

Type

Canal del Campo de Cartagena


Start of the Campo de Cartagena canal
CCMICAR-01
Gate
MICAR-01
Off-take
MICAR-02
Off-take
MICAR-03
Off-take
CCMICAR-04
Gate
MICAR-04
Off-take
MICAR-05
Off-take
MICAR-06
Off-take
MICAR-07
Off-take
MICAR-08
Off-take
CCMICAR-05
Gate
MICAR-09
Off-take
MICAR-10
Off-take
CCMICAR-06
Gate
CCMICAR-07
Gate
MICAR-11
Off-take
CCMICAR-08
Gate
Canal de la Pedrera
CCMIPED-01
MIPED-01
MIPED-02
MIPED-03
MIPED-04
MIPED-05
MIPED-06

Gate
Off-take
Off-take
Off-take
Off-take
Off-take
Off-take

P/G

Description

km

G
G
G
P

Initial Gate
Off-take 5 Fuensanta and Estafeta
Off-take 5 Palacete
Off-take 6 Santo Domingo
Gate Canal Pedrera
Off-take 7 Campo Salinas
Off-take 8 San Miguel

Off-take 9 Las Canadas


Off-take 10 San Miguel
Off-take 11 Campo Salinas
Gate Tunel San Miguel
Off-take 12 San Miguel
Off-take 13 Campo Salinas
Gate La Rambla La Fayona (start)
Gate La Rambla La Fayona (end)
Off take 14 Villamartin

Gate Canada
La Estacada

0.000
0.200
1.170
2.540
2.840
4.485
5.970
6.550
8.050
9.390
9.590
10.480
12.630
12.780
14.433
14.579
16.540
17.444

P
G
G
G
P
G
P

G
G
P
G
G
G

Dynamic model is based on Eqs. (1) and (2). Note that Eq. (1)
is linear, but Eq. (2) is not. Thus, it has to be linearized so that
the system model can be expressed in the form of Eq. (10). For
simplicity, we have assumed that the terms hup (t) and hdn (t)
of Eq. (2) are constants whose value is estimated.
4.3. Risks identication
A number of potential risks can be encountered when the IC is
operating. Table 2 shows the risks that have been considered in this
example. Some have been provided by the INOCSA technical staff.
Initial impacts (II) are expressed on the parameters Z ={Z1 , Z2 }, with
Z1 being the cost (euros/day) and Z2 the variation on reference level
in reaches (m).
Risk R1 is stated as Inadequate fresh water quality. If this risk
occurs, an impact of 2000 euros/day is incurred; the probability is
constant over time: 0.1. R2 is associated with failure at gates. The
probability of this risk depends on the canal operation: the wear
and tear of the gate increases the likelihood of deterioration. Thus,
a function  1 that depends on the control variables that operate
the gates (u) and the time, has been included. As the use of the
gates and the time of operation is increasing, the value returned
by  1 will be greater. R3 depends on the water level and on the

Starting of the canal La Pedrera


Off-take 1P Santo Domingo
Off-take 2P Santo Domingo y Mengoloma
Off-take 3P Santo Domingo
Off-take Riegos Levante 1
Off-take 4P Santo Domingo
Off-take Riegos Levante 2 and 3

0.000
0.770
3.740
4.260
5.260
6.440
6.680

time. In the same way, a function  2 depending on hi and t has been


included. If farmers take un-authorized, R4 may appear. This risk is
well-known usually appears in drought season. P4 is modelled as
a normal distribution with mean 0.5 and deviation 0.3 in months
June and July. During the rest, this probability is 0. Fig. 4 shows the
random probability of P4 over time.
R4 has impacts on Z2 , an increasing of 15% on the initial level
reference. R5 states the possibility of water logging. The probability
depends on the season. The impact on water level reference consist
of an increasing on the initial water level reference. This amount
is based on the rainfall forecast of Murcia RF(t)(l/m2 ). R6 has been
established as the possibility of the water strategy for the plant
being modied by changes in the government, being only 0.1 for the
last quarter. R7 and R8 are external risks related to incentives for IC
plants from government and events of force majeure, respectively.
Both of them have an impact on cost and a constant probability.
4.4. Mitigation plan
There are six mitigation actions and therefore, uM =
{uM1 , . . . , uM6 }, where uM5 and uM6 are real and the rest are
boolean. A description of the actions used to mitigate risks is
shown in Table 3. The third column represents functions f and g

Table 2
Risk description (case study).
Rr

Description

Impacts

Pr (t)

