You are on page 1of 14

Journal of Process Control 19 (2009) 506519

Contents lists available at ScienceDirect

Journal of Process Control


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

Smith predictor based robust fractional order control: Application to water


distribution in a main irrigation canal pool
V. Feliu-Batlle a,*, R. Rivas Prez b, F.J. Castillo Garca a, L. Sanchez Rodriguez c
a

Escuela Tcnica Superior de Ingenieros Industriales, Universidad de Castilla-La Mancha, Ave. Camilo Jos Cela S/N, Ciudad Real 13071, Spain
Department of Automatica and Computer Science, Havana Polytechnic University, CUJAE, Marianao, C. Habana 19390, Cuba
c
Escuela Universitaria de Ingenieros Tcnicos Industriales, Universidad de Castilla_La Mancha, Campus Tecnolgico de la Antigua Fbrica de Armas, Toledo 45071, Spain
b

a r t i c l e

i n f o

Article history:
Received 29 December 2006
Received in revised form 7 May 2008
Accepted 13 May 2008

Keywords:
Robust fractional order control
Variable time delay process
Smith predictor based control system
Water distribution control in a main
irrigation canal pool

a b s t r a c t
This paper proposes a new methodology to design fractional integral controllers combined with Smith
predictors, which are robust to high frequency model changes. In particular, special attention is paid
to time delay changes. These controllers show also less sensitivity to high frequency measurement noise
and disturbances than PI or PID controllers. This methodology is applied to design controllers for water
distribution in a main irrigation canal pool. Simulated results of standard PI and PID controllers plus a
Smith predictor, and the controller developed in this paper are compared when applied to the dynamical
model of a real main irrigation canal pool showing that our controller exhibits better and more robust
features than these. Moreover our controller is compared with other more complex control techniques
as predictive control and robust H1 controllers, exhibiting better or similar performances than these.
 2008 Elsevier Ltd. All rights reserved.

1. Introduction
Water constitutes one of the most precious resources of the
earth. However in many cases it is being consumed as if it existed
in limitless quantities. Then it is important to manage the water resources in an effective way and to minimise the losses [1].
At present a lot of water is wasted in most of the networks of
open irrigation main canals because of lack of an efcient control.
It is estimated that irrigation water users can cut their water consumption by 1050% by using water more efciently [1]. In this
context, automatic control is considered as a powerful tool for
improving efciency in water distribution open irrigation canal
systems [2,3].
Irrigation canals are systems distributed over long distances,
with signicant time delays and dynamics that change with the
operating conditions [4,5]. A typical open irrigation canal consists
of several pools separated by undershot gates that are used for regulating the water distribution from one pool to the next one.
Dynamics of a main irrigation canal pool has traditionally been
modelled by the Saint-Venant equations, which are nonlinear
hyperbolic partial differential equations (e.g. [4,6]). Nowadays different methods exist for the solution of the Saint-Venant equations, all of them exhibiting large mathematical complexities
* Corresponding author. Tel.: +34 926 295364; fax: +34 926 295361.
E-mail addresses: Vicente.Feliu@uclm.es (V. Feliu-Batlle), rivas@electrica.cujae.edu.cu (R. Rivas Prez).
0959-1524/$ - see front matter  2008 Elsevier Ltd. All rights reserved.
doi:10.1016/j.jprocont.2008.05.004

[25]. These equations are also very difcult to use for prediction
and control [2,3]. Often, an equivalent rst order system plus a delay is used to model the canal pool dynamic behavior [3,5,7]. This
model has the strong drawback that its parameters may experience
large changes when the discharge regime varies. The variations in
discharge regimes of irrigation canal pools depend on many
parameters such as the pool length, bed slope, cross section, roughness, initial water prole, gate opening magnitude, etc. [8].
Experiments developed on canal dynamics identication conrm that the dynamic parameters of a main irrigation canal pool
such as time constant, time delay and static gain, exhibit wide variations when the discharge regimes change in an operation range
[4,8]. Then any controller to be designed for a main irrigation canal
pool has to be robust to variations in some parameters of the linearized model [4].
Different strategies have been proposed for control of water distribution in main irrigation canal pools [24,9]. PI controllers have
often been used, and sometimes PID controllers too. Some modern
approaches try to improve the robustness of these classical controllers [10,11] when applied to main irrigation canal pools.
The use of Smith predictor in control systems of a main irrigation canal pool has been proposed by different authors, to overcome the time delay that characterizes these systems [1214].
However it is well known that small modeling errors can cause
instability in Smith predictor based control systems if the controller is not properly designed [15]. Conditions for robust stability
of these systems have been proposed in [1618], and several

507

V. Feliu-Batlle et al. / Journal of Process Control 19 (2009) 506519

modications of the basic Smith predictor scheme in [19,20]. A recent survey on control of systems with delay and the use of predictors is [21].
In the last years, fractional operators have been applied with
satisfactory results to model and control processes with difcult
dynamical behavior [22,23]. Recently, different works have appeared about the application of fractional controllers in a main irrigation canal pool [24,25]. An interesting feature of fractional order
controllers is that they exhibit some advantages when designing
robust control systems in the frequency domain for processes
whose parameters vary in a large range. In this paper these characteristics are explored in order to design robust controllers to solve
the problem of effective water distribution control in a main irrigation canal pool whose dynamic parameters vary in a wide range. In
particular, this paper is focused on the design of fractional integral
controllers combined with a Smith predictor, which show to be
in some cases more robust to changes in the time delay and
unmodelled high frequency dynamics than other controllers like
PI and PID. Time delay is the parameter more determinant in the
stability of a closed-loop control system of a main irrigation canal
pool [3,9].
This paper is organized as follows. An introduction to fractional
order operators and controllers is presented in Section 2. A model
of the irrigation canal pool to be controlled is proposed in Section
3. Section 4 outlines the principles that justify the robustness properties of the proposed controller. Section 5 develops a numerical
procedure for tuning the parameters of fractional integral controllers with a Smith predictor. Section 6 denes the cases in which
our fractional controller is more robust than PI and PID controllers
always combined with Smith predictors. Section 7 compares the
designed controller with several more complex controllers by carrying out simulations using the dynamic model of a real main irrigation canal pool. Finally some conclusions are drawn in Section 8.

where ak, bk (k = 0, 1, . . ., n) constant coefcients of the FODE; bk


(k = 0, 1, . . ., n) the orders of the fractional derivatives represented
by arbitrary real numbers.
The conventional PID controller which involves proportional
plus integral plus derivative actions based on the error signal:
1
ut K  et K i  D1
t et K d  Dt et;

can be generalized to a PI Dl fractional controller involving an integrator of order k and a differentiator of order l [22,26]. The equak

tion, in the time domain, of such a controller has the form:


l
ut K  et K i  Dk
t et K d  Dt et;

0 6 k; l 6 1:

2.2. Digital implementation


If a fractional system has to be simulated or a fractional controller has to be implemented, expression (1) has to be approximated
by discrete realizations. They can be obtained by two ways [27]:
(1) to approximate the fractional operator by a standard transfer
function in the frequency range of interest and then apply any
habitual discretization technique (Tustin operator, e.g.), (2) to
approximate numerically the fractional operator.
In the second way, numerical approximations of the fractional
derivative/integral operator are often implemented by using the
following numerical generalization:


a
d f t
lim T a Df ta ;
a 
dt tkT T!0

where a 2 R, Df(kT)jt=kT = f(kT)  f(kT  T); T the period of discretization. Formally this operator is expressed as

Da 

1  z1
T

a

and can be implemented by


2. Fractional calculus and fractional order control systems
a

Fractional calculus is a 300-years-old topic. The theory of fractional order derivatives was developed mainly in the 19th century.
However, applying fractional order calculus to dynamic systems
control is just a recent focus of interest [22].

