Physically Consistent Global Spatiotemporal Network for Precipitation Nowcasting

Abstract

Precipitation nowcasting imposes stringent requirements on the modeling of spatiotemporal dynamics. Although traditional methods can generate precipitation nowcasting based on physical operators, they exhibit significant limitations in capturing the intricate dynamical processes of precipitation, such as advection, merging, and intensity variation. In contrast, contemporary deep learning methods demonstrate enhanced capability in extracting precipitation features directly from data, thereby partially mitigating the deficiencies of traditional methods. However, existing nowcasting methods rely heavily on data-driven approaches and lack explicit physical constraints. The precipitation feature relationships learned by these models do not remain conserved under intensity variations, which can result in the spurious creation or removal of strong radar echoes. Moreover, the inferred motion of precipitation is discontinuous, causing abrupt jumps in the echo displacement. These shortcomings of physical consistency ultimately distort the spatiotemporal structure of the predictions and reduce the nowcasting accuracy. To address these challenges, this article proposes a physically consistent global spatiotemporal network for precipitation nowcasting (PCGS). Specifically, the method deeply integrates physically coherent features to guide the learning process rather than relying on physical constraints as external losses or postprocessing. PCGS includes a physical-prior-guided optical flow estimation module (POFEM) and a physics-constrained attention mechanism (PCAM). POFEM encodes mass-conservation-consistent bounded-divergence constraints and temporal continuity as differentiable operations, providing physically consistent motion priors. PCAM embeds fluid constraints into the backbone to form a physically biased feature selection and fusion module. PCGS not only enables the model to accurately capture the essential dynamical evolution of precipitation under complex spatiotemporal conditions, but also enhances its capacity to model precipitation spatial patterns and intensity. To evaluate the performance of the proposed framework, we conduct extensive experiments on different datasets. Experimental results demonstrate that our network achieves performance that is superior or at least comparable to several state-of-the-art methods.