You are on page 1of 15

Computers & Operations Research 32 (2005) 1337 – 1351

www.elsevier.com/locate/dsw

A chance-constrained multi-period model for a special


multi-reservoir system
M.N. Azaiez∗;1 , M. Hariga, I. Al-Harkan
Industrial Engineering Department, King Saud University, P.O. Box 800, Riyadh 11421, Saudi Arabia

Abstract

We develop a model for optimal multi-period operation of a multi-reservoir system for a basin operating
under a conjunctive use of ground and surface water. The in1ows to the main reservoir as well as the
irrigation demands are stochastic. The ground stock su2ers from severe overdrafts increasing the risk of the
total depletion of the aquifer in addition to the quality degradation and the threat of seawater intrusion. We
treat the uncertainties in the in1ows through chance constraints and penalties of failure to release the planned
amounts of surface water from the main reservoir. However, we re1ect uncertainties in irrigation demands
by opting for de4cit irrigation and using adequate production functions to estimate the expected crop yields.
We attempt in the model to avoid large de4cits except perhaps for periods where crop yields are relatively
insensitive to water shortage. The objective is to maximize the total expected pro4t of the entire region. We
illustrate the model through an example partially based on some hypothetical data.
? 2003 Published by Elsevier Ltd.

Keywords: Reservoir management; De4cit irrigation; Stochastic supply; Stochastic demand; Chance constraints

1. Introduction

The basin of interest operates under a conjunctive use of ground and surface water. Surface water
is imported through a water canal. However, as many regions su2ering from water de4cits are also
connected to the same canal, this canal became unable to ful4ll all demands simultaneously, espe-
cially during dry periods. Therefore, the supply of imported water may be considered as stochastic.
The aquifer su2ers from severe overdrafts increasing the risk of total water depletion and seawater


Corresponding author. Mechanical Engineering Department, King Saud University, IE Program, P.O. Box 800, Riyadh
11421, Saudi Arabia. Tel.: +966-1-4676826; fax: +966-1-4676652.
E-mail address: mazaiez@ksu.edu.sa (M.N. Azaiez).
1
Currently on leave from Ecole SupFerieure de Commerce, Sfax, Tunisia.

0305-0548/$ - see front matter ? 2003 Published by Elsevier Ltd.


doi:10.1016/j.cor.2003.11.009
1338 M.N. Azaiez et al. / Computers & Operations Research 32 (2005) 1337 – 1351

intrusion in coastal areas, in addition to the continuous quality degradation of groundwater. There-
fore, the local water authorities in the region of interest (the decision-makers in this study) plan to
make imported water as the main source of water supply, while reducing considerably groundwater
pumping. This plan is expected to occur for several seasons, until the ground stock is suHciently
replenished from natural and arti4cial recharges. Thus, groundwater withdrawal will be applied only
under emergency bases leading to more water de4cit. Imported water is stored in a main reservoir.
Then, releases are made to local reservoirs, where water is allocated among crops. The local water
authorities decide on the size of surface water releases from the main reservoir as well as the aggre-
gate amount of groundwater pumping (if at all) to be made at each period. Over-irrigation of crops
is avoided and the excess of allocated water to crops if any is lost from the system (say injected in
the aquifer).
The problem is to determine at each irrigation period in a year-plan:

• The size of releases from the main reservoir.


• The size of groundwater pumping for the entire region.
• The allocation of the total quantity of water among local reservoirs.
• The allocation of water among the di2erent crops at each local reservoir.

This work extends the study in [1] from a single-period to a multi-period model. The bene4t of
this extension is obvious as crop yield responses to water de4cit widely vary according the crop
development stages. Consequently, an operating policy should take advantage of this variation so
that low irrigation levels would be allocated to crops in periods of relatively low sensitivity to water
de4cit, while higher irrigation levels should be assigned during critical periods of a crop growth. In
addition, the model gives a higher priority in water allocation to high-value crops.
We organize the remaining of the paper as follows: We present a literature review in Section 2.
Then, we discuss the model in Section 3. Next, we provide an illustration for the applicability of
the model in Section 4. Finally, we o2er some concluding remarks in Section 5.

2. Literature review

There is a large number of published studies in reservoir management and operations models
[2–7], to name only a few. The work in [8] provides a state-of-the-art review of reservoir management
models.
With the exception of few models, such as the one in [9], most of reservoir management models
consider only the supply or the demand as stochastic. This is due to the added complexities by
treating both demand and supply as stochastic. Also, as for the present study, most of the models
developed for reservoir operations for irrigation consider a single decision-maker approach (e.g.,
[2,6,7,10]). The study in [3], however, develops a model for irrigation reservoir and farm manage-
ment with a multi-decision maker approach, namely a reservoir manager and individual irrigation
farmers.
A lot of e2ort has been devoted to the conjunctive use management of ground and surface water
([11–17], to name only a few). Many of the conjunctive use models use simulation combined
with optimization, resulting in the so-called simulation-management models [15,16]. The work in
M.N. Azaiez et al. / Computers & Operations Research 32 (2005) 1337 – 1351 1339