Internal risks
Operation and Maintenance
Inadequate fresh water quality.
R1
Failure in gates due to wear and tear
R2
Seepage losses
R3

II11 = 2000/II12 = 0
II21 = 400/II22 = 0
II31 = 10/II32 = 0

0.1
0.1 +  1 (u, t)
0.1 +  2 (hi , t)

External risks
Politics and Weather
R4
R5
R6
R7
R8

II41 = 0/II42 = + 0.15hi (t)


II51 = 0/II52 = RF(t)
II61 = 250/II62 = 0
II71 = 2000/II72 = 0
II81 = 6000/II82 = 0

P4 (t)
P5 (t)
P6 (t)
0.01
0.01

Farmers, water demand varies from forecast


Rainfall changes water level of canal, producing water logging of adjacent lands
Changes in politics modify the strategy
State policies provide incentives for IC systems
Uninsured events of force majeure

794

A. Zafra-Cabeza et al. / Journal of Process Control 21 (2011) 787799

Probability of R4 (P4(t))
1

0.9

0.8

Probability

0.7

0.6

0.5

0.4

0.3

0.2

0.1

30

60

90

120

150

180

210

240

270

300

330

360

Days
Fig. 4. Random probability of R4 according N (0.5,0.3) over time.

that model the reduction of impacts and cost of execution, respectively. The fourth column is the period of validity of the action
(D = Daily, W = Weekly, B = Biyearly, Y = Yearly). That means if an
action is executed, it not will be reassessed until past the time of
validity. For example, an insurance contract may be executed every
180 days (if estimated) and a water analysis may be undertaken
every week.
Some functions f and g depend proportionally on the impacts.
For example, f11 = 0.7II1 uM1 means that the execution of this action
reduces the impact by 70%. Mitigation actions are carried out
depending on the setting of the execution frequency. The risk-based
structure with links between risks and actions is shown in Fig. 5.
Values from terms RAra in Eq. (4) are obtained from this gure.
Some example are: RA11 = 1, RA12 = 1, RA13 = 0 or RA45 = 1; that is,
risk R1 is mitigated by actions A1 and A2 and A4 is mitigated by
A5 . A5 states the modication of water level reference as a consequence of A4 or A5 . Note that this actions assumes the increasing
of water level, II26 and reduces the cost. A6 models an insurance
contract. The value of uM6 means the cost of the contract and f16
the insurance coverage.
4.5. Results and discussion
As a consequence of boolean variables, the problem is set as a
mixed integer quadratic problem. Results have been obtained using
the commercial solver Cplex [13].

Higher controller. Terms Jint and Jext in Eq. (5) have been dened
as:
2
5 


Jint (uM , t) =

(Y int,c (t + k|t)) ,

(20)

k=1 c=1

Jext (uM , t) =

2
5 


(Y ext,c (t + k|t)) ,

(21)

k=1 c=1

where
Y ext,1 =

8


REr1 (u, t + k)

(22)

REr2 (u, t + k)

(23)

r=1

Y ext,2 =

8

r=1

For risks R4 and R5 , the rainfall forecast for the city of Murcia
and the discharge from farmers have been considered during 2009.
Fig. 6 shows the rainfall forecast in the top panel; the initial level is
represented by a dotted green line. This level is modied by R2 and
R3 , giving rise to an actual level reference shown by the solid red
line (this is because action A6 is executed). Note that in the summer
season the level is increased as farmers may demand more water
as a result of drought.

Table 3
Mitigation actions description (case study).
Aa

Description

f1a , g1a on Z1 (cost)

PV

A1
A2
A3
A4
A5
A6

Periodic water analysis


Control weed growth
Appropriate monitoring or control over devices
Lining Irrigation Canal
Modify set-points of water levels (uM5 R)
Insurance policy (uM6 R)

f11 = 0.7II1 uM1 , g11 = 250uM1


f12 = 0.3II1 uM2 , g12 = 1500uM2
f13 = II1 uM3 , g13 = 250uM3
f14 = 0.95II1 uM4 , g14 = 2700uM4
f15 = 0, g15 = 0
f16 = 225uM6 , g16 = uM6

W
B
W
Y
D
B

A. Zafra-Cabeza et al. / Journal of Process Control 21 (2011) 787799

795

Irrigation canal

External risks
(Politic and weather)

Internal risks
(opearation and maintenance)

R2

R1

A1

A2

R3

R6

R7

A5

A4

A3

R5

R4

R8

A6

Fig. 5. Risk-based structure for the case study.

