Abstract Accurate short‐term forecasting of thunderstorm gusts remains a major challenge in the Beijing–Tianjin–Hebei region due to limited temporal resolution, the lack of vertical velocity and thermodynamic data, and the difficulty of coupling multi‐scale features. To address these issues, we propose a deep learning model suite called Thunderstorm Gusts Swin Transformer U‐Net (G‐Net), designed to integrate multi‐source meteorological data and capture multi‐scale gust structures. G‐Net includes: (a) a data fusion strategy incorporating 16 key surface and upper‐air variables with 3D vertical profiles; (b) a hybrid Swin Transformer‐U‐Net architecture to extract both global and local features; and (c) a progressive spatiotemporal decoupling architecture, including three functionally distinct model variants: G‐Net1H (1 hr forecast) and G‐Net2H (2 hr forecast) with a Feature Bridging Layer (FBL) designed for minute‐level average wind predictions, and G‐NetG for gust mapping. Results show that G‐Net1H surpasses the baseline models ConvLSTM, TG‐TransUNet, and PredRNN‐V2 in CSI and FAR across all wind speed thresholds within the 0–1 hr forecast. The introduction of spatiotemporal supplementary information and the efficient integration of the FBL enable G‐Net2H to achieve improved 2‐hr forecasts, reducing mean absolute and root mean square errors and enhancing overall predictive performance. G‐NetG further outperforms traditional gust factor methods and deep learning baselines. Evaluation during the 2024 thunderstorm season demonstrates that G‐Net provides the most accurate 0–2 hr forecasts, improving both the timeliness and spatial precision of weather warnings.
Li et al. (Wed,) studied this question.