Neural Networks for Rainfall Forecasting by Atmospheric Downscaling

Abstract

Several studies have used artificial neural networks (NNs) to estimate local or regional precipitation/rainfall on the basis of relationships with coarse-resolution atmospheric variables. None of these experiments satisfactorily reproduced temporal intermittency and variability in rainfall. We attempt to improve performance by using two approaches: (1) couple two NNs in series, the first to determine rainfall occurrence, and the second to determine rainfall intensity during rainy periods; and (2) categorize rainfall into intensity categories and train the NN to reproduce these rather than the actual intensities. The experiments focused on estimating 12-h mean rainfall in the Chikugo River basin, Kyushu Island, southern Japan, from large-scale values of wind speeds at 850 hPa and precipitable water. The results indicated that (1) two NNs in series may greatly improve the reproduction of intermittency; (2) longer data series are required to reproduce variability; (3) intensity categorization may be useful for probabilistic forecasting; and (4) overall performance in this region is better during winter and spring than during summer and autumn.