Fig. 7 shows the following costs:


1. With risks but no mitigation (dashed blue line) where impacts
are considered but no actions are executed to reduce them.
2. With mitigation (solid red line) where mitigation actions are
executed to reduce impacts.
The no mitigation case (dashed blue line) is computed considering the accumulative impacts on cost day to day. The mitigation
line takes into account the reduction of the impact and the cost of
the actions when they are carried out. Note how the no mitigation
option reects the highest cost. As expected, the proposed cost is
lower than the no mitigation line.
The mitigation actions to be executed to reach the optimum
costs, are shown in Fig. 8; notice that they are undertaken depending on the period of validity shown in Table 3. The mitigation actions
are chosen by considering the probabilities of risk with time.

Actions A1 and A3 are weekly executed. Action A2 is carried out


every 180 days. The frequency of A4 is one year; it is only carried
out at the end of the period due to the increasing of probability P3 .
A5 sets the increasing/decreasing on the water level reference due
to risks R4 and R5 . uM6 is the cost of the insurance contract (g16 =
uM6 ). This action takes the value 5000 because a constraint has been
added imposing that the cost of insurance cannot be higher than
5000 euros.
Lower controller
If the reference changes, the higher controller sends the modications to the DMPC at the lower level. The DMPC controller has
a sample time of 1 min and the control horizon Nc has been set
to 5. The prediction horizon for the agent i is equal to the control
horizon plus the delay i of the reach, that is, Np (i) = Nc + i . Note
that the prediction horizon has to be big enough to compensate
the transport delay in the calculations.

Rainfall forecast in Murcia


40

mm

30

20

10

30

60

90

120

150

180

210

240

270

300

330

360

Water level reference in a reach


3.5

Initial reference
Reference + risks

3.4

3.3
3.2
3.1
3
2.9
2.8

30

60

90

120

150

180
Days

210

240

270

300

330

360

Fig. 6. Top panel: rainfall forecast in Murcia. Lower panel: level reference in one reach by considering risks R2 , R3 and action A6 .

796

A. Zafra-Cabeza et al. / Journal of Process Control 21 (2011) 787799


4

Accumulated internal risks costs

x 10
15

Risks Mitigation
No mitigation

Euros

10

30

60

90

120

150

180

210

240

270

300

330

360

270

300

330

360

Accumulated external risks costs

x 10
2

Euros

1.5
1
0.5
0
0

30

60

90

120

150

180
Days

210

240

Fig. 7. Optimization of the cost by considering risks.

Several simulations have been performed for the DMPC controller for a one-day period. In these simulations, all the reaches
begin with a water level of 3.0 m and there is a change of reference
for all the reaches to 3.40 m at time k = 0. This change is originated
at the higher control level as a function of the risk mitigation policy.
In particular, the change of reference corresponds to the day 150 in

Fig. 6, where the evolution of the references during a 1-year period


is depicted.
The simulation shown in Fig. 9 corresponds to the nominal case,
that is, the simulation was performed without disturbances. It can
be seen how the reference is followed for all the reaches. In Fig. 10
non measurable disturbances are considered. In particular, we have

A1

MITIGATION ACTIONS

A2

1
0.5
0

A3

1
0.5

A4

A5

0.4
0.2
0

A6

4000
2000
0

30

60

90

120

150

180

210

240

270

Time (days)
Fig. 8. Mitigation actions to be undertaken to reduce risks impacts.

300

330

360

A. Zafra-Cabeza et al. / Journal of Process Control 21 (2011) 787799


3.5

levels (meters)

3.4
3.3

h1

3.2

h2
h

3.1

h4

2.9
h6

2.8

h7

2.7
2.6
2.5
0

200

400

600

800

1000

1200

1400

k (minutes)
Fig. 9. Levels in reaches for day 150 for the nominal case.

focused on the effect of the farmers activity. For this reason, we


have only considered disturbances which decrease the water level
in the reaches. As it can be seen, there is a steady state error due
to the effect of the disturbances. In Fig. 11 we show the case in