Fractional calculus is a generalization of integration and differentiation to non-integer (fractional) order fundamental operators
represented as a Dat where a and t are the limits and aa 2 R the
order of the operation. Several denitions of this operator have
been proposed [22]. One of the most used denitions of the general
fractional integro-differential operator is the RiemannLiouville
(RL) denition:
n

f t

1
d
Cn  a dt n

f s
t  san1

ds;

where n  1 < a < n; n an integer; C() the Eulers gamma function; t > a. The Laplace transform of the RL fractional derivative/
integral (1) under zero initial conditions for order a, (0 < a < 1) is given by

Lfa Dt a f t; sg sa Fs:

A fractional order system may be represented by a typical nterm linear fractional order differential equation (FODE) in the
time domain:

an Dbt n yt



bn Dbt n ut

a1 Dbt 1 yt



a0 Dbt 0 yt

b1 Dbt 1 ut

b0 Dbt 0 ut;

f t T a

 
ta
T 
X
a
f t  jT;
1j
j
j0

where [] means the integer part, and the combinatorial function
has been generalized in the following sense:

 

a
l

2.1. Basic concepts

a Dt

a Dt

aa  1 . . . a  l 1
l!

The limit value of expression (8) when T ? 0 is called the


GrundwaldLetnikov (GL) denition of the discretized fractional
operator [22]. Provided that T is small enough, this expression is
a numerical approximation that gives sufciently accurate results
in most of the cases. This discrete operator may be approximated
by FIR or IIR discrete lters. Expression (8) can be truncated to a
xed number of N + 1 terms of this sum (0 6 j 6 N) in order to
 
a 0, and it is
get a FIR lter. This can be done because lim
l
l!1
called the short memory approximation. This approximation leads
usually to FIR lters of very large order (often N > 100).
There are several methods to approximate numerically the fractional operator by IIR lters. IIR lters are better suited for microprocessor implementations of fractional controllers than FIR as
they lead to transfer functions of much lower order. A popular
technique is the Al Alaoui operator, which is a mixed scheme of
the Euler and the Tustin operators, whose generating function
can be expressed by (e.g. [28]):
r

D z

8
7T

(

r
CFE

r )
1  z1
1 z1 =7

p;q

8
7T

r

Pp z1
;
Q q z1

10

where P, Q polynomials of degrees p and q respectively in the variable z1; CFE continued fraction expansion.

508

V. Feliu-Batlle et al. / Journal of Process Control 19 (2009) 506519

2.3. Robustness properties


In the last years it has been an extensive research effort in
developing fractional PID controllers designed in the frequency domain with enhanced robustness properties [29,30]. Often the
robustness feature is designed for a particular range of frequencies
(e.g. the locally at phase condition) in order to obtain robustness
to changes in the plant gain or other specic parameters, e.g. [23].
Some robustness conditions have been obtained for fractional state
space models by using some matrix Lyapunov inequalities [31]. Finally we mention a recent work where the fractional controller has
been designed from a QFT loop shaping point of view [32].
Our paper studies the properties of a very simple fractional integral controller that is embedded in a Smith predictor control
scheme. Moreover we propose a methodology to design these controllers. Robustness features in the frequency domain are explored,
and are applied to design robust controllers for effective water distribution control in main irrigation canal pools whose dynamic
parameters vary in a wide range. Next we present the property that
we will exploit along this paper.
Let us have a plant with transfer function G(s). Assume we want
to full the next closed-loop specications: crossover frequency
(xc) and phase margin (/m). Then the simplest controllers able
to achieve this have only two parameters to be tuned (two degrees
of freedom), being the PD controller (RPD(s) = K(1 + Tds)) the most
often used one. However there are other controllers with different
structures and two degrees of freedom that can achieve these
specications.
Property 1. The minimum-phase controller (all poles and zeros
placed in the complex left half plane) of two degrees of freedom able to
provide a phase (0 6 / 6 p /2) and a gain (g) at frequency (xc) (in
order to achieve the above specications) that exhibits the magnitude
Bode diagram with the smallest slope at high frequencies is of the
form:

RFD s Ksa

where 0 6 a

2/

6 1; K

xac

11

We denote this controller as the FD controller.


This property is easily justied in Fig. 1 where the Bode diagrams of controller (11) and a standard PD (RPD(s) = K(1 + Tds))
are compared for the case of / = 30 and g = 1. Parameters K and
Td of the PD are also designed to verify the desired /and g values.

This plot shows that controller (11) gives exactly the phase needed
while the PD gives this phase at frequency xc but ends with a
phase of / = 90  30 at high frequencies. Moreover the magnitude Bode diagram of the PD controller exhibits a positive slope
of 20 dB/dec while (11) exhibits a slope of 6.7 dB/dec. Then the frequency response of the open-loop plant G(s)R(s) exhibits smaller
magnitude at high frequencies with the FD than with the PD controller, which means more robustness at high frequencies, attenuating better (or amplifying less) the effects of high frequency
unmodelled dynamics and noises. We mention that a phase lead
controller of two degrees of freedom (RPL(s) = (s + a)/(s + b)) can
also achieve the phase and gain specications exhibiting a slope
of 0 dB/dec at high frequencies, which is smaller than the slope
of (11). But often these specications lead to non-minimum-phase
controllers that make unstable the closed-loop system. This problem can be precluded by the use of the complete phase lead controller (RPL(s) = K(s + a)/(s + b)) which has three degrees of freedom.
3. Irrigation canal pool dynamic model
We have studied the Aragons Imperial Main Canal (AIMC)
belonging to the Ebro Hydrographic Confederation in Zaragoza,
Spain. This canal is 108 km long, and has a trapezoidal cross section
and a design head discharge of 30.0 m3 s1. It has nine pools of different length separated by undershot gates. Data used in this paper
was obtained from the rst pool which is 8.0 km. long, it has a variable depth between 3.7 and 3.1 m, a variable width between 15.0
and 30.0 m, and the mentioned discharge of 30.0 m3 s1 in all its
extension. This canal pool is operated by using undershoot gates
and distant downstream end control method. The available measurements are the downstream end water levels and the gate
positions.
In downstream end controlled irrigation canal pool i, the controlled variable is the downstream end water level yi(t), the manipulated variable is the upstream gate position ui(t), and the
fundamental disturbance variable is the unknown offtake discharge qi(t) as sketched in Fig. 2.
In order to control the water distribution in a main irrigation canal pool, it is not necessary to know the water level variations
along the whole pool. It is enough to measure it in some specic
points that depend on the canal operation method to be used [3].
In this case, since the water distribution is done by gravity offtakes,
a good distribution is obtained by maintaining a constant water level at the offtake. Considering this, a linear model with concentrated parameters and a time delay adequately characterizes the
dynamical behavior of a main irrigation canal pool in the measurement points [24].
Experiments based on the response to a step like input were
carried out in order to obtain a mathematical model that describes
the dynamic behavior of a main irrigation canal pool under study.
In this test the downstream gate was kept in a xed position, the
upstream gate was excited with a step signal and the downstream
end water level was measured with a level sensor. After recording
the response and applying a standard identication procedure we
obtained the transfer function:

Gs

ys
K

ess ;
us T 1 s 1T 2 s 1

12

where K static gain; T1, T2 time constants; s time delay. This


identication procedure also provided with a model of the canal
pool disturbance and noise, given by expression:

Gd s
Fig. 1. Frequency responses of equivalent PD and FD controllers.

ys 0:8806s2 0:02378s 0:00007123

;
vs
s2 0:01987s 0:00000692

where v(t) is a white noise.

