PstpNet: Dual-Branch Physics-Guided U-Net With Multiscale Spatiotemporal Disentanglement for Precipitation Nowcasting

Abstract

Precipitation nowcasting remains challenging because precipitation systems exhibit strong nonlinear evolution, rapid development, and complex long-range spatiotemporal dependencies. Existing deep learning methods are still limited in characterizing precipitation dynamics and improving model interpretability. To address these limitations, we propose PstpNet, a physics-informed nowcasting model. Inspired by the continuity equation, PstpNet constructs an advection-residual decomposition framework, in which a motion branch and a residual branch separately model advective transport and nonconservative intensity changes. Multiscale feature fusion is further introduced to enhance the representation of precipitation structures at different spatial scales. In addition, the TM module embeds historical temporal evolution information into the physical decomposition process, while the physics-guided fusion module adaptively coordinates advection and intensity-residual processes, thereby improving spatiotemporal consistency and physical interpretability in long-lead forecasting. Experimental results on the KNMI-NL50 and SEVIR datasets show that PstpNet achieves the best or near-best performance across threshold-based metrics, with strong high-intensity detection performance at the 10 mm/h rainfall threshold on KNMI-NL50 and the VIL 181 threshold on SEVIR.