797

which the agents can measure the disturbances. In this case the
response is better than in Fig. 10 although slower than the one
shown in Fig. 9. Finally, in Fig. 12 we show a comparison between
the absolute effects of the disturbance (hd3 ) and the manipulated
variable (hu3 ) in the level of the reach 3, expressed in meters per
minute. In other words, hd3 is the absolute value in the variation
of water level due to farmers and hu3 represents how the manipulated variable tries to correct the deviation of the water level with
respect to its reference.
Note that even if the effect of the disturbance is a decrement of
the water level, it has been depicted with a positive sign, so that
its magnitude can be compared with the effect of the manipulated
variable in a simpler way. Likewise, note that we have modelled the
disturbances as a train of steps of one hour length with a random
amplitude.
It should be noted that these results have been obtained with
an average number of 5 communications per agent and sampling
time; in other words, each agent makes up to 5 proposals a minute
in order to get a cooperative solution with the rest of the agents.
Note that, given the random component of the distributed MPC
algorithm, it is not possible to make a deterministic comparison
with a centralized MPC in order to evaluate its loss of performance. For this reason, we have performed different simulations in
order to determine the average of the loss of performance. We per-

3.5
3.4
3.3

levels (meters)

3.2
3.1
3
h1
h2
h3
h4
h5
h6
h7

2.9
2.8
2.7
2.6
2.5

200

400

600

800

1000

1200

1400

k (minutes)
Fig. 10. Water levels in the case of non measurable noise.

3.5
3.4
3.3

levels (meters)

3.2
3.1
3
h1
h2
h3
h4
h5
h6
h7

2.9
2.8
2.7
2.6
2.5

200

400

600

800

1000

k (minutes)
Fig. 11. Water levels in the case of measurable noise.

1200

1400

798

A. Zafra-Cabeza et al. / Journal of Process Control 21 (2011) 787799


3

1.5

x 10

hd3

Absolute level increment


(meters/min)

hu3

0.5

200

400

600

800

1000

1200

1400

k (minutes)
Fig. 12. Absolute changes in the level of reach 3 due to its manipulated variable (U3 ) and the farmers activity (D3 ).

formed twenty different simulations and obtained approximately a


7% increment in the cumulated cost of the simulations with respect
to a centralized MPC scheme.
5. Summary and conclusions
This paper describes a control-based methodology for decisionmaking in irrigation canals to address prevention and control
problems in the plant. The objective is to optimize the operation
of the system, taking into account explicitly modelled risks that
can be identied prior to the planning.
MPC is particularly meaningful for the given problem given
because of its favorable properties, such as ease of constrainthandling, extension to multivariable case, time delays inclusions
or changing objectives. The extension to the distributed approach
has improved the results; the system has been divided into agents
that exchange information about the control variables.
Risk modelling involves risk identication, assigning probabilities, and devising a strategic plan to mitigate risks; therefore,
getting information from weather forecasts, failures in operations
and trained personnel to generate these models is crucial to the
success of this approach. The presented approach provides recommendations on the actions to undertake in order to mitigate risks
that could appear. The procedure can be considered as a helpful
tool to assist experts in evaluating different scenarios providing
a denitive set of mitigation actions and values of control variables. Finally, it is worthwhile to mention that the control policy
at the lower level can be implemented in a distributed fashion that
requires a small amount of communication between the nodes in
order to get a cooperative solution.
Acknowledgement
This research has been supported by the European 7th Framework STREP Project Hierarchical and distributed model predictive
control (HD-MPC), contract number INFSO-ICT-223854 and DPI2008-05818.
References
[1] A. Alessio, A. Bemporad, Decentralized model predictive control of constrained
linear systems, in: Proceedings of the 2007 European Control Conference, 2007,
pp. 28132818.

[2] S. Baggett, P. Jeffrey, B. Jefferson, Risk perception in participatory planning for


water reuse, Desalination 187 (2006) 149158.
[3] S. Baines, K. Sinha, K. Kirschchenheiter, Risk mitigation by managing water
treatment system interfaces in the power industry, in: International Water
Conference, 2004.
[4] M.M. Baron, M.E. Pat-Cornell, Designing risk-management strategies for critical engineering systems, IEEE Transactions on Engineering Management 46 (1)
(1999) 87100.
[5] O. Begovich, V.M. Ruiz, G. Besancon, C.I. Aldana, D. Georges, Predictive control with constraints of a multi-pool irrigation canal prototype, Latin American
Applied Research 37 (2007) 177185.
[6] C.P. Buyalski, D.G. Ehler, H.T. Falvey, D.C. Rogers, E.A. Serfozo, Canal Systems
Automation Manual, vol. 1, U.S. Department of Interior, Bureau of Reclamation,
1991.
[7] E.F. Camacho, C. Bordons, Model Predictive Control, 2nd edition, SpringerVerlag, London, 2004.
[8] E. Camponogara, Controlling networks with collaborative nets, PhD Thesis,
Carnegie Mellon University, September 2000.
[9] V.T. Chow, Open-Channel Hydraulics, Mc-Graw-Hill Book, 1859.
[10] J.M. Coron, B. dAndrea Novel, G. Bastin, A lyapunov approach to control irrigation canals modeled by saint-venant equations, in: Proceedings del ECC,
1999.
[11] H. El Fawal, D. Georges, G. Bornard, Optimal control of complex irrigation systems via decomposition-coordination and the use of augmented lagrangian,
in: Proceedings of the 1998 International Conference on Systems, Man and
Cybernetics, San Diego, California, 1998.
[12] R. Fletcher, S. Leyffer, Numerical experiences with lower bounds for miqp
branch-and-bound, SIAM Journal on Optimization 8 (2) (1998) 604616.
[13] ILOG, http://www.ilog.com/products/cplex/, 2006.
[14] D. Jia, B. Krogh, Distributed model predictive control, in: Proceedings of the
2001 American Control Conference, 2001, pp. 45074512.