13

V. Feliu-Batlle et al. / Journal of Process Control 19 (2009) 506519

509

Fig. 2. Schematic representation of an open irrigation main canal with undershot gates.

We denote as K0, s0, T10, T20 the nominal values of the parameters of model (12). We consider that T1 is the dominant time constant (the larger one associated to the dynamics of the canal pool),
while T2 is the smaller time constant that represents the motor + gate dynamics, which is much faster than the canal pool dynamics,
and it is nearly invariant with respect to discharge regimes. The
estimated nominal values of the model parameters are
K0 = 0.053, T10 = 1500s, T20 = 20s, and s0 = 360 s.
As the obtained dominant time constant T10 is about 75 times
larger than the secondary time constant T20, model (12) is simplied for control design purposes to

K
Gs
ess :
T1s 1

tuned in order to full specications (a) and (b). All this suggests
that these three specications can be attained by a PI controller
of the form:

Rs K p

15

arranged according to the standard control scheme of Fig. 3.


The robustness of this controller to changes in the time delay
(maximum deviation of the time delay from the nominal value that
keeps stable the closed-loop system) can be easily obtained:

s^max  s0

14

Experiments reported in previous works on identication of canal pool dynamic behavior showed that all these parameters may
exhibit wide variations when the discharge regimes change across
the gates in the operation range [33]. In our particular canal pool,
we consider variations in the time delay, smin 6 s 6 smax, because
this is the parameter more determinant in the stability of the
closed-loop control system of main irrigation canal pools [3,9].
Moreover we will also study the effects of variations in the secondary time constant in the simulation section.

1 Tds
;
s

/m

xc

16

It has to be noticed that any controller R(s) of Fig. 3 that fulls


specications (a) and (b) will exhibit the same time delay stability
margin (16), independently of its particular form. Then a different
control structure has to be used in order to improve the robustness
to changes in the time delay. Next a control structure based on the
Smith predictor is proposed, which is shown in Fig. 4. In this case
the closed-loop transfer function assuming a nominal time delay
s0 and a real time delay s is

Ys M c sCs Md sDs;

17

where

Mc s

18

In this Section our Smith predictor control scheme is dened


and the effects of time delay variations are studied.

RsG0 sess
;
1 RsG0 s1  es0 s ess

Md s

1 RsG0 s1  es0 s
;
1 RsG0 s1  es0 s ess

19

4.1. Smith predictor control scheme

and

4. The Smith predictor based control scheme

Assume we want to design a controller for the main canal pool


(12) (or (14)) with the next specications: (a) phase margin (/m),
(b) crossover frequency (xc), and (c) zero steady state error. The
last specication implies that the controller must include an integral term. Moreover the controller needs two parameters to be

Gs G0 sess ;
0

20

being G (s) the rational part of the model. Moreover expression (13)
yields D(s) = Gd(s)v(s).
In the scheme of Fig. 4 (see expression (18)) the time delay stability margin depends on the particular form of any controller R(s)

Fig. 3. Standard control system of AIMC rst pool.

510

V. Feliu-Batlle et al. / Journal of Process Control 19 (2009) 506519

Fig. 4. Smith predictor based control system of AIMC rst pool.

that fulls specications (a) and (b), in opposition to what happens


in the control scheme of Fig. 3 where any controller R(s) which fulfils these conditions exhibits the same time delay stability margin.
Then a proper design of such controller R(s) in a Smith predictor
based control scheme may increase such stability margin.
4.2. Effects of a detuned Smith predictor
When ss0 the Smith predictor is detuned and the closed-loop
characteristic equation is obtained from expression (18):

 If s = 0, thenW(jx, 0, s0) = 2  cos(s0 x) + jsin(s0x) and its phase


is bounded: 30 6 \W(jx, 0, s0) 6 30.
 Function W exhibits a positive phase at low frequencies in the
interval 0 6 s < s0, while its phase is negative at low frequencies
in the interval s0 < s < 1.
 In the case that 0 6 s 6 s0, the rst change of the phase of W
from positive to negative values is produced at a frequency:

xc

p
:
s0 s

23

The open-loop frequency response to be taken into account for


the closed-loop stability analysis is

Fig. 5 draws the frequency response of W(jx, s, s0) for several


values of d = s  s0. It shows that, while the magnitude plot never
surpasses 3 (10 dB), the phase decreases quickly when the frequency surpasses a given value that depends on d.

Hjx; s; s0 RjxG0 jxWjx; s; s0 ; Wjx; s; s0

4.3. The FI controller

1 RsG0 s1  es0 s ess 0:

jxs0

1e

jxs

21

22

where the term W(jx, s, s0) represents the effects introduced by the
non-nominal time delay. Some features of this term are
 Its magnitude is bounded: 0 6 jW(jx, s, s0)j 6 3.
 In the case of nominal delay: W(jx, s0, s0) = 1.

The FD controller (11) is not able to full the third specication


of zero steady state error to step commands. A natural modication
of the previous fractional controller in order to achieve this is to include an integral term. It leads to the fractional integral controller:

RFI s

RFD s Ksa K

b;
s
s
s

Fig. 5. Frequency response of W(jx, s, s0) for several values of d.

b 1  a P 0:

24

V. Feliu-Batlle et al. / Journal of Process Control 19 (2009) 506519

Property 2. Controller (24) is the controller of two degrees of freedom


that simultaneously achieves desired phase margin, crossover frequency and zero steady state error to step commands, and exhibits the
magnitude Bode diagram with the smallest slope at high frequencies.
Controller (24) is obtained from controller (11) by adding a pole at the
origin in order to achieve the steady state condition. A similar
procedure can be used to obtain a PI controller from a PD one, and to
make any other controller of two degrees of freedom exhibit zero
steady state error to steps commands. Then the above property easily
follows from previous Property 1.

511

phase at high frequencies facilitates the fullment of (25) and,


therefore, increases the stability margin dm too. As (24) has the
property of adding less phase than other controllers at high
frequencies, it reduces the robustness of this controller.
The two before effects are contradictory and one will dominate
the other depending on the plant and controller parameters.
Section 6 will study the circumstances that cause the preeminence of each effect, making the FI controller more or less
robust than the well known PI and PID controllers. Previously a
simple controller design procedure is proposed in Section 5 that
will be used in the calculations of Section 6.

