A gated spatiotemporal fusion network for lightning forecasting based on weather foundation models

Lightning is a significant natural hazard that poses considerable risks to both human safety and industrial operations. Accurate, fine-scale lightning forecasting is crucial for effective disaster prevention. Traditional forecasting methods primarily rely on numerical weather prediction (NWP), which demands substantial computational resources to solve complex atmospheric evolution equations. Recently, deep learning-based weather prediction models—particularly weather foundation models (WFMs)—have demonstrated promising results, achieving performance comparable to NWP while requiring substantially fewer computational resources. However, existing WFMs are unable to directly generate lightning forecasts and struggle to satisfy the high spatial resolution required for fine-scale prediction. To address these limitations, this paper investigates a fine-scale lightning forecasting approach based on WFMs and proposes a dual-source data-driven forecasting framework that integrates the strengths of both WFMs and recent lightning observations to enhance predictive performance. Furthermore, a gated spatio-temporal fusion network (gSTFNet) is designed to address the challenges of cross-temporal and cross-modal fusion inherent in dual-source data integration. gSTFNet employs a dual-encoding structure to separately encode features from WFMs and lightning observations, effectively narrowing the modal gap in the latent feature space. A gated spatiotemporal fusion module is then introduced to model the spatiotemporal correlations between the two types of features, facilitating seamless cross-temporal fusion. The fused features are subsequently processed by a deconvolutional network to generate accurate lightning forecasts. We evaluate the proposed gSTFNet using real-world lightning observation data collected in Guangdong from 2018 to 2022. Experimental results demonstrate that: (1) In terms of the ETS score, the dual-source framework achieves a 50% improvement over models trained solely on WFMs, and a 300% improvement over the HRES lightning forecasting product released by the European Centre for Medium-Range Weather Forecasts (ECMWF); (2) gSTFNet outperforms several state-of-the-art deep learning baselines that utilize dual-source inputs, clearly demonstrating superior forecasting accuracy.