[13] uses lattice programming to characterize the behavior of optimal policies for a conjunctive
use model with arti4cial recharge. The work in [14] develops a nonlinear multi-objective model
where two con1icting objectives were considered, namely cost minimization of water allocation and
minimization of saltwater intrusion in the aquifer. The work in [17] uses convex programming to
formulate the problem of conjunctive use with arti4cial recharge, where the aquifer su2ers from
overdrafts and de4cit irrigation is suggested.
De4cit irrigation is gaining interest in irrigation planning. A number of models have been devel-
oped in the literature, where crops are deliberately under irrigated in order to save water and increase
the irrigated area and possibly the pro4t. The work in [18] develops a multiplicative formula for
crop yield as a function of applied irrigation water. This formula is widely used in de4cit irrigation.
The work in [19] analyzes experimental data on crop yield response to water, and empirically de-
rives yield response factors, Ki . These factors are sensitivity indices for water stress in the speci4ed
growth stage i of the crop. The study in [20] compares three dated water-production function models
and concludes that a simple heuristic multiplicative form is applicable over a wide range of stress
conditions. It also 4nds out that the additive production function is appropriate for a reasonably
large variety of crops. Studies [6] and [21] develop procedures to assess crop yields as a function
of applied water, in terms of the ratio of actual to maximum evapotranspiration.
Applications of stochastic optimization and chance-constrained programming techniques in agri-
culture can be found in a number of studies [9,22–25].
The work in [1] analyzes the problem of the current study only for a single-period. It shows that
the formulated nonlinear stochastic problem is concave with linear constraints. It also proposes an
iterative procedure that generates an optimal operating policy.

3. The model

The main reservoir receives a stochastic quantity of water from the north of the country at the
beginning of each irrigation period. The decision-makers allocate an amount of water to each local
reservoir after possibly allowing for some emergency groundwater withdrawal from the aquifer. Any
amount of water remaining in the main reservoir at the end of a given period will be maintained
(e.g. through chlorinating and inspections to be kept in acceptable quality) for use in the next pe-
riod. The decision-maker will also maintain surplus water at each local reservoir if the obtained
total stock of water exceeds the local demand. This action, however, would be unlikely to hap-
pen due to the constant de4cit of irrigation water in the region. If a shortage occurs at a given
period for a given crop, then this would be re1ected in the yield response of the crop to wa-
ter de4cit for that period. The problem is to identify the optimal operating policy that maximizes
the expected total pro4t of the entire region. An operating policy will specify at each period of a
year-plan:

• The size of water release from the main reservoir to local reservoirs.
• The size of emergency groundwater withdrawal (if any).
• The share of each local reservoir (including possibly some groundwater).
• The irrigation level per hectare assigned to each crop in every local area.
1340 M.N. Azaiez et al. / Computers & Operations Research 32 (2005) 1337 – 1351

Note that irrigation demands of crops are uncertain at the decision phase and it is possible that an
assigned irrigation level exceeds the actual demand. Since it is not bene4cial to over-irrigate crops,
the excess of water should not be applied in irrigation and would be lost. For instance it could be
injected in the aquifer.
The expected pro4t would be the expected total return over the year minus the expected total
cost. The expected total return is the sum over all crops in all local areas of the expected yield of
each crop multiplied by the average price of the crop. The costs considered in this model are:

• Acquisition cost, proportional to the amount of imported water.


• Holding costs at the main and local reservoirs considered as linear.
• Penalty costs for withdrawing groundwater, proportional to the size of withdrawals.
• Penalty costs for failure to release the planned amount of water from the main reservoir as a
result of a shortage in the main reservoir storage, also proportional to the amount of shortage.

3.1. Assumptions

For simplicity in the presentation of the model, we will neglect the losses of water due to evap-
oration and deep percolation during delivery to irrigated areas. We will also neglect transportation
costs of water delivery within the region of interest, as these are minor. In addition, we will assume
no initial stock of water at the di2erent reservoirs. We should mention however that the extension
of this model to account for all the factors above is straightforward.
We will assume that production function of the crop yield in terms of applied irrigation water is
concave and suHciently di2erentiable. Justi4cations of this assumption are based on the real behavior
of a large variety of crops and can be found in several studies including [1,21,26].
Finally, we assume that the in1ows to the main reservoir (respectively the irrigation demand of
each crop) at the di2erent periods of the year are independent random variables.

