Regularized Back-Propagation Neural Network for Rainfall-Runoff Modeling

Abstract

In this study, we applied regularized back-propagation neural network (BPNN), which made use of a performance function different from normal BPNN, to predict daily flow. On the other hand, Broyden-Fletcher-Goldfarb-Shanno (BFGS) -algorithm-based BPNN was also used to compare its prediction performance with that of regularized BPNN. From 1979 to 1998, precipitation and stream flow data in Xitiaoxi watershed for 20 years were collected. All these data were divided into 2 sets: one was the training set (1979-1988), and the other was the testing set (1989-1998). The mean absolute error (MAE), mean square error (MSE) and coefficient of efficiency (CE) were used to evaluate the performance of these two algorithms. The results indicated that regularized BPNN could enhance generalization ability and avoid over fitting effectively, and it outperformed BFGS-algorithm-based BPNN during training and testing stages. From this study, it could be found that regularized BPN is appropriate for rainfall-runoff modeling due to its simple structure and high accuracy.