Control Engineering Practice 21 (2013) 870886

Contents lists available at SciVerse ScienceDirect

Control Engineering Practice

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

SCADA system with predictive controller applied to irrigation canals

Joa~ o Figueiredo a,n, Miguel Ayala Botto b, Manuel Rijo c
vora, R. Roma~ o Ramalho, 59, 7000-671 E
vora, Portugal
CEM/IDMEC, Universidade E
IDMECInstituto Superior Tecnico, Technical University of Lisbon, R. Rovisco Pais, 1049-001 Lisboa, Portugal
c
vora, Apartado 94, 7002-554 E
vora, Portugal
NuHCC, Universidade E
a

a r t i c l e i n f o

a b s t r a c t

Article history:
Received 3 May 2011
Accepted 28 January 2013
Available online 25 March 2013

This paper applies a model predictive controller (MPC) to an automatic water canal with sensors and
actuators controlled by a PLC network (programmable logic controller), and supervised by a SCADA
system (supervisory control and data acquisition). This canal is composed by a set of distributed subsystems that control the water level in each canal pool, constrained by discharge gates (control
variables) and water off-takes (disturbances). All local controllers are available through an industrial
network managed by the SCADA system, where the centralized predictive controller runs.
In this paper a complete new platform connecting the SCADA supervisory system and the MATLAB
software (named SCADAMATLAB platform) is built, in order to provide the usual SCADA systems with
the ability to handle complex control algorithms. The developed MPC-model presents a novelty in the
control of irrigation canals as it allows the use of industrial PLCs to implement high complex
controllers, through the new developed SCADAMATLAB platform.
Experimental results demonstrate the reliability and effectiveness of the proposed strategy in
real-life typical situations, including gate malfunctioning and extreme water off-take conditions.
& 2013 Elsevier Ltd. All rights reserved.

Keywords:
Predictive Control
Automation
Supervisory control
Multivariable control

1. Introduction
Irrigation is the largest water user in the world, using up to
85% of the available water in developing countries (Plusquellec,
Burt, & Wolter, 1994). Irrigation canal automation can decisively
contribute to attain a necessary exibility in water management.
In the last decades, the advances in modern computational
technology and industrial communication allow the combination
of unsteady open channel ow simulation models with real-time
control algorithms (Burt & Piao, 2004;, Clemmens, Bautista,
Whalin, & Strand, 2005). This approach has allowed signicant
advances in the engineering of canal control and automation. The
present study follows this line of research.
The main goal in canal control for agricultural purposes is to
minimize the water waste when supplying water to farmers.
Since the off-takes are, in most cases, gravity fed, the requirement
of being able to supply water has traditionally been converted
into set-point regulation of water levels.
The design of a control strategy that is able to handle exible
water delivery schedules while simultaneously dealing with the
overall water canal physical constraints, is of key importance in
order to reduce inefciency and, consequently, water waste.

Corresponding author. Tel.: 351964245394; fax: 351266745394.

E-mail address: jg@uevora.pt (J. Figueiredo).

0967-0661/$ - see front matter & 2013 Elsevier Ltd. All rights reserved.
http://dx.doi.org/10.1016/j.conengprac.2013.01.008

The available research on canal control algorithms can be

grouped into four main categories: heuristics, PID-type, optimal
and predictive (Malaterre, Rogers, & Schuurmans, 1998). Nowadays, most of the automated water canals are controlled with PItype decentralized controllers. Despite the resulting improvement
in comparison with manual canal operation, such decentralized
control is usually difcult to tune and can hardly cope with
physical constraints. However, mainly due to their robustness and
ease to implementation in the eld, the classical PI-controllers are
still actual in canal control research mainly with the focus of
enhancing the tuning of the controller gains (Litrico, Malaterre,
Baume, Vion, & Ribot, 2007). A huge collection of this type of
controllers, as well as heuristic ones, can be found in the literature
(Guenova, Litrico, & Georges, 2004; Litrico, Fromion, Baume, &
Rijo, 2003;, Ooi & Weyer, 2008). Examples of optimal linear
controllers for water canals control can be found in (Malaterre,
1998; Weyer, 2003; Feliu-Battle, Perez, & Rodriguez, 2007).
Most of previously referred controllers propose decentralized
solutions that neither consider the multivariable nature of the
problem nor the natural and physical constraints of the system.
Besides, they often disregard water off-takes known in advance,
and only focusing on the downstream water level controls in each
pool. Fig. 1 explains the main nomenclature in the eld of canal
control (upstream and downstream).
Considering the control of canals there are two main
approaches (Burt, 1987): upstream control (Fig. 2a) and downstream control (Fig. 2b), each referring to the location from which

J. Figueiredo et al. / Control Engineering Practice 21 (2013) 870886

Fig. 1. Main variables in canal control.

Fig. 2. Main strategies on canal control: upstream and downstream controllers.

information is needed by the controller in relation to the actuator

(usually the gate opening/closing).
Under upstream control the adjustments on the gate opening/
closing are based on information from upstream (Fig. 2a); thus
the upstream control is appropriate for canal systems that are
supply-oriented.
Under downstream control the adjustments on the gate opening/closing are based on information from downstream (Fig. 2b).
Downstream control transfers the off-take demands to the
upstream water supply source, thus it is appropriate for
demand-oriented delivery systems.
Upstream control canals are only efcient when they operate
with rigid water delivery methods. Nevertheless, the major part of
water distribution networks work with exible water delivery
schedules and, in these contexts, the above mentioned control
strategy implies great operational losses.
Most of the few automated water canals use DMPIC (downstream water-level SISO PI-controller) to control the pools

871

downstream water levels. Similarly, UMPIC (upstream waterlevel SISO PI-controller) is frequently used to control the pools
upstream water levels. Since both DMPIC and UMPIC are decentralized systems, a local single controller is usually applied to
each gate. Besides, since there is no information interchanged
between each local controller, this type of control conguration is
not able to handle disturbance rejection in an efcient way. In
Almeida, Figueiredo, and Rijo (2002), (Ratinho, Figueiredo, and
Rijo (2002) there are shown experimental results of DMPIC and
UMPIC applied to the experimental canal that is here under study.
An alternative to overcome these constraints could be the design
of a MPC (model predictive controller) (Garcia & Morari, 1989;
Maciejowski, 2002;, Dougherty & Cooper, 2003). MPC is an optimization based control strategy that uses a process model, called the
prediction model, to foretell future process behaviour for a given
prediction horizon. Besides, MPC is particularly suited to deal with
time-variant constraints (process and physical). The design of an MPC
for water canals control has been addressed by several researchers
(Malaterre & Rodellar, 1997; Malaterre & Baume, 1998; Baena, 2003;
Overloop, Weijs, & Dijkstra, 2008; Xu, Overloop, & Giesen, 2011;
Cabeza, Maestre, Ridao, Camacho, & Sanchez, 2011).
In this paper, a centralized MIMO model based on discretized
Saint-Venant equations is applied to model the dynamical characteristics of the whole water canal. This dynamic model integrates all the hydraulic structures, such as gates, off-take valves
and water level constraints, as well as other physical constraints
of the particular experimental canal. The integration of this
centralized prediction model into the MPC scheme will enable
the design of a realistic controller that is able to consider water
off-takes known in advance, while guaranteeing the overall
physical constraints satisfaction.
The developed MPC controller is tested on an experimental
water canal. The canal setup is an automated based-PLC system
with a master-slave network of PLCs (programmable logic controller), which is supervised by a SCADA system.
The major contribution of the present study is the development of a complete new platform connecting the SCADA supervisory system and the MATLAB software (named SCADAMATLAB
platform) in order to provide usual SCADA systems with the
ability to handle complex control algorithms.
The paper is structured in the following way: In Section 2 the
hardware and control architecture of the canal prototype is
characterized. In Section 3 the whole canal model is presented.
In Section 4 the MPC is developed for the canal under study.
Finally Section 5 shows the experimental comparative results
between the MPC-model and the traditional decentralized PIcontrollers. The analytical developments of the canal model
(assembly of the several linearized sub-structurescanal pools,
off-take valves and gates) are presented in Appendix A.