B. Ohran, P.D. Christodes, J.F. Davis, A two-tier


[15] J. Liu, D. Munoz
de la Pena,
architecture for networked process control, Chemical Engineering Science 63
(2008) 53495409.

E.F. Camacho, Distributed model pre[16] J.M. Maestre, David Munoz


de la Pena,
dictive control based on a cooperative game, Optimal Control Applications and
Methods, doi:10.1002/oca.940.
[17] I. Mareels, E. Weyer, S.K. Ooi, M. Cantoni, Y. Li, G. Nair, System engineering for
irrigation systems: successes and challenges, Annual review in control, IFAC 29
(2005) 191204.
[18] R.R. Negenborn, P. van Overloop, T. Keviczky, B. de Schutter, Distributed model
predictive control of irrigation canals, Networks and Heterogeneous Media 4
(2) (2009) 359380.
[19] H. Plusquellec, K. McPhail, C. Polti, Review of irrigation system performance
with respect to initial objectives, Irrigation and Drainage Systems 4 (1990)
313327.
[20] B. De Schutter, R.R. Negenborn, H. Hellendoorn, Multi-agent model predictive
control of transportation networks, in: Proceedings of the 2006 IEEE International Conference on Networking, Sensing and Control (ICNSC 2006), Ft.
Lauderdale, Florida, April, 2006, pp. 296301.
[21] A.A. Radwan, D.T. Shaw, Performance of integrated municipal water systems
during drought, ASCE Journal of Water Resources Planning and Management
120 (4) (1994) 531545.
[22] A. Rantzer, Dynamic dual decomposition for distributed control, in: American
Control Conference, 2009, 2009.

A. Zafra-Cabeza et al. / Journal of Process Control 21 (2011) 787799


[23] R. Scattolini, Architectures for distributed and hierarchical model predictive
control a review, Journal of Process Control 19 (2009) 723731.
[24] H. Scheu, J. Busch, W. Marquardt, Nonlinear distributed dynamic optimization based on rst order sensitivities, in: Proceedings of the American Control
Conference, Baltimore, Maryland, USA, July, 2010.
[25] A.N. Venkat, J.B. Rawlings, S.J. Wright, Stability and optimality of distributed
model predictive control, in: Proceedings of the 44th IEEE Conference on Decision and Control, and the European Control Conference ECC 2005, Seville, Spain,
2005, pp. 66806685.
[26] S.N. Wan, G. Kuczera, Incorporating demand uncertainty in water supply headworks simulation, Water Resources Research 29 (2) (1993) 469478.
[27] A. Zafra-Cabeza, M.A. Ridao, E.F. Camacho, Applying risk management to combined heat and power plants, IEEE Transactions on Power Systems 23 (2008)
938945.

799

[28] A. Zafra-Cabeza, M.A. Ridao, E.F. Camacho, A mixed integer quadratic programming formulation of risk management for reverse osmosis plants, Dessalination
268 (2011) 4654.
[29] A. Zafra-Cabeza, M.A. Ridao, E.F. Camacho, K.G. Kempf, D.E. Rivera, Managing
risk in semiconductor manufacturing: a stochastic predictive control approach,
Control Engineering Practice 15 (8) (2007) 969984.
[30] Y. Zhou, R.S.J. Tol, Economic analysis of domestic, industrial and agricultural
water demands in China, Water Supply 5 (6) (2005) 8593.
[31] X. Zongxue, K. Jinno, A. Kawamura, S. Takesaki, K. Ito, Performance risk analysis
for fukuoka water supply system, Water Resources Management 12 (1998)
1330.