ConCast: a CBAM-enhanced SimVP with temporal consistency regularization for precipitation nowcasting

Precipitation nowcasting is essential for weather warning and rapid-response decision-making, yet existing deep spatiotemporal models often struggle to emphasize meteorologically salient echo structures and to keep their predictions coherent across future frames. We address these issues with ConCast, an enhanced SimVP for radar-based precipitation nowcasting that couples convolutional block attention module (CBAM) refinement with temporal consistency regularization (TCR). The CBAM stage strengthens channel-wise and spatial feature selection so that the network attends to informative precipitation patterns, whereas an auxiliary cosine consistency term constrains the relationship between successive predictions during training. On the Shanghai-Radar and SEVIR datasets, ConCast improves forecasting quality over representative baselines and produces sharper and temporally steadier echo fields. Ablation studies show that the two components act on different aspects of the problem, with attention sharpening spatial discrimination and the consistency term improv-ing sequence coherence, and that their gains do not overlap. Because ConCast adds only marginal overhead to SimVP, attention refinement combined with TCR is a practical option for nowcasting.