3.2. Notations

T number of irrigation periods in a year


n number of local reservoirs
wt (random variable) in1ow to the main reservoir during period t, 1 6 t 6 T
Wt (random variable) cumulative in1ow to the main reservoir up to period t, 1 6 t 6 T
Wt expected value of Wt ; 1 6 t 6 T
fWt (:) p.d.f. of Wt ; 1 6 t 6 T
FWt (:) C.D.F of Wt ; 1 6 t 6 T
It ending water stock level in the main reservoir at period t; 0 6 t 6 T .
(Note that I0 = 0)
It; i ending water stock level in reservoir i at period t; 1 6 i 6 n; 0 6 t 6 T .
(Note that I0; i = 0)
C0 unit cost of imported water
ht unit (holding) cost of maintaining the remaining water in the main reser-
voir at the end of period t (to be used at the next period), 1 6 t 6 T
M.N. Azaiez et al. / Computers & Operations Research 32 (2005) 1337 – 1351 1341

ht; i unit (holding) cost of maintaining the remaining water in reservoir i at


the end of period t (to be used at the next period), 1 6 i 6 n; 1 6 t 6 T
t unit penalty cost for withdrawing groundwater (including pumping cost)
at period t; 1 6 t 6 T
pt unit penalty cost of failure to release the planned amount of water from
the main reservoir as a result of a shortage in the storage level during
period t; 1 6 t 6 T
t predetermined probability of failure to release the planned amount of
water from the main reservoir as a result of a shortage in the storage
level during period t; 1 6 t 6 T
Ji number of crops grown in local area i for the year of interest, 1 6 i 6 n
Ai; j land in hectares of crop j grown in local area i; 1 6 j 6 Ji ; 1 6 i 6 n
Pi; j average unit price of crop j grown in local area i; 1 6 j 6 Ji ; 1 6 i 6 n
Yi;Mj maximum yield per hectare (if full irrigation level is applied) of crop
j grown in local area i; 1 6 j 6 Ji ; 1 6 i 6 n
Yi; j (:) yield function per hectare of crop j grown in local area
i; 1 6 j 6 Ji ; 1 6 i 6 n
Ki; j; t yield response factor for period t of crop j grown in local area
i; 1 6 j 6 Ji ; 1 6 i 6 n; 1 6 t 6 T
AETi; j; t (:) actual evapotranspiration (for a given level of irrigation) at period t of
crop j grown in local area i; 1 6 j 6 Ji ; 1 6 i 6 n; 1 6 t 6 T
PETi; j; t potential evapotranspiration (for a given level of irrigation) at period t
of crop j grown in local area i; 1 6 j 6 Ji ; 1 6 i 6 n; 1 6 t 6 T
Di; j; t (random variable) irrigation demand at period t of crop j grown in local area
i; 1 6 j 6 Ji ; 1 6 i 6 n; 1 6 t 6 T
fi; j; t (:) p.d.f. of Di; j; t
st (decision variable) amount of released water from the main reservoir at period t; 1 6 t 6 T
St (decision variable) cumulative amount of released water from the main reservoir up to
period t; 1 6 t 6 T
gt (decision variable) amount of emergency withdrawal of groundwater at period t; 1 6 t 6 T
Gt (decision variable) cumulative amount of emergency withdrawal of groundwater up to pe-
riod t; 1 6 t 6 T
Rt; i (decision variable) cumulative amount of allocated water to reservoir i up to period
t; 1 6 t 6 T; 1 6 i 6 n (from the main reservoir and possibly sup-
plemented by groundwater)
IRt; i; j (decision variable) irrigation level per hectare allocated to crop j in local area i at period
t; 1 6 j 6 Ji ; 1 6 i 6 n; 1 6 t 6 T
The situation is depicted in Fig. 1.

3.3. Model formulation

3.3.1. Model constraints


Note that for every t; 1 6 t 6 T , the ending stock level at the main reservoir at period t is the
ending stock level at period t − 1 plus the in1ow minus out1ow during period t. It is therefore given
1342 M.N. Azaiez et al. / Computers & Operations Research 32 (2005) 1337 – 1351

w1 w2 w3

I0 I1 I2 I3
1 2 3

p2
g1 p1 g2 g3 p3

r1 r2 r3

r1,1 r2,1 r3,1


I 0,1 I1,1 I 2,1 I 3,1
1,1 r1, 2 2,1 r2, 2 3,1 r3, 2
I 0, 2 I1, 2 I 2, 2 I 3, 2
1,2 2,2 3,2

o11 o12 o 21 o 22 o 31 o 32

A12 r112 A 21 r121 A 22 r122


A11 r111 A13 r113

Fig. 1. Graphical representation of the reservoirs system.

by
It = I t − 1 + w t − s t ; 1 6 t 6 T: (1)
The size of release should not exceed the storage level at the main reservoir for a given period t.
This is given by
st 6 It −1 + wt ; 1 6 t 6 T: (2)
To avoid keeping track of the level of the reservoir at each period, we can easily show that (2) can
be equivalently substituted by
St 6 Wt ; 1 6 t 6 T: (3)
Note that (3) is a chance (soft) constraint as the cumulative in1ow, Wt , is not controlled. Therefore,
we may change (3) by the following constraint:
Pr(St 6 Wt ) 6 1 − t ; 1 6 t 6 T: (4)
In the expression given by (4), Pr stands for probability, and t is a predetermined constant denoting
the acceptable risk-level set by the decision-makers to violate the chance constraint.
To keep the size of the problem as small as possible, we may ignore the decision variables st
and wt in the formulation, as these variables can be easily obtained from the cumulative variables
St and Wt .
M.N. Azaiez et al. / Computers & Operations Research 32 (2005) 1337 – 1351 1343

