Development of effective and efficient rainfall‐runoff models using integration of deterministic, real‐coded genetic algorithms and artificial neural network techniques

Abstract

Many researchers have reported about the problems in modeling low‐magnitude flows while developing artificial neural network (ANN) rainfall‐runoff models trained using popular back propagation (BP) method and have suggested the use of alternative training methods. This paper presents the results of a new approach employing real‐coded genetic algorithms (GAs) to train ANN rainfall‐runoff models, which are able to overcome such problems. The paper also presents a new class of models termed gray box models that integrate deterministic and ANN techniques for hydrologic modeling. Daily rainfall and streamflow data from the Kentucky River watershed were employed to test the new approach. Many standard statistical measures were employed to assess and compare various models investigated. The results obtained in this study demonstrate that ANN rainfall‐runoff models trained using real‐coded GA are able to predict daily flow more accurately than the ANN rainfall‐runoff models trained using BP method. The proposed approach of training ANN models using real‐coded GA can significantly improve the estimation accuracy of the low‐magnitude flows. It was found that the gray box models that are capable of exploiting the advantages of both deterministic and ANN techniques perform better than the purely black box type ANN rainfall‐runoff models. A partitioning analysis of results is needed to evaluate the performance of various models in terms of their efficiency in modeling and effectiveness in accurately predicting varying magnitude flows (low, medium, and high flows).