Suitability of different neural networks in daily flow forecasting

Abstract

Alternative forms of neural networks have been applied to forecast daily river flows on a continuous basis with the purpose of understanding how recent architectures like ANFIS, GRNN and RBF compare with traditional FFBP when monsoon-fed rivers involving significant statistical bias are involved. The forecasts are made at a location called Rajghat along river Narmada in India. Division of yearly data into four seasons and development of separate networks accordingly was found to be more useful than a single network applicable for the entire year. When a variety of error criteria were viewed together the most satisfactory network for all seasons was the radial basis function, which showed better performance then FFBP, GRNN and ANFIS. The FFBP network was found to be equally acceptable as the RBF in seasons other than the monsoon. Generally the peak flows were more satisfactorily modeled by the RBF than FFBP, GRNN and ANFIS. The relatively simpler handling of data non-linearity in FFBP was more attractive than complex ones of ANFIS and GRNN. The representative statistical model, namely response surface method, yielded highly unsatisfactory results compared to any ANN model involved in this study, confirming that the complexity of ANNs is really necessary to model daily river flows.