Application of bootstrap-based neural networks for monthly rainfall forecasting in Western Jilin Province, China

Abstract

Rainfall forecasting is an important pre-requisite for effectively managing and planning water resources. This study developed a generalized regression neural network (GRNN) combined with a bootstrap approach for rainfall forecasting, and the forecasting results were compared with the autoregressive model and single GRNN model. The test was performed in western Jilin Province, China with a 53-year (1957–2010) monthly rainfall time series. To obtain the good performance of GRNN model, the number of input neurons was decided by the analysis of Bayesian information criterion, and the appropriate spread was selected considering the performance of the training and testing phases. mean absolute error, root mean square error, coefficient of efficiency and R2 are employed to evaluate the performances of the forecasting models. The results showed that the bootstrap-based GRNN model performed better than single GRNN and AR models in forecasting monthly rainfall and the proposed method can improve the prediction accuracy of monthly rainfall time series, while generating uncertainty estimates of the rainfall forecasting.