4.4. Robustness of the FI controller with a Smith predictor


5. Algebraic procedure of design
First a sufcient stability criterion for the closed-loop system is
given.
Lemma. Consider a plant of the form (12) or (14) being controlled by
a FI controller (24) combined with a Smith predictor tuned to s0,
0
(ss0). Let us denote as xe the frequency such that jG (jxe)RFI(jxe)j =
i
1/3. Let us denote as xs s; s0 , 1 6 i 6 n(s, s0) the set of frequencies
that verify j Hjxis ; s; s0 j 1, where n is the number of them and its
dependence on parameters s and s0 has been made explicit. Then (1)
at least one frequency xis s; s0 exists, (2) all these frequencies verify
xis s; s0 6 xe , (3) a sufcient condition for the stability of the closedloop system is

\G0 jxis \RFI jxis \Wjxis ; s; s0 > p;

1 6 i 6 ns; s0 :
25

Proof. If we consider that (a) lim j Hjx; s; s0 j 1 because


RFI(jx)

always

introduces

x!0

pole

in

the

origin,

and

lim j Wjx; s; s0 j 1, (b) lim j Hjx; s; s0 j 0, and (c) function

x!0

x!1

i
s

jH(jx, s, s0)j is continuous, then at least one frequency x

s; s0

exists such that j Hjxis ; s; s0 j 1, and the rst part of the lemma
is proven.

G0 jxc Rjxc ejp/m ;

28

where now the phase margin /m is expressed in radians. This equation holds for any controller R(s). For the case of our FI controller it
gives

G0 jxc ejp/m

jxc b xbc jpb

e2 ;
K
K

29

which yields conditions for the phase and magnitude that allow to
calculate the parameters of this controller:

b
K

p \G0 jxc  /m ;

b
c

j G0 jxc j

30
31

Condition (28) can also be applied to a PI controller RPI s K i Ks

s
yielding:

We have that:

jHjx; s; s0 j 6 3jG0 jxRFI jxj;

In this Section we develop an algebraic procedure to calculate


the parameters of controller (24) in a very simple way. This controller must full specications (a)(c) of Section 4.1.
Specications (a) and (b) in the frequency domain lead to two
equations that can be written in a compact form in the complex
plane:

26

because jWj 6 3, "x, as was stated in Section 4.2. Moreover combining the denition of xe with inequality (26), and the fact that
0
jG (jx)RFI (jx)j is always a descendent function for transfer functions (12) and (14) with a controller of the form (24), it easily follows that jH(jx, s, s0)j 6 1, " x P xe. Therefore it is not possible
to have frequencies xis s; s0 P xe and the second part of the lemma is proven.
The third part of the lemma easily follows from the Nyquist
criterion. Expression (25) is a version for multiple frequencies of
the well known positive phase margin stability condition, and
guarantees that the Nyquist plot does not encircle the point (1, 0),
yielding to the sufcient stability condition.
Then the stability margin dm = sm  s0 of the time delay is given
by

sm maxs; s j condition 25 is verified 8 s0 ; s0 6 s0 6 s: 



jxc ej/m
;
0
G jxc
 j/ 
je m
K I 0
;
G jxc
K i R

32
33

where R and I stand for the real and imaginary part of a complex
number.
This frequency domain algebraic design procedure can be easily
extended to other design cases:
1. A Smith predictor based controller with a delay different from the
0
nominal value. Controllers can be designed substituting G (jx)
0
by G (jx)W(jx, s, s0) in expressions (28)(33).
2. A FI controller in a standard closed-loop scheme (see Fig. 3). A controller that provides the desired frequency specications can be
0
0
obtained by substituting G (jx) by G (jx)ejxs in expressions
(28)(33).

27
6. Application to rst order systems with delay
A consequence of Property 2 is that (24) is the controller of two
degrees of freedom that fulls conditions (a)(c) of Section 4.1, and
makes the magnitude of H(jx, s, s0) decrease fastest (smaller slope)
at high frequencies, making thus frequency xe be minimum. Then
the frequency interval where condition (25) has to be veried
decreases and we may expect stable closed-loop systems for larger
values of d. However, controller (24) provides with less phase at
high frequencies than other controllers like a PI. Adding more

In this Section we study the cases where a FI controller is more


robust to changes in the time delay than a PI or PID controller. We
consider the quite common case of rst order plants with a delay.
As we are not able to provide with an analytical demonstration, we
will carry on the comparison through extensive calculations and
simulations. Lemma of Section 4 and the design procedure of Section 5 are used to reduce computations.

512

V. Feliu-Batlle et al. / Journal of Process Control 19 (2009) 506519

A plant of the form (14) can be normalized by scaling the time t


by T1 (tn = t/T1) and the plant output y by K(yn = y/K). The normalized transfer function is

Gn s

1  sn s
;
e
s1

sn s=T 1

34

The common specication of a design phase margin /m = 60 is


assumed. The time delay margin depends on the normalized crossover frequency xcn (second design specication). From now on we
will use the normalized closed-loop time constant Tcn = 1/xcn instead of xcn because some graphics and expressions become simpler. This parameter varies from 0.05 s, which means a closed-loop
system 20 times faster than the open-loop one (without taking into
account the delay), to 1.5 s, which means closed-loop system 1.5
times slower than the open-loop one (which is quite uncommon
as we usually want to make the closed-loop system faster than
the open-loop one), in intervals of 0.05 s. The range of variation
of the normalized time delay used in the Smith predictor s0n is
from 0 s to 10 s in intervals of 0.1 s, which means considering from
zero delay to a delay ten times larger than the main time constant.

Fig. 6. Time delay stability margin obtained using the FI controller dFI
m .

6.1. FI controller versus PI controller


Fig. 6 shows the time delay stability margin obtained using the
FI controller dFI
m . This gure shows an abrupt change in the
robustness which basically divides the whole area into two regions: a region or high robustness where the robustness dFI
m slowly
increases as the nominal time delay s0n increases, and a region of
low robustness where dFI
m decreases as s0n increases. This pattern
can also be noticed in the PI and PID controllers. Fig. 7 shows the
border line between the before two regions in the cases of FI, PI
and PID controllers.
After extensive calculations in the whole dened search space
(0.05 6 Tcn 6 1.5, 0 6 s0n 6 10) three regions were found: a region
where the FI is more robust than the PI, a region where the PI is
more robust than the FI, and a third region where both controllers
have about the same robustness (normalized time delay stability
margins differ in less than 0.2 s). These regions can be approximated by
PI
1 FI better than PIdFI
m > dm :

s0n 6 0:3021T 2cn 2:5551T cn ;


35

PI
2 PI better than FIdFI
m < dm :

s0n P 0:3021T 2cn 2:5551T cn ;


36

s0n 6 1:3844T 4cn 2:0312T 3cn  0:0322T 2cn 2:5551T cn ;


37
3 PI similar to

FIdFI
m

dPI
m

s0n > 1:3844T 4cn 2:0312T 3cn  0:0322T 2cn 2:5551T cn :


38
6.2. FI controller versus PID controller
Improvements using a PID controller are studied next. PID has
three parameters to be adjusted in (4) (K, Ki, Kd). As we have now
an extra parameter, we try to improve the robustness properties
while keeping the same nominal behavior by adequately tuning
these three parameters. We run an optimization process that maximizes the time delay stability margin dPID
m while keeping the
same phase margin, crossover frequency and zero steady state error to step commands. The algebraic procedure of Section 5 was
modied in order to deal with PID controllers. As expected, larger
stability margins are obtained than with PI in all the cases (Tcn, s0n),
but this improvement is signicant only in a small region. This

Fig. 7. Border line between the high and low robustness regions in the cases of FI, PI
and PID controllers.