2. Facility description
The experimental automatic canal is located in the University
of Evora, Portugal (Fig. 3).
It consists of a closed-loop water canal and an instrumental
platform that integrates electro-mechanical sensors and actuators,
a network of PLCs and a SCADA supervisory system. First presenting the automatic canal, it has four pools with a nominal capacity
of 0.090 m3/s ow for a uniform water depth of 0.700 m. The canal
main geometric properties are illustrated in Fig. 4 and Table 1.
The water in this experimental canal ows in a closed circuit,
regarding water savings. The return to the storage reservoir is
guaranteed by a traditional irrigation canal.
The pools of the automatic canal are divided by three sluice
gates. At the canal end there is an overshot gate. Immediately

872

J. Figueiredo et al. / Control Engineering Practice 21 (2013) 870886

Fig. 3. The automatic canal (right) discharges the ow to the traditional irrigation
canal (left).

Fig. 4. Geometry of a canal pool.

Table 1
Main geometric characteristics of the experimental canal.

Pool
Pool
Pool
Pool

1
2
3
4

through an industrial MODBUS network (RS 485) (ModBus, 2012).

All this interconnected equipment is illustrated in Fig. 6. The main
geometric properties are shown in Table 2, where the mnemonics
account for SiTjSensor Si in Pool Tj.
Resuming Fig. 6, it illustrates the complete setup that consists
of two main structures, named Research Building and On
eld, that communicate using both protocols: Modbus (RS485)
and RS232.
The Research Building is the intelligent framework where
the controller actions are computed (SCADAMATLAB platform).
The SCADAMATLAB platform develops an extremely effective
role allowing the system manager to continuously compare the
actual hydraulic state of the canal with its optimal hydraulic state
(from model simulations) and to take appropriate actions.
The On eld structure comprises the canal infrastructure and
its instrumental platform consisting of 5 interconnected slavePLCs (MODBUS network), several sensors (water levels and water
ows from off-takes) and actuators (motorized gates).
The MPC control algorithm is implemented using the active
sets solver from MATLAB (Matlab, 2008a; 2008b; Bemporad,
Morari, & Ricker, 2005) as it is explained in Section 4.3. The
MPC algorithm is too complex to be implemented directly on the
SCADA system (Axeda, 2002), as usual SCADA supervisors do not
handle heavy mathematical computations.
The communication between SCADA and MATLAB is performed using the DDE protocol (dynamic data exchange). This
communication protocol permits the exchange of data between
two independent running software programs (Client and Server).
In the developed application the MATLAB software is the Client, as
it initiates the communication, and the SCADA software is the
Server, as it responds to Clients requests. For the assignment
client/server the choice was based on the communications main
ux. In this implementation, the main communication ux has
the direction MATLABSCADA, so the SCADA was dened as the
server. This strategy permits to design and test very different
complex controllers without changing the plant setup. The SCADA
system works as a main interface, updating inputs and outputs,
assuring the communication between the controller (MATLAB)
and the plant (PLC).

3. Mathematical model

Total length (m)

Slope

40.37
34.87
35.14
26.55

0.0016
0.0014
0.0019
0.0004

upstream of each gate there is an off-take valve which is equipped

with a ow-meter. Counterweight-oat level sensors are distributed along the canal. Electro-valves control the ow in the offtakes. All the discharge gates are electro-actuated and they are
instrumented with position sensors. Fig. 5 illustrates the existing
equipment in the neighbourhood of each discharge gate.
The ow within the automatic canal is regulated by an electrovalve located at the exit of an high reservoir (head of the automatic
canalrepresented by MC1 in Fig. 6), simulating a real load
situation. This high reservoir is lled with the recovered water
pumped from a low one, which collects the ow from the
traditional canal. All actuators and sensors in the canal are
connected to local PLCs (programmable logic controllers) that are
responsible for the sensor data acquisition and for the computed
control actions sent to the actuators. All local PLCs are connected

The mathematical model of the system under study is

obtained by assembling all the linearized hydraulic equations
from the canal main structures: pools, off-take valves and gates.
The development of the linearized models for these
sub-structures is shown in Appendix A.
Considering the standard space state notation for a MIMO system
(multiple input, multiple output) one obtains the general equations:
A1  X k 1 B1  X k B2  U kperturbations B3  U kcontrol

Y k C output  X k

After small manipulation one gets the nal formulation:

X k 1 Adynamics  X k Bperturbations  U kperturbations Bcontrol
 U kcontrol

Y k C output  X k
where
Adynamics A1
1  B1

J. Figueiredo et al. / Control Engineering Practice 21 (2013) 870886

873

Fig. 5. Schematics of the electromechanical equipment in each gate off-take.

Fig. 6. Schematics of the complete facility.

Bperturbations A1
1  B2
Bcontrol A1
1  B3

The nal number of states will depend on the number of sections

considered in the discretization. In this paper, the experimental canal
under study has four pools (Fig. 4) and each pool is divided into 11

874

J. Figueiredo et al. / Control Engineering Practice 21 (2013) 870886

sections, conning 10 intervals of equal dimensions within each pool,

as illustrated in Table 3:
The corresponding space-state vector, without the consideration of the 3 discharge-gates, that are separating the 4 canal pools,
is a 88-coordinate vector, (22 coordinates for each canal pool) and
it is given by:
2
dQ k1 1 dyk1 1 dQ k2 1 dyk2 1    dQ k11 1 dyk11 1 ,
k1
X
4 |{z}
Pool1-10 int ervals-11 sections-112 states

3
7
   dQ k22 1 dyk22 1   7
5
|{z}

dQ k12 1 dyk12 1 dQ k13 1

dyk13 1

Pool 2-10 intervals-11 sections-112 states

Considering now the model of the discharge-gates that

connect the 4 canal pools, the nal space-state vector will be a
85-coordinate vector (3), because the coordinatesdQ 11 , dQ 22 and
dQ 33 , will vanish as they become redundant when making the
ow continuity between the four canal pools.
h
k1
dyk1 1 dQ k2 1 dyk2 1    dQ k10 1 dyk10 1 dyk11 1
Xk 1 Q 1

dQ k12 1 dyk12 1 dQ k13 1

dyk13 1



dQ k21 1 dyk21 1

a ow-sensor located at the head of the canal. This output set gives
13 output variables as indicated in Eq. (10).
2
3
dQ koff-take1
6
7
6 dQ k 1
7
6
off-take1 7
6
7
6 dQ k
7
6
off-take2 7
6
7
6 dQ k 1
7
6
off-take2 7
k
6
7
U perturbations 6
8
k
7
6 dQ off-take3 7
6
7
6 dQ k 1
7
6
off-take3 7
6
7
6
7
k
6 dQ off-take4 7
4
5
1
dQ koff-take
4
2