We also have the hard constraints:

S1 ¿ 0; St ¿ St −1 ; G1 ¿ 0; Gt ¿ Gt −1 ; 2 6 t 6 T: (5)

In addition, we have the balance constraints stating that the total allocated water at each period
(released water from the main reservoir plus possibly pumped groundwater) equals the total in1ow
to the di2erent local reservoirs. These constraints can be equivalently formulated as
n

Rt; i − St − Gt = 0; 1 6 t 6 T: (6)
i=1

At each period, and for each local reservoir, the irrigation water to be allocated among crops cannot
exceed the storage level of the reservoir. These hard constraints can equivalently be formulated as
Ji
t 

Rt; i − Ai; j IRi; j; t  ¿ 0; 1 6 t 6 T; 1 6 i 6 n: (7)
t  =1 j=1

Finally, we have the following (hard) constraints:

Rt; i ¿ Rt −1; i ; 1 6 i 6 n; 2 6 t 6 T; (8)

IRi; j; t ¿ 0; 1 6 i 6 n; 1 6 j 6 Ji ; 1 6 t 6 T: (9)

3.3.2. Cost function


We have the following components of the expected total cost over a year of operation:

1. Expected cost of imported water is C0 WT .  ∞


2. Expected holding cost at the main reservoir is Tt=1 ht St (Wt − St )fWt (Wt ) dWt .

3. Penalty cost for pumping groundwater is Tt=1 (t − t+1 )Gt , where T +1 = 0.
4. Expected penalty cost for violating the chance constraints is
 T  St
pt (St − Wt )fWt (Wt ) dWt :
t=1 0

Note that this cost can be interpreted as the cost it takes to supplement for the shortage St − Wt ,
in case of insuHcient storage level It . For the basin of study, the only alternative source for
imported water is groundwater. Therefore,
 we willbe taking  i pt =  t .
5. Holding cost at local reservoirs is Tt=1 ht; i (Rt; i − tt  =1 Jj=1 Ai; j IRi; j; t  ). Therefore, the expected
total cost over a year, ETC(St ; Gt ; Rt; i ; IRi; j; t ; 1 6 i 6 n; 1 6 j 6 Ji ; 1 6 t 6 T ), is the sum of all
costs above.

It is not diHcult to show the following result:

Proposition 3.1. The expected total cost, ETC, is convex.


1344 M.N. Azaiez et al. / Computers & Operations Research 32 (2005) 1337 – 1351

3.3.3. Return function


Assuming that all input factors other than water are kept at their optimum levels (water is con-
sidered as the only limiting factor), the yield function of a given crop will depend on the irrigation
level assigned to the crop as well as the irrigation demand of the crop. It is shown in a number
of studies, that the yield function in terms of applied irrigation is a concave function. It can be
approximated by
  
T
AETi; j; t (IRi; j; t ; Di; j; t )
M
Yi; j (IRi; j; t ; Di; j; t ; 1 6 t 6 T ) = Yi; j 1 − Ki; j; t 1 − :
t=1
PET i; j; t

In this expression, Ki; j; t is the yield response factor of crop j at period t, re1ecting the sensitivity
of the crop to water shortage at this period. The work in [18] provides yield response factors of a
large variety of crops at their di2erent growth stages.
Also, PETi; j; t is the potential evapotranspiration of crop j at period t, when full irrigation level
is applied. Finally, AETi; j; t is the actual evapotranspiration of crop j at period t, when IRi; j; t is the
irrigation level applied, and Di; j; t is the actual irrigation demand. We have the following properties:
AETi; j; t (IRi; j; t ; Di; j; t ) is non-decreasing in IRi; j; t concave and suHciently di2erentiable with
AETi; j; t (IRi; j; t ; Di; j; t )
06 6 1;
PETi; j; t
AETi; j; t (IRi; j; t ; Di; j; t )
=1 for IRi; j; t ¿ Di; j; t :
PETi; j; t
Note that in case the allocated irrigation level IRi; j; t exceeds the actual demand Di; j; t , the applied
level would be exactly Di; j; t , while the remaining water would be lost from the system (say injected
in the aquifer) to avoid the negative e2ects of over-irrigating crops.
The expected total return is given by

ETR(IRi; j; t ; 1 6 i 6 n; 1 6 j 6 Ji ; 1 6 t 6 T )
Ji
 

n  T
 ∞
AETi; j; t (IRi; j; t ; Di; j; t )
M
= Ai; j Pi; j Yi; j 1 − Ki; j; t 1 − fi; j; t (Di; j; t ) dDi; j; t :
i=1 j=1 t=1 IRi; j; t PETi; j; t

From the assumptions above, it is not diHcult to show the following result:

Proposition 3.2. ETR is a concave function.

