EXAMINATION OF RAINFALL PREDICTABILITY BY MACHINE-LEARNING LSTM MODEL TRAINED WITH AMEDAS DATA

Abstract

 We evaluated rainfall predictability of a data-driven machine learning model, LSTM (Long short-term memory) model, which is one of the deep learning architecture and excels at learning time-series data. We configured the LSTM model for rainfall prediction and trained the model with AMeDAS ground observed data in the Kyushu region for providing every 1-hour forecast.

 In many observational sites of AMeDAS, the LSTM model predicted accurately more than persistence method (PER) and Mesoscale model (MSM) did, which is one of the physically-based numerical weather models by Japan Meteorological Agency. The three models were compared their accuracy in terms of RMSE, threat score, and prediction accuracy of beginning of rainfall. The features of the LSTM model were as follows.

 1. The prediction accuracy was higher in the eastern side and inland of Kyusyu than in the northern, southern and the western region near seacoast. That is possibly because the LSTM effectively learned the pattern that the precipitation moves from west to east, from north to south and from south to north. As a result, the better prediction was found in cold and warm frontal rain which were patterned by westerly wind.

 2. Difficulty found in outlier events such as unprecedented torrential rain which is lack of learning data.

 3. Among the various meteorological elements, only the rainfall played the important role in prediction accuracy possibly because the AMeDAS observational density was not high enough to capture cold outflow by rainfall, warm-air convergence precedent precipitation as well as surface wind were affected much by topography.