3
dQ k1 1 dQ k1
6
7
6 dak 1 dak 7
1 7
1
6
6
7
k1
k 7
U kcontrol 6
6 da2 da2 7
6
7
6 dak3 1 dak3 7
4
5
dQ k12 1 dQ k12

dyk22 1

39
>
>
>
6a
7>
>
6 11 a12 a13 a14
7>
>
6
7>
>
6 a21 a22 a23 a24
7>
>
6
7>
>
6
7>
>
a11 a12 a13 a14
6
7>
>
6
7>
>
>
6
7
>
a21 a22 a23 a24
6
7>
>
>
6
7>
&
6
7>
>
6
7>
>
6
7>
>
a14
a11 a12
6
7>
>
>
6
7>
>
>
6
7
a21 a22
a24
=
6
7>
 
 
@Q j
@Q j
6
7
A1 6
 @y
1  @y
7 85
i
j
6
7>
r
r
>
6
7>
a11
a12
a13 a14
>
6
7>
>
6
7>
>
6
7>
a
a
a
a
21
22
23
24
>
6
7>
>
>
6
7
>
&
6
7>
>
6
7>
>
6
7>
&
>
6
7>
>
6
7>
>
&
6
7>
>
6
7>
>
6
7>
a
a
a
a
11
12
13
14
>
6
7>
>
6
7>
>
a21 a22 a23 a24 5>
4
>
>
>
;
1
|{z}
2

85

dQ k23 1 dyk23 1 dQ k24 1

dQ k34 1

dyk34 1

k1
dQ 35

dyk24 1

dyk35 1





dQ k32 1 dyk32 1

dQ k43 1

dyk43 1

dyk33 1

dQ k44 1

dyk34 1

6
The state matrix [A1] is given by Eq. (7):
The row with the explicit partial differential equations refers
to the discharge-gate model.
The perturbations vector U kperturbations (ow at the off-takes) and
the control vector (main electro-valve represented by MC1 in Fig. 6
and the four gates, C1 to C4 in Fig. 6) U kcontrol are shown in
Eqs. (8) and (9), respectively. Finally the output matrix Coutput is
shown in Eq. (10). The composition of this output matrix is related
to the selected output vector, which is composed by 12
water level-sensors, distributed along the canal, and additionally

10

J. Figueiredo et al. / Control Engineering Practice 21 (2013) 870886

Table 2
Geometric characterization of the experimental canal.

Pool 1

Pool 2

Pool 3

Pool 4

Distance to monovar-MC1 [m]

Section type

Monovar-MC1
Section 1
Section 2
S1T1
S2T1
Transition 1
Transition 2
S3T1
Gate-C1
S1T2
S2T2
Transition 3
Transition 4
S3T2
Gate-C2
S1T3
S2T3
Transition 5
Transition 6
S3T3
Gate-C3
S1T4
S2T4
Transition 7
Transition 8
S3T4
Section 3
Section 4
Transition 9
Gate-C4

0.0
0.6
1.6
2.5
21.4
39.1
39.7
40.7
40.7
41.7
58.3
74.1
74.8
75.7
75.7
76.7
93.4
109.1
109.7
110.7
110.7
111.7
123.8
135.6
136.2
137.1
137.1
138.0
138.0
145.9

Rectangular
Rectangular
Trapezoidal
Trapezoidal
Trapezoidal
Trapezoidal
Rectangular
Rectangular
Trapezoidal
Trapezoidal
Trapezoidal
Trapezoidal
Rectangular
Rectangular
Trapezoidal
Trapezoidal
Trapezoidal
Trapezoidal
Rectangular
Rectangular
Trapezoidal
Trapezoidal
Trapezoidal
Trapezoidal
Rectangular
Rectangular
Trapezoidal
Trapezoidal
Rectangular
Rectangular

Table 3
Geometric characteristics of the canal intervals.
Pools

Intervals1y10

Pool
Pool
Pool
Pool

4.04 m
3.49 m
3.51 m
2.66 m

4. Predictive controller
Among the several approaches to predictive control, this paper
follows the MPC methodology (model predictive control) rst
referred by Garcia and Morari (1989). Fig. 7 illustrates the aim of
the predictive control strategy, where Hp is the prediction horizon
and Hc is the control Horizon. Usually one has Hc r Hp and Hp Z1.
Additionally, as it can be seen in Fig. 7, after the control horizon, it
results Duk i 0, i A Hc ; Hp .
The Model ((3) and (4)) is used to form predictions of future
states (Maciejowski, 2002):
k HP 9k
k HP 19k
k HP 19k
k HP 19k
x^
Ad x^
Bc u^ c
Bp u^ p
k9k
k HP 19k
Ad Hp xk Ad Hp 1 Bc u^ c    Bc u^ c
Ad Hp 1 Bp ukp

Ad Hp 2 Bp u^ p

k 19k

   Bp u^ p

k 19k
k 19k
k 19k
C y  x^
d^
y^

where
Ad Adynamics
Bp Bperturbations

k HP 19k

Future

Past

Element

1
2
3
4

875

11

12

Hp

Output prediction y k + j

Hc

K-1 K K+1

K+Hc K+Hp

Fig. 7. Basics of a predictive controller.

Bc Bcontrol
d output disturbance

13

4.1. Cost function

The implementation of a predictive controller involves typically the minimization of a cost function (Maciejowski, 2002).
In this study, focusing on the primary purpose of testing the
new SCADAMATLAB platform in canal control, it is selected a
whole quadratic cost function, relaying on the advantages that QP
(quadratic programming) formulations offer, regarding LP (linear
programming) or mixed QP/LP formulations (Maciejowski, 2002;
Fletcher, 1987; Bemporad, Morari, Dua, & Pistikopoulos, 2002).
The selected cost function is presented in Eq. (14):
02 k H 9k 3 2 k H 31T 02 k H 9k 3 2 k H 31
a
a
ry a
ry a
y^
y^
B6
7 6
7C B6
7 6
7C
6
76 ^ 7C Q B6
76 ^ 7C
^
^
Jk B
@4
5 4
5A @4
5 4
5A
k

H
k

H
p
p
k Hp 9k
k Hp 9k
ry
ry
y^
y^
02
3 2
31T 02
3 2
31
k9k
k9k
rkuc
rkuc
u^ c
u^ c
B6
7 6
7C
B6
7 6
7C
6
76
7C W B6
76
7C
^
^
^
^
B
@4
5 4
5A
@4
5 4
5A
k Hp 1
k Hp 1
k Hp 19k
k Hp 19k
ruc
ruc
u^ c
u^ c
2
6
6
4

Du^ k9k
c
^

Du^ kc Hp 19k

3T 2
7 6
7 R6
5 4

Du^ k9k
c
^

Du^ ck Hp 19k

3
7
7
5

14

where the superscript ^ species the estimated values.

