In the present study, two different Artificial Neural Networks (ANN) were developed for Dharoi reservoir, Gujarat, one for flood control operation and another for conservation operation. In recent years, ANNs are increasingly being used to predict water resources variables particularly in the operation of reservoirs, the most important elements of complex water resources systems, constructed for spatial and temporal distribution of water.

The 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 modeling and analysis methods involve allocating storage capacity and streamflow between multiple uses and users. The models developed to provide operating rules for reservoirs are classified as simulation models, optimization models and a 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.

An ANN can represent any arbitrary non-linear 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 depend the output neurons (for example, release).

Generally sigmoid function is applied as an 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.

Seven different combination 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 highest coefficient of correlation and sum of squared errors, was selected.

Download the report here: