Application of artificial neural network for optimal operation of a multi-purpose multi-reservoir system, II: optimal solution and performance evaluation

Abstract

This papers forms the second part of series on application of artificial neural network (ANN) for optimal operation of a multi-purpose multi-reservoir system. Optimal operating policies of a reservoir system are derived using Discrete differential dynamic programming (DDDP)-based ANN model. In ANN model development a feed-forward network with delta learning rule and back propagation algorithm is used. Neural networks have been trained using supervised learning approach. Water supply for irrigation, municipal and industrial use have been selected as objective of operation and other purposes are treated as binding constraints. Minimization of the sum of square of penalties incurred due to deviation of release from the target, is selected as the objective function. Damodar Valley (DV), a multi-purpose four reservoir system in India is used for this study. With different combination of input data, five types of ANN models are developed. Simulation has been done with 5 years (out of 1000 years) generated monthly inflow sequence as well as three types of observed historical monthly inflow sequences: maximum annual inflow year, 75% dependable inflow year and minimum annual inflow year. For simulation, total 360 monthly networks are trained and stored. ANN model: in which initial storage, current period’s inflow and previous period’s inflow are considered as input and optimal final state as output, yields lowest objective function value. Performances of the said model is computed based on modern reliability parameters, i.e., reliability, resiliency and vulnerability.