The cost function V penalizes deviations from the estimated
k i9k
output y^
in relation to the values of the reference trajectory
k ijk
(rst term), deviations on the estimated control actions
r
k ijk
k ijk
u^ c
(second term),
in relation to the reference value r uc
k ijk
and changes on the estimated control actions Du^
(third
term). Q, W and R are weight matrices that inuence the outcome
of optimization. These weights determine the response of the
system and the selection of different values in the Q, W and R
matrices allows the system to reach different objectives. In fact it
is possible by varying the matrices coefcients to allocate more
weight to a certain variable in relation to other factors.
The present paper follows a simple and efcient methodology
where the matrices Q, W and R are diagonal matrices (symmetric
positive-denite) with constant weights designed for the whole
prediction horizon (Maciejowski, 2002). These weights were
calculated observing two main restrictions:

876

J. Figueiredo et al. / Control Engineering Practice 21 (2013) 870886

Minimization of the economic costs of the used energy;

Stabilization of the closed loop system.

Table 4
steady-state water depth in each canal section.
Pool

Beyond the fullling of these general restrictions the calculated weights verify additionally the following specic restrictions of the problem under study:

i) weights of output variablesdepending on the selected controller (upstream/downstream), all the weights in a controlled
pool are zero except for the controlled water depth, that has a
weight equal to 1. This implies that all the other measured
depths in the pool have freedom to vary in order to accomplish
the optimization of the built functional;
ii) weights of input variableshere two different weights were
considered; weights that can differentiate the input type (it
was assumed that all input variables have equal relevance);
and weights that penalize the change in inputs (it was
assumed for all input changes the weigh 0.1. This small factor
avoids too aggressive actions in the inputs).

4.2. MPC constraints

In this study, the optimization of the cost function is constrained by the allowable amplitudes of both the input variables
(9) and the outputs (4), (6) and (10). Usually these constraints are
represented by a set of equations:
2 Ha 3 2
3 2 Ha 3
ymin
ymax
yHa
6
6
7 6
7
6  7 6  7
6
7 6  7
6 Hp 7 6
7
7
H
6 y
6 y p 7
7
yHp 7
6 min 7 6
6 max 7
6
7
6 0
6 0
7 6
7
6 uc
7 6 u0c 7
7
7 6
min
6
6 ucmax 7
7 6
7
6  7 6  7 6  7
15
6
7r
7
7r6
6 Hp 1 7 6
6 Hp 1 7
p 1 7
6 uc
7 6 uH
ucmax 7
7 6
c
6 min 7 6
6
7
7 6
6
7 6
Du0cmax 7
Du0c 7
6 Du0cmin 7 6
7
7 6
6
6
7 6
7
6
7
6  7 4  5 6  7
4
4
5
5
H
1
H
1
Hp 1
p
Ducmax
Duc p
Ducmin
These equations after being specied for the problem under
study, they have to be written in terms of the increments in the
control variables, Dukc (Maciejowski, 2002).
Considering rst the output variables (water depths) these
cannot be negative nor exceed the limit value of the canal height
(0.9 m). To evaluate the canal water depth in the steady-state
regime, the Saint-Venant equation (Saint-Venant, 1871) is solved
for the particular case: (i) constant ow and (ii) time-constant
water-depth dependent on the position. Hence, this equation
becomes (with subscript r indicating steady-state regime):
!
@ Qr2
@y
g  Ar  r g  Ar  JI 03
@x Ar
@x
3g  Ar 

@yr Q r 2
@y

 B0 2  m  yr  r g  Ar  IJ3
|{z} @x
@x Ar 2
|{z}
@A
@Q 2 =A

@y

1
2
3
4

Section [m]
1

10

11

0.54
0.56
0.54
0.59

0.55
0.56
0.55
0.59

0.55
0.57
0.55
0.60

0.56
0.57
0.56
0.60

0.57
0.57
0.56
0.60

0.57
0.58
0.57
0.60

0.58
0.58
0.58
0.60

0.58
0.59
0.58
0.60

0.59
0.59
0.59
0.60

0.59
0.60
0.59
0.60

0.60
0.60
0.60
0.60

Table 5
maximal deviations in water depth in each canal section.
Pool

1
2
3
4

Sec. [m]
1

10

11

0.36
0.34
0.36
0.31

0.35
0.34
0.35
0.31

0.35
0.33
0.35
0.30

0.34
0.33
0.34
0.30

0.33
0.33
0.34
0.30

0.33
0.32
0.33
0.30

0.32
0.32
0.32
0.30

0.32
0.31
0.32
0.30

0.31
0.31
0.31
0.30

0.31
0.30
0.31
0.30

0.30
0.30
0.30
0.30

computed with the restrictions that all water depths cannot be

negative nor exceed the limit value of the canal height (0.9 m). The
constraint limits for increments in the water-depth for each pool/
section are presented in Table 5 (see dened sections in Table 3):
In order to calculate the constraints on the input variables, it is
needed to analyze the constraints on the allowable amplitudes
and increments. The considered input variables are supplied by
industrial equipment (main electro-valve and four motorized
gates) which operation intervals are specied by the suppliers.
Considering the time period of 10 s (as it will be justied in
Section 4.3), the maximal allowable increments for the electrovalve, within one sample period, are [  10 L/period; 10 L/
period], and for the gates are [ 30 mm/period; 30 mm/period].
Finally the input amplitudes are constrained by the actuators
specications to the following ranges: servo-valve [  30 L/s;
10 L/s]; gates [0; 800 mm].

4.3. MPCProblem solving

Fig. 8 presents a schematic conguration of the MPC applied to
the studied water canal.
Departing from the recursive formulation of the state space
model (11) and using the approach from Maciejowski (2002), the
MPC problem is formulated as the computed controller action Duc
that minimizes the cost function:
Jk Dukc T wDukc Dukc T G
Subjected to the constraints:
2
3
3
F
Fuk1
f
c
7
6
7 k 6
g 5
4 GY 5Duc r 4 GCxk Uuk1
c
Z
z

17

18

@A

@yr
IJ

@x
1V r 2  Br =g  Ar

16

The solution of this equation for each section gives the

stationary water depth in each pool/section.
Table 4 shows the solution of Eq. (16) when the particular
conditions of a constant ow (Q0.03 m3/s) and a downstream
time-constant water depth (y0.6 m) are considered. The maximal
increments in the water-depth for each pool/section, are easily

As already referred in Section 4.1, this is a quadratic programming problem (QP), which assures the convexity of the problem,
because J(k)Z0, as a result of the design of X40. This fact facilitates
the use of some available optimization algorithms (Matlab, 2008c).
In terms of computational burden, the constrained optimization
problem ((17) and (18)) shows a major shortcoming when controlling
systems with fast dynamics because the dimensions of the involved
matrices are usually very high (Maciejowski, 2002). Methods to fasten
MPC controllers are discussed in Wang and Boyd (2010).

