Performance Evaluation of Artificial Neural Networks for Runoff Prediction

Abstract

Spring runoff prediction in the Red River Valley, southern Manitoba, Canada, is an important issue because of the devastating effect of the flood of 1997 in that area. Increasing the accuracy of the prediction process is a practical necessity. This study looks at the artificial neural networks (ANN) technique and compares it to linear and nonlinear regression techniques. The advantages and disadvantages of the three modeling techniques are discussed. To fill the predictive accuracy evaluation gap left by the mean squared error and the mean relative error, a modified statistic, namely, pooled mean squared error, is developed and explained. The aim of this work is to show the applicability of ANN for runoff prediction and to evaluate their performances by comparing them with traditional techniques. In this study, according to the accuracy of results, the ANN models show superiority in most of the cases. However, in some situations, the performance of the other two techniques was comparable.