A comparative study of artificial neural network (MLP, RBF) and support vector machine models for river flow prediction

Abstract

This study investigates the applicability of multilayer perceptron (MLP), radial basis function (RBF) and support vector machine (SVM) models for prediction of river flow time series. Monthly river flow time series for period of 1989–2011 of Safakhaneh, Santeh and Polanian hydrometric stations from Zarrinehrud River located in north-western Iran were used. To obtain the best input–output mapping, different input combinations of antecedent monthly river flow and a time index were evaluated. The models results were compared using root mean square errors and the correlation coefficient. A comparison of models indicates that MLP and RBF models predicted better than SVM model for monthly river flow time series. Also the results showed that including a time index within the inputs of the models increases their performance significantly. In addition, the reliability of the models prediction was calculated by an uncertainty estimation. The results indicate that the uncertainty in the SVM model was less than those in the RBF and MLP models for predicting monthly river flow.