J. Figueiredo et al. / Control Engineering Practice 21 (2013) 870886

877

Fig. 8. MPC controller applied to a water canal.

Fig. 9. SCADA main graphical user interface.

In the present case the computational time constraint is not so

critical as hydraulic systems present commonly slow dynamics.
As it is explained at the end of the present chapter, the implementation of the developed MPC controller used a sampling
period of 10 s.
Three different optimization algorithms were compared in
terms of computational burden (Silva, 2007): (i) Active sets
(Bemporad et al., 2005); (ii) Augmented Lagrangian (Luenberger,
2003) and (iii) barrier method (Luenberger, 2003). Several tests
were performed with the optimal solution near zero and distant
from zero. The results showed that although the Augmented
Lagrangian method nds usually the best solution, it requires
more time to converge than the Active Sets. The Barrier method
shown to be the most time consuming method.
For the present study the selected algorithm for the constrained optimization problem was the Active Sets, which is
available in the Matlab MPC toolbox (Matlab, 2008b; Bemporad
et al., 2005). This method offered a good compromise between the
accuracy of the solution and the required convergence time.

To solve the optimization problem ((17) and (18)) there are

some non measured states that are needed for the computations.
In this case this difculty is overcome using state observers based
on the Kalman lter, once the system observability is assured
(Maciejowski, 2002). The Matlab MPC toolbox was used for the
implementation of the Kalman lter (Matlab, 2008b).
Referring the selected prediction and control horizons, it can
be said that, in general, the prediction horizon (Hp) must be large
enough to hold all the transient response of the system.
A practical way to determine this maximal value is to enlarge Hp
until no different response is seen in the output. On the other
hand the control horizon (Hc) is normally small when compared to
Hp, and it veries usually HcA[3,10] (Bemporad et al., 2005). With
these both initial approximations for Hp and Hc, several tests
were performed until the best parameters were found. The obtained
values were: Hp 36 and Hc 10.
Finally referring the selected time step (sampling period), this
choice is closely related with the canal under study. There is one
main maximal time constraint that is: it has to be shorter than the

878

J. Figueiredo et al. / Control Engineering Practice 21 (2013) 870886

time that takes a wave to traverse the channel (Baena, 2003).

Hydraulic systems have additionally minimal time restrictions on
the sampling period, mainly by the restrictions on the safe
operation of the actuators (gates). In Madeira and Torrado
(2005) this problem was studied for the canal under consideration and it came 10 s for the optimal sampling period.

5. Experimental results
In Fig. 9 it is shown the developed SCADA main graphical user
interface.
The experimental results obtained with MPC are compared
with DMPIC and UMPIC, already referred in Section 1. Both these
decentralized PI-controllers (DMPIC and UMPIC) were developed
and tuned for the experimental canal in Almeida et al. (2002),
Ratinho et al. (2002). In the following section a brief characterization of these PI-Controllers is presented.
5.1. Decentralized PI-controllers
The most commonly used controller in hydraulic systems is
the PID. In most cases, the controller is reduced to a PI due to
difculties in tuning properly the derivative gain, KD (Ratinho
et al., 2002).
The strategy adopted in Almeida et al. (2002), Ratinho et al.
(2002) to determine the proportional and integral gains was to
compare the frequency response of the state space model, which
relies on the linearized Saint-Venant equations, with a simple

model with the assumption of each canal pool treated as a

reservoir (reservoir model). Recalling Fig. 2 it can be seen that
this simple methodology is able to supply good approximations
without considering time delays in the model despite the slow
characteristic dynamics of hydraulic systems because both typical
control actions: UMPIC and DMPIC are collocated control strategies, where sensors (water depth meters) and actuators (discharge gates) are near each other. This physical proximity
between sensors and actuators eliminates the need for assuming
time delays in the model, while maintaining the simplicity of the
problem.
Considering each Canal pool as a reservoir, with H its water
depth, Qin and Qout the upstream and downstream discharges and
ASup the supercial area of the pool, one reaches the following
equation, in Laplace domain (where s is the Laplace variable):
H

1
Q Q out
sASup in

19

Considering the water depth as the controlled variable and the

discharge Qout as the control action, Eq. (19) can be rearranged
according to the standard PI-Controller formulation (Ratinho
et al., 2002):


KI
us K P
es
20
s
Leading to:


KI
Href H
Q out K P
s

21

Fig. 10. Bode plots. (a) both reservoir and nite differences models for the 1st canal pool; (b) lowest resonance frequency and highest frequency for similar behaviour of
both models.

J. Figueiredo et al. / Control Engineering Practice 21 (2013) 870886

Combining Eqs. (19) and (21), one obtains the nal 2nd order
controlled model:
H

K P =ASup s K I =ASup
s
H 
Q
s2 K P =ASup s K I =ASup ref s2 K P =ASup s K I =ASup in
22

Following the standard notation for 2nd order models, where

on is the undamped natural frequency and x the damping ratio,

one can easily compute the KP and KI gains through the equations:
K P 2on xA

23

K I o2n A

24

879

interval [20 L/s, 80 L/s], the downstream discharge Qout (gate

opening) was the control action and the water depth H (level)
was the controlled variable. The interrelated behavior of these
three variables is shown in each one of the following Figs. 1113.
Analysing the data, it can be seen that the best experimental
results were obtained with the PI-controller C, reaching similar
performance in the controlled variable (level) with signicantly
reduced control effort (gate opening).
In the following sections, for the experimental comparisons
between MPC controller and PI-controllers, it was used the PIcontroller C (KP  0.6 and KI  0.006).
5.2. Performance indicators

The undamped natural frequency is related with the systems

lowest resonance frequency (or) by on or =n where n is an
integer that is chosen according to the condition that both
systems have similar behavior.
For the studied canal, the Bode plot (Fig. 10a) shows clearly
that both models have a similar response until a certain value of
frequency. Fig. 10b identies the frequency of the lowest resonance peak as well as the value where the differences between
both models become signicant.
It can be seen by analyzing the Bode plots above that any value
of n has to be greater or equal than 4. Therefore, with the chosen
values of n 4, 6 and 8 it results the correspondent proportional
and integral gains ((23) and (24)) shown in Table 6.
The following illustrated validation tests (Figs. 1114), correspond
to the rst pool, where a water depth regulator was implemented, as
this controller is the most usual in irrigation systems, as it can provide
previously accorded discharges without any signicant delay. Similar
tests were done for the other individual canal pools.
In all presented tests the considered water depth setpoint was
700 mm, the upstream discharge Qin (ow) was considered the
disturbance, with a step-varying amplitude of 20 L/s within the

Several performance indicators are available to evaluate the

controllers (Bos, 1997). In this paper the following well known
performance indicators are used:
X
25
2 Integrated absolute error : IAE
9yr y 9;

Table 6
Chosen values of n and resulting proportional and integral
gains, KP and KI.

where T is the total test time, yi is the controlled water level in pool
i with reference ryi, Dai is the gate aperture variation in pool i, and n
is the total number of pools. The value 16.5 is used to balance both
terms in the index. This indicator is particularly useful since it gives
a general evaluation of each controllers performance.
The experimental results that will be presented in the next
Sections 5.3 to 5.6 use the same cost function and constraints
Eqs. (17) and (18) but with different scenarios: different

Controller

KP

KI

A
B
C

4
6
8

 1.2
 0.8
 0.6

 0.024
 0.010
 0.006

Maximum absolute error :

Integrated gate movement :

Maximum gate aperture :

MAE max9yr y 9;

IAW

9Da9;

MAW max9ar a 9;

26

27

28

where ra is the gate nominal aperture. The rst two indicators

refer to water levels, while the last two refer to the gate movements. High water level deviations and great gate movements
should be avoided in any water canal. To this view, the following
performance indicator is also monitored:



n Z T 
X

yi r yi  
IAE
IAW
29
Dai  dt
b
16:5
16:5
0
i1

Fig. 11. PI-controller Aresults for discharge-step disturbances of 0.020 m3/s (ow).

