Estimation of Prediction Intervals for Artificial Neural Network-Based Rainfall-Runoff Modeling

Abstract

AbstractIn this study, the uncertainty associated with artificial neural network (ANN) for rainfall-runoff modeling on daily and monthly scales was evaluated by Prediction Intervals (PIs) for two watersheds, West Nishnabotna River basin in the United States and Lighvanchai River basin in Iran. Upper Lower Bound Estimation (LUBE) method was applied to construct the PIs. Furthermore, the Bootstrap technique, as a benchmark model, was applied to evaluate the uncertainty of ANN. In the LUBE method, the ANN is trained by minimizing the objective function via the genetic algorithm optimization method, and the objective function contains the width and coverage criteria of PIs evaluation. PIs coverage probability and PIs width values, respectively were up to 20% higher and 30% lower in the LUBE method compared to the Bootstrap method. Moreover, the CWC measure was considerably lower for LUBE method than the Bootstrap method. Also, the Lighvanchai basin modeling showed more accurate results than the West Nishnabotna River basin, which is due to four well defined regular seasons of the Lighvanchai basin.KeywordsRainfall-runoffPrediction IntervalArtificial neural networkBootstrapUpper Lower Bound Estimation