Wind gusts pose an increasing threat to sustainable development, damaging resilient infrastructure (SDG 9), disrupting clean energy systems (SDG 7), and endangering community safety (SDG 11). However, the reliability of early warning systems remains limited by systematic biases in Numerical Weather Prediction (NWP) models and insufficient uncertainty quantification, particularly in regions with sparse monitoring networks. To address these challenges in sustainable disaster risk reduction, this study proposes a physics-informed deep learning framework—the Physics-Informed Spatial Attention Network (PISA-Net). The model integrates high-resolution WRF-UPP forecasts as a physical prior within a Transformer-based architecture, enabling effective bias correction and spatial dependency learning under data-sparse conditions. A hybrid probabilistic learning objective is employed to simultaneously improve deterministic gust predictions and provide calibrated uncertainty estimates. Evaluated on 61 extratropical cyclone events in the northeastern United States, PISA-Net substantially outperforms baseline NWP and conventional deep learning models, reducing the mean absolute error and root mean square error to 1.75 m/s and 2.26 m/s, respectively. In addition, the resulting 95% prediction intervals are well calibrated and offer reliable risk-based guidance. By improving both the accuracy and credibility of wind gust forecasts, PISA-Net provides a practical decision-support tool for infrastructure maintenance, wind farm operations, and public safety planning. This work demonstrates the potential of physics-informed deep learning to strengthen sustainable early warning systems in data-sparse regions.
Li et al. (Sun,) studied this question.