Selecting most suitable members for neural network ensemble rainfall forecasting model

Abstract

Neural network ensembles are more accurate than a single neural network because they have higher generalization ability. To increase the generalization ability the members of the ensemble must be accurate and diverse. This study presents a method for selecting the most suitable members for an ensemble which uses genetic algorithms to minimize the error function of the ensemble ENN-GA. The performance of the proposed method is compared with the performance of two widely used methods, bagging and boosting. The models developed are trained and tested using 41 years rainfall data of Colombo and Katugastota Sri Lanka. The results show that the ENN-GA model is more accurate than Bagging and Boosting models. The best performance for Colombo was obtained by ENN-GA with 14 members with RMSE 7.33 and for Katugastota by ENN-GA with 12 members with RMSE 6.25.