IMPROVEMENT OF DEEP LEARNING RAINFALL FORECAST AND DETAILED EVALUATION

Abstract

 Short-term rainfall prediction using deep learning has gained attention recently. However, the evaluation of forecast accuracy has been limited, and comparisons with operational models are rare. In this study, we improved an existing forecasting method for rainfall prediction in Japan up to 6 hours ahead and comprehensively evaluated the results. The proposed method outperforms operational forecasts for precipitation exceeding 50 mm h^-1 and 5 mm h^-1, and the results indicated that the possibility that the model learn the physical phenomena, which the operational model could not consider. However, even during the winter season, there are events caused by low-pressure-induced rainfall and other similar characteristics to summer rainfall, which can be predicted. These findings suggest the need for further discussions on constructing an accurate model and building the dataset appropriately.