The chance-constrained optimization program is

max P = ETR − ETC


s:t: (4) to (9): (10)

Now set t = FW−t1 (t ), where FW−t1 (:) is the inverse function of the C.D.F of Wt , for all 1 6 t 6 T .
Then, the deterministic equivalent of (10) is the same optimization program (10), except that the
M.N. Azaiez et al. / Computers & Operations Research 32 (2005) 1337 – 1351 1345

chance constraints (4) are replaced by (4) , where (4) are given by

St 6 t ; 1 6 t 6 T: (4 )

In other words, the deterministic equivalent of (10) is (11), given by

max P = ETR − ETC


s:t: (4) and (5) to (9): (11)

It is clear from Propositions 3.1 and 3.2 that (11) is a concave program with linear constraints.
Therefore, any local optimum of (11) is a global optimum.
Concerning the size of the problem, reasonably T shall not exceed 26, considering biweekly
irrigation decisions in order to incorporate all seasonal and annual crops, and with overlapping growth
stages. In this case, assuming 10 local areas, each with 10 crops, the problem will have 2912 decision
variables and 3014 constraints (including 2600 non-negativity constraints for the irrigation levels of
crops that are not needed as optimality will force these levels to take non-negative values). In any
case, the size of the problem is considered as moderate and can be eHciently solved using Lingo
8.0. Therefore, the problem is within the computational capacities of existing commercial packages.

4. Illustration

To illustrate the model applicability, we will use the same illustration as in [1]. However, we will
analyze the problem from a dynamic framework rather than a static approach.
We consider three local areas (n = 3), each having a total land of 1000 hectares (ha). We assume
that the selected crops to be grown in these areas are given in Table 1.
We consider 5 periods in a year (i.e., T = 5), corresponding to the following 4ve growth stages:
establishment stage, vegetative stage, 1owering stage, yield formation stage, and ripening stage. The
yield response factors of the di2erent crops in each of the growth stages are given in the following
table (Table 2).
Note the higher the yield response factors, the more sensitive the crop to water de4cit.

Table 1
Distribution of selected crops in the three local areas

Crop Area (ha)

1 2 3

Sorghum 600 300 —


Wheat 400 — 500
SaPower — 700 500
Total 1000 1000 1000
1346 M.N. Azaiez et al. / Computers & Operations Research 32 (2005) 1337 – 1351

Table 2
Yield response factors of the crops

Crop Yield response factors

t=1 t=2 t=3 t=4 t=5

Sorghum 0.2 0.2 0.55 0.45 0.2


Wheat 0.2 0.2 0.65 0.55 0.2
SaPower 0.2 0.3 0.55 0.65 0.2

Table 3
In1ow and unit demand distributions

Distribution t=1 t=2 t=3 t=4 t=5

In1ow, wt Distribution N (300; 150) N (700; 300) N (900; 300) N (600; 300) N (200; 80)
Mean 300 700 900 600 200
Variance 22 500 90 000 90 000 90 000 6400

Demand for sorghum Distribution G(2; 40) G(2; 20) G(2; 6:67) G(2; 6:67) G(2; 40)
Mean 0.05 0.1 0.3 0.3 0.05
Variance 0.035 0.07 0.21 0.21 0.035

Demand for wheat Distribution G(2; 20) G(2; 20) G(2; 5) G(2; 6:67) G(2; 20)
Mean 0.1 0.1 0.4 0.3 0.1
Variance 0.07 0.07 0.28 0.21 0.07

Demand for saPower Distribution G(2; 20) G(2; 10) G(2; 4) G(2; 5) G(2; 20)
Mean 0.1 0.2 0.5 0.4 0.1
Variance 0.07 0.14 0.35 0.28 0.07

4.1. The demand and supply distributions

The amount of water to be imported from the north to the region of interest depends on rainfalls
in the north as well as on the demand of water in the other regions that are connected to the canal.
As in [1], we will assume that the stochastic in1ow, wt , to the main reservoir for each period t
can adequately be represented by a normal distribution. We will also use the gamma distribution
to represent demands for irrigation water in the three local areas considered in this illustration. The
details are provided in Table 3 above, where all volumes are in 103 m3 , all demands are per hectare.
In addition, G(.; .) stands for the gamma distribution.

4.2. The costs

We will assign the same costs as in [1]. Therefore, the holding costs (for maintaining water for
the next period) are 25 monetary units (MU) in the main reservoir and 30 MU in each of the local
reservoirs per thousand of cubic meters. The purchasing cost is 90 MU per thousand of cubic meters.
M.N. Azaiez et al. / Computers & Operations Research 32 (2005) 1337 – 1351 1347

Table 4
Yield data of the crops

Crop "t

t=1 t=2 t=3 t=4 t=5 YM

Sorghum 0.6 0.6 0.3 0.4 0.6 16


Wheat 0.7 0.7 0.4 0.4 0.7 7
SaPower 0.8 0.75 0.4 0.3 0.75 12

We will analyze the problem for a number of shortage costs ranging from 100 to 1; 000 MU per
thousand of cubic meters.

4.3. The returns

We also set the returns in MU per ton to be, respectively, 150, 200, and 300 for sorghum, wheat,
and saPower.

4.4. The yield functions

Among the alternative yield functions displayed in the literature, we select the additive yield
function given in Section 3.3.3. This yield function provides reasonable approximations of the real
expected yields. See for instance [20].
The fraction of the actual to the maximum evapotranspiration depends on the irrigation level
to be applied compared with the irrigation demand of the crop. This dependence is concave and
approximately linear (see for instance [1,20,21]). If no irrigation water is applied for a given period t,
then AET will result from rainfalls only and the fraction of AET=PET will have some value "t (with
0 ¡ "t ¡ 1) that can usually be estimated empirically. For the purpose of this illustration, we will
assign values for "t arbitrarily (but re1ecting the sensitivity of each crop to the shortage in arti4cial
irrigation). If, however, enough water is applied to satisfy full water requirement of the crop, then
AET will simply coincide with PET . Therefore, the fraction AET=PET will be approximated by

(1 − "t )IRt
AETt  + "t if IRt 6 Dt ;
= Dt
PETt 
1 otherwise:
The data for "t and the maximum yield in tons per hectare, Y M are given by Table 4.
The probability of violating the chance constraints is selected to be 10%. The optimal policy
for selected penalty costs are given in the following table where GW stands for groundwater, SW
for surface water, and allocation water for crops (in 103 m3 per hectare) is given in vectors of 5
components representing, respectively, the 5 irrigation stages.

4.5. Analysis of the results

The optimal operating policy is displayed in Table 5 for several penalty values for groundwater
pumping. For one extreme value ( = 100 MU), comparable to the unit cost of imported water,
1348 M.N. Azaiez et al. / Computers & Operations Research 32 (2005) 1337 – 1351

Table 5
Optimal operating policies for selected values of penalty costs

 Pro4t Total GW Total SW Water allocation


(MU) (MU) pumped released
Sorghum Wheat SaPower

100 6 639 435 2066 2000 [0.09,0.13,0.44,0.40,0.10] [0.09,0.09,0.43,0.36,0.15] [0.12,0.18,0.57,0.68,0.14]


400 5 958 653 1736 2000 [0.08,0.12,0.42,0.35,0.09] [0.08,0.08,0.39,0.31,0.09] [0.10,0.19,0.61,0.59,0.11]
600 5 692 660 1275 2000 [0.07,0.10,0.37,0.32,0.07] [0.08,0.07,0.34,0.29,0.06] [0.09,0.16,0.50,0.55,0.10]
800 5 454 413 1174 2000 [0.07,0.10,0.35,0.30,0.06] [0.08,0.07,0.38,0.26,0.05] [0.09,0.16,0.53,0.50,0.08]
1000 5 312 136 884 2000 [0.07,0.10,0.40,0.30,0.06] [0.05,0.05,0.29,0.23,0.04] [0.07,0.12,0.49,0.44,0.08]

Table 6
Behavior of pro4t and pumping size as the penalty cost increases

Penalty cost,  100 400 600 800 1000

% of expected pro4t 100 90 86 82 80


% of pumping level 100 84 62 57 43

no de4cit is expected to occur as the ground stock is large enough compared with the irrigation
requirement of one year. For the other extreme (=1000 MU), some important de4cit is observed and
re1ected in the small size of groundwater pumping. This large penalty cost is beyond what growers
can a2ord to pay for irrigation water. Clearly, as the penalty cost increases, the level of pumping
decreases and the expected pro4t also decreases. It is also important to notice that the surface water
assigned is always the same and coincides with the upper limit imposed by the con4dence level
for meeting the chance constraints. Therefore, one can argue for relaxing this level. This will be
discussed later. Now, to analyze the optimal policies with more care we need to display two tables
to investigate the behavior of pro4t and pumping size on one hand and the allocation process of
irrigation levels on the other hand.
In Table 6 below, the expected pro4t and the pumping sizes will be expressed as a percentage of
the case when  = 100 MU.
From Table 6, it is clear that some substantial saving in groundwater resources is possible at some
moderate reductions in expected pro4ts. For instance, about 40% of groundwater can be saved at
the expense of reducing the pro4t by less than 15%. Similarly, about 60% in groundwater saving
can be achieved for a 20% reduction in expected pro4t.
In Table 7 below, we represent irrigation levels allocated to each crop for the di2erent irrigation
periods as a percentage of the mean demand of the crop for the same periods. We use a vector
representation for each crop for the respective irrigation periods.
As the pumping size decreases (or equivalently  increases) less water is allocated to crops. The
critical periods (3 then 4 for sorghum and wheat, and 4 then 3 for saPower) keep preserving the
highest shares, while the periods when crops are less sensitive to water shortage (1,5, and 2) are the
most a2ected by the de4cit. SaPower and sorghum that have higher economic values receive more
share of water. These two crops, regardless of the size of groundwater withdrawal, keep receiving
M.N. Azaiez et al. / Computers & Operations Research 32 (2005) 1337 – 1351 1349

Table 7
Irrigation levels as a percentage of the mean demand

 Sorghum Wheat SaPower

100 [176;130;147;133,193] [93;87;107;121;151] [104;90;114;171;139]


400 [158;119;139;117;179] [81;77;97;102;87] [104;97;122;148;111]
600 [144;101;122;106;131] [75;72;86;98;61] [91;79;100;139;99]
800 [135;96;118;99;127] [77;73;94;85;54] [87;81;106;124;84]
1000 [140;102;132;99;129] [53;52;73;77;45] [67;62;99;109;79]

at least 100% of their mean demand during their respective most critical periods (i.e., periods 3
and 4), as displayed in Table 7. Even for less critical periods, sorghum keeps having approximately
no de4cit. However, wheat can be irrigated even by less than 50% of its average demand.
From the above comments, it is clear that the optimal allocation policies assign water while taking
into account the sensitivity of crops to water de4cit during the di2erent growth stages as well as the
economic values of the crops.
We will next investigate the behavior of optimal policies as we relax the con4dence levels for
meeting the chance constraints.

4.6. Sensitivity analysis of the con9dence levels

There are two types of groundwater withdrawals at each period t: a planned amount (expressed
by gt ) and a supplemental amount for de4cit in the planned release size of the surface water
(expressed by st − wt ). When we select a given con4dence level for not violating chance constraints,
we are assigning upper limits for the size of releases from the main reservoir over the di2erent
periods. The higher this con4dence level, the less surface water we can use, and thus the more
groundwater we plan to pump. However, the supplemental groundwater decreases as the con4dence
level increases. Therefore, we should tradeo2 the pumping sizes for planned against supplemental
groundwater by appropriately selecting the con4dence level. We will brie1y discuss the behavior
of the optimal policies as this con4dence level varies. We will mainly focus on the variation in
pro4t and groundwater pumping. We will estimate the supplemental groundwater through the worst
case scenarios. Thus, for a given con4dence level 1 − , we estimate the supplemental amount of
groundwater to be  times the planned release size from the main reservoir. The optimal policies
discussed above are found for  = 10%. All the analysis below is made by comparison with the
results obtained for this level. When  takes the values of 20% and 30%, respectively, the highest
expected pro4ts (obtained when  = 100 MU) increase roughly by 4%, while the lowest expected
pro4ts (obtained when  = 1000 MU) increase, respectively, by 7.4% and 8.9%. Also, when the
penalty cost for groundwater withdrawal becomes signi4cant (say when  = 800 and 1000 MU),
the planned groundwater is reduced to 32% and 9%, respectively, for  = 20%, and to 20% and
3%, respectively, for  = 30%. In addition, the worst case of total pumped groundwater (planned
and supplemental) is reduced to 50% and 47%, respectively, for  = 20%, and to 70% and 69%,
respectively, for  = 30%. Therefore, it is clear that the selection of the con4dence level (of not
violating chance constraints) plays an important role in the optimal operating policies. The above
analysis suggests relaxing the con4dence level of 90%.
1350 M.N. Azaiez et al. / Computers & Operations Research 32 (2005) 1337 – 1351

5. Conclusions

This study deals with a special situation of a basin operating under conjunctive use of ground and
surface water, where the aquifer su2ers from severe overdrafts as a result of constant water de4cit in
the surface supply of a multi-reservoir system. The impacts of these overdrafts are important as total
depletion of the aquifer may occur in the near future, in addition to the water quality degradation and
the serious threat of seawater intrusion. We suggest (in conformity with the decision-makers plans)
restricting the groundwater use to “emergency withdrawals”, while opting for de4cit irrigation in
order to allow for a suHcient replenishment of the aquifer from natural and arti4cial recharge in the
next several seasons. We investigate the impact of this policy on the optimum expected pro4t of the
entire region as well as the groundwater withdrawal size. We consider the problem in a multi-period
framework, extending the single-period model investigated in [1].
An operating policy consists on determining at each period of a year plan:

• The size of water to release from the main reservoir.


• The size of groundwater pumping.
• The share of each local reservoir from the total amount of surface and ground water made available.
• The irrigation level of each crop at each local area.

We approach this sequential decision problem through a chance-constrained optimization model. We


show that its deterministic equivalent is a concave program with linear constraints and a moderate
size. For illustration, we use the same application as in [1] but in a multi-period context. We
show that some substantial saving in the groundwater resources may be achieved at the expense of
relatively low reductions in expected pro4ts when an eHcient irrigation policy is applied. Such a
policy takes into account the di2erent sensitivities of crops to water shortage in the di2erent growing
stages, as well as the economic value of crops.
We also conduct some sensitivity analysis that provides guidelines on selecting an appropriate
con4dence level for not violating the chance constraints and shows how the optimal policies may
signi4cantly change with this level.