result is consistent with the statements made in Section 4 because


a PID exhibits a magnitude Bode diagram with a larger slope at
high frequencies than a PI, then yielding an only marginal robustness improvement.
Again the space of search is divided into three regions: the region where the FI is more robust than the PID, the region where
the PID is more robust than the FI (which is larger now), and a third
region where both controllers have about the same robustness
(normalized time delay stability margins differ in less than 0.2 s).
These regions can be approximated by
PID
1 FI better than PIDdFI
m > dm :

s0n > 1:3795T 3cn  1:8114T 2cn 1:2724T cn ;


s0n 6 0:3021T 2cn 2:5551T cn ;
2

PID
PID better than FIdFI
m < dm :
2
0n P 0:3021T cn 2:5551T cn ;
3
2
0n 6 1:5279T cn 3:2443T cn 3:1839T cn ;
FI
PID
PID similar to FIdm  dm :
3
2
0n > 1:5279cn 3:2443T cn 3:1839T cn ;
and 0n 6 1:3795T 3cn  1:8114T 2cn 1:2724T cn :

39
40

s
s

41

43

42

44

V. Feliu-Batlle et al. / Journal of Process Control 19 (2009) 506519

513

the process, as the Smith predictor based control schemes keep


the delay invariant.
Fig. 8 shows regions (35)(38) (FI versus PI), and includes the
rectangular design area dened above. Fig. 9 does the same for
the regions (39)(44) (FI versus PID). Fig. 8 shows that FI is more
robust than PI in most of the design area. Fig. 9 shows that FI is
more robust than PID in an important section of the design area,
and in most of the remaining area it exhibits similar robustness
than PID. These results are the principal motivation for using our fractional controller combined with an Smith predictor in a main irrigation canal pool.
7. Case study: a main irrigation canal pool

Fig. 8. Comparison of the time delay stability margins obtained with the FI and PI
controllers.

Comparing regions (39)(44) with (35)(38) we notice that the region where the FI controller is better than the PID shrinks by its
lower bound: the upper bound (35) remains the same but now appears a lower bound (39). This reduction has little effect on the design criteria as it happens mostly when Tcn > 0.7, which is a zone
where its is quite unlikely that we place our closed-loop system
speed of response specication. The region where the FI controller
is worse than the PID has grown (see border (42)).
6.3. Application to a main irrigation canal pool
Dynamics of a main irrigation canal pool are usually characterized by large main time constants compared to their time delays. In
some cases they are even approximately modelled by a transfer
function of the form G(s) = ess/(As) (see e.g. [34,35]), which implies an innite main time constant. This means that ratio sn (see
(34)) is usually small (sn < 1). Moreover, we usually may want
our closed-loop system be faster that the open-loop one. Reasonable improvements can range from Tcn = 1 (the same speed as the
open-loop system) to Tcn = 0.1 (ten times faster than the open-loop
system). Lower values of Tcn may saturate actuators. These
improvements are achieved only in the main time constant of

In this section our controller is applied to the closed-loop control of our main irrigation canal pool. Robustness to changes in
the time delay of six control schemes are compared: (a) a PI standard controller in the scheme of Fig. 3, (b) a PI standard controller
with Smith predictor (Fig. 4), (c) our fractional controller with
Smith predictor (Fig. 4), a PID standard controller with Smith predictor (Fig. 4), a predictive controller, and a H1 robust controller.
All these controllers are designed to achieve the same frequency
specications (xc and /m) for the simplied plant (14) when nominal values K0, T10 and s0.
In our main irrigation canal pool the time delay may experience
variations in the range smin = s0/2 6 s 6 2s0 = smax. The controller
must remain stable for delays up to smax, but we want a controller
that keeps the closed-loop system stable for larger values in order
to cope with some emergency situations. Control systems will be
compared from six points of view: (a) dynamic performance of
the controlled systems in the case of nominal plant dynamics,
where the controllers are designed with the same given frequency
specications, (b) amplitudes of control signals, (c) maximum
deviation of the time delay that keeps the closed-loop system stable (robustness to time delay changes), (d) behavior of the dynamic
response in the normal work range dened above [smin, smax], (e)
robustness to changes in the secondary time constant (T2), which
implies changes in the high frequency range of the canal model,
(f) sensitivity to measurement noise and perturbations.
7.1. Design specications
Open-loop canal pool settling time is obtained as Ts 
3T10 + s0 = 4860 s. The closed-loop control system is designed to
be nearly four times faster i.e. exhibit a settling time T cs  1300 s.
A crossover frequency of xc  3=T cs 0:0023 rad/s and the standard phase margin /m = 60 are chosen as design specications.
7.2. Standard PI controller
A PI controller is designed for the scheme of Fig. 3. The procedure of Section 5 applied to the whole transfer function
0
(G (jx)ejxs) with model (14) yields a negative value in expression
(32) resulting a controller:

PIs

0:00411  16555s
;
s

45

which makes unstable the closed-loop system. Then there is not


possible to nd a PI controller that fulls the above frequency domain specications.
7.3. Design of PI, FI and PID controllers with the Smith predictor

Fig. 9. Comparison of the time delay stability margins obtained with the FI and PID
controllers.

The design crossover frequency has to be modied in the case of


a Smith predictor. Now expression T cs  3T c10 s0 has to be considered in order to achieve the desired settling time. This yields a

514

V. Feliu-Batlle et al. / Journal of Process Control 19 (2009) 506519

closed-loop main time constant of T c10  313 s, and a corresponding crossover frequency of xc = 0.0032 rad/s. The PI controller is
obtained by using expressions (32), (33):

PIs s

0:1961 350:7s
;
s

46

and the FI controller from (30), (31):

FIs s

6:4
:
s0:46

47

A PID that fulls the same /m and xc than the two previous controllers and maximizes the time delay stability margin is obtained
from an optimization procedure:

PIDs s

0:121 572:7s  62260s2


:
s

48

Fig. 10 plots the Bode diagrams of the open-loop system with


controllers (46)(48). The responses of these three controllers to
unity step commands C(s) are drawn in Fig. 11. The region of errors
lower than 5% of the desired nal value, which denes the settling
time, is also drawn in this gure and in the following ones. This gure shows that the overshot is 18.7% (PI), 9.6% (FI) and 6.9% (PID),
and the settling time is 2059 s (PI), 1601 s (FI) and 1864 s (PID). The
rise time is approximately the same for all controllers. Fig. 12 plots
the control signals u(t) of all controllers and shows that the maximum value of the control signals given by the PI and the PID are
respectively about 4.3% and 38.6% larger than the maximum of
the signal given by the FI.
Then the FI gives a faster response (less settling time) than the
PI and PID controllers while generating a control action of less
amplitude (all controllers having been designed for the same frequency specications). Moreover the FI response is more damped
than the PI response and slightly less damped than the PID response. This suggests that the FI controller provides with a more
efcient control action than the PI and PID. In turn Fig. 11 shows
that the steady error converges to zero more slowly when using
the FI than PI or PID controllers. This may be a severe drawback
when a precise control is required, like in the case of servos. But
in the case of controlling a main irrigation canal pool, achieving
an error of less than 5% is accurate enough.
We have also included in Fig. 11 the response of the complete
plant (12) (second order with delay plant) using the FI controller
(47) plus Smith predictor. This response is very close to the one obtained using the rst order with delay plant (14), supporting the
validity of the proposed simplication.

