Optimization of neural network structure for radial basis function network for simulation of hydrological processes

The results of optimization of neural network structures of Radial Basis Function Network (RBFN) model to predict sediment yield as a function of daily rainfall and stream flow during rainy season for a watershed area in India, has been presented in this paper. Various combinations of shuffled data set of different time lags were developed and compared for their ability to make a reliable sediment yield prediction. Measured data with 736 patterns of input-output vector was divided into three sets: 391 patterns for training, 215 patterns for testing and the remaining 130 patterns for validation. The root mean square error (RMSE) and Correlation coefficient (R) were determined to ascertain the model performance. The momentum rate, number of nodes at hidden layer, linear coefficient, learning rule and transfer functions were optimized based on lowest RMSE and highest R values. The RBFN model yielded better performance at 8 numbers of hidden layer node, momentum rate of 0.40, linear rate of 0.60, Delta learning rule and TanH transfer function. The RBFN model performed satisfactorily with R and NSE values of 0.98 and 0.97 during calibration, and 0.91 and0.91 during validation periods, respectively.