PhyTGNet: A Physics‐Prior‐Guided Interpretable Forecasting Method for Thunderstorm Gusts in the Beijing–Tianjin–Hebei Region
- 1Faculty of Information Science and Engineering Ocean University of China Qingdao China
- 2Institute of Urban Meteorology China Meteorological Administration Beijing China
Abstract
The sudden and localized nature of thunderstorm gusts makes pure data‐driven models struggle to maintain physical plausibility and interpretability, leading to nonphysical false alarms and spatial displacement in complex terrain. To address this, we propose PhyTGNet, a deep learning framework with embedded physics‐inspired guidance that achieves 500 m resolution and 1–12 hr lead‐time thunderstorm gust forecasting for the Beijing–Tianjin–Hebei region. First, we construct a thunderstorm gust physical prior model (TGPP) based on ERA5, learning physical–statistical relationships between environmental dynamical and thermodynamic conditions, terrain factors, diurnal variability, and thunderstorm gust occurrence, to generate an interpretable environmental prior field. Next, we design a Physics‐Guided Attention Block (PGAB) embedded into the decoder of PhyTGNet with Swin‐Unet as the backbone. Through a gating mechanism, the PGAB dynamically weights physical priors and corrects spatial displacements caused by numerical model biases. This allows PhyTGNet to reconcile high‐resolution data‐driven representations with physically meaningful environmental knowledge. Comparative experiments show that PhyTGNet outperforms U‐Net, DeepLabV3+, and TG‐TransUnet across the 1–12 hr lead‐time range. At lead times of 1, 3, and 6 hr, critical success index increases by 31.5%, 33.8%, and 18.1%, respectively, compared with TG‐TransUnet, while reducing false alarms and improving spatial localization. The model reduces false alarms in environmentally unfavorable regions over complex terrain. Ablation experiments and case studies verify that TGPP priors can stably identify high‐risk environments, whereas the PGAB refines gust positioning at fine scales. This study provides a physics‐inspired and interpretable framework for thunderstorm gust forecasting, demonstrating potential for urban disaster prevention and mitigation applications.
Related
- Station-specific thunderstorm forecasting using climate predictors with seasonal decomposition analysis in Bangladesh’s most vulnerable regions
- An interpretable model for efficient and robust runoff forecasting via causal-linear attention and adaptive KAN decoding
- Is Seismic Forecasting Possible with Physics-based AI?
- FAST-ML: A Hybrid Physics-Machine Learning Framework for Tropical Cyclone Intensity Forecasting
- Fusing physics and AI enhances accuracy and diagnostic capabilities in air quality forecasting.
BibTeX
@article{yan2026phytgnet,
title = {{PhyTGNet: A Physics‐Prior‐Guided Interpretable Forecasting Method for Thunderstorm Gusts in the Beijing–Tianjin–Hebei Region}},
author = {Zhaoyu Yan and Mingxuan Chen and Lu Yang and Lei Han and Linye Song and Xinze Li},
journal = {Journal of Geophysical Research: Machine Learning and Computation},
year = {2026},
month = oct,
doi = {10.1029/2026jh001422},
url = {https://doi.org/10.1029/2026jh001422},
note = {CC BY},
}
RIS
TY - JOUR
TI - PhyTGNet: A Physics‐Prior‐Guided Interpretable Forecasting Method for Thunderstorm Gusts in the Beijing–Tianjin–Hebei Region
AU - Zhaoyu Yan
AU - Mingxuan Chen
AU - Lu Yang
AU - Lei Han
AU - Linye Song
AU - Xinze Li
PY - 2026
DA - 2026/10/28
JO - Journal of Geophysical Research: Machine Learning and Computation
DO - 10.1029/2026jh001422
UR - https://doi.org/10.1029/2026jh001422
AB - The sudden and localized nature of thunderstorm gusts makes pure data‐driven models struggle to maintain physical plausibility and interpretability, leading to nonphysical false alarms and spatial displacement in complex terrain. To address this, we propose PhyTGNet, a deep learning framework with embedded physics‐inspired guidance that achieves 500 m resolution and 1–12 hr lead‐time thunderstorm gust forecasting for the Beijing–Tianjin–Hebei region. First, we construct a thunderstorm gust physical prior model (TGPP) based on ERA5, learning physical–statistical relationships between environmental dynamical and thermodynamic conditions, terrain factors, diurnal variability, and thunderstorm gust occurrence, to generate an interpretable environmental prior field. Next, we design a Physics‐Guided Attention Block (PGAB) embedded into the decoder of PhyTGNet with Swin‐Unet as the backbone. Through a gating mechanism, the PGAB dynamically weights physical priors and corrects spatial displacements caused by numerical model biases. This allows PhyTGNet to reconcile high‐resolution data‐driven representations with physically meaningful environmental knowledge. Comparative experiments show that PhyTGNet outperforms U‐Net, DeepLabV3+, and TG‐TransUnet across the 1–12 hr lead‐time range. At lead times of 1, 3, and 6 hr, critical success index increases by 31.5%, 33.8%, and 18.1%, respectively, compared with TG‐TransUnet, while reducing false alarms and improving spatial localization. The model reduces false alarms in environmentally unfavorable regions over complex terrain. Ablation experiments and case studies verify that TGPP priors can stably identify high‐risk environments, whereas the PGAB refines gust positioning at fine scales. This study provides a physics‐inspired and interpretable framework for thunderstorm gust forecasting, demonstrating potential for urban disaster prevention and mitigation applications.
SN - 2993-5210
ER -