880

J. Figueiredo et al. / Control Engineering Practice 21 (2013) 870886

Fig. 12. PI-controller Bresults for discharge-step disturbances of 0.020 m3/s (ow).

Fig. 13. PI-controller Cresults for discharge-step disturbances of 0.020 m3/s (ow).

Fig. 14. Superposition of results from controllers B and C for discharge-step disturbances of 0.020 m3/s (ow).

references (Sections 5.3 and 5.6) and different disturbances

(Sections 5.4 and 5.5).
In the following sections the comparison between the controllers performance is easily accomplished on a quantitative
basis, by analyzing the values of the correspondent indicators.
5.3. Upstream water level controlMPC versus UMPIC
The performance of MPC and the UMPIC are compared when
controlling the pools upstream water levels. The results obtained

for a water off-take of 10 L/s extracted from the 3rd off-take valve
of the canal, for the time interval t [200 s, 400 s], are presented
in both Fig. 15 and Table 7.
The values presented in Table 7 refer to the following notation:
S#T# indicates sensor S# in pool T#. Thus, S1T2 refers to the
upstream water level measured by sensor 1 in pool 2. Likewise,
the notation G# is used to indicate gate number #.
The high oscillation of the water level deviations captured by
the sensors is caused by the fact that these sensors are located
near the gates. The water that ows under the gates produces

J. Figueiredo et al. / Control Engineering Practice 21 (2013) 870886

881

Fig. 15. Results of pools upstream water level control (1st column: MPC; 2nd column: UMPIC; rows: upstream water levels in pools 2, 3 and 4, respectively).

Table 7
Upstream water level controlperformance indexes.
IAE (nm)
S1T2
S1T3
S1T4

MPC
UMPIC
MPC
UMPIC
MPC
UMPIC

2590
7172
2361
5733
1948
3563
IAW (nm)

G1
G2
G3

MPC
UMPIC
MPC
UMPIC
MPC
UMPIC

545
372
268
181
176
80

MPC
UMPIC

1408
1632

MAE (nm)
14
26
14
26
8
14
MAW (nm)
206
70
84
53
32
23

disturbance waves that are captured by these sensors. Comparing

the results it can be concluded that MPC has a better performance
leading to smaller water level deviations than the UMPIC controller (cf. Performance indicator IAE in Table 7). Furthermore,
the performance index, b, is smaller for the MPC than it is for
the UMPIC controller which is a clear indication that MPC
outperforms UMPIC.
5.4. Gate malfunctioning
The performance of MPC and DMPIC is compared under the
scenario of a gate malfunctioning. This experiment is also useful to
show the advantage of considering a centralized MIMO controller
instead of decentralized PI-controllers. Whereas decentralized PIcontrollers divide the system into distinct parts without regarding
its interconnections, centralized MIMO controllers make use
of cross information along the canal sensors to produce the best
control action. This experiment was conducted considering the
3rd gate xed at its nominal position, with a water off-take of 10 L/s

882

J. Figueiredo et al. / Control Engineering Practice 21 (2013) 870886

Fig. 16. Gate malfunctioning results (1st column: MPC; 2nd column: DMPIC; rows: upstream water levels in pools 1, 2 and 3, respectively).

in the 3rd off-take valve, between the time instants t150 s and
t415 s. The results obtained with both controllers are presented
in Fig. 16 and Table 8.
Both controllers experienced difculties in minimizing the
water level deviations. However MPC produces better results
than the DMPIC. For instance, notice how MPC is able to maintain
the water level deviation in section S3T3 in a reasonable interval
for different ows in the water off-take with gate G3 xed
(malfunction). The outperformance of MPC in relation to DPMIC
can be explained by the centralized nature of the MPC controller
that can control the different pools with the common objective of
reducing the water level deviations in all pools. In the decentralized architecture the other pools cannot help pool 3 in reacting to
errors originated by the malfunction of its gate (gate 3).
The MPC is able to keep the water level deviations smaller
while making approximately the same use of the gates as the
DMPIC controller (see indicators IAE and IAW from Table 8).
5.5. Severe water off-takes
The performance of MPC is compared with DMPIC when
severe water off-takes are considered. In this experiment, rapidly

varying water off-takes up to 83% of the nominal ow are

extracted from the off-take valves following the prole given in
Fig. 17.
The results obtained with MPC and with the DMPIC are
depicted in Fig. 18 and the corresponding performance indicators
are presented in Table 9.
The MPC is able to keep the water level deviations in a smaller
range than the DMPIC (cf. performance indicator MAE), at the
expense of a slight increase in gate usage (cf. Performance
indicator IAW). As the difference in gate usage between DMPIC
and MPC is small, this has a negative effect on the overall
performance of the DMPIC.
A maximum water level deviation of 28 mm was measured for
DMPIC, compared to 12 mm that was registered for the MPC.
5.6. Water level setpoint tracking
The problem of water level setpoint tracking through time is
here analysed. This type of disturbance is more likely to occur on
rigid water distribution methods than on exible ones. For
example the water levels setpoint can change to higher values
during Summer time, due to the increase in water usage, while it

J. Figueiredo et al. / Control Engineering Practice 21 (2013) 870886

Table 8
Gate malfunctioningperformance indexes.

S3T1
S3T2
S3T3

G1
G2
G3

MPC
DMPIC
MPC
DMPIC
MPC
DMPIC

IAE (nm)

MAE (nm)

3134
4076
2469
3826
4745
8978

11
14
9
15
16
28

IAW (nm)

MAW (nm)

MPC
DMPIC
MPC
DMPIC
MPC
DMPIC

163
150
157
139
0
0

MPC
DMPIC

948
1313

60
73
69
76
0
0

883

The MPC was able to handle water off-takes (whether known

or not known in advance), water level reference tracking and
process failure, while obeying the canal constraints.
The dynamical model of the water canal was obtained based
on the Saint-Venant equations, which proved to be suitable for
being integrated in the predictive controller. Besides, authors
believe that this control setup can be extended to other similar
water distribution networks since the model used is mostly based
on canal geometric characteristics.
The obtained results show that the MPC controller has good
robustness properties when the water canal is subjected to severe
external disturbances. This type of modern facilities (water
canals) managed by intelligent control strategies can improve
tremendously the performance of water distribution networks, in
the near future.

