Stable Attention Response for Reliable Precipitation Nowcasting

arXiv cs.AI · 18 hours ago

Researchers developed HARECast, a framework that improves precipitation nowcasting by stabilizing attention mechanisms in neural networks, which addresses instability in attention responses across different samples that was previously overlooked. The method applies group-wise regularization to reduce fluctuations in head-wise attention-response energy and was evaluated on SEVIR and MeteoNet benchmarks. This approach achieves state-of-the-art results and can be applied to both single-modality and multimodal weather forecasting architectures.