Estimation of Reservoir Storage Using Artificial Neural Network (ANN)

  • P. SatishEmail author
  • H. Ramesh
Conference paper


The rapid growth in population increases water demand thus resulting in scarcity of water which is due to improper management rather than lack of resources. Reservoir is the most important source for surface water. So, reservoir storage plays a crucial role in efficient reservoir management. Artificial neural network (ANN) is capable of simulating reservoir storage capacity. So, in the present work five different feed forward back propagation ANN models by varying number of hidden layer neurons were developed for estimation of Harangi reservoir storage, Karnataka, India. The first 2 years (2010–12) data was used for supervised training and remaining data (2013–14) was used in prediction. The predictive accuracy using the statistical parameters like correlation coefficient (R) and mean absolute percentage error (MAPE) were found within the acceptable limit. Result shows that, ANN model with five hidden neurons (i.e., network architecture of 6-5-1) is performing well compared to all other models for prediction of reservoir storage estimation.


Artificial neural network Reservoir Feed forward back propagation Prediction Storage 

1 Introduction

Extreme spatial and temporal variation of rainfall results in availability of surface water unpredictable. To cope this, reservoirs are constructed to store the water for future purpose and here comes the necessity of estimation of that storage for many reasons like reservoir operating policy, to know the capacity of reservoir, flood control and many more. This needs real-time physical parameters on continuous basis. Artificial neural network is a computational model which works same as brain and having structure similar to biological neural network. It has a capability of correlating the input and output of any system.

Better understanding of input and output variables from statistical analysis is required prior to network modeling which results in coherent network design. A proper training-based neural network model is able to acquire the physical relation between the variables and may generate better results than conventional prediction techniques. Suitable ANN modeling is always advantageous in complex systems when compared with conventional modeling techniques [1]. Artificial neural networks have been accepted as a potentially useful tool for modeling various linear and complex nonlinear systems. In the hydrological forecasting and prediction context, ANNs have also proven to be a productive alternative to conventional methods for rainfall and stream flow forecasting, Groundwater modeling and reservoir operation [2]. Feed forward network having input, output and one hidden layer with sigmoid activation function to the monthly inflow data series and their performances were compared with an autoregressive integrated moving average time-series model. Jain et al. [3] concluded that the high flows are modeled better through neural networks. Optimal operating policies of a reservoir have been derived in deterministic and stochastic frame work as well as using ANN with five different network models by varying input values. Out of five the network which contains the current storage, current inflow and previous inflow as input variables and future storage as output (target) variable gives best result [4]. For important planning, design and management activities, we need accurate forecasted variables which affect the performance of the system. So, we need to choose suitable model for forecasting [5]. Forecasting of a river stage contains major input information for water resources systems planning and management. If such a forecasting is done on a continuous basis in the given year would help in providing a warning of the impending flood during high river stage and would assist in controlling reservoir outflows during the low river stage [6].

For this study, the most acceptable learning approach, i.e., supervised learning in training and feed forward back propagation ANN network model was used. Three years hydrologic data of Harangi reservoir was used. First two years (2010–12) of the data is used for training and remaining untrained available data (2013–14) is used to compare predicted data. Well accepted range of statistical parameters ensures the best model fit for this work.

2 Study Area

The focused area is Harangi reservoir. It is geographically lies between latitude 12°-29′-30″ North and longitude 75°-54′-20″ East. Near Hudgur village, Somwarpet taluk in Kodagu district in the Indian state of Karnataka. The reservoir is formed by a masonry dam built across the river Harangi. This is a major tributary to river Cauvery and has its origin in Pushpagiri hills, Coorg district situated on its western part and separated from Bhagamandala, the source of Cauvery. The Location of Harangi reservoir is shown in below Fig. 1.
Fig. 1

Location of Harangi reservoir

2.1 Data Used

Data used in this work, contains hydrologic data including initial storage, inflow, irrigation sluice, river sluice, spillway, and evaporation loss for three years (2010–2012, 2013–2014) on daily basis of the Harangi reservoir, Hudgur village, Somwarpet taluk, Kodagu district of Karnataka, India. Data for the study was collected from the Water resources department, Government of Karnataka.

3 Theory and Methodology

Artificial neural network is a computational model which works same as brain and having structure similar to biological neural network. It has a capability of correlating the input and output of any linear and complex nonlinear system. The purpose of the neural network is to unfold the given task similar to brain would, although several neural networks are more abstract. Neural networks are still several orders of magnitude less complex than the human brain and closer to the computing power of a worm.

In feed forward back propagation network Input vector will propagate from input layer to the output layer through hidden layers. The output of the network is then compared with the desired output, and an error value is calculated for each of the neurons in the output layer. The error values are then propagated backwards from output layer to the input layer. Using this error values the weights are subjected to change to minimize the loss function.

To calculate the loss function gradient, a known desired output for each input is necessary. So, it is therefore usually considered to be a supervised learning method. It is a generalization of the delta rule to multilayered feed forward networks, made possible by using the chain rule to iteratively compute gradients for each layer.

