Prediction of runoff and sediment yield for Damodar and Mayurkashi basin using artifical neural network and regression analysis

The estimation of runoff and sediment yield is needed in various hydrological studies. Usually a stream gauge is unavailable at the site of interest. Therefore new information technology tools like artificial neural network (ANN) and traditional methods like multiple linear regression analysis with readily available catchment and climatic characteristics can provide a practical solution. In the first ANN model the neural network was trained with catchment area and annual rainfall as input and estimated runoff as the desired output. The best ANN architecture was selected on the basis of mean absolute error, mean square error and maximum correlation coefficient. The ANN architecture 2-4-1 (two four and one neurons in the input, hidden and output layers, respectively) was found best in training and testing. In the same way in the second ANN model network were trained with catchment area and annual rainfall as input and estimated sediment yield as the desired output. The ANN architecture 2-6-1 was found as best in this case. Same data set, which was used in ANN, was used for developing Multiple Regression (MREG) models and also for validation. The results of MREG equations that relate the runoff and sediment yield with area of catchment and annual rainfall have been compared with the results obtained through artificial neural network (radial basis function). Performance evaluation studies indicated that prediction ability of ANN model is better than multiple non linear regression model. This is certainly an advantage over traditional method. Thus, the results established that ANN can predict runoff and sediment yield more accurately as compared to the conventional methods.