Estimating runoff using feed-forward neural networks in scarce rainfall region

Abstract

This work depicts the use of two neural network techniques: (i) feed-forward back propagation network (FFBPN) and (ii) radial basis function network (RBFN), to forecast runoff as a function of rainfall, temperature and loss due to evapotranspiration. For model architecture, the criteria for evaluation are convergence of mean square error and coefficient of determination. For Tan-sig function in FFBPN, 4-2-1, 4-3-1, 4-5-1, and 4-9-1 architectures are taken into consideration for computation of performance. For Tan-sig function, the best model architecture is found to be 4-2-1 which possess MSE training value 0.000808, MSE testing value 0.004119, RMSE training value 0.028392, RMSE testing value 0.064165 and coefficient of determination for training 0.9914 and testing 0.9255. FFBPN performs the best among four networks with model architecture 4-5-1 using Log-sig transfer function. Similarly for the case of LRN, with Tan-sig function the best model architecture is found to be 4-5-1 which possess MSE training value 0.000483, MSE testing value 0.001025, RMSE training value 0.02316, RMSE testing value 0.03085 and coefficient of determination in training and testing as 0.9925, 0.9611, respectively. Overall results found that LRN performs best as compared to FFBPN for predicting runoff in the watershed. This result will help for planning, design and management of hydraulic structures in the vicinity of the watershed.