TR(BR)-5/99-2000 : Application of artificial neural networks in flood studies of Ajay river basin

Abstract

In recent years there has been a growing interest in artificial neural networks (ANN) which operate in a manner analogous to that of biological neurons system. The ANN technique offers several advantages over conventional computing methods. Most important among these is ANN's ability to generalise a solution to a problem from a set of example. The ability of ANN to model nonlinear events advocates their use in hydrology to model various hydrological events which are dominantly nonlinear in nature. In the study a comprehensive review of ANN technique is presented. The study shows how and why such a system works. A case study has been performed for Ajay river basin to demonstrate the applicability of ANN techniques in model development.

In the case study, an Artificial Neural Networks (ANN) model is developed for Ajay river basin up to Sarath gauging site of Jharkhand (South Bihar) for forecasting floods which are generally sudden, flashy and of short duration and require solutions in quick succession. Artificial Neural Network (ANN) is important, where the time required to generate solutions is critical and is also capable of modelling nonlinear relationship between rainfall-runoff process as compared to other methods which assume the linear relation. The ANN model is developed utilising the available limited rainfall and runoff data of the Ajay river basin up to Sarath. The model is used to forecast 6-hr ahead runoff at Sarath. In the process of model development the 1-hourly rainfall-runoff data are divided into two groups. One part is used to find the network connection weights to represent the relationship between rainfall and runoff. The back propagation algorithm is used to optimise the network connection weights. The developed model is then validated on the second set of data. The performance of the model on training and test data has been evaluated cases on root mean square error (r.m.s.e.) index. The result shows that ANN based rainfall-runoff model can develop a nonlinear relationship for rainfall-runoff process even when the training data are limited and contain noise components. The trained ANN model is then applied to problems other than those used for training for validation. The validated model can be used to issue 6-hr ahead forecast at Sarath. However the performance of the model can be significantly improved if more rainfall-runoff events, with low noise components are made available for the study.