Real-time nowcasting of sudden heavy rainfall using artificial neural network and multi-parameter phased array radar

Abstract

Sudden localized heavy rainfall events, capable of disrupting daily life and damaging infrastructure, are becoming more frequent. Their nowcasting (very short-term forecast) requires higher spatiotemporal (4D) resolution than conventional radars, and effective 4D methods to extrapolate the vertical development of convective systems. This study evaluates the performance of a new system that generates 10-minute lead-time precipitation nowcasts in real time, which are used by a publicly available smartphone application to issue heavy rainfall warnings. Dense 4D observations from new Multi-Parameter Phased Array Weather Radars (MP-PAWR) in Saitama, Osaka, and Kobe (Japan) are extrapolated using an Artificial Neural Network (ANN4D), which has demonstrated high performance in forecasting the sudden onset of precipitation in Saitama, prior to 2020. The study demonstrates that, despite using the same ANN4D instance, the system generates reliable nowcasts, generalizes well to new locations and years, and that performance is enhanced by a post-ANN4D procedure for mitigating false rainfall predictions. ANN4D outperforms a 4D Eulerian model (TREC4D) in predicting convective rainfall onset, while TREC4D is more efficient for well-developed rainfall. The study identifies minor issues, like the need to expand ANN4D's vertical range, and highlights the next major step: integrating ANN4D with TREC4D to exploit their complementarity.