Tropical Cyclone Precipitation Nowcasting Based on Flow Matching Model With Numerical Wind Field Constraints

Abstract

Tropical cyclones (TCs) making landfall often produce extreme precipitation with complex spatiotemporal patterns, posing major challenges to both deep learning models and numerical weather prediction (NWP) systems. This study introduces Rainflow, a conditional flow matching (CFM) based generative model for high‐resolution TC precipitation nowcasting. For the North Atlantic data set, Rainflow predicts precipitation at 0.01 ° ${}^{\circ}$ spatial and 10‐min temporal resolution over 6‐hr lead times. By modeling the residual transition from observed to future precipitation sequences, Rainflow jointly captures precipitation motion and intensity evolution within a unified framework. To improve physical consistency, 500 and 850 hPa wind fields from NWP forecasts are incorporated as dynamic constraints. Experiments show that Rainflow substantially outperforms the operational High‐Resolution Rapid Refresh (HRRR) system and the state‐of‐the‐art NowcastNet, especially for extreme precipitation intensities. Additional experiments on Western Pacific TC cases further demonstrate the potential generalizability of Rainflow under a shorter 3‐hr nowcasting setting.