Sustainable Water Resources Management (2020) 6:60

https://doi.org/10.1007/s40899-020-00411-w

ORIGINAL ARTICLE

Application of artificial neural network for optimal operation

of a multi‑purpose multi‑reservoir system, I: initial solution
and selection of input variables
Safayat Ali Shaikh1

Received: 4 April 2019 / Accepted: 28 May 2020

© Springer Nature Switzerland AG 2020

Abstract
In this two-part series research paper, the effect of different input variables for optimal operation of a multipurpose multi-
reservoir system using artificial neural network (ANN) models has been thoroughly investigated. As an alternative to multi-
reservoir stochastic model, dynamic programming-based ANN models are developed. Due to nonavailability of any specific
criterion to select the number and type of input variables for such ANN model, an exhaustive study with various combina-
tions of input variables for single-reservoir ANN models is done. Damodar Valley, a multi-purpose multi-reservoir system
in India, is used for this study. Results obtained from discrete dynamic programming, an optimization technique, are used
as training data for ANN models. With different combinations of input data, five types of ANN models are developed. ANN
models are simulated with generated inflow sequence as well as different types of observed historical inflow sequences. For
each simulation, 240 monthly networks are trained and stored. Four different stochastic dynamic programming (SDP) models
of the system are developed considering one reservoir at a time approach. Linear regression analysis is performed between
the results obtained from ANN models with results of the best SDP model. Objective function values as well as different
regression parameters are compared to select preferred ANN model and its input variables.

Keywords  Artificial neural network · Discrete dynamic programming · Stochastic dynamic programming · Reservoir
operation

Introduction a review of different dynamic programming (DP) models

for various water resources problems. Yeh (1985) reviewed
Due to random nature of inflow into a reservoir, the final in detail about the different reported works in the field of
storage state is a stochastic function of initial storage state optimal operation of reservoir. Wurbs (1993) reviewed dif-
and release from a reservoir. The volume of release from a ferent simulation, optimization and network flow models in
reservoir also depends on the type and magnitude of imposed different decision situation. Labadie (2004) has presented a
demands as well as constraints on storage and release. More- comprehensive review of various optimization techniques
over, monthly variations in releases within a year also indi- like dynamic programming, linear programming, nonlinear
cate the dependence of current period’s release on previous programming, heuristic programming methods using evolu-
period’s storage and inflow. All these factors make optimal tionary and genetic algorithms, artificial neural network and
water allocation from a reservoir very complex one. So far, fuzzy rule-based methods, applied to the optimal operation
dynamic programming (Bellman 1957) is used in the fields of multi-reservoir systems. In spite of being the most effec-
of long-term and short-term reservoir operation for water tive tool, DP suffers from computational difficulty, which is
allocation and power generation. Yakowitz (1982) presented popularly known as the ‘Curse of Dimensionality’ (a term
used for excessive high-speed memory and computational
time requirement of a multidimensional DP problem). This
* Safayat Ali Shaikh disadvantage becomes more acute when the system has mul-
[email protected] tiple reservoirs and the problem is treated in a stochastic
1
Department of Civil Engineering, Aliah University, Kolkata, framework.
India

13
Vol.:(0123456789)
 60   Page 2 of 14 Sustainable Water Resources Management (2020) 6:60

In recent past, various alternative regression simulation- where reservoir storage, inflows to the reservoir and soil
based approaches have been tried with various degrees of moisture in the irrigated area have been treated as state vari-
success. Noteworthy among one of them is ANN-based ables. Vedula and Kumar (1995) presented a model con-
models. Although the ANN model building process is not sisting of two modules with three state variables: reservoir
yet consolidated, it has practical benefits (Maier and Dandy storage, seasonal inflow and rainfall. Transition probabilities
2000). In many respects, ANNs are similar to regression- of seasonal inflow and rainfall together with results from
based models in hydrology, except that they do not require module 1 have been used as input in module 2. Celeste et al.
specification of a mathematical form. ANN is particularly (2008) solve a monthly operation model by a stochastic pro-
useful in performing classification and pattern recognition gramming approach and then use its information to guide the
for process governed by complex nonlinear interrelation- daily operation, which is solved by deterministic optimiza-
ships. Nevertheless, ANN may be viewed as an alternative tion. This stochastic–deterministic procedure is applied to
tool to conventional optimization models (ASCE 2000a, b). a Ishitegawa Dam in Japan, and the results show that it pro-
vides operating rules relatively similar to the ones obtained
when perfect forecasts are used for the monthly inflows.
Literature survey Application of stochastic DP (SDP) to multi-reservoir
system is restricted to a few cases, due to excessive compu-
Over the last 60 years, researchers are engaged to develop tation time and storage requirement. Some kind of approxi-
reservoir operating policies for multi-reservoir system. mation is necessary to make the problem computationally
Among various mathematical programming techniques, feasible, and in that respect the single-reservoir models are
dynamic programming was used most widely. However, useful. Some reported applications of SDP to multi-reservoir
application of this technique to multi-reservoir system operation problems include works of Paudyal et al. (1990),
suffered due to computational complexities, particularly Ravikumar and Venugopal (1998) and Teixeira and Marina
when the operation problem is treated as stochastic. To (2002). Paudyal et al. (1990) used incremental DP (IDP) to
alleviate these dimensionality problems, various approxi- determine a set of best configurations using critical period
mate DP solution algorithms were applied by researchers, inflows and then SDP technique to determine the steady-state
with various degrees of success. So far, reported works of long-term operation policy which maximized the expected
multi-reservoir operation solved in stochastic framework are value of the annual energy generation for determining the
approximate and case specific. As no two-reservoir system is optimal configuration of a two-reservoir system. They con-
similar in nature, conclusions drawn for a particular system sidered the spatial correlation of the two inflow series and
are unlikely to be tenable for other system. constructed joint transition probabilities to be used in the
Since 1990s, a different style of problem-solving SDP model. Ravikumar and Venugopal (1998) developed an
approach has been started in the field of water resources and optimal release policy using SDP where inflow was treated
hydrology, through the application of artificial intelligence so as to follow first-order Markov chain model for Krishna-
techniques. Different genetic algorithm and neural network giri reservoir system in India. Teixeira and Marina (2002)
models have been applied to solve the reservoir operation developed optimal operation of a two-reservoir in-parallel
problems. Although a few applications are reported , ANN system using SDP where a Markov chain process is used to
models appear to be a promising solution alternative for estimate inflows.
reservoir operation problems. The present study deals with Some of the important applications of ANN in single-res-
the development of models for multi-reservoir operation ervoir operation are the works of Raman and Chandramouli
problem using DP and ANN techniques. These models are (1996), Jain et al. (1999), Cancelliere et al. (2004), Chan-
developed for and applied to an existing multi-reservoir sys- dramouli and Deka (2005) and Chaves and Chang (2008).
tem in India. Raman and Chandramouli (1996) derived four operating
In single-reservoir operation problems, SDP has been policies by using four different models: (1) a stochastic
successfully applied by many researchers: Esmaeil-Beik dynamic programming (SDP), (2) a simulation-based stand-
and Yu (1984), Huang et al. (1991), Vedula and Majumdar ard operating policy (SOP), (3) a dynamic programming
(1992), Vedula and Kumar (1995) and Celeste et al. (2008). based-neural network (DPN) and (4) multiple linear regres-
Esmaeil-Beik and Yu (1984) have studied the practical sion procedure (DPR). It is established that the performance
implication of using current period’s inflow, previous peri- of the DPN model is better than the other three models.
od’s inflow and forecasted inflow as second-state variables. Jain et al. (1999) determined optimal operating policies by
Huang et al. (1991) studied four types of SDP models, based linear and nonlinear regression model as well as by ANN.
on different state space formulation and different assump- It is established that ANN is best suited for inflow predic-
tions regarding correlation structure of inflows. Vedula and tion as well as optimal operation. Cancelliere et al. (2004)
Majumdar (1992) have derived optimal operating policy introduced a neural network approach to derive the operating

13
Sustainable Water Resources Management (2020) 6:60 Page 3 of 14  60

rules of an irrigation supply reservoir. Chandramouli and are developed for and applied to an existing multi-reservoir
Deka (2005) derived an ANN-based decision support model system in India.
(DSM) for optimal operation of a reservoir. The DSM based
on ANN outperforms the regression-based approach. Chaves
and Chang (2008) proposed an evolving ANN model which System under study
gave improved operation performance of the reservoir when
compared with its current operational strategy. The system considered in this study is Damodar Valley (DV)
The reported applications in multi-reservoir optimal system (Fig. 1) in India. There are four reservoirs: Konar,
operation using ANN are very few. Some important works Tilaiya, Panchet and Maithon, and a barrage: Durgapur, in
are done by Chandramouli and Raman (2001) and Chan- this system. Konar with Panchet and Tilaiya with Maithon
dramouli et al. (2002), Wang et al. (2010), Fayaed et al. are connected in parallel, whereas Panchet and Maithon are
(2013) and Coerver et al. (2018). Chandramouli and Raman connected in series. Conservation storage capacity of Konar,
(2001) developed a dynamic programming-based neural Tilaiya, Panchet and Maithon is 220.50 Mm3 , 141.6 Mm3 ,
network model for determination of optimal release policy 222.05 Mm3 and 570.95 Mm3 , respectively. About 82 per-
of Parambikulam Aliyar Project (three-reservoir project) cent of the total rainfall occurs during the southwest mon-
in Tamil Nadu, India. In their model, storage, inflow and soon period (June to September). During the post-mon-
demand have been considered as an input and release as soon (October–November) and pre-monsoon (April–May)
output. Chandramouli et al. (2002) developed a dynamic months, the amount of rainfall is about 8 percent and 7 per-
programming-based neural network model to determine cent, respectively. The estimated average annual runoff at
release policies of a three-reservoir system supplying water different reservoir sites is 555.1 Mm3(Konar), 431.7 Mm3
only for irrigation. Results of the proposed model have been (Tilaiya), 4539.2 Mm3(Panchet) and 2615.0 Mm3(Maithon).
compared with that of a standard operating policy. Wang Due to high temperature, humidity and wind, almost 50
et al. (2010) have developed optimal operating policies of percent of the total annual evaporation takes place during
Yellow River reservoir system in China by using neural net- March to June. On the basis of land pans, mean annual evap-
work and particle swarm optimization technique. There are oration losses from the Konar, Tilaiya, Panchet and Maithon
eight input (inflow, storage and demand) elements and three are estimated as 1496 mm, 1773 mm, 1547 mm and 1389
output nodes consisting of releases from three reservoirs of mm, respectively, as per data collected from the Office of the
the system. Fayaed et al. (2013) have derived optimum oper- Manager (Reservoir Operation), Damodar Valley Corpora-
ating policies of a Sg. Langat Reservoir in Malaysia using tion, Maithon Dam, Jharkhand, India.
SDP-ANN model. Coerver et al. (2018) have determined
the release policies of 11 reservoirs in Central Asia , the
USA and Vietnam using artificial neural network capable
of mimicking fuzzy logic.
Reported applications of ANN to reservoir operation
problems are less in number compared to the DP models.
Defining the network structure and identifying the input vec-
tors is a critical task, for which no generalized algorithm
exists. For this reason, the model study should be done in
an exhaustive manner. Because of its strong generalization
power, a properly trained ANN can provide accurate output
for an unknown input. However, proper training requires
proper input data set and it is preferable to use some type of
optimization model to generate this input data set. In some
reported works, the results of DP are used as the input data
set. However, as the operation problem is essentially sto-
chastic, performance of an ANN model should be compared
with the results of a stochastic optimization model. Due to
its inherent power to relate system variables with the out-
put, ANN appears to be a viable alternative to existing opti-
mization techniques, from practical solution point of view.
With this ability of ANN in mind, the present study deals
with the development of models for multi-reservoir opera-
tion problem using DP and ANN techniques. These models Fig. 1  Damodar Valley Reservoir System

13
 60   Page 4 of 14 Sustainable Water Resources Management (2020) 6:60

It is estimated that en route losses (losses due to seepage (DDP) model development, (b) explicit stochastic dynamic
and evaporation from releases) are at the rate of 5 percent programming model development, (c) single-reservoir ANN
of the flows during July to October and at 10 percent during model development, (d) regression analysis between best
November to June. In this system study, reservoirs Konar, ANN model output with best SDP model output and (e)
Tilaiya, Panchet and Maithon are termed as Reservoir 1, selection of preferred ANN model and its input data.
Reservoir 2, Reservoir 3 and Reservoir 4, respectively. So,
in July to October about 95% of release and in November Discrete dynamic programming model
to June about 90% of release from Reservoir 1 along with
uncontrolled catchment area contribution of Reservoir 3 In this model, development of discretized version of DP as
would come to Reservoir 3 as inflow. Similarly, in July to proposed by Bellman (1962) is used. DV system has been
October about 95% of release and in November to June about decomposed into four individual reservoirs maintaining the
90% of release from Reservoir 2 along with uncontrolled integrity of the system configuration and continuity of flow
catchment area contribution of Reservoir 4 would come to between the reservoirs and optimal policies are derived using
Reservoir 4 as inflow. From an historic inflow record of 44 discrete dynamic programming (DDP). In this model, final
years (October 1961–Sep 2004), it is found that there is a storage state is considered as decision variables and assumed
wide variation in the average annual inflow. The monthly that invertibility condition holds good. For this model, system
mean values show that the months of high inflows are July, dynamics, system constraints, performance function, objec-
August, September and October. Appreciable amount of tive function and recursive equations would be as follows:
skewness is also observed in these four inflow series, indi-
cating a non-normal distribution. Regarding the serial co- System dynamics
relation structure of the inflows, moderate to strong correla-
tions are observed in most of the months. xt+1 (1) = xt (1) + yt (1) − ut (1) − evt (1) (1)
As per Damodar Valley Reservoirs Regulation Manual
(1995), the multiple purposes of the DV system are water xt+1 (2) = xt (2) + yt (2) − ut (2) − evt (2) (2)
supply for irrigation, municipal and industrial use, flood
control, hydroelectric power generation and fulfillment of xt+1 (3) = xt (3) + yt (3) + ut (1) − ent (3) − ut (3) − evt (3) (3)
mandatory requirement. For sustenance of aquatic life and
ecology in the streams, a minimum quantity of water is to be
xt+1 (4) = xt (4) + yt (4) + ut (2) − ent (4) − ut (4) − evt (4) (4)
released from different reservoirs for different reaches. For
reach 1 (reach between Reservoir 1 to Reservoir 3), reach In Eqs. ((1) to (4)), xt (i) , ut (i) , yt (i) , evt (i) and ent (i) represent
2 (reach between Reservoir 2 to Reservoir 4) and reach 3 the volume of storage, the volume of release, the volume
(downstream of Reservoir 3 and Reservoir 4 to upstream of inflow, the amount of evaporation loss to the ith reser-
of Durgapur barrage), quantity of water for municipal and voir during time period t and the en route loss in the reach
industrial (M&I) use is considered to be adequate to meet between (i − 2) th reservoir and ith reservoir.
this mandatory requirement. However, as there is no provi-
sion of M&I supply for reach 4 (downstream of barage), an System constraint
estimated quantity of 2.1 m 3 /s is to be released from the
barrage to maintain the river ecology in that reach. 𝐂𝐦𝐢𝐧 ≤ 𝐱t ≤ 𝐂max (on storage) (5)

𝐮t,min ≤ 𝐮t ≤ 𝐮t,max (on release) (6)

Methodology
Terms 𝐂min and 𝐂max represent vectors of minimum and
In this study, water supply for irrigation, municipal and maximum storage capacities of the four reservoirs. The
industrial use is selected as the objective of operation, while term 𝐮t,min is the vector of minimum mandatory releases,
other purposes (mandatory requirement for sustenance of and 𝐮t,max is the vector of maximum permissible releases.
aquatic life and ecology in the streams, flood control and
hydroelectric power generation) are treated as binding con- Performance function
straints on the system variables. Considering water supply
as the main objective, there are some target values to be In the optimization model, a loss function is used in the form
achieved and the efficiency of the system depends on the of a penalty function to get its performance. The penalty is
degree of achievement. Formulation of the present problem associated with the deviation of release from the target, and
(initial solution of multi-reservoir system and input selec- the objective is to minimize the total penalty incurred dur-
tion) is done in five steps: (a) discrete dynamic programming ing the years of operation. If Dt and ut represent the target

13
Sustainable Water Resources Management (2020) 6:60 Page 5 of 14  60

release and corresponding reservoir release, respectively, After solving for Reservoir 1 and Reservoir 2, Reservoir
during stage t, then the generic form of the single stage 3 and Reservoir 4 are subsequently optimized assuming the
loss function to be used in the DP recursive equation can common demand ( Dct ) on Reservoir 3 and Reservoir 4 is
be expressed as shared by the ratio of 0.6:0.4, respectively. Then, assume
demand of Reservoir 3 can have target Dt (3) = 0.6Dct .
Lt (1) = [(Dt (1) − ut (1))∕Dt (1)]2 (for reservoir 1) (7) The resulting optimal trajectory and release sequences are
denoted as {xt (3) } and {ut (3)} , respectively. Now, Reser-
Lt (2) = [(Dt (2) − ut (2))∕Dt (2)]2 (for reservoir 2) (8) voir 4 is considered with Dt (4) = Dct − ut (3) and optimized.
The resulting optimal trajectory and release sequences are
denoted as {xt (4) } and {ut (4)} , respectively.
Lt (3) = [(Dt (3) − ut (3))∕Dt (3)]2 (for reservoir 3) (9)
For analysis of DV system, a calendar year has been
divided into twelve periods, i.e., each period represents
Lt (4) = [(Dct − ut (3) − ut (4))∕Dct ]2 (for reservoir 4) (10) a calendar month. Inflow data of all 12 periods of 44
years (October 1961–September 2004) have also been
where Lt is the single-stage loss function, Dt (1) and Dt (2) collected.Optimal trajectories (considering one reservoir at
represent the water supply target levels for Reservoirs 1 and a time optimization) are determined and stored for 44-year
2, respectively, and Dct represents the combined water supply historical inflow data. After obtaining optimal trajectories
target level for Reservoirs 3 and 4. Units of Dt and Dct are in period-wise ( t ), numerical values of 44 sets of initial stor-
million cubic meter. age (𝐱t ) , inflow of current period (𝐲t ) , previous period stor-
age (𝐱t−1 ) , inflow of previous period(𝐲t−1 ) , release (𝐮t )and
Objective function final storage (𝐱t+1 ) are stored in matrix form. So in the said
∑
N matrix initial storage state , inflow of current period , previ-
min Lt (ut ) (11) ous period storage , inflow of previous period , release and
ut
t=1 final storage are the elements of column1, column 2, column
3, column 4, column 5 and column 6, respectively. This data
set is termed as training data set I, and its size is 12 number
Recursive equation
[44 × 6] . Similarly from the 41 years historical data (elimi-
Vt (𝐱ti ) = min k
{Lt + Vt+1 (𝐱t+1 )} i = 1, 2, … , I, nating inflow data of maximum annual inflow year, 75%
dependable annual inflow year and minimum annual inflow
K
𝐱t+1

t = 1, 2, … , T (12) year from 44 years historical data) optimal trajectories are

determined and like training data set I, 12 number [41 × 6]
where VN+1 (𝐱N+1
k
) = 0 for all k and 𝐱ti , 𝐱t+1
k
, 𝐮t are determined set of data are stored. This data set is termed training data
from equations. As the system is invertible for known initial set II.
and final storage ( 𝐱t (i) and 𝐱t+1 (i) ) the average area and the
corresponding evaporation loss can be computed directly, Explicit stochastic dynamic programming model
without any iteration.
In deterministic analysis, the reservoir operation problem
Determination of initial trajectories is solved for a particular known inflow sequence. Actually
during any stage, inflow to the reservoir is random in nature.
First, all the four reservoirs of DV system have been decom- Thus, the final state is not a deterministic function of the
posed into four numbers of individual reservoir maintain- initial state and the release made during that stage. Hence,
ing the continuity of the system (Eq.1 to Eq. 4), and then it cannot be computed uniquely from the system dynamics.
two upstream reservoirs and subsequently two downstream If statistical analysis on the historical inflows shows strong
reservoirs are optimized. Reservoir 1 is optimized using correlation, it has to be considered in the model and should
system dynamics Eq.  1, respective system constraint as not be neglected.
mentioned in Eq. 5 and Eq. 6, performance function Eq. 7 One important thing about the serial correlation structure
in conjunction with objective function Eq. 11 and respec- of the inflows is that when the inflows are serially corre-
tive recursive Eq. 12. The resulting optimal trajectory and lated and assumed to be a first-order Markov process, then
release sequences are denoted as {̄xt (1)} and {̄ut (1)} , respec- in the stochastic dynamic programming (SDP) formula-
tively. The next optimization problem is solved for reser- tion, for a single-reservoir case, the state space becomes
voir 2 in a similar manner, and the resulting optimal trajec- two-dimensional containing the storage and inflow vari-
tory and release sequences are stored as {̄xt (2)} and {̄ut (2)} , able, thus increasing the computational burden. In SDP
respectively. formulation for serially correlated inflows sequences, the

13
 60   Page 6 of 14 Sustainable Water Resources Management (2020) 6:60

second state variable, i.e., the inflow forecast for the current value of Lt  , a set of possible values of Lt is obtained and
period, may be replaced by the observed inflow during the the expected value of the systems performance is consid-
previous period, if reliable forecast is not available. In this ered in the objective function. Accordingly, the objective
regard, Esmaeil-Beik and Yu [1984] found that the simulated function becomes
average losses almost doubled when previous stage inflow [N ]
was used as a state variable rather than inflow of the cur- ∑
J(R) = E Lt (13)
rent stage. As they have mentioned, the increased loss may t=1
be interpreted as the expected penalty for decisions made
without advanced information about the current inflow. On A solution to this stochastic optimization problem is a
the other hand, a risk is involved in using predicted inflow sequence of feasible decisions u that minimizes J(R) .
based on available streamflow forecast models. Thus, the
choice between these state variables has serious practical
Transition probability
implications. So major issues in the explicit stochastic model
development are that inclusion of inflow as the second state
The expected value of the systems performance is com-
variable and representation of the correlation structure of
puted from the inflow transition probabilities. The random
the inflows.
inflows to the reservoir may either be serially correlated
Based on these issues, four different models (with recur-
or serially independent. When these inflows are serially
sive equation as recommended by Loucks et al. (1981),
correlated, i.e., the current period’s inflow is dependent on
Huang et al. (1991) and Esmaeil-Beik and Yu (1984)) are
previous period’s inflows, such dependence can be mod-
developed. These models are:(1) SDP-M1 in which current
eled by a Markov chain. Assuming a first-order Markov
period’s inflow is second state variable and correlated inflow
chain, the inflow transition probabilities can be defined as
sequence, (2) SDP-M2 in which current period’s inflow is
second state variable and independent inflow sequence, (3) j
pij = Pr[yt+1 = yt+1 |yt = yit ] (14)
SDP-M3 in which previous period’s inflow is second state
variable and correlated inflow sequence and (4) SDP-M4 This transition probability pij specifies the conditional prob-
in which storage is only the state variable and independent ability that the next inflow state during stage t + 1 is at state
inflow sequence. yt+1 , given that the current inflow state is at state yit during
j

In the stochastic models, the expected value of objec- stage t.

tive function is optimized. This expected value of objective
function at each stage is based on state transition probabili-
Recursive equation
ties. Transition probabilities are computed from probability
distribution of inflows. For computation of transition prob-
Here, Lt represents the value of the system performance
abilities, 1000-year synthetic inflow sequences are gener-
in stage t (single-stage loss function) corresponding to an
ated using AR(1) model. So system dynamics, system con-
initial storage volume xti , and inflow ykt and a final storage
straints, performance function and objective function would
volume xt+1 , then considering the inflows to be serially cor-
j
be same for all SDP models, whereas transition probability
related, the discretized version of the recursive equation can
and recursive equations would be different and these are as
be written as Loucks et al. 1981 and Huang et al. (1991):
follows:
⎡ � ⎤
Vt (xti , ykt ) = min ⎢L + L (x
j
, yl
) p ⎥
xt+1 ⎢ ⎥
t t+1 t+1 t+1 kl
System dynamics j
(15)
⎣ l
yt+1 ⎦
j
System dynamics is same as given t = N, N − 1, … , 1; ∀xti , ykt , xt+1 feasible.
{ } in Eqs. ((1) to (4)). How-
ever, as the inflow sequence yt is stochastic, xt+1 is not a
The superscripts i, j, k, l denote the discretization levels for
deterministic function of xt and ut and cannot be computed
the corresponding variables. In Eq. (15), the term pkl is the
uniquely from the system dynamics. It is only possible { to} transition probability specifying the conditional probability
compute a set of xt+1 values for a known possible set of yt
of inflow ylt+1 in stage t + 1 when the inflow in stage t is ykt .
values.
Stochastic DP model with Eq. (15) is termed as SDP-M1.
If there is no co-relation between two consecutive
Objective function
inflows, i.e., current inflow is independent of previous
inflow, then inflow transitional probability becomes uncon-
As the final stage becomes uncertain, single-stage loss
ditional probability, and then recursive Eq. (15) becomes
function Lt also becomes uncertain. Instead of a single

13
Sustainable Water Resources Management (2020) 6:60 Page 7 of 14  60

⎡ ⎤ reservoirs on the downstream reservoirs is indirectly taken

�
Vt (xti , ykt ) = min ⎢L + L (x
j
, yl
) p ⎥ into account.
xt+1 ⎢ ⎥
t t+1 t+1 t+1 k
j
⎣ l
yt+1 ⎦ (16) Equations (15) to (18) are solved, and a steady-state solu-
j tion is obtained for each model and presented in Table 1. In
t = N, N − 1, … , 1; ∀xti , ykt , xt+1 feasible this table, the term ‘Expected Cost’ (termed as EC) repre-
sents the minimum expected yearly loss obtained from the
where the term pk is the unconditional probability Pr[yt = yit ] .
steady-state solution. The term ‘Average Cost’ (termed as
Stochastic DP model with Eq. (15) is termed as SDP-M2.
AC) indicates the average value of the objective function
Since at the beginning of stage t the actual value of yt is
computed from 44 years of release sequence obtained from
unknown, an inflow forecast is to be used. If reliable forecast
each model. Model SDP-M1 yielded the values of ‘Expected
is unavailable, the net inflow during the preceding stage, yt−1
Cost’ and ‘Average Cost’ much less than all other models
(known at the beginning of stage t) could be used as a state
for all the reservoirs. Hence, it can be concluded that SDP-
variable instead of yt . Accordingly, the backward recursive
M1 is a the best model. Subsequently, optimal trajectories
equation may be expressed as Esmaeil-Beik and Yu 1984;
for (a) 5-year generated inflow (obtained from 1000-year
⎧ ⎫ generated data) and (b) historical monthly inflow sequence
⎪� � j
� ⎪ correspond to observed (i) maximum annual inflow year,
Vt (xti , ykt−1 ) = min ⎨
l
Lt + Vt+1 (xt+1 , yt ) pkl ⎬
j
xt+1 ⎪ l (ii) 75% dependable annual inflow year and (iii) minimum
It ⎪
⎩ ⎭ annual inflow year which are derived using SDP-M1.
j
t = N, N − 1, … , 1; ∀xti , ykt−1 , xt+1 feasible
(17) Single‑reservoir ANN model
where pkl is the conditional probability of inflow ylt during In this study, a three-layer fully connected feed-forward
stage t, given that the inflow in stage t − 1 is ykt−1 . Stochastic network architecture of type ( p − 10 − 10 − 1 ), indicating p
model with Eq. 17 is termed as SDP-M3. number input variables in input layer, two hidden layers with
For independent inflow sequence the above Eq. (17) sim- ten nodes each and one output layer with one output vector,
plifies to; as shown in Fig. 2 is selected. Selection of two hidden lay-
⎧ ⎫ ers is rather arbitrary. In most of the reported cases, either
⎪� � j
� ⎪ one or two hidden layers have been used. Using activation
Vt (xti ) = min ⎨ Lt + Vt+1 (xt+1 ) pk ⎬ functions, tan-sigmoid, log-sigmoid and pure-line, used in
(18)
j
xt+1 ⎪ l ⎪
⎩
It
⎭ the first hidden layer, second hidden layer and output layer,
j respectively, different networks are developed by varying
t = N, N − 1, … , 1; ∀xti , ykt−1 , xt+1 feasible
the number of nodes in a layer from 8 to 12. Each network
Stochastic model with Eq. 18 is termed as SDP-M4. is trained with the same data set, and the network which
At the beginning, one reservoir at a time approach is yielded minimum mean square error (MSE) is selected. On
applied to two upstream reservoirs. In this way, both the this basis, 10 numbers of nodes in a hidden layer are fixed.
upstream reservoirs are operated individually in a stochas-
tic optimized way. A forward search (simulation) is made Selection of input vector and target vector
with the operating policies for generated inflow sequences
to determine the corresponding releases. With these releases For the ANN model of stage t, the following system vari-
and generated inflows, the net inflows to the downstream ables are considered to be influencing the release from the
reservoirs are computed. The operation of the downstream reservoir: initial storage (xt ) , inflow (yt ) and demand (Dt ) of
reservoirs is then optimized separately with these net the current period t, as well as initial storage (xt−1 ) and inflow
inflow sequences. Thus, the effect of release from upstream (yt−1 ) of the previous period (t − 1) . No specific assump-
tion is made about the selection of a particular variable as

Table 1  Comparison of Reser- SDP-M1 SDP-M2 SDP-M3 SDP-M4

objective function values
voir EC AC EC AC EC AC EC AC

1 33.075 33.560 52.645 39.034 62.325 56.33 76.53 6 64.323

2 62198 71296 75573 77329 76001 79232 88020 97200
3 28.752 66.493 38.195 70.061 45.002 70.381 42.225 60.232
4 10.557 33.041 42.514 52.922 66.028 86.345 98.025 116.25

13
 60   Page 8 of 14 Sustainable Water Resources Management (2020) 6:60

input, and a number of feasible combinations are analyzed. 𝐛m (k + 1) = 𝐛m (k) − 𝛾𝐬m (20)
Release utduring stage t is considered as the target variable.
However, since the DDP models used final storage xt+1 as where 𝛾 = learning rate and sm i
≡ 𝛿F̂∕𝛿nm i
 , W = weight
the control variable, the same is also used as the target vari- matrix, b = bias vector, a = input vector.
able, separately. It is to be noted here that demand (Dt ) is not In supervised training with back-propagation ( Leven-
included as an input variable, as for any particular month, berg–Marquardt) algorithm, every input–output pattern
(Dt ) remains the same over the years. Hence, introduction of is presented to a network and modification of neuron con-
the same will only add a constant to the input vector. nection weights is done by (a) a forward pass and (b) a
Based on these considerations, five different types of backward pass. For a particular pattern vector i , the actual
input/output combinations are proposed, as shown in Table 2 response of the network ( ai ) is compared with the desired
for training data set I and training data set II. For any such response ( aci ). The error signal (ES) is the difference between
model, size of the input matrix depends on the number of actual response and desired response and is calculated as
variables selected as input and the number of years of DDP follows:
results used.
1 ∑∑ c
PN
ES = [a − ai ]2 (21)
2 q i=1 i
Training
where q is the total number of output nodes and PN is the
Training of those models has been done with supervised
number of patterns.
environment with back-propagation algorithm. In back-
Initialization of network is done by using Nguyen–Wid-
propagation algorithm, the network weights and biases are
row methods. Free parameters like learning rate, epoch
updated in the direction in which the performance function
and momentum factor have been taken 0.1, 1500 and 0.3,
decreases most rapidly, i.e., in the direction of negative gra-
respectively, and goal (acceptable value of the mean square
dient. Mathematically, it can be expressed as follows:
error) is taken as 1 × 10−6 . Input and target data (12 num-
𝐖m (k + 1) = 𝐖m (k) − 𝛾𝐬m (𝐚m−1 )T (19) ber, training data set I of size [44 × 6] and training data
set II of size [44 × 6] as mentioned in ‘Discrete dynamic

Table 2  Input and target vectors ANN Input vectors Target vectors Training data set I Training data set II
for ANN models
Input size Target size Input size Target size

M1 xt , yt xt+1 [2 × 44] [1 × 44] [2 × 41] [1 × 41]

M2 xt , yt ut [2 × 44] [1 × 44] [2 × 41] [1 × 41]
M3 xt , yt , yt−1 xt+1 [3 × 44] [1 × 44] [2 × 41] [1 × 41]
M4 xt , yt , yt−1 ut [3 × 44] [1 × 44] [3 × 41] [1 × 41]
M5 xt , yt , xt−1 , yt−1 xt+1 [4 × 44] [1 × 44] [4 × 41] [1 × 41]

Fig. 2  Functional diagram of
single-reservoir ANN model

13
Sustainable Water Resources Management (2020) 6:60 Page 9 of 14  60

Table 3  Comparison for A Reservoir 1 Reservoir 2

different ANN models for
Reservoir 1 and Reservoir 2 N Regression parameter Regression parameter
N m b r OFV-1 OFV-2 m b r OFV-1 OFV-2

Case 1
M1 1.094 − 45.35 0.956 1331 0.880 0.585 0.832 583122
M2 0.987 − 30.41 0.906 1340 0.869 1.809 0.843 626766
M3 1.008 − 17.48 0.945 1189 1297 0.957 − 6.62 0.935 557953 582115
M4 1.055 − 23.57 0.942 1450 0.882 10.82 0.917 624547
M5 0.794 3.781 0.787 1511 0.943 − 19.99 0.810 586611
Case 2
M1 1.171 − 46.47 0.977 259 0.910 13.670 0.968 376690
M2 1.089 − 26.52 0.999 270 1.155 − 35.87 0.993 437645
M3 1.026 − 1.330 0.986 257 235 0.911 15.728 0.969 390234 374227
M4 0.965 5.680 0.998 262 0.929 15.342 0.981 384138
M5 1.053 − 25.88 0.887 262 0.977 0.4406 0.979 378749
Case 3
M1 1.061 − 28.90 0.988 6.20 1.160 − 39.15 0.994 9282
M2 1.027 − 13.87 0.997 7.28 0.997 − 3.541 0.998 8596
M3 1.037 − 9.975 0.999 6.14 6.17 1.007 − 8.310 0.994 9141 8579
M4 1.021 − 10.86 0.998 6.26 1.022 − 9.666 0.994 8706
M5 1.007 0.0261 0.999 6.43 1.217 − 33.35 0.984 9386
Case 4
M1 0.953 4.1107 0.994 7.18 1.011 − 12.29 0.993 11400
M2 0.948 3.7351 0.995 5.12 0.928 1.6566 0.985 15266
M3 1.032 20.044 0.996 4.49 2.98 1.78 − 186.05 0.975 1993 2891
M4 0.982 − 4.026 0.996 7.08 0.955 − 2.465 0.989 14480
M5 1.548 − 164.8 0.989 3.37 1.045 − 37.22 0.956 14686

programming model’ section) are standardized to keep the are simulated with historical monthly inflow sequence cor-
minimum and maximum values in the default interval of responding to observed (ii) maximum annual inflow (termed
[0, 1]. Now, standardized data set is partitioned into three as Case 2), (iii) 75% dependable inflow (termed as Case 3)
parts : 80% for training, 10% for validation and 10% for and (iv) minimum annual inflow (termed as Case 4). For all
testing to prevent over-fitting during training. Using the simulation at the beginning (i.e., at t = 1 ), it is assumed that
partitioned data as input and target as required by differ- the reservoir is full (i.e., storage = Cmax).
ent models (mentioned in Table 2), 12 trained networks are (i) Simulation for Case 1: At the beginning (i.e.,t = 1 ) of
developed for every model and stored. It is to be noted here simulation, if the trained network is obtained using training
that for a particular model 12 trained networks are obtained. data set I and it corresponds to ANN M1 (refer Table 1),
So, for a particular reservoir a total of (5 × 12) = 60 monthly then the input data set would be ( Cmax , y1 ). For this data set,
networks are trained and stored. So for all reservoir using the first trained network yields x2 . It is checked whether this
training data set I total 4 × (5 × 12) = 240 trained networks x2 is within the permissible storage limits and then release
are obtained and stored. Similarly, using training data set II u1 is calculated from the system dynamics. Next, the second
total 4 × (5 × 12) = 240 trained networks are obtained and trained network is taken and ( x2 , y2 ) is applied as input to
stored. Now, totally 2 × 240 = 480 networks are ready for it to obtain x3 . The process continues to obtain all releases
simulation. during the entire simulation period of five years. After simu-
lation, surplus and deficit of each period are squared and
Simulation summed up to get objective function value and termed as
OFV-2. In this way, using ( Cmax , y1 ) values of respective res-
Trained networks obtained using training data set I are ervoirs simulation of all other reservoirs can be done. In the
simulated with (i) 5-year generated inflow and are termed similar way as per input data set of all other models, final
as ‘Case 1,’ and networks obtained using training data set II storage/release can be simulated.

13
 60   Page 10 of 14 Sustainable Water Resources Management (2020) 6:60

Fig. 3  Trajectories for Reservoir 1 for different flow conditions

Fig. 4  Trajectories for Reservoir 2 for different flow conditions

13
Sustainable Water Resources Management (2020) 6:60 Page 11 of 14  60

Table 4  Comparison for A Reservoir 3 Reservoir 4

different ANN models for
Reservoir 3 and Reservoir 4 N Regression parameter Regression parameter
N m b r OFV-1 OFV-2 m b r OFV-1 OFV-2

Case 1
M1 0.616 89.74 0.611 125.9 0.668 157.2 0.744 127.2
M2 0.611 78.15 0.521 130.1 0.761 62.6 0.812 129.3
M3 0.762 48.17 0.742 210.8 125.2 0.708 79.56 0.794 88.2 125.0
M4 0.782 17.38 0.721 132.7 0.807 26.20 0.796 128.2
M5 0.296 169.7 0.266 214.5 0.507 145.5 0.667 314.3
Case 2
M1 0.400 170.4 0.419 275.8 0.478 217.6 0.470 319.3
M2 0.543 130.9 0.512 264.2 0.819 22.40 0.770 345.8
M3 0.562 114.3 0.565 268.5 243.5 0.989 − 42.1 0.867 191.4 314.5
M4 0.903 − 9.23 0.848 263.5 0.697 67.08 0.629 518.0
M5 0.351 175.4 0.317 274.9 0.860 45.12 0.838 353.9
Case 3
M1 0.906 − 3.89 0.872 21.69 0.939 0.345 0.989 2.76
M2 0.827 18.46 0.834 23.90 1.085 − 96.1 0.956 1.76
M3 0.859 4.60 0.858 17.51 19.72 1.092 − 99.87 0.961 0.81 1.17
M4 0.882 − 4.50 0.881 19.99 1.113 − 114.1 0.935 3.23
M5 0.823 15.17 0.824 22.11 0.943 6.268 0.992 3.63
Case 4
M1 1.184 − 43.71 0.936 7.91 0.956 0.040 0.945 10.6
M2 1.118 − 28.86 0.914 10.1 0.944 24.88 0.965 5.95
M3 0.998 − 20.67 0.957 2.32 4.82 0.961 − 12.9 0.964 0.29 2.59
M4 0.993 − 11.67 0.922 6.60 0.992 − 35.8 0.928 3.16
M5 0.362 127.5 0.382 8.40 1.040 − 182 0.811 3.51

(ii) Simulation for Case 2, Case 3 and Case 4 : Here, of regression, r. Theoretically, if m = 1 , b = 0 and r = 1 ,
simulation is same as that of generated inflow, the only dif- then the ANN result can be treated as exactly equal to the
ference is the trained network obtained using training data SDP-M1 result.
set II. However, in actual cases these values will be different
than the theoretical values. Hence, closeness of the actual
Regression analysis values to the theoretical values is examined. Sometimes, it
is observed that m and r values are close to unity, but b has a
Regression analysis is performed on all simulation results positive or negative value. This indicates that the regression
with the corresponding results of SDP-M1 model. It is to line does not pass through the origin and there is a posi-
be noted here that any simulation yields both final stor- tive or negative shift in the value of the dependent variable,
age state xt+1 and release ut for period t. Hence, regression with respect to the independent variable. In other words,
is done i) between final storage of ANN model and final independent variable and predicted dependent variable will
storage of SDP-M1 model. Such regression analysis yields differ by a constant value. Results of the regression analy-
three parameters: slope of the regression line, m; intercept sis, objective function values: OFV-1 (sum of the square of
of the regression line with the ordinate, b; and coefficient deviation of release from target as obtained from SDP-M1
model) and OFV-2 (sum of the square of deviation of release
from target as obtained from ANN models) are presented in
Tables 3 and 4.

13
 60   Page 12 of 14 Sustainable Water Resources Management (2020) 6:60

Fig. 5  Trajectories for Reservoir 3 for different flow conditions

Model selection criteria yielded comparatively better values of these parameters. So,
for this case, ANN-M3 may be considered as the preferred
Out of the five models developed, a particular model is model.
selected as the preferred one after examining objective func- Figure 3 shows that storage trajectories obtained from
tion value and regression parameters. Usually, the model ANN-M3 and SDP-M1 model are very close and similar,
that yields the lowest objective function value (OFV-2) is although during 6th period to 18th period and 39th to 52nd
considered as the preferred model. Closeness of the mod- period, ANN-M3 resulted higher storage trajectory, indicat-
el’s result with the SDP-M1 result is also examined through ing less releases during these periods. This may be partly
the regression parameters. If it is observed that objective attributed due to impose of penalty both for deficit and for
function values (OFV-2) of more than one ANN model are surplus. It is also observed that for Case 2, Case 3 and Case
almost equal, then the model that yields the highest r value 4, ANN-M3 yielded the best results in terms of the objec-
is selected as the most preferred model. Again, if r values of tive function values and these values are 235, 6.17 and 2.98,
more than one models are equal, then the model that yields respectively, which are lowest among other values. It may
m value close to 1 and lowest absolute b value is taken as the be noted here that for Case 4, OFV-1 is higher than OFV-2,
most preferred model. which is feasible, because in SDP-M1 the expected value of
the objective function is minimized, which is computed over
Results and discussion a long period of operation. In this case, the average value of
the objective function is actually the objective function value
For Reservoir 1, Table 3 shows the regression parameters for 1 year only, which may be more than the expected value.
for Case 1; the objective function value obtained from SDP- Figure 3 also shows the closeness of trajectories obtained
M1 (OFV-1) is 1189. Among all other ANN models, ANN- from ANN-M3 and SDP-M1. Based on the above observa-
M3 yielded the lowest objective function value (OFV-2) of tions, ANN-M3 may be considered as the preferred model
1297. Observing the m, b, r values, it is found that ANN-M3 for Reservoir 1.

13
Sustainable Water Resources Management (2020) 6:60 Page 13 of 14  60

Fig. 6  Trajectories for Reservoir 4 for different flow conditions

For Reservoir 2, it is that for Case 1, ANN-M3 yielded Conclusion

the lowest OFV-2 value and is equal to 582115. Also, r
and m values of ANN-M3 are higher than the other mod- As no standard rules are available to identify the input
els. Hence, ANN-M3 is preferred. For Case 2, Case 3 and variables for an ANN model, exhaustive search is to be
Case 4 also, ANN-M3 yielded the lowest values of OFV-2 as made with different possible combinations of candidate
2891, 374227 and 8579, respectively. Hence, ANN-M3 3 is variables. For DV system, it is found that the model which
the preferred model for Reservoir 2. Figure 4 shows that for included the current period’s storage and inflow as well as
Case 1, Case 2 and Case 3, ANN-M3 and SDP-M1 trajecto- previous period’s inflow as input variable yielded better
ries are very close to each other. In case of Case 4, SDP-M1 results. Based on the performances of the ANN models,
releases more water than ANN-M3. So for Reservoir 2, it in terms of average value of the objective function and
can be concluded that ANN-M3 is the preferred model. results of regression analysis with SDP-M1 models, it may
Table 4 shows that for Reservoir 3, ANN-M3 produced be concluded that ANN model is a viable alternative to
lowest OFV-2, 125.2, 243.5, 19.72 and 4.82 for Case 1, Case SDP model. Hence, in single-reservoir ANN models, for
2, Case 3 and Case 4, respectively. For Reservoir 4, ANN- all reservoirs, ANN-M3 has emerged as the most preferred
M3 also yielded lowest OFV-2 for 125., 314.5, 1.17 and 2.59 one and trajectories obtained can be taken as initial solu-
for Case 1, Case 2, Case 3 and Case 4, respectively. Hence, tion. So for preliminary analysis, prior to more detailed
ANN-M3 is selected as the preferred model for both Reser- stochastic optimization study, ANN-M3 can be useful.
voir 3 and Reservoir 4. Figure 5 and Fig. 6 show that ANN- ANN-M3 has xt  , yt as well as yt−1 in the input variables
M3 and SDP-M1 trajectories are erratic in nature. This may and xt+1 in the output variable. The presence of yt−1 indi-
be due to the variable demand placed on Reservoir 3 and cates that serial correlation of inflows has influence on
Reservoir 4. In these plots, xt indicates non-dimensionalized
′
the system output. This selection and inflow behavior is
storage, obtained by taking the ratio of the actual storage to of course system dependent and applicable to the present
the maximum available storage volume. The negative value system under study.
of xt indicated storage below the dead storage level.
′

13
 60   Page 14 of 14 Sustainable Water Resources Management (2020) 6:60

References inflows. J Water Resour Plan Manag Div Am Soc Civ Eng
117(1):28–36
Jain SK, Das A, Srivastava DK (1999) Application of ANN for reser-
ASCE Task Committee on Application of Artificial Neural Networks
voir inflow prediction and operation. J Water Resour Plan Manag
in Hydrology (Rao Govindaraju) (2000a) I: Preliminary concept.
Div Am Soc Civ Eng 125(5):263–271
J Hydrol Eng 5(2):115–123
Labadie JW (2004) Optimal operation of multireservoir systems: state-
ASCE Task Committee on Application of Artificial Neural Networks in
of-the-art review. J Water Resour Plan Manag Div Am Soc Civ
Hydrology (Rao Govindaraju) (2000b) II: Hydrologic application.
Eng 130(2):93–111
J Hydrol Eng 5(2):124–137
Loucks DP, Stedinger JR, Haith DA (1981) Water resources systems
Bellman R (1957) Dynamic Programming. Princeton University,
planning and analysis. Prentice-Hall, New Jersey
Princeton, New York
Maier HR, Dandy GC (2000) Neural networks for the prediction and
Bellman R (1962) Dynamic programming. Princeton University,
forecasting of water resources variables: a review of modeling
Princeton
issues and applications. Environ Model Softw 15(1):101–124
Cancelliere A, Giuliano AA, Rossi G (2004) A neural networks
Paudyal GN, Shrestha DL, Bogardi JJ (1990) Optimal hydropower sys-
approach for deriving Irrigation reservoir operating rules. Water
tem configuration based on operational analysis. J Water Resour
Resour Manag 16(1):71–88
Plan Manag Div Am Soc Civ Eng 116(2):233–246
Celeste AB, Koichi S, Kodata A (2008) Integrating long- and short-
Raman H, Chandramouli V (1996) Deriving a general operating policy
term reservoir operation models via stochastic and deterministic
for reservoirs using neural network. J Water Resour Plan Manag
optimization: case study in Japan. J Water Resour Plan Manag Div
Div Am Soc Civ Eng 122(5):342–347
Am Soc Civ Eng 134(5):440–448
Ravikumar V, Venugopal K (1998) Optimal operation of South Indian
Chaves P, Chang FJ (2008) Intelligent reservoir operation system
irrigation systems. J Water Resour Plan Manag Div Am Soc Civ
based on evolving artificial neural networks. Adv Water Resour
Eng 124(5):264–271
31(6):926–936
Teixeira AS, Marina MA (2002) Coupled reservoir operation-irri-
Chandramouli V, Deka P (2005) Neural network based decision
gation scheduling by dynamic programming. J Irrig Drain Eng
support model for optimal reservoir operation. Water Resour
128(2):63–73
Manag 19(4):447–464
Vedula S, Kumar DN (1995) An integrated model for optimal reser-
Chandramouli V, Raman H (2001) Multireservoir modeling with
voir operation for irrigation of multiple crops. Water Resour Res
dynamic programming and neural networks. J Water Resour
32(4):1101–1108
Plan Manag Div Am Soc Civ Eng 127(2):89–98
Vedula S, Majumdar PP (1992) Optimal reservoir operation for irriga-
Chandramouli V, Kuppusamy KA, Manikandan K (2002) Study
tion of multiple crops. Water Resour Res 28(1):1–9
on water sharing in a multi-reservoir system using dynamic
Wang Y, Chang J, Huang Q (2010) Simulation with RBF neural net-
programming and neural network model. J Water Resour Dev
work model for reservoir operation rules. Water Resour Manag
18(3):425–438
24(12):2597–2610
Coerver HM, Ruten MM, van de Giesen NC (2018) Deduction of
Wurbs WA (1993) Reservoir-system simulation and optimization
reservoir operating rules for application in global hydrological
models. J Water Resour Plan Manag Div Am Soc Civ Eng
models. Hydrol Earth Syst Sci 22:831–851
119(4):455–472
Damodar Valley Reservoirs Regulation Manual (1995) Central Water
Yakowitz S (1982) Dynamic programming applications in water
Commission. Government of India, New Delhi, India
resources. Water Resour Res 18(4):673–696
Esmaeil-Beik S, Yu YS (1984) Optimal operation of multipurpose
Yeh WW-G (1985) Reservoir management and operations models: a
pool of Elk City Lake. J Water Resour Plan Manag Div Am Soc
state-of- the-art review. Water Resour Res 21(12):1797–1818
Civ Eng 110(1):1–14
Fayaed S, El-Shafie A, Jafar O (2013) Integrated artificial neu-
Publisher’s Note Springer Nature remains neutral with regard to
ral network (ANN) and stochastic dynamic programming
jurisdictional claims in published maps and institutional affiliations.
(SDP) model for optimal release policy. Water Resour Manag
27(10):3679–3696
Huang W-C, Harboe R, Bogardi JJ (1991) Testing stochastic dynamic
programming models conditioned on observed or forecasted

13