2021-11-06[Event Report]Data Assimilation Seminar November 5, 2021
Prof. Philippe Baron presented a study of 3D recurrent neural networks for very short-term prediction of torrential rains with a Multi-Parameter Phased-Array Radar (MP-PAWR). His work is based on the extrapolation MP-PAWR observations to predict in real-time short-lived convective rains with a convGRU3D based model. He had promising results for 10 min lead-time but low performance for 20 min lead-time. He also highlighted the strong variation of the performance with respect to the types of rain. In this work, he obtained similar results as 3D-RNN of Osaka Univ which uses different training strategy
and architecture. As a next step, he will focus on improving high intensity prediction accuracy as well as on reducing small rain prediction artifacts.
This project is very practical in terms of emulating heavy physical models with limited computation resources using neural networks.