TR(BR)-6/99-2000 : Application of artificial neural networks (ANN) in reservoir operation

Abstract

Reservoirs, the most important elements of complex water resources systems, are constructed for spatial and temporal redistribution of water in quantity and quality. Ever increasing water demands and the difficulties associated with building new surface storage facilities envisage more efficient operation of existing reservoirs such as, improved coordination of reservoir operations and the effective use of stream flow and demand forecasts. Systems analysis has proved to be a potential tool in the planning and management of the available water resources. Reservoir system management practices and associated modelling and analysis methods involve allocating storage capacity and stream flow between multiple uses and users. The models developed to provide operating rules for reservoirs are classified as simulation models, optimization models and combination of these two models. Simulation models are used to study the reservoir system with different operating rules whereas optimization models are used to optimize the operation by considering the inflows, demands, reservoir characteristics, evaporation rates, etc., as constraints. Simulation models can also provide near optimized releases by repeated runs of different operating policies.

In recent years, Artificial Neural Networks (ANN) are increasingly being used to predict water resource variables. An ANN can represent any arbitrary nonlinear function given sufficient complexity of the trained network. Feed forward networks are generally used in ANN models. This type of ANN consists of three types of layers, namely an input layer, hidden layer(s) and an output layer. The input layer consists of number of neurons (for example, reservoir storage and inflow) on which depends the output neurons (for example, release). Generally sigmoid function is applied as activation function to provide the output. These networks are trained mostly by back propagation algorithm. The input and output neuron values are normalized between 0 and 1 before the training.

In the present study, two different neural network models were developed for Dharoi Reservoir, Gujarat: one for flood control operation and the other for conservation operation. Seven different combinations of input variables were trained for both flood control and conservation operation. The coefficient of correlation and the sum of squared errors for different network structures were compared and the combination, which gave the highest coefficient of correlation and small sum of squared errors, was selected.

The floods of 10 July 1977, 22 June 1980 and 23 July 1982 were used to evaluate the trained neural network for flood control operation. The floods were moderated as per the policy adopted in the training of the neural network and the end reservoir storage in all three floods were below revised HFL (193.60 m). So the trained neural network model can be used effectively to moderate the floods.

Two neural network models were developed for conservation operation: one with actual release for 10 daily duration and other with simulated release for monthly duration. The coefficient of correlation and the sum of squared errors were 0.609 and 5242 for neural network model with actual release for the evaluation data set. The coefficient of correlation and the sum of squared errors were 0.934 and 2134 for neural network model with simulated release for the evaluation data set. The neural network trained with the simulated release can be used to decide the release from the reservoir for conservation purposes.