[A-TT29] Machine Learning Techniques in Weather, Climate, Ocean, Hydrology and Disease Predictions

[ATT29-02] Study of systematic errors on the very short-term prediction of heavy localized precipitation obtained with a 4D neural network.

*Philippe Baron^1,2、Kohei Kawashima²、Hiroshi Hanado¹、Seiji Kawamura¹、Takeshi Maesaka³、Shinsuke Satoh¹、Tomoo Ushio² (1.National Institute of Information and Communications Technology (NICT), Koganei, Japan、2.Electrical Engineering Dept., Osaka University, Japan、3.National Research Institute for Earth Science and Disaster Resilience (NIED), Tsukuba, Japan)

キーワード：Machine learning, Nowcast, Radar, Phased-Array Weather Radar, Torrential rain

Sudden heavy rainfalls occur more and more frequently in Japan during the summer. They are concentrated in restricted areas (typically of 5x5 km²) and can cause severe damage to infrastructure, often with casualties. The extrapolation of precipitation measured by radars is the conventional method for carrying out their nowcast in real time, i.e., short-term prediction on small spatial scales. However, the limit of predictability of such storms with current operational nowcasts is less than 10 minutes. This is due to the short lifetime (<15 minutes) of the individual convective cells causing the precipitation, as well as the limitations of nowcast models to properly account for the 3D nonlinear evolution of the cells.
A neural network (NN) for real-time nowcasting has been developed. It successfully predicts the onsets of sudden storms on meso-γ-scale (2-20 km) where conventional approaches fail. It uses the Multi-Parameter Phased-Array Weather Radar (MP-PAWR) operating in Saitama prefecture (Japan) which provides dense 3D observations of individual convective cells every 30 sec. The NN is a recurrent neural network enhanced with 3D spatial convolutions and it is trained with the method developed for Generative Adversarial Networks (GAN). In this study we present the latest results obtained with the model with a special focus on systematic errors found in the nowcasts. Methods to mitigate them will be discussed.