ResistNet: Quantