MODELING OF IRRIGATION AND DRAINAGE CANALS

FOR CONTROLLER DESIGN

By J. Schuurmans,1 A. J. Clemmens,2 Member, ASCE, S. Dijkstra,3

A. Hof,4 and R. Brouwer5

ABSTRACT: Schuurmans et al. (1995a) presented a model for the design of water level controllers for irrigation
and drainage canals that describes the essential characteristics of the processes relevant for canal control (such
as water movements and control structures). This paper evaluates the accuracy of this model in two ways: (1)
by comparing its frequency response with a model based on a finite difference approximation of the linearized
St. Venant equations, and (2) by comparing simulation results with data from field experiments. Using these
results, we characterize the accuracy of the model and discuss how these results can be taken into account in
controller design.

INTRODUCTION Under manual control, we include all control methods used

by those who operate the structures, including remote manual
To perform their functions efficiently, irrigation and drain- control. Although operators are able to adapt to unforeseen
age canals must be managed appropriately; that is, the control changes in the behavior of the canal, their performance is lim-
structures need to be operated in such a way that the canals ited by fatigue and loss of concentration, to compensate for
execute their functions adequately. In most cases, this require- which they need to work in shifts. Furthermore, one operator
ment can be translated into a water level control problem. can operate a limited number of control structures at a time.
The main function of an irrigation canal is to deliver water As a result of these factors, the costs of operation are relatively
in an accurate and flexible way (Burt 1987). The delivery is high.
said to be accurate if the actual supply matches the intended Gates that function as automatic water level controllers have
supply, and it is said to be flexible if the delivery meets the been developed for the control of (1) the water level just up-
changing water requirements of the users. This main function stream of the gate (for instance, weirs), (2) the water level just
can be translated into a water level control problem consisting downstream of the gate, and (3) the difference in water level
of two parts. between the upstream and downstream sides of the gate
First, the water levels in the canals located just upstream of (Ankum 1995). A control system that involves only gates is
the offtakes and control structures need to be controlled within simple and reliable. However, the technical possibilities are
a sufficiently small range. A prerequisite for allowing the flow limited, for only the water levels in the close vicinity of the
rates through the offtakes and control structures to be main- gates can be controlled. Consider for instance the AVIS gates
tained at any desired rate (that is, with sufficiently small var- (Goussard 1987). The AVIS gate controls the water level just
iations around the desired rate) is that the water levels at the downstream of the gate within fixed limits. Therefore, a canal
aforementioned locations are kept within a well-defined range. with AVIS gates can deliver water in an accurate and flexible
Second, the water levels preferably should be controlled by way. However, the offtakes in existing canals are usually lo-
adjusting the control structures located at the upstream ends cated at the downstream ends of the reaches, and the canal
of the reaches of the canal. This requirement guarantees that embankments are parallel to the bottom. In order to apply the
the delivery matches the demand; for instance, if a water level AVIS gates, the offtakes need to be relocated at the upstream
in one reach approaches the maximum allowable level, the ends of the reaches and the embankments need to be raised.
inflow is apparently higher than necessary, and thus the inflow The costs involved in adapting an irrigation canal so that AVIS
(the flow through the control structure at the upstream end of gates can be applied are relatively high in developed countries;
the reach) should be decreased. One control strategy that sat- application of these gates is therefore often not economically
isfies this latter requirement is the downstream control strategy feasible (Burt 1987).
(Burt 1987), in which a control structure is adjusted based on Since the 1960s, electronic devices have been used to con-
downstream information. trol water levels. They offer the possibility of automating the
Several possibilities exist to realize a water level control operation of control structures without having to adapt the
strategy: (1) manual control, (2) automatic gates, and (3) au- ‘‘hardware,’’ such as the canal embankments and offtake struc-
tomatic electronic devices. tures. In the following, we shall refer to the logic of electronic
1 control devices as controllers.
Res. Assoc., Univ. of Twente, Dept. of Electr. Engrg., P.O. Box 217,
7500 AE Enschede, The Netherlands. (Paper written when writer was Practicing control engineers of one company or institution
with Delft Univ. of Technol., Dept. of Civ. Engrg., The Netherlands.) usually apply one specific type of controller. Examples are
2
Supervisory Researcher and Hydr. Engr., U.S. Water Conservation ELFLO (Rogers et al. 1995), BIVAL (Chevereau and Schwartz-
Lab., 4331 East Broadway, Phoenix, AZ 85040.
3
Bemeth 1987), and Dynamic Regulation (Rogier et al. 1987).
Asst. Prof., Dept. of Mech. Engrg., Delft Univ. of Technol., P.O. Box The parameters of these controllers are usually tuned during
5048, 2600 GA Delft, The Netherlands.
4 operation or by trial and error on a simulation model. Fur-
PhD Candidate, Dept. of Civ. Engrg., Delft Univ. of Technol., P.O.
Box 5048, 2600 GA Delft, The Netherlands. thermore, the design of these controllers has been mainly heu-
5
Prof., Dept. of Civ. Engrg., Delft Univ. of Technol., P.O. Box 5048, ristic.
2600 GA Delft, The Netherlands. A systematic design method for controllers requires a suit-
Note. Discussion open until May 1, 2000. To extend the closing date able model of the processes most relevant for control. These
one month, a written request must be filed with the ASCE Manager of include, among others, canal behavior (resulting from water
Journals. The manuscript for this paper was submitted for review and
movements), control structures, and actuators.
possible publication on October 30, 1997. This paper is part of the Jour-
nal of Irrigation and Drainage Engineering, Vol. 125, No. 6, November/ It is well known that the flow through a canal with control
December, 1999. qASCE, ISSN 0733-9437/99/0006-0338–0344/$8.00 structures can be described quite accurately by the St. Venant
1 $.50 per page. Paper No. 16867. equations in combination with algebraic discharge equations
338 / JOURNAL OF IRRIGATION AND DRAINAGE ENGINEERING / NOVEMBER/DECEMBER 1999
for the flow through the structures (such as offtakes and con-
trol gates). Many researchers propose controllers based on a
nonlinear model of the canal system. Examples are methods
based on solving the St. Venant equations backwards; that is,
the controllers compute the control actions by inverting the
discretized St. Venant equations (Wylie 1969; Falvey and Lun-
ing 1979; Chevereau 1991; and Liu et al. 1992). These con-
trollers may show numerical instabilities, making them unre-
liable (Bautista et al. 1995). Apart from this problem, it is FIG. 2. Backwater Profile
difficult to get insight into the influence of model parameter
variations on closed loop behavior. To clarify this, consider
In the following, we present control-oriented models for
the effect of an uncertain resistance coefficient (which is used
these processes.
in the St. Venant equations): How does the controlled system
behave if this coefficient changes 10% from its assumed
Water Movements
value? With a complicated nonlinear model, one is not able to
answer this latter question until one has performed an exten- Schuurmans et al. (1995a) analyzed and modeled the re-
sive simulation analysis. sponse of water levels to changes in flow (water movements).
To obtain simple, reliable control systems, systematic design Here, we summarize the results described in that paper.
methods have been developed in linear control theory (Åström The water profile in controlled canals is usually a backwater
and Wittenmark 1984; Doyle et al. 1992). These design meth- profile, as shown in Fig. 2. We refer to the part of the reach
ods require a model that consists of linear ordinary differential affected by backwater as the backwater part, and the part un-
equations or linear difference equations. It is clear that line- affected by backwater as the part with uniform flow.
arized canal models are by no means as accurate as the non- From the characteristics of the St. Venant equations, it is
linear models (based on the Saint Venant equations). However, known that a disturbance, created somewhere in a reach, re-
it appears that in most control applications the model just sults in two wave movements. Defining c as the critical celer-
needs to capture the ‘‘essential’’ dynamics in order to serve ity and V as the mean velocity, one wave travels with velocity
as a basis for the design of simple, reliable controllers. We V 1 c and one travels with velocity V 2 c (Abbott 1979;
believe that in many cases the essential canal dynamics can French 1986). During travel, the waves deform. Waves that
be captured by a simple linear model. travel with velocity V 2 c damp essentially after traveling over
This paper evaluates the accuracy of a simplified model of a certain distance. How quickly waves deform depends on the
the water movements presented by Schuurmans et al. (1995a). flow conditions. In parts with uniform flow (where the depth
The results of these evaluations are useful for controller de- of flow is practically at normal depth), the waves deform con-
sign. The paper is structured as follows: First, we briefly sum- siderably faster than in backwater-affected parts. An explana-
marize the results described in Schuurmans et al. (1995a) and tion for this is that in the backwater part the mean velocity is
extend the water movements model with models of other rel- relatively small, and thus there is less friction from the channel
evant parts of a canal system: the control gates and signal boundary; according to the St. Venant equations, the friction
converters. Then the accuracy of the model is evaluated from from the channel boundary is the only cause for damping.
the response of a different model and field data. Using these When a wave arrives at a boundary (a control structure, for
results, we characterize how the model error depends on the instance), part of the wave is reflected. As a consequence, in
model inputs and canal physics and briefly discuss how these leveled reaches that are sufficiently short, waves can travel up
results can be used for controller design. Finally, we present and down several times before they damp out. If in such a
the conclusions. reach (with length L) a wave is initiated from one of the
boundaries, it returns after a period (Tr), which can be ap-
SIMPLIFIED CANAL SYSTEM MODEL proximated by
Fig. 1 shows some of the most relevant processes for down- L L
Tr ' 1 (1)
stream control of a canal. The controller manipulates the ac- V1c c2V
tuators that move the control gates (i.e., a control action).
Other structures may be used as well, but gates are most com- Suppose that the flow rate at the upstream boundary (Q(0, t))
mon. Sensors measure water levels and some of the load dis- would vary in a sinusoidal way with period Tr, while the flow
turbances, such as offtake flows. The measured signals are rate at the downstream boundary (Q(L, t)) would remain con-
sampled, that is, converted into digital signals by analog-to- stant, for instance, according to
digital (AD) converters before they enter the control system.
The digital control signals from the controller are converted
into analog signals by a digital-to-analog (DA) converter.
Q(0, t) = sin S D
2kp
Tr
t 1 Q0, Q(l, t) = Q0 (2)

with Q0 the initial flow rate. Then waves would add to each
other, causing a standing wave, or, in other words, resonance.
The resonance frequencies (in radians) can be approximated
by
2pk
vr(k) = , k = 1, 2, 3, . . . (3)
Tr
Hence, there are infinitely many resonance frequencies, each
a multiple of the lowest one.

Finite Difference Model

FIG. 1. Block-Diagram Showing Some of Most Important Pro- One way to obtain a linear model of the water movements
cesses in Controlled Canal is to approximate the partial differentials in the linearized St.
JOURNAL OF IRRIGATION AND DRAINAGE ENGINEERING / NOVEMBER/DECEMBER 1999 / 339
Venant equations by finite differences, whereby the canal is The parameter k is given by

U
divided into nodes and branches, and for each branch the dif-
ferentials are approximated by finite differentials. This mod- 4 P0 dR
k=11 (7)
eling technique is described in more detail in Balogun et al. 3 B0 dY 0

(1988); Reddy (1990); and Malaterre (1994). We shall refer to where B is the flow top width (m), P the wetted perimeter (m),
such a model as a finite difference (FD) model. If the number R the hydraulic radius (m), and Y the depth of flow (m). The
of nodes chosen is sufficiently high, the FD model can be very subscript zero refers to steady-state conditions.
accurate for ‘‘small’’ variations of its inputs around their Eqs. (5) and (6) ‘‘assume’’ that waves travel with velocity
steady-state values. However, the FD model contains many V0(1 1 k)/2. In many cases, this velocity is 1–2 times the
parameters whose importance is unclear, such as how the mean velocity V0, which is usually considerably lower than V0
model’s response changes with changing model parameters. (A 1 c0. Hence, the effect of wave deformation is taken into
parameter is defined here as a physical parameter, such as account by assuming a ‘‘lower’’ velocity of the wave than the
Manning’s coefficient, the length of a pool, or its bed slope.) theoretical velocity V0 1 c0.
For instance, the FD model contains Manning’s coefficient, If the whole reach is affected by backwater (Lu = 0), the
which in leveled reaches can often be changed 100% without model given by (4) can be applied. If the backwater affects
affecting the model’s largest time constant (its main dynam- only part of the reach, the flow q(Lu, t) is related to the up-
ics). Manning’s coefficient does change the travel time of grav- stream inflow q(0, t) via (5). Hence
ity waves, but in leveled reaches these waves hardly affect the
largest time constant. The model described in Balogun et al. q(Lu, t) = q(0, t 2 T(Lu))
(1988) is an example. It already contains 18 parameters when
This means that the inflow at x = Lu at time t is taken to be
four nodes are used. Therefore, only the main characteristics
equivalent to the inflow q(0, t), delayed with T(Lu) time units.
of the dynamics of water movements were identified and de-
From here on, we shall ignore the dependence of T on x and
scribed by the model presented in Schuurmans et al. (1995a).
speak of T instead of T(Lu).
Integrator Delay Model The parameters of this model (A and T) can be computed
rather accurately from the geometry of the channel and the
In the backwater part, the dynamics are complicated: waves steady-state values of the flow rate(s) and water depths. If the
move up and down and reflect against the boundaries. How- geometry changes over the distance, it may be necessary to
ever, at low frequencies, the water level ‘‘integrates’’ flow var- compute the parameters (A and T) for small parts and sum
iations in the backwater part. In other words, the backwater them thereafter. If geometrical data are lacking or hard to ob-
part can be considered to behave as an integrator or reservoir tain, it may be more practical to obtain the parameters by
for ‘‘low’’ frequencies. Referring to Fig. 3, the dynamics in fitting the model’s response to measurements: using measure-
the backwater part are approximated by ments of the flow rates at the boundaries, the model’s response
dh(x, t) can be fitted to measured water levels by adjusting the param-
A = q(Lu, t) 2 q(L, t) (4) eters.
dt
where h is the variation of water level or depth of flow around Control Structures and Actuators
steady state (m), q the variation of flow rate around steady The flow through control structures can usually be described
state (m3/s), A the value of surface area of the backwater part quite accurately by an algebraic nonlinear discharge relation,
(m2), Lu the length of the part with uniform flow (m), x the based on the Bernoulli equation and a mass balance; the dis-
distance from the upstream boundary, and t the time (s). Here- charge relations can often be described as
after, we shall speak of ‘‘the water level h’’ and ‘‘the flow rate
q’’ to refer to variations of these variables around their steady- Q = Q(Ug, H1, H2) (8)
state values.
In the part with uniform flow, waves with velocity V 2 c where Ug is the adjustable variable, such as the position of a
dampen relatively quickly (Schuurmans et al. 1995a). There- gate, H1 the water level upstream of the structure, H2 the water
fore, the influence of these waves is neglected, and it is as- level downstream of the structure, and Q the flow rate (m3/s).
sumed that only upstream flow changes can influence the flow. Linearization of this equation for a change in flow rate q gives
To describe the effect of the upstream inflow q(0, t) on the q = cgug 1 c1h1 1 c2h2 (9)
downstream flow in a part with uniform flow, a ‘‘pure’’ delay
model is assumed, as follows: with h1, h2, q and ug the variations of H1, H2, Q, and Ug around
their steady-state values, respectively, and with cg, c1, and c2
q(x, t) = q(0, t 2 T(x)) (5) given by
with T(x) given by Schuurmans et al. (1995a):

T(x) =
2x
(6)
cg = S D
­Q
­Ug 0
; c1 = S D
­Q
­H1 0
; c2 = S D
­Q
­H2 0
(10a–c)

(1 1 k)V0 The water levels h1 and h2 can be obtained as outputs of a

water movement model, while q is an input for the water
movement model.
Although there are many types of actuators, we consider
one of the most common actuators here: an electromotor drives
the position (Ug) of the control gate with a constant speed (Vm)
during a desired time (am). This may seem to limit the possi-
bilities of control, but the gate speed can be pulse modulated;
keeping this in mind, the speed of gate movement can be ap-
proximated by Papageorgiou and Messmer (1985):
dUg
' Vmam, uamu # 1 (11)
FIG. 3. Meaning of Notations dt

340 / JOURNAL OF IRRIGATION AND DRAINAGE ENGINEERING / NOVEMBER/DECEMBER 1999

This model shows a nonlinearity, since there is a maximum
gate speed (uamu # 1). In reality, the relation from the activation
time (am) to the position (Ug) shows even more nonlinearities.
First, there is a minimal gate movement or dead zone to avoid
inexecutable small gate changes. Second, there is hysteresis:
If the actuator reverses movement, the gate movement does
not immediately ‘‘follow.’’ This can be caused by loose con-
nections in the actuator (i.e., gear lash) or friction between the
gate and the construction. Third, the gate can be moved within
a limited range only. FIG. 4. Longitudinal View of Reaches One and Two during Ini-
For controller design, the actuator model as given above can tial Conditions. Vertical Scale is Different from Horizontal Scale
be used, but without the limitation on the activation time. to Show Details

Converters TABLE 1. Parameter Values of Test Canal One

The DA conversion is usually performed by a zero-order- 100 m (reach # 1),

hold (ZOH) mechanism. In this mechanism, the digital signal Length of reaches, L 1,200 m (reach # 2)
is held constant between two sample times. The modeling of (1) (2)
a continuous time system with a ZOH on the inputs and a Manning’s resistance coefficient, n 0.014 s/m1/3
sampler on the output is done most naturally in the discrete Bed slope, I 0.002 m/m
time domain. There exists a method, known as ZOH sampling, Bottom width, Bb 1.0 m
to transform continuous time linear models (with a ZOH and Side slope, n 1.5 horizontal/vertical
a sampler) into discrete-time models (Åström et al. 1984). Note: Cross-sections are trapezoidal.
Here we present the result of this transform, applied to the
model of the water movements, given by (4) and (5). If the
delay time T is exactly an integer t times the sample period model and a finite difference (FD) model for the two upstream
Ts, it can be shown that ZOH sampling of this model gives reaches of Test Canal One and compare the frequency re-
sponses of these models, whereby each model describes the
Ts relation between the flow rates q1 and q2 and the water levels
Dh(k) = [q1(k 2 t) 2 q2(k)] (12) h1 and h2.
A
The models were derived for a steady-state condition in
with k the discrete time level (k = 0, 1, 2, . . .), Dh(k) = h(k) which the flow rate was equal to 0.8 m3/s, while the down-
2 h(k 2 1) the water level change (m), q1 and q2 the inflow stream water levels were at target level (a depth of 0.9 m).
and outflow of the pool, respectively, and t = T/Ts the discrete
time delay. If the delay t is not precisely an integer times the ID Model
sample period Ts, the discrete time model is slightly different.
However, (12) can be used as an approximation whereby the The continuous time ID model is given by
parameter t is obtained by rounding off the quotient T/Tsx to dhi(t)
the nearest integer. Ai = qi(t 2 Ti) 2 qi11(t) 2 di(t) (13)
dt
EVALUATION OF ACCURACY where Ai is the backwater surface of reach i and Ti the delay
time for reach i. These parameters were computed as follows:
In this section we evaluate the accuracy of the model, as The water profile in the backwater part was assumed to be
described above, whereby we focus on the integrator delay given by a horizontal line, while the water profile in the part
(ID) model. We first derive an ID model for a hypothetical with uniform flow was assumed to be parallel to the bottom
canal and compare its frequency response to that of an FD at a depth equal to the normal depth of flow. These assump-
model. The flow in this canal is subcritical, allowing us to tions are reasonable for this particular canal, as can be seen in
evaluate the accuracy of the ID model in ‘‘normal’’ circum- Fig. 4: the water profiles satisfy the description as given above
stances. In particular, we investigate the resonance phenomena rather well. The backwater surface is then determined over the
occurring in this canal in detail. Then we derive a model for part from the downstream end up to the point where the two
an irrigation canal in the Maricopa Stanfield Irrigation and lines intersect. The delay time was computed with (6) using
Drainage District in Arizona and compare this model’s re- the length of the part with uniform flow for x and the normal
sponse with field measurements. depth to compute the mean velocity of flow V and the param-
eter k. This leads to A1 ' 340 m2, A2 ' 756 m2, T1 = 0 s, and
Linear Model of Test Canal One T2 = 528 s. Here, the subscript refers to the reach number.
Test Canal One is a hypothetical irrigation canal, defined by
the ASCE Task Committee on Canal Control Algorithms in FD Model
Kacerek et al. (1995). Although the actual test canal consists The FD model was obtained using the method described in
of eight reaches, we consider only the two upstream reaches Balogun et al. (1988), whereby each reach was divided into
here. The reason for choosing these two reaches is that these 20 nodes.
cover two different possibilities. Reach one is completely af-
fected by backwater, whereas reach two has a part with uni- Comparison of Frequency Responses
form flow, as shown in Fig. 4. In this figure, the variables h1
and h2 are water levels, q1 and q2 are the flow rates through Fig. 5 shows Bode diagrams of the models; the titles ‘‘qi
the control gates, and d1 and d2 are offtake flow rates; recall –> hj’’ indicate the relation from the flow rate (qi) through the
that all variables denote variations around their steady-state ith gate to the water level (hj) in the jth reach. The Bode
value. The geometrical data and other relevant parameters of diagram gives detailed information on the steady response
the two reaches are given in Table 1. (which occurs after the transient response) of a system to si-
In this example, we formulate an integrator-delay (ID) nusoidal signals; since each periodic signal can be written as
JOURNAL OF IRRIGATION AND DRAINAGE ENGINEERING / NOVEMBER/DECEMBER 1999 / 341
 is constant in such a pool. For pools with trapezoidal cross
sections (which usually is the case), the ID model error can
be expected to increase if the backwater surface A differs more
from its nominal value (the value that is assumed in the ID
model). When it concerns a sloping reach, the effects of the
nonlinearities are even more pronounced. For instance, the de-
lay time T can then easily vary 100%. Although in such cases
the ID model is by no means as accurate as a nonlinear model,
it can still be useful for controller design. For instance, it is
possible to link the variations of the flow rate with the delay
time T and the backwater surface A. By doing so, one obtains
a quasilinear model that can be used as a basis for an adaptive
controller, that is, a controller whose settings change with
changing flow rate. Alternatively, the model errors can be
quantified and used in the design of a so-called robust con-
troller, that is, a controller that is stable and performs ade-
quately in the face of model errors. Details on the design of
adaptive and robust controllers (using the ID model) can be
found in Schuurmans (1997). For a practical of these tech-
FIG. 5. Bode Diagrams of Test Canal One niques, see Schuurmans et al. (1997).
So far, the accuracy of the ID model has been evaluated by
comparing its response to that of the FD model. In the next
a sum of sinuses (the Fourier series), the Bode diagram pro-
section, the accuracy of the model is evaluated using field data.
vides full information on the steady response to any periodic
signal. For instance, for the relation between q1 and h1, at a
Experimental Model Verification on WM Canal
frequency of 0.03 rad/s, the magnitude is (according to Fig.
5) approximately 0.15 s/m2, and the phase is approximately The accuracy of a model containing an integrator delay
187 ('0.14 rad). This means that if q1(t) would vary accord- model for the water movements has been verified experimen-
ing to q1(t) = sin(0.03t), the water level would show the steady tally on a lateral irrigation canal in the Maricopa Stanfield
response (after the transient response): h1(t) = 0.15 sin(0.03t Irrigation and Drainage District in Arizona. This canal consists
1 0.14). of eight reaches, with lengths varying from 100 to 2 km. Fig.
From the FD model’s response, resonance can be observed 6 shows a longitudinal view of this canal. The graph is to
in reach one. The resonance peak gain (Rp) is approximately scale, although the vertical scale is different from the horizon-
1 s/m2, while the resonance frequencies are approximately tal scale to show details.
0.07, 0.14, 0.21, . . . rad/s [0.07*k rad/s (k = 1, 2, 3, . . .)]. The WM Canal takes water from the Santa Rosa canal. Off-
These resonance frequencies are in good agreement with the take structures are located at the downstream ends of the
frequencies predicted by (3). To show this, we first compute reaches. Therefore, the water level at the downstream end of
the cross-sectional area (Ac) and the width at surface level (B), each reach is controlled. The canal is currently operated man-
assuming the depth is everywhere equal to the target depth (Yt) ually, but there are plans to operate it automatically (Clem-
of 0.9 m (using data from Table 1): mens et al. 1995). In fact, several closed-loop experiments
have been conducted on this canal; the tests and results of
B = Bb 1 2nYt ' 3.7 m, Ac = (Bb 1 nYt)Yt ' 2.12 m2. these tests are described in Schuurmans and Liem (1995b) and
It follows that c0 = =(gAc /B) ' 2.37 m/s and V0 = Q/Ac ' Liem (1994).
0.38 m/s. Since the length (L) of reach one is 100 m, (3) gives
Linear Model of WM Canal
2pk 2pk
vr(k) ' ' Inspection of the water profiles in the WM Canal, as shown
L L 100 100 in Fig. 6, leads to the conclusion that all reaches but the first
1 1
V0 1 c c 2 V0 0.38 1 2.37 2.37 2 0.38 one have a part with uniform flow. In fact, flow is supercritical
in the parts with uniform flow and shows a hydraulic jump
' 0.07k(rad/s) (14) when entering the backwater part. The integrator delay model
For reach one, where the backwater extends over the entire for this particular canal becomes
length, the ID model is inaccurate in the neighborhhood of the
resonance frequencies, but is rather accurate for lower fre-
quencies. This indicates that the main dynamics (the low fre-
quent dynamics) in leveled reaches are indeed affected by the
backwater surface only. This explains, for instance, why Man-
ning’s coefficient in the FD model can be changed 100% with-
out affecting the main dynamics. In reach two, a large portion
is affected by uniform flow; consequently, no resonance can
be observed in the relation between q2 and h2. In this reach,
the ID model is rather accurate over all frequencies. This again
shows that the main canal dynamics can be described by a
much simpler model than the FD model.
Due to the simplicity of the ID model, it is also possible to
get a grip on the nonlinear effects of the ‘‘true’’ dynamics.
For instance, the dynamics (at low frequencies) of a leveled
canal pool with a square canal cross section can be predicted
rather accurately by the ID model, since the backwater surface FIG. 6. Longitudinal View of WM Canal

342 / JOURNAL OF IRRIGATION AND DRAINAGE ENGINEERING / NOVEMBER/DECEMBER 1999

 dhi(t)
Ai = qi(t 2 Ti) 2 qi11(t) (15)
dt
where hi is the variation of water level number i (m), qi the
variation of flow rate through gate i (m3/s), Ai the backwater
surface of reach i (m2), and Ti the delay time of reach i (s).
The offtake structures are not modeled here, as they were not
used at the time the experiments were conducted. The control
gates were orifices. The flow through these (sharp-crested)
gates was modeled by linearization of
Q = BmUgÏ2g(H1 2 mUg (16)
3
with Q is the flow rate (m /s), B the width under the gate (m),
m the contraction coefficient, Ug the opening height under the
gate, and H1 the upstream water level. For gate i, the linearized
discharge equations can then be written as
q = cg,iui 1 c1,ihi (17) FIG. 7. Water Level Variations (h1–h8) in Response to Gate
Opening Variations (u1–u8). Solid Line: Simulated Levels; Dots:
where ui is the variation of the opening of gate number i (m) Measured Levels
and cg,i and c1,i are constants that follow from linearization and
the initial values of flow variables, such as the flow rate Q0.
The physical parameters, such as the gate widths and resis- These weirs protect the canal from overtopping by passing
tance coefficient, were gathered from design drawings. Using excess flows downstream. The results of these experiments
this data, the parameters Ai were determined by assuming the show that the ID model is accurate if the input changes are
water profile in the backwater part to be horizontal. The pa- relatively small. Indeed, the experiments were performed in
rameters Ti were computed using (6), where x was taken as such a way that the input changes were sufficiently small. In
the length of the uniform part. (All physical parameters of the practice, the input changes may not be small, and therefore
WM canal (geometrical data, roughness coefficients, etc.), as the model’s response may differ substantially from the actual
well as the computer code to generate the ID model parameters response. As we stated before, these errors can be reduced by
Ai and Ti from these data, can be obtained via electronic mail adapting the model parameters on-line, or, alternatively, the
by sending an email to [email protected].) For sharp-crested model error can be quantified and used to design a robust
orifices, the contraction coefficient (m) is usually close to 0.63 controller. Furthermore, the model error can be reduced by
(Bos 1976); therefore, we used m = 0.63. manipulating the flow rates through the control gates as control
inputs instead of the gate positions. By doing so, the nonlin-
earities introduced by the control gates are no longer part of
Experiments and Results
the model. To achieve this, low-level flow rate controllers can
To test the accuracy of the model, five experiments were be used to manipulate the control gates so as to maintain the
conducted that are described here and in Ellerbeck (1995). In flow rates at the desired levels, as in Schuurmans (1997a,
each of these experiments, the openings for one or two gates 1997b).
were changed while the variations of the water levels in two
adjacent reaches were recorded, whereby the gate changes CONCLUSIONS
were made small to avoid nonlinear effects. For comparison,
we simulated the response of the water levels using the model For the design of water level controllers in irrigation and
as described above, whereby the model parameters were com- drainage canals, a simple linear model (the ID model) has been
puted for the steady-state conditions as recorded at the begin- presented in Schuurmans et al. (1995a). This paper has ana-
ning of the experiment. lyzed the accuracy of this model and discussed how the in-
Fig. 7 shows the gate opening variations around their initial accuracies can be taken into account in controller design.
values (u1 . . . u8) and the water level variations (h1 . . . h8) For a hypothetical canal, the accuracy of the ID model was
around the initial levels. The variables u and h are located as evaluated using a finite difference (FD) model as a reference.
shown in Fig. 6: h1 is the water level (deviation) in the first It appeared that, when resonance occurs, the ID model’s ac-
reach, u1 is the gate opening (change) of the most upstream curacy is limited up to the first resonance frequency. Other
gate, and so on. The water level responses of the linear model than that, the frequency responses of both models match well.
are plotted as solid lines, and the measured water level re- The accuracy of the ID model was further verified success-
sponses of the WM Canal are plotted as dots. Two graphic fully using field measurements of an irrigation canal in Ari-
windows next to each other belong to one experiment. Hence, zona. However, the input variations were deliberately kept
the upper graphs belong to one experiment, the next two to small to avoid nonlinear effects.
another experiment, and so on. In the first experiment, gate u2 In practical situations, the input variations may not be small,
was changed, followed by a change to gate u3. Both of these and the model’s response will differ from the actual response:
changes affected water level h2, but only gate u2 affected wa- nonlinearities are not modeled by the ID model (nor by the
ter level h1. The observed water levels are well predicted by FD model). There are two ways in which the nonlinearities
this linear model. can be taken into account in controller design. First, it is pos-
For the remaining experiments, only one gate change was sible to update the model and control parameters on-line; this
made, which affects the water levels both upstream and down- results in an adaptive controller. Second, it is possible to quan-
stream. It can be seen that the response of the linear model is tify the model errors and use these in connection with the ID
reasonably accurate, apart from the simulated water level h5, model to design a robust controller.
which deviates considerably from the measured level after ap- In conclusion, the model is rather accurate under the con-
proximately 75 min. The difference resulted from water flow ditions that (1) input variations are small (around one steady-
over the weirs of the check structure, which was not modeled. state value), and (2) (in leveled reaches) the frequency contents
JOURNAL OF IRRIGATION AND DRAINAGE ENGINEERING / NOVEMBER/DECEMBER 1999 / 343
of the input are concentrated below the (lowest) resonance Ellerbeck, M. B. (1994). ‘‘Model predictive control for an irrigation ca-
frequency. nal,’’ MSc thesis, Lab. for Measurement and Control, Delft, The Neth-
erlands.
Falvey, H. T., and Luning, P. C. (1979). ‘‘Gate stroking.’’ Rep. REC-
ACKNOWLEDGMENTS ERC-79-7, U.S. Bureau of Reclamation, Washington, D.C.
French, R. H. (1986). Open-channel hydraulics. McGraw-Hill, New York.
We would like to thank the water master G. Sloan for this cooperation Goussard, J. (1987). ‘‘NEYRTEC automatic equipment for irrigation ca-
and assistance during the experiments on the WM canal in the Maricopa nals.’’ Proc., Planning, Operation, Rehabilitation and Automation of
Stanfield Irrigation and Drainage District. Furthermore, we would like to Irrigation Water Delivery Sys., ASCE, New York.
thank the following organizations for their financial support: the U.S. Kacerek, T., Clemmens, A. J., and Sanfilippo, F. (1995). ‘‘Test cases for
Water Conservation Lab, the Netherlands Organization for Scientific Re- canal control algorithms with examples.’’ Proc., First Int. Conf. on
search (NWO), the Delft University of Technology (Universiteitsfond Water Resour. Engrg., San Antonio, Tex.
Delft and College van Bestuur) and the Royal Dutch Institute of Engi- Liem, G. R. (1994). ‘‘Controller design for irrigation canals,’’ MSc thesis,
neers (KIvI). Lab. for Measurement and Control, Delft University of Technology,
Delft, The Netherlands.
Liu, F., Feyen, J., and Berlamont, J. (1992). ‘‘Computational method for
APPENDIX. REFERENCES regulating unsteady flow in open channels.’’ J. Irrig. and Drain.
Abbott, M. B. (1979). Computational hydraulics: Elements of the theory Engrg., ASCE, 118(10), 674–689.
of free surface flows. Pitman Publishing Limited, London. Malaterre, P. O. (1994). ‘‘Modélisation, analyse et commande optimale
Ankum, P. (1995). ‘‘Flow control in irrigation and drainage.’’ Rep. No. LQR d’run canal d’irrigation,’’ PhD thesis, Ecole Nationale du Génie
65, Water Mgmt. Dept., Delft, The Netherlands. Rural des Eaux et des Forêts (ENGREF), Paris.
Papageorgiou, M., and Messmer, A. (1995). ‘‘A continuous-time and dis-
Åström K. J., and Wittenmark, B. (1984). Computer controlled systems:
crete-time design of water flow and water level regulators.’’ Automa-
Theory and design. Prentice-Hall International, London.
tica 1985, 21(6), 649–661.
Balogun, D., Hubbard, O. S., and De Vries, J. J. (1988). ‘‘Automatic
Reddy, J. M. (1990). ‘‘Local optimal control of irrigation canals.’’ J.
control of canal flow using linear quadratic regulator theory.’’ J. Hydr.
Irrig. and Drain. Engrg., ASCE, 116(5), 616–631.
Engrg., ASCE, 114(1), 75–102. Rogers, D. C., et al. (1995). Canal systems automation manual, Vol. 2,
Bautista, E., Clemmens, A. J., Strelkoff, T. S. (1995). ‘‘Inverse compu- 1st Ed., U.S. Dept. of the Interior, Bureau of Reclamation, Denver,
tational methods for open-channel control.’’ Proc., First Int. Conf. on Colo.
Water Resour. Engrg., San Antonio, Tex. Rogier, D., Coeuret, C., and Brémond, J. (1987). ‘‘Dynamic regulation
Bos, M. G. (1976). ‘‘Discharge measurement structures,’’ Rep. No. 4, on the Canal de Provence.’’ Proc., Planning, Operation, Rehabilitation
Lab. for Hydr. and Catchment Hydrol., Wageningen, The Netherlands. and Automation of Irrig. Water Delivery Sys., ASCE, New York.
Burt, C. M. (1987). ‘‘Overview of canal control concepts.’’ Proc., Plan- Schuurmans, J. (1997). ‘‘Control of water levels in open-channels,’’ PhD
ning, Operation, Rehabilitation and Automation of Irrig. Water Deliv- thesis, Dept. of Civ. Engrg., Delft University of Technol., Delft, The
ery Sys., ASCE, New York. Netherlands.
Chevereau, G., and Schwartz-Bemeth, S. (1987). ‘‘BIVAL system for Schuurmans, J., Bosgra, O. H., and Brouwer, R. (1995a). ‘‘Open-channel
downstream control.’’ Proc., Planning, Operation, Rehabilitation and flow model approximation for controller design.’’ Appl. Math. Mod-
Automation of Irrigation Water Delivery Sys., ASCE, New York. elling, 19, 525–530.
Chevereau, G. (1991). ‘‘Contribution à l’etude de la regulation dans les Schuurmans, J., and Liem, G. (1995b). ‘‘Controller design for the WM
systemes hydrauliques à surface libre,’’ PhD thesis, Institut National Canal.’’ Proc., First Int. Conf. on Water Resour Engrg., San Antonio,
Polytechnique de Grenoble, Grenoble, France. Tex.
Clemmens, A. J., Strand, R. J., and Schuurmans, J. (1995). ‘‘Application Schuurmans, J., Schuurmans, W., Berger, H., Meulenberg, M., and Brou-
of automatic control of MSIDD WM lateral canal.’’ Proc., First Int. wer, R. (1997). ‘‘Control of water levels in the Meuse river.’’ J. Irrig.
Conf. on Water Resour. Engrg., San Antonio, Tex. and Drainage Div., ASCE, 123(3), 180–184.
Doyle, J. C., Francis, B. A., and Tannenbaum, A. R. (1992). Feedback Wylie, E. B. (1969). ‘‘Control of transient free-surface flow.’’ J. Hydr.
control theory. Macmillan, New York. Div., ASCE, 95(1), 347–361.

344 / JOURNAL OF IRRIGATION AND DRAINAGE ENGINEERING / NOVEMBER/DECEMBER 1999