Appendix A
Canal-pool model
The dynamic behaviour of hydraulic systems can be obtained
from the mechanical relations established for the control volume
(Fig. 20).
The set of equations describing a hydraulic system is very well
known and referred as the Saint-Venant equations (Saint-Venant,
1871; Litrico, Fromion, Baume, Arranja, & Rijo 2005).
The Saint-Venant equations are simultaneous space and time
dependent. For their discretization the implicit implementation of
the Preissmann scheme is adopted (Mantecon, Gomez, & Rodellar,
2002). Considering the linearization methodology applied to a
point, near to the system steady-state, it is valid the following
approximation:

Fig. 17. Water off-take prole.

can be reduced to lower values during Winter season. The

simulation of this real situation is made considering a staircaselike change on the reference of the 3rd downstream water level
(S3T3). The experimental results for the MPC are shown in Fig. 19
(left column). The performance indicators for the MPC as well as
for the DMPIC are presented in Table 10. This test clearly shows
that MPC is better than the DMPIC since it is able to closely follow
the water level reference, while the DMPIC shows a signicant
delay. Notably the MPC is able to make slightly less use of the
system gates (smaller IAW) while keeping water levels deviations
(IAE) smaller than the DMPIC.

Q x,t Q r x dQ x,t

30

yx,t yr x dyx,t

31

where Q x,t is the variable discharge ow; yx,t the variable

water depth; Q r x the discharge ow at steady-state; yr x the
water depth at steady-state; dQ x,t the discharge ow increment; dyx,t the water depth increment.
Applying the implicit Preissmann scheme to the Saint-Venant
equations (Mantecon et al., 2002) and considering a single pool, it
results:
a11  dQ ki 1 a12  dyki 1 a13  dQ kj 1 a14  dykj 1
b11  dQ ki b12  dyki b13  dQ kj b14  dykj

32

a21  dQ ki 1 a22  dyki 1 a23  dQ kj 1 a24  dykj 1

b21  dQ ki b22  dyki b23  dQ kj b24  dykj

33

where
6. Conclusions
k1

This study presented experimental results of a model predictive controller (MPC) applied to a water canal located at the
University of Evora, Portugal. A wide range of real-life typical
situations were covered in the experiments showing the effectiveness of the control approach over other classical water canal
control strategies, namely two methods of decentralized PI
control.

a21

1
2  y  V i

2  Dt
Dxi
k1

a22

V 2  B i

Dxi

r  y

g  Ai

g  Ai

r  y

Dxi

1
2  y  V i

2  Dt
Dxi

 fi r

k1

k1

a23

k1

k1

g  Ai

k1

g  Ai

4
 fj r

 ji r

884

J. Figueiredo et al. / Control Engineering Practice 21 (2013) 870886

Fig. 18. Severe water off-take results (1st column: MPC; 2nd column: DMPIC; rows: upstream water levels in pools 1, 2 and 3, respectively).

k1

a24
b21

V 2  B i

r  y

Dxi

k1

g  Ai

1
2  1y 

2  Dt
Dxi

r  y

Dxi
k1
V i r

k1

b22 

V 2  B i

k1

V 2  B i

g

k1
Ai

 fi r

k1

k1

b24

 jj r

k1

r  1y g  Ai r  1y g  Ai
 j i r

Dxi
Dxi
4

1
2  1y  V i

2  Dt
Dxi

b23

k1

g  Ai

k1

k1

g  Ai

 fj r

4
k1

 j j r
r  1y g  Ai r  1y g  Ai


Dxi
Dxi
4


@J
P4=3
2 2
 jQ j
@Q i
K s  A10=3
 


@J
2
2 @P 5  B


ji
 Ji 
@y i 3
P @y
A

fi

and a13  a11, a14 a12, b12 a12, b13 b12, b14 b12.
The indexes i and j refer to contiguous sections within a pool,
whereas k refers to time instants; yA[0,1] is the space discretization coefcient. A bar on top of a generic variable, e.g. s, stands for
k1
sx,ti
1=2ski skj , and is related to the time discretization
according to the Preissmann scheme.

Off-take-valve model
J

4=3

K s 2  A10=3

 Q  jQ j

For off-take-valves, considering sections h and j as represented

in Fig. 21, and assuming that the approximation dyh dyj, one

J. Figueiredo et al. / Control Engineering Practice 21 (2013) 870886

obtains, under the state-space notation, A1 xk 1 B1 xk B2 ukd :

2

a11
6
6 a12
6
6 a13
4
a14

9
8
3T > @Q k 1 > 2
>
h
a21 >
b11
>
>
>
>
>
> 6
7 > k1 >
=
a22 7 < @yh
6 b12
7
6
6 b13
a23 7
@Q kj 1 >
>
>
4
5 >
>
>
>
>
>
>
a24 >
b14
;
: @yk 1 >
j

3T 8 k 9
@Q >
b21 >
>
> "
> h>
>
> k >
7 >
b11
b22 7 < @yh =
7

k
@Q
>
b23 7
b21
j >
>
5 >
>
>
>
> k>
: @y ;
b24 >
j

9
#8 k
a11 < Q off-take =
1
;
a21 : Q koff-take
|{z}

885

Table 10
Water level setpoint trackingperformance Iindexes.
IAE (mm)
S3T3

MPC
DMPIC

udisturbance

930
3591

Notice that in udisturbance the knowledge of

is required.
Two situations can occur: either the water off-takes are known in
1
advance or they are not and the approximation Q koff-take
Q koff-take is
assumed.

9
21

IAW (mm)

34
1
Q koff-take

MAE (mm)

G3

MPC
DMPIC

246
256

MPC
DMPIC

1303
1991

MAW (mm)
125
88

Gate model
The linearization and discretization of the dynamic relations
for hydraulic gates, considering the notation illustrated in Fig. 22,
yields to:
9
8
2
3T > @Q k 1 > 2
3T 8 k 9
@Q >
>
>
>
0
0
>
>
>
>
> h >
> h>
>
>
>  
> k >
6 @f =@y 7 >
7 >
= 6
< @ykh 1 >
6
6 @f j =@yi r 7 < @yh =
j
i r7
@f
6
7
6
7
6

u
k
k

1
6
7
7
1
1
@Q j >
>
>
@Q j
@ai r control
>
>
>
>
4
4
5 >
5
>
>
>
>
>
>
> k>
>
>
>
>
@f j =@yi r >
@f j =@yi r : @y ;
;
: @yk 1 >

FG

y+

y
x
x

FR

Fp
Q+

Q
x
x

Fig. 20. Hydraulic control volume.

35
where a represents the gate aperture.
Notice that the partial derivatives depend on the considered
gate characteristics and type, which can be normally written as
Table 9
Severe water off-takesperformance indexes.

S3T3

13
22
13
23
12
28

IAW (mm)
G1
G2
G3

MPC
DMPIC
MPC
DMPIC
MPC
DMPIC

868
506
767
615
1170
715

MPC
DMPIC

3855
4374

Fig. 21. Considered sections for the linearization of off-take-valves.

Gate

MAW (mm)
87
97
66
94
170
181

Fig. 22. Considered sections for the linearization of gates.