Fig. 10. Bode plots of the open-loop system with PI, FI and PID controllers.

Fig. 11. Responses of PI, FI and PID control systems to unity step commands C(s).

Fig. 12. Control signals u(t) of PI, FI and PID controllers.

7.4. Robustness to delay changes


First, we must locate the design point (Tcn, s0n) in Figs. 8 and 9. It
easily follows that Tcn = 1/(T10xc) = 0.2089 and s0n = s0/T10 = 0.24.
This point is marked with a p in these gures and lies inside
the design region. Moreover these two gures show that this point
lies inside the region where the FI controller is more robust than
the PI and PID controllers.
If we carry on a numerical procedure similar to the one used to
calculate Figs. 8 and 9, the time delay stability margins are obFI
tained for the three controllers: dPI
m 0:2931, dm 0:3698 and

0:3526.
The
fractional
controller
improves
the robustness
dPID
m
of the PI in 26 % and the robustness of a PID in 4%. The maximum
time delay allowed by these controllers is given by s0max s0
dm T 10 which yields: s0max 2:22s0 for the PI, s0max 2:54s0 for the
FI and s0max 2:47s0 for the PID.
Figs. 1315 show the responses of the PI, FI and PID controllers
with a Smith predictor to unity step commands for time delay values of smin, smax and 2.5s0 respectively. Control signals are also
plotted. Figs. 13 and 14 show that the FI provides with more
damped and faster (less settling time) responses in the working
range [smin, smax] than the other controllers (these results were obtained for the extremes of the working interval and remain true for

V. Feliu-Batlle et al. / Journal of Process Control 19 (2009) 506519

515

intermediate values too), while the amplitude of the control signal


is slightly smaller. Fig. 15 shows the previously established result
that the PI and the PID controllers are unstable for 2.5s0 while
the FI remains stable.
7.5. Comparison with other robust control schemes

Fig. 13. Responses of PI, FI and PID control systems with a Smith predictor to unity
step commands for values smin.

Next our fractional controller is compared with other more


complex control schemes: a Model Based Predictive controller
(MPC) and an H1 controller (this last one is specially designed to
be robust). Fig. 16 shows the temporal responses of these controllers when time delay changes (s0/2 6 s 6 2s0). Both control
schemes have been tuned using the Robust Control Toolbox of MATLAB [36,37]. Tables 1 and 2 resume the design parameters used for
the MPC and H1 tuning respectively.
The MPC presents higher robustness than our fractional controller plus Smith predictor, nevertheless its temporal response with
nominal time delay is too slow (Ts = 4300 s). The rapidity of the
temporal response can be improved decreasing the sampling period. In fact with a sampling period of 1 s a settling time Ts = 2100 s
is achieved. The drawback of this method is that often the sampling frequency can not be increased much because of technological limitations of the installed sensors.
The H1 controller presents a similar temporal response than the
fractional controller plus Smith predictor in the nominal case
(Ts  1900 s). It shows a much better behavior than the FI at
smax = 2s0. Nevertheless it exhibits a slightly lower robustness to

Fig. 14. Responses of PI, FI and PID control systems with a Smith predictor to unity
step commands for values smax.

Fig. 16. Temporal responses of H1 and MPC with time delay changes (s = s0/2,
s = s0 and s = 2s0).

Table 1
Design parameters of the Predictive control
Sample time
Prediction horizon
Optimizer solver
Maximum number of iterations

60 s
2000
ActiveSet
200

Table 2
Design parameters of the H1 control

Fig. 15. Responses of PI, FI and PID control systems with a Smith predictor to unity
step commands for values 2.5s0.

Sample time
Algorithm
Weighting function

60 s
DGKF [38]
Ds
DK K s DT 0 DK 0 T 0 0:5Ds0 s2K 0 T 0 DK 0 T 0 K 0 DT 0 20 s2
Wy 0 0 0
K 0 T 0 DT 0 s1

516

V. Feliu-Batlle et al. / Journal of Process Control 19 (2009) 506519

changes in the time delay being its maximum stable value


s = 2.4s0. The drawback of this controller is that its design can be
hardly automated. H1 controller tuning methods require an iteration process to calculate the controller transfer function. Moreover,
several adequate weighting functions must be selected to obtain
controllers that provide the system with a robust behavior. The order of the resulting controller is usually high.
7.6. Robustness to changes in the secondary time delay
Next the secondary time constant T2 T1 is considered. As
changes in T2 inuence the high frequency dynamics of the plant,
we check if the property of the FI controller of making the magnitude of the open-loop system decrease faster at high frequencies
implies any improvement in the robustness to this parameter. Very
small values of T2 do not unstabilize the closed-loop system (transfer function (12) approaches (14)). Then we study how much we
can increase the value of T2 before the system becomes unstable.
Simulations for different values of T2 show that the closed-loop
system with the PI controller becomes unstable for values
T2 P 635 s and the PID for values T2 P 1050 s while the system
with the FI controller never becomes unstable.

Fig. 18. Control signals of PI, FI and PID control systems with a Smith predictor to
step commands of amplitude 3 in presence of the modelled measurement noise and
perturbations.

7.7. Sensitivity to measurement noise and perturbations


Sensitivity to output measurement noise of PI, FI and PID controllers is compared. The water level sensor has a sampling period
of T = 60 s. A disturbance modelled by a noise with an uniform
energy spectrum in the range [0.5fn, fn] where fn = 0.5/T is the
Nyquist frequency colored by transfer function (13) has been
considered in the simulations. The command signal is a step of
amplitude 3. Then the noise is a very noticeable component of
the measured signal. This high noise/signal ratio is not unusual
in main irrigation canal pools as often changes in the command
signal are of short amplitude, and the control is of incremental nature. Fig. 17 plots the responses of the canal pool with the PI, FI and
PID controllers, and Fig. 18 plots the control signals generated by
these controllers. They show that the FI controller deals better with
this noise as the amplitude of the noisy component of its output is

Fig. 17. Responses of PI, FI and PID control systems with a Smith predictor to step
commands of amplitude 3 in presence of the modelled measurement noise and
perturbations.

noticeably smaller than in the other cases. Control signals of Fig. 18


show this more clearly.
Fig. 19 shows the behavior of PI, FI and PID controllers when a
step disturbance of amplitude 1 (D(s) in Fig. 4) is produced. This
disturbance may model an unpredicted water offtake. The responses are shown for three different time delays: s = s0/2, s = s0
and s = 1.5s0. This gure shows that the response of the fractional
controller is usually more damped and converges faster to the error
band of 5% than the responses of the other two controllers. For
example, in the nominal case the settling times are: 3076.3 s for
the PI, 3008.5 s for the PID and 2723.4 s for the FI.
8. Discussion and conclusions
The design of Smith predictor based controllers robust to high
frequency unmodelled dynamics and disturbances has been studied. Special attention has been paid to the stability robustness to
changes in the process time delay.
Such robustness is associated to a fast descent of the magnitude
Bode diagram at high frequencies. We found that the controller
that provides with the maximum descent of such magnitude diagram in the open-loop transfer function at high frequencies while
satisfying the desired closed-loop specications is of fractional
order.
In particular, among all the controllers having two parameters
to be tuned, the proposed fractional integral controller (24) is the
one that provides with the smallest magnitude diagram slope at
high frequencies.
A simple methodology to design these FI controllers combined
with Smith predictors has been proposed. Robustness of these controllers to time delay changes in rst order plus time delay plants
(14) has been compared with the robustness of PI controllers, and
PID controllers optimized to exhibit the largest time delay margin
(both controllers combined with a Smith predictor). Though the FI
controller provides with the smallest magnitude Bode diagram
slope at high frequencies, there are regions where the PI and the
PID are more robust than the FI. This is explained by the fact that
the mismatching in the time delay can not be considered only as
a high frequency modeling error, but this error also inuences at
medium frequencies.
However there is a very important region where the FI is more
robust than PI, and even than PID which is a more complex con-

V. Feliu-Batlle et al. / Journal of Process Control 19 (2009) 506519

517

Fig. 19. Responses of PI, FI and PID control systems with a Smith predictor to step disturbances D(s) of amplitude 1 for the cases of time delays s = s0/2, s = s0 and s=1.5s0.

troller having its robustness time delay margin maximized by a


proper tuning of its three parameters. Many control system designs
of real main canal pools are placed in this region. So our FI controllers are well suited for robust Smith predictor based control of a
main irrigation canal pool.
As dynamic parameters of a main irrigation canal pool change
through time, and canal pool dynamic responses are inuenced
by several disturbances, the behavior of several controllers has
been compared under these circumstances for the case of a real
main canal pool that we have modeled and parametrically identied. These comparisons showed:
 The FI controller performed better than the PI and PID controllers in the case of the nominal plant in terms of the output
response and the amplitude of the control signal.
 The dynamic response to a step command of the plant controlled
with the FI remains better than responses with the PI and the
PID for any delay included in the working conditions range.
 An improvement of about the 20% in the stability margin of the
time delay was obtained in our example by using FI instead of PI
controllers. However we have achieved improvements over 40%
in this margin in other canal pools [24].
 More complex control schemes as H1 or MPC do not necessarily
improve the robustness to changes in time delay. Often these
controllers lead to worst temporal response in the nominal
delay case if the robustness to the time delay must be made
comparable to what can be achieved with our FI.
 The robustness of the Smith predictor based closed-loop control to changes in the secondary time constant of the canal
pool is improved if a FI controller is used instead of a PI or a
PID.

 The FI controller exhibits less sensitivity to disturbances than PI


and PID. FI control actions require less abrupt motions of the
gate motors than the other controllers in all the cases: with
and without disturbances. This may be an advantage in preventing the ageing of these actuators.
The proposed methodology can be viewed as a particular application of the loop shaping design used in QFT (quantitative feedback theory). Loop shaping design has the drawbacks that (a) it
requires optimization algorithms that often can not nd the globally optimum solution; (b) it needs some error bounding transfer
functions that often can not be systematically chosen. Then usual
approaches are based on manual designs or automated designs
based on particular controller structures. We mention that, from
this point of view, fractional controllers can play an important role
in QFT as they allow a loop shaping with interesting robustness
properties by designing a new controller parameter: the value of
the fractional derivative. This allows us to design open-loop frequency responses by tuning few parameters (b and one gain in
our controller) with similar features to what can be obtained in
the standard QFT by tuning more parameters in high order controllers, and with more complex optimization procedures.
Then our FI controller allows to design robust controllers with a
minimum design effort that can be easily automated according to
(30), (31). Its robustness is larger than what can be achieved with
PI and PID controllers and similar (may be slightly lower) to what
can be achieved with an H1, in the cases of many irrigation main
canal pools. But the design effort needed for this last controller is
much larger. This makes our controller especially well suited for
implementing adaptive robust controllers (like gain scheduling
control schemes) where the controller has to be redesigned in real

518

V. Feliu-Batlle et al. / Journal of Process Control 19 (2009) 506519

time as the parameters of the dynamics of the process change,


while it is wanted to get some robustness properties in the tuned
controller in order to face abrupt changes in the plant dynamics
or errors in the real time parameter estimation algorithm (parameter identication subsystem of the adaptive control system). This
kind of control systems is of great interest for irrigation main canal
pools as they are time varying processes.
Fractional controllers can be digitally implemented in very
straightforward ways by using both IIR or FIR lters ([27,28],
e.g.). Once the parameters of the very simple controller (24) are designed, its digital implementation is obtained automatically. The
rst practical implementation of a simple fractional controller in
a prototype canal pool was done in [25]. Moreover a version of this
controller has been implemented in a Siemens PLC (a Simatic 300)
[39]. In both applications, as well as in our simulations, the discretized operator (8), (9) combined with the short memory approximation was used. We also mention that there are some recent
developments that allow to implement fractional controllers by
analogic physical devices that reproduce their frequency responses
over ranges of 6 or more decades [40].
The next objective of our research is generalizing the application of our FI to the control of a main canal with multiple pools.
We mention that modern canal control systems may be more complex than a PI or a Smith predictor based controller, and multivariable controllers may be proposed. But often these control systems
include series of simple PI controllers as the lowest control level
which are coordinated by other more complex controllers at
upper levels ([41], e.g.). Although centralized H1 or MPC controllers may provide with a better performance than decentralized
controllers, the last ones offer some advantages: (a) the design
and implementation of these control systems are simpler, (b) the
impact of communication failures or electrical power breakdown
(which are common due to lightning, vandalism, and rodent bites,
e.g. [42]) is more limited if the control system is implemented locally (or there were local controllers at a low level coordinated by a
high level multivariable controller), (c) communication systems
and electronic devices are cheaper. Substituting these PI controllers by our fractional controllers at such low level control may improve the global control system because local control robustness is
increased having an impact on the overall control system performance. We nally mention that our design methodology can be extended to design other more complex fractional PID controllers.
Acknowledgments
The authors would like to acknowledge the support provided by
the Program of International Cooperation of the Universidad de
Castilla-la Mancha, Spain. The authors would also like to thank
the journal editor and reviewers for their revision and fruitful comments and recommendations, which have made it possible to improve upon the original paper.
References
[1] S. Postel, Last Oasis: Facing Water Scarcity, Norton, New York, USA, 1992.
[2] P.O. Malaterre, D.C. Rogers, J. Schuurmans, Classication of canal control
algorithms, Journal of Irrigation and Drainage Engineering 124 (1) (1998) 310.
[3] P.I. Kovalenko, Automation of Land Reclamation Systems, Kolos, Moscow,
1983.
[4] X. Litrico, V. Fromion, J.P. Baume, C. Arranja, M. Rijo, Experimental validation of
a methodology to control irrigation canals based on Saint-Venant equations,
Control Engineering Practice 13 (11) (2005) 13411454.
[5] S.K. Ooi, M.P.M. Krutzen, E. Weyer, On physical and data driven modeling of
irrigation channels, Control Engineering Practice 13 (4) (2005) 461471.
[6] M.H. Chaudhry, Open-Channel Flow, Prentice-Hall, Englewood Clifs, NJ, USA,
1993.
[7] R. Rivas Perez, J.R. Peran Gonzalez, B. Pineda Reyes, S. Perez Pereira S, Distributed
control under centralized intelligent supervision in the Gira de Melena
Irrigation System, Hydraulic Engineering in Mexico 18 (2) (2003) 5368.

