HINDCAST OF TORRENTIAL RAINFALL EVENTS IN KAGOSHIMA CITY USING LSTM

Abstract

In this study, we constructed a long short-term memory (LSTM) model to hindcast/forecast the torrential rainfall events occurred in Kagoshima city in July 2018 and July 2020. In this model, meteorological observation data drived from the ground stations of Japan Meteorological Agency (JMA) was used as training data. Our LSTM model could hindcast the time variation of the rainfall event; however, it tended to underestimate the observed rainfall event, and in some cases the start times of the precipitation events were late. To further improve the prediction accuracy of our model, it was necessary to improve the reproducibility of the magnitude of the precipitation amplitude. Therefore, we selected training data using cluster analysis in order to remove training data that has negative impact on the hindcast/forecast of torrential rainfall events. This removal significantly improved the accuracy of our LSTM model compared to that without the removal. In addition, our model hindcasts the torrential rainfall events mentioned above more accurately compared with MesoScale model derived from JMA and regional climate model. However, the late forecast of start times of precipitation was no improved by data selection using cluster analysis.