Enhancing power grid resilience during tropical cyclones: Deep learning-based real-time wind forecast corrections for dynamic risk prediction

Tropical cyclones (TCs) pose severe risks to power transmission systems, yet conventional Numerical Weather Prediction (NWP) models lack the resolution to resolve sub-kilometer wind dynamics critical for infrastructure risk assessment. This study introduces a Real-time Wind Forecast Correction (RWFC) model, a deep learning framework that dynamically refines mesoscale NWP forecasts during TCs by assimilating multi-source observational data. The RWFC integrates Convolutional Neural Network and Bidirectional Long Short-Term Memory (CNN-BiLSTM) to capture spatiotemporal wind-terrain interactions, with a custom loss function balancing accuracy and conservative bias for proactive risk mitigation. Validated during Typhoon Hagupit (2020) in Zhejiang Province, China, the RWFC reduced wind speed and direction mean absolute errors (MAE) by 78 % (6.47 to 1.41 m/s) and 50 % (53.57° to 26.79°), respectively, compared to raw NWP forecasts. By interpolating corrections from sparse observational sites, it achieved province-scale MAE reductions of 56 %, demonstrating robust generalizability. When applied to Zhejiang’s transmission grid, RWFC lowered the number of projected high-risk towers by 98 %, enabling precise, terrain-sensitive risk predictions. The framework bridges NWP’s physical rigor with deep learning’s adaptive capacity, offering a scalable solution for enhancing grid resilience during evolving TCs. This work advances real-time disaster management by transforming coarse forecasts into actionable, high-resolution risk insights for critical infrastructure.