[8] J.P. Baume, J. Sau, P.O. Malaterre, Modeling of irrigation channel dynamics for
controller design, in: Proceedings of IEEE International Conference on Systems,
Man and Cybernetics (SMC98), San Diego, California, USA, 1998.
[9] B.T. Wahlin, Performance of model predictive control on ASCE test canal
1, Journal of Irrigation and Drainage Engineering 130 (3) (2004) 227
237.
[10] A.J. Clemmens, J. Schuurmans, Simple optimal downstream feedback canal
controllers: theory, Journal of Irrigation and Drainage Engineering 130 (1)
(2004) 2634.
[11] X. Litrico, V. Fromion, Tuning of robust distant downstream PI controllers for
an irrigation canal pool: (I) theory, Journal of Irrigation and Drainage
Engineering 132 (4) (2006) 359368.
[12] M.J. Shand, Automatic downstream control systems for irrigation Canals, PhD
Thesis, University of California, Berkeley, USA, 1971.
[13] J.L. Deltour, F. Sanlippo, Introduction of the Smith predictor into dynamic
regulation, Journal of Irrigation and Drainage Engineering 124 (1) (1998) 47
52.
[14] X. Litrico, D. Georges, Robust continuous-time and discrete-time ow control
of a dam-river system: (II) controller design, Applied Mathematical Modelling
23 (11) (1999) 829846.
[15] Z.J. Palmor, Y. Halevi, On the design and properties of multivariable dead time
compensators, Automatica 19 (1983) 255264.
[16] Z.J. Palmor, Y. Halevi, Robustness properties of sampled-data systems with
dead time compensators, Automatica 26 (1990) 637640.
[17] A. Bhaya, C.A. Desoer, Controlling plants with delay, International Journal of
Control 41 (1985) 813830.
[18] C. Santacesaria, R. Scattolini, Easy tuning of Smith predictor in presence of
delay uncertainty, Automatica 29 (1993) 15951597.
[19] Z.J. Palmor, The Control Handbook. Time Delay Compensation: Smith Predictor
and its Modications, CRC Press and IEEE Press, 1996.
[20] J.E. Nomey-Rico, E.F. Camacho, Smith predictor and
modications: a
comparative study, in: Proceedings of European Control Conference,
Germany, 1999.
[21] J.E. Nomey-Rico, E.F. Camacho, Prediccin Para Control: Una Panormica del
Control de Procesos con Retardo, Revista Iberoamericana de Automtica e
Informtica Industrial 3 (4) (2006) 525.
[22] I. Podlubny, Fractional Differential Equations, Academic Press, San Diego, USA,
1999.
[23] A.A. Oustaloup, O. Cois, P. Lanusse, P. Melchior, X. Moreau, J. Sabatier, The
CRONE approach: theoretical developments and major applications, in:
Proceedings of the 2nd Workshop of the International Federation of
Automatic Control on Fractional Differentiation and its Applications, Porto,
Portugal, 2006.
[24] V. Feliu Batlle, R. Rivas Perez, F.J. Castillo Garcia, Fractional robust control to
delay changes in main irrigation canals, in: Proceedings of the 16th IFAC World
Congress, Prague, Czech Republic, 2005.
[25] V. Feliu Batlle, R. Rivas Perez, L. Sanchez Rodriguez, Fractional robust control of
main irrigation canals with variable dynamic parameters, Control Engineering
Practice 15 (6) (2007) 673686.
[26] I. Podlubny, Fractional-order systems and PIkDl-controllers, IEEE Transactions
on Automatic Control 44 (1) (1999) 208214.
[27] B.M. Vinagre, I. Podlubny, A. Hernandez, V. Feliu, Some approximations of
fractional order operators used in control theory and applications, Journal of
Fractional Calculus and Applied Analysis 3 (3) (2000) 231248.
[28] I. Petras, S. Grega, L. Dorcak, Digital fractional order controllers realized by PIC
microprocessor: experimental results, in: Proceedings of International
Carpathian Control Conference, High Tatras, Slovak Republic, 2003.
[29] R.S. Barbosa, J.A. Tenreiro Machado, I.M. Ferreira, Controller tuning based
on Bodes ideal transfer function, Nonlinear Dynamics 38 (14) (2004) 305
321.
[30] C.A. Monje, A.J. Calderon, B.M. Vinagre, Y. Chen, V. Feliu, Fractional PIk
controllers: some tuning rules for robustness to plant uncertainties, Nonlinear
Dynamics 38 (14) (2004) 369381.
[31] H. Ahn, Y. Chen, I. Podlubny, Robust stability checking of a class of linear
interval fractional order system using Lyapunov inequality, in:
Proceedings of the 2nd Workshop of the International Federation of
Automatic Control on Fractional Differentiation and its Applications, Porto,
Portugal, 2006.
[32] N.V. Paluri, K. Rambabu, N. Kubal, V. Vyawahare, QFT loop shaping using a
fractional integrator-real-time application to a coupled-tank system, in:
Proceedings of the 2nd Workshop of the International Federation of
Automatic Control on Fractional Differentiation and its Applications, Porto,
Portugal, 2006.
[33] R. Rivas Perez, V. Feliu Batlle, L. Sanchez Rodriguez, Robust system
identication of an irrigation main canal, Advances in Water Resources 30
(2007) 17851796.
[34] X. Litrico, V. Fromion, Simplied modeling of irrigation canals for
controller design, Journal of Irrigation and Drainage Engineering 130 (5)
(2004) 373383.
[35] M. Cantoni, E. Weyer, Y. Li, S.K. Ooi, I. Mareels, M. Ryan, Control of large-scale
irrigation networks, Proceedings of the IEEE 29 (1) (2007) 7591.
[36] J. Hu, C. Bohn, H.R. Wu, Systematic Hinf weighting function selection and its
application to the real-time control of a vertical take-off aircraft, Control
Engineering Practice 8 (3) (2000) 241252.
[37] R.W. Beaven, M.T. Wrigth, D.R. Seaward, Weighting function selection in the
Hinf design process, Control Engineering Practice 4 (5) (1996) 625633.

V. Feliu-Batlle et al. / Journal of Process Control 19 (2009) 506519


[38] J.C. Doyle, K. Glover, P.P. Khargonekar, B.A. Francis, State space solutions to
standard H2 and Hinf control problems, IEEE Transaction on Automatic
Control 34 (6) (1989) 831847.
[39] V. Feliu Batlle, R. Rivas Perez, L. Sanchez Rodriguez, F. Castillo Garcia, A. Linarez
Saez, Robust fractional order PI controller for a main irrigation canal pool, in:
Proceeding of 17th IFAC World Congress, Seoul, Korea, 2008.
[40] G.W. Bohannan, Analog fractional order controller in a temperature control
application, in: Proceedings of the 2nd Workshop of the International

519

Federation of Automatic Control on Fractional Differentiation and its


Applications, Porto, Portugal, 2006.
[41] A.J. Clemmens, B.T. Wahlin, Simple optimal downstream feedback canal
controllers: ASCE test case results, Journal of Irrigation and Drainage
Engineering 130 (1) (2004) 3546.
[42] D.C. Rogers, D.G. Ehler, H.T. Falvey, E.A. Serfozo, P. Voorheis, R.P. Johansen, R.M.
Arrington, L.J. Rossi, Canal Systems Automation Manual, 1st rst ed., vol. 2, US
Department of the Interior, Bureau of Reclamation, Denver, 1995.

You might also like