Tan-sigmoid and pure line functions have been used as activation function in artificial neural networks more often. Basic artificial neural network architecture is as shown below Fig. 2.
Fig. 2

Basic artificial neural network architecture

The methodology used in this work starts with the collection of the reservoir data and preprocessing of that data. Using toolbox called “nntool” in MATLAB was chosen for this work. Data was divided into three sets (i.e., Training, Testing & Validation). These data sets were imported to MATLAB and developed different multilayered feed forward back propagation networks by varying number of hidden neurons and Levenberg-Marquardt as a training algorithm which is highly recommended supervised algorithm with six inputs such as Initial storage, Inflow, Irrigation sluice, River sluice, Spillway and Evaporation loss, and final storage as single Output/Target. Supervised training needs better distinction in input and output parameters. In this case it comes from the well established water balance Eq. (1) gave the clear understanding of relation between input and output parameters.
$$S_{t + 1} = S_{t} + I_{t} - E_{t} - R_{t} ,$$
where St+1 = Future Storage; It = Current Inflow; St = Current Storage; Et = Evaporation loss; t = Time increment; Rt = Outflow (i.e., irrigation, reservoir/power sluice and spillway).
Tan-sigmoid and pureline functions are used in input layer and hidden layer respectively as an activation function which helps in proper training of the network and gives the desired results. First 2 years (2010–12) data set used in training and remaining one year (2013–14) data set used in testing/prediction. Accuracy of network was found using the statistical parameters like correlation coefficient (R) and mean absolute percentage error (MAPE). Flowchart of methodology of artificial neural network was shown below Fig. 3.
Fig. 3

Flowchart of methodology of artificial neural network

4 Results and Discussion

Five feed forward back propagation models with different network architecture are developed by toolbox in MATLAB named “nntool”. The results were obtained by taking 2 years (2010–2012) of the data for training and the remaining data (2013–2014) on daily basis for prediction. The output results were checked for accuracy of prediction using statistical parameters like, correlation coefficient (R) and mean absolute percentage error (MAPE).

The analysis of all five networks in feed forward back propagation network (FFBN) was done after training, validation and testing. The results including some statistical parameters [i.e., correlation coefficient (R) and mean absolute percentage error (MAPE)] are presented in Table 1.
Table 1

Network results including some statistical parameters

S. No.

Feed forward back propagation network (FFBN)

No. of hidden neurons

Correlation coefficient (R)

Mean absolute percentage error (MAPE) (%)





















The bold values are implies that the network architecture corresponding to those values gave best results and suitable in this work

Based on the above results (Table 1), network with five hidden neurons of the feed forward back propagation performing good with correlation coefficient of 0.9882 and mean absolute percentage error (MAPE) of 5.225%. The observed and predicted storage of best network (6-5-1) is plotted as shown in Fig. 4.
Fig. 4

Plot of observed and predicted storage (FFBN with 5 hidden neurons) of Harangi reservoir for June’13–May’14

From Fig. 4 it is observed that the predicted storage is nearly equal to observed storage throughout the year. Hence the feed forward back propagation network model with five hidden neurons is best suitable for present study.

Proper understanding of input and output parameters results in better training which leads to the good results/prediction.

The predicted storage will be useful for many purposes. Such as, to know the capacity of reservoir, reservoir operation policy and flood controlling and management studies.



The authors are thankful to Water resources department, Karnataka for providing all the necessary data. Head of the department, all the teachers, family and friends who are helped us to complete this work.


  1. 1.
    Mustafa, M.R., Isa, M.H., Rezaur, R.B.: Artificial neural networks modeling in water resources engineering: infrastructure and applications. World Acad. Sci. Eng. Technol. 6(2), 341–349 (2012)Google Scholar
  2. 2.
    Chang, F.J., Chang, Y.T.: Adaptive neuro-fuzzy inference system for prediction of water level in reservoir, Adv. Water Resour. 29(1), 1–10 (2006).
  3. 3.
    Jain, S.K., Das, A., Srivastava, D.K.: Application of ANN for reservoir inflow prediction and operation. J. Water Resour. Plann. Manage. 125(5), 263–271 (1999)CrossRefGoogle Scholar
  4. 4.
    Safayat Ali, S.: Optimal operation of single reservoir using artificial neural network. Int. J. Civ. Eng. Technol. 6(6), 124–132 (2015)Google Scholar
  5. 5.
    Othman, F., Naseri, M., Reservoir inflow forecasting using artificial neural network. Int. J. Phys. Sci. 6(3), 434–440 (2011).
  6. 6.
    Thirumalaiah, Deo M.C.: River Stage forecasting using artificial neural networks. J. Hydrol. Eng. 3(1), 26–32 (1998)Google Scholar

Copyright information

© Springer Nature Singapore Pte Ltd. 2019

Authors and Affiliations

  1. 1.Applied Mechanics & HydraulicsNational Institute of Technology KarnatakaSurathkal, MangaluruIndia

Personalised recommendations