Application of Neural Networks to Runoff Prediction

Abstract

In this paper, a new method to forecast runoff using neural networks (NNs) is proposed and compared with the fuzzy inference method suggested previously by the authors (Fujita and Zhu, 1992). We first develop a NN for off-line runoff prediction. The results predicted by the NN depend on the characteristics of training sets. Next, we develop a NN for on-line runoff prediction. The applicability of the NN to runoff prediction is assessed by making 1-hr, 2-hr and 3-hr lead-time forecasts of runoff in Butternut Creek, NY. The results indicate that using neural networks to forecast runoff is rather promising. Finally, we employ an interval runoff prediction model where the upper and lower bounds are determined using two neural networks. The observed hydrograph lies well between the NN forecasts of upper and lower bounds.