Fig. 19. Water level setpoint tracking (left: MPC; right: DMPIC).

Section l

S3T2

6612
11147
5067
14691
5636
16032

Section j

MPC
DMPIC
MPC
DMPIC
MPC
DMPIC

Section h

S3T1

MAE (mm)

Section i

IAE (mm)

886

J. Figueiredo et al. / Control Engineering Practice 21 (2013) 870886

Qf(yi,yj,a), with Q being the ow crossing the gate section, yi and

yj the water levels immediately before and after the gate, respectively. Specic equations for slice gates and overshoot gates can be
found in literature (Quintela, 2005).
References
Almeida, M., Figueiredo, J., Rijo, M. (2002) Scada conguration and control modes
implementation on an experimental water supply canal. In: Conference
proceedings 10th IEEE Mediterranean conference on control and
automationMED2002. Lisbon: Portugal.
Axeda Supervisor Wizcon for windows and internet 8. 2user guide (2002): Axeda
systems.
Baena, J. (2003) Predictive control systems in irrigation canalsformulation and
numeric simulation. Ph.D. Thesis (in Spanish). Universitat Politecnica
Catalunya.
Bemporad, A., Morari, M., Dua, V., & Pistikopoulos, E. (2002). The explicit linear
quadratic regulator for constrained systems. Automatica, 38, 320.
Bemporad, A., Morari, M., Ricker, N. (2005) Model predictive control toolbox for use
with MATLAB, user guide version 2. The MathWorks.
Bos, M. (1997). Performance indicators for irrigation and drainage. Irrigation and
Drainage Systems, 11(2), 119137.
Burt, C. (1987). Overview of canal control concepts. In: D. Zimbelman (Ed.),
Planning, operation, rehabilitation and automation of irrigation water delivery
systems (pp. 2937). New York: American Society of Civil Engineers.
Burt, C., & Piao, X. (2004). Advances in PLC-based irrigation canal automation.
Irrigation and Drainage, 53, 2937.
Cabeza, A., Maestre, J., Ridao, M., Camacho, E., & Sanchez, L. (2011). A hierarchical
distributed model predictive control approach to irrigation canals: a risk
mitigation perspective. Journal of Process Control, 21, 787799.
Clemmens, A., Bautista, E., Whalin, B., & Strand, R. (2005). Simulation of automatic
canal control systems. Journal of Irrigation and Drainage Engineering, 131(4),
324335.
Dougherty, D., & Cooper, D. (2003). A practical multiple model adaptive strategy
for multivariable model predictive control. Control Engineering Practice, 11(6),
649664.
Feliu-Battle, V., Perez, R., & Rodriguez, L. (2007). Fractional robust control of main
irrigation canals with variable dynamic parameters. Control Engineering
Practice, 15(6), 673686.
Fletcher, R. (1987). Practical methods of optimization ((2nd ed.)). Wiley.
Garcia, C., & Morari, M. (1989). Model predictive control: theory and practice: a
survey. Automatica, 25(3), 308323.
Guenova, I., Litrico, X., Georges, D. (2004) Modelling and robust PID control of an
irrigation canal pool. In: Proceedings modelling, identication and control. Spain.
Litrico, X., Fromion, V., Baume, J., Rijo, M. (2003) Modelling and PI controller design
for an irrigation canal. In: Proceedings European control conference. Cambridge: UK.

Litrico, X., Fromion, V., Baume, J., Arranja, C., & Rijo, M. (2005). Experimental
validation of a methodology to control irrigation canals based on Saint-Venant
equations. Control Engineering Practice, 13(11), 14251437.
Litrico, X., Malaterre, P., Baume, J., Vion, P., & Ribot, B. (2007). Automatic tuning of
PI controllers for an irrigation canal pool. Journal of Irrigation and Drainage
Engineering, 133(1), 2737.
Luenberger, D. (2003). Linear and nonlinear programming ((2nd ed.)). Prentice-Hall.
Maciejowski, J. (2002). Predictive control with constraints ((1st ed.)). Prentice-Hall.
Madeira, C., Torrado, M. (2005) Flow and level control in a network of canals
and reservoirsnal BSc project (in Portuguese). Instituto Superior
TecnicoTechnical University Lisbon.
Malaterre, P., Rodellar, J. (1997) Multivariable predictive control of irrigation
canals. Design and evaluation on a 2-pool model. In: Proceedings internatinal
workshop on regulation of irrigation canals. pp. 230238.
Malaterre, P. (1998). Pilote: linear quadratic optimal controller for irrigation
canals. ASCE Journal of Irrigation and Drainage, 124(4), 187194.
Malaterre, P., Baume, J. (1998) Modeling and regulation of irrigation canals:
existing applications and ongoing researches. In: Proceedings IEEE international
conference on systems, man and cybernetics. Vol. 4: pp. 38503855.
Malaterre, P., Rogers, D., & Schuurmans, J. (1998). Classication of canal control
algorithms. Journal of Irrigation and Drainage Engineering, 124(1), 310.
Mantecon, J., Gomez, M., & Rodellar, J. (2002). A simulation-based scheme for
simulation of irrigation canal control Systems. Simulation, 78(8), 485493.
Matlab 7.6.0.324 (2008a): The MathWorks Inc.
MatlabModel Predictive Control Toolbox (2008b): The MathWorks Inc.
MatlabOptimization Toolbox (2008c): The MathWorks Inc.
ModBus (2012), /http://www.modbus.orgS. Latest access: July 2012.
Ooi, S., & Weyer, E. (2008). Control design for an irrigation channel from physical
data. Control Engineering Practice, 16(9), 11321150.
Overloop, P., Weijs, S., & Dijkstra, S. (2008). Multiple model predictive control on a
drainage canal system. Control Engineering Practice, 16(5), 531540.
Plusquellec, H., Burt, C., Wolter, H. (1994) Modern water control in irrigation. World
Bank technical paper no. 246irrigation and drainage series. Washington DC.
Quintela, A. (2005) Hidraulica. (9th ed.). Fundac- a~ o Calouste Gulbenkian.
Ratinho, T., Figueiredo, J., Rijo, M. (2002) Modelling, control and eld tests on an
experimental irrigation canal. In: Conference proceedings 10th IEEE Mediterranean conference on control and automationMED2002. Lisbon: Portugal.
Saint-Venant, A. (1871). Theorie du mouvement non-permanent des eaux, avec
application aux crues des rivie res et a lintroduction des marees dans leur lit.
Academie Scientic Paris, 73, 148154.
Silva, P. (2007) Predictive control of an irrigation canal (in Portuguese). M.Sc. Thesis.
Instituto Superior Tecnico; Lisbon.
Wang, Y., & Boyd, S. (2010). Fast model predictive control using online optimization. IEEE Transactions on Control Systems Technology, 18(2), 267278 2010.
Weyer, E. (2003) LQ control of an irrigation channel. In: Proceedings 42nd
Conference on decision and control (pp. 750755). Maui, Hawaii: USA.
Xu, M., Overloop, P., & Giesen, N. (2011). On the study of control effectiveness and
computational efciency of reduced Saint-Venant model in model predictive
control of open channel ow. Advances in Water Resources, 34(2), 282290.