Forecastformer: Spatiotemporal Decoupled Transformer and Diffusion Models for Severe Convective Storm Nowcasting

Nowcasting of severe convective storms is crucial yet challenging for detecting potential hazardous weather events within upcoming hours. While recent advances in deep learning-based nowcasting methods utilizing deterministic modeling have demonstrated improved accuracy in low-intensity precipitation forecasting, they remain constrained by gradual dissipation, blurring effects, and intensity or location errors in high-reflectivity echoes due to the rapid development and highly nonlinear characteristics of convective-scale storms. To address this limitation, we propose Forecastformer, a novel nowcasting model that combines a mesoscale evolution network with a convective-scale denoising generative network (CSDGNet). The proposed framework employs spatial–temporal decoupled Transformer architectures and residual diffusion models to systematically decompose and model the deterministic motion of mesoscale echo systems and the local stochastic variations in convective-scale echoes. Evaluated on radar reflectivity datasets from Jiangsu Province and Shanghai, the experimental results demonstrate that Forecastformer effectively captures the intensity and morphology of severe echoes while enhancing prediction sharpness. Comparative analyses reveal significant improvements in both quantitative metrics and visual quality over state-of-the-art approaches, particularly in maintaining structural integrity and reducing spatial displacement errors for severe convective storms.