[M-IS09] Effects of lightning, severe weather and tropical storms

[MIS09-P01] Machine Learning Prediction of Precipitation in Metro Manila, Philippines

★Invited Papers

*Akira Noda¹, Yukihiro Takahashi², Hisayuki Kubota², Kenichi Fukui³, Mitsuteru Sato² (1.Department of Earth and Planetary Sciences, Graduate School of Science, Hokkaido University., 2.Graduate School of Science, Hokkaido University, 3.The Institute of Scientific and Industrial Research, Osaka University)

Keywords:extreme rainfall, machine learning, nowcasting

It is difficult to accurately predict the occurrence, spatiotemporal changes, and absolute amount of torrential rainfalls. One of the reasons is that it is very hard to detect the highly localized heavy rainfall area and the rapid development of the associated thunderclouds using the conventional weather observation instruments and networks. In Manila, Philippines, where the meteorological observation infrastructure is relatively weak, large-scale disasters such as heavy rains and floods caused by monsoons and typhoons occur every year and there is an urgent need to establish cost-effective weather forecasting technology. Thus, we developed the automatic weather and lightning observation system called P-POTEKA and deployed it in Metro Manila. So far 35 P-POTEKA units have been installed in Metro Manila and they continue the acquisition of meteorological data every minute. While AMeDAS (The Automated Meteorological Data Acquisition System) in Japan is deployed with the average interval of 17 km, P-POTEKA in Metro Manila is deployed with the average interval of 2-3 km, which is the observation network with the world’s highest spatial and temporal resolution suitable for monitoring extreme weather. Using the rainfall and other weather data obtained by 35 P-POTEKA units, 50×50 grid data was created by interpolating data using gaussian process regression. By training these sequential rainfall and weather data on a machine learning model called as ConvLSTM (Convolutional Long-Short Term Memory), the distribution and the hourly rainfall amount from the present to 1 hour later with the 10 minutes intervals are predicted using the observation data in the past 1 hour. The root mean squared error (RMSE) and R2 score of the observed rainfall data and prediction data were calculated in order to evaluate the performance of the machine learning method. It is found that the rainfall prediction using ConvLSTM is relatively accurate up to 30 minutes from the present, but the prediction accuracy is getting worse after 30 minutes from the present. Although this model predicted well the weakening of rainfalls, the model could not accurately predict the occurrence or the intensification of rainfalls. At the presentation, the details of the machine learning method using ConvLSTM and the prediction results of precipitation in Metro Manila will be shown.