Acknowledgements

This work is supported by the Research Center of the College of Engineering at King Saud
University, research grant No. 419/16. In addition, the 4rst author was partially supported from
a grant by ICARDA. He also extends his thanks to Professor M.S. Mattoussi from “FacultFe des
Sciences Economique et de Gestion, Tunis III” for his assistance.

References

[1] Azaiez MN, Hariga M. A single-period model for conjunctive use of ground and surface water under severe overdrafts
and water de4cit. European Journal of Operational Research 2002;133(3):653–66.
[2] Dudley JN. A single decision-maker approach to irrigation reservoir and farm management decision making. Water
Resource Research 1988;24(5):633–40.
[3] Dudley JN. Volume sharing of reservoir water. Water Resource Research 1988;24(5):641–8.
M.N. Azaiez et al. / Computers & Operations Research 32 (2005) 1337 – 1351 1351

[4] Wang D, Adams BJ. Optimization of real-time reservoir operations with Markov decision processes. Water Resource
Research 1986;22(3):345–52.
[5] AdigVuzel R, Cokunoglu O. A descriptive decision process model for hierarchical management of interconnected
reservoir systems. Water Resource Research 1984;20(7):803–11.
[6] Vedula S, Mujumdar PP. Optimal reservoir operation for irrigation of multiple crops. Water Resource Research
1992;28(1):1–9.
[7] Vedula S, Nagesh Kumar D. An integrated model for optimal reservoir operation for irrigation of multiple crops.
Water Resource Research 1996;32(4):1101–8.
[8] Yeh WWG. Reservoir management and operation models: a-state-of-the-art review. Water Resource Research
1985;21(12):1797–818.
[9] Eiger G, Shamir U. Optimal operation of reservoirs by stochastic programming. Engineering Optimization
1991;17:293–312.
[10] Haouari M, Azaiez MN. Optimal cropping patterns under water de4cit. European Journal of Operational Research
2001;130(1):133–46.
[11] Tsur Y. The stabilization role of groundwater when surface water supplies are uncertain: the implications for
groundwater development. Water Resources Research 1990;26(5):811–8.
[12] Tsur Y, Graham-Tomasi T. The bu2er value of groundwater with stochastic surface water supplies. Journal of
Environmental Economics and Management 1991;21:201–24.
[13] Knapp KC, Olson LJ. The economics of conjunctive groundwater management with stochastic surface supplies.
Journal of Environmental Economics and Management 1995;28:340–56.
[14] Emch PG, Yeh W-G. Management model for conjunctive use of coastal surface water and ground water. Journal of
Water Resources Planning and Management 1998;124(3):129–38.
[15] Yang S-L, Hsu N-S, Hsu S-K, Yeh WW-G. Model development for conjunctive use planning in Taiwan. Water
Resources and the Urban Environment Proceedings of the Annual Water Resources Planning and Management
Conference, ASCE Reston, VA, 1998. p. 597– 602.
[16] Basagaoglu H, Morino MA, Shumway RH. Delta-form approximating problem for a conjunctive water resource
management model. Advances in Water Resources 1999;23(1):69–81.
[17] Azaiez MN. A model for conjunctive use of ground and surface water with opportunity costs. European Journal of
Operational Research 2002;143(3):611–24.
[18] Stewart JI, Hagan EM. Functions to predict e2ects of crop water de4cits. Journal of Irrigation and Drainage Division
ASCE 1973;99(IR4):421–39.
[19] Doorenbos J, Kassam AH. Yield response to water. Irrigation and Drainage Paper 33, F.A.O., Rome, 1979.
[20] Rao NH, Sharma PBS, Chander S. A simple dated water-production function for use in irrigated agriculture.
Agricultural Water Management 1988;13:25–32.
[21] Mannocchi F, Mecarelli P. Optimization analysis of de4cit irrigation systems. Journal of Irrigation and Drainage
Engineering Division ASCE 1994;120(3):484–503.
[22] Du2uaa SO. A chance-constrained model for the operation of the Aswan high dam. Engineering Optimization
1991;17:109–21.
[23] Darby-Dowman K, Barker S, Audsley E, Parsons D. A two-stage stochastic programming with recourse model for
determining robust planting plans in horticulture. Journal of the Operational Research Society 2000;51:83–9.
[24] Loucks D. An evaluation of some linear decision rules in chance constrained models for reservoir planning and
operation. Water Resources Research 1975;11:777–82.
[25] Lambert DK. Calf retention and production decisions over time. Western Journal of Agricultural Economics 1989;
14:9–19.
[26] Hargraves GH, Samani ZA. Economic consideration of de4cit irrigation. Journal of Irrigation and Drainage
Engineering Division ASCE 1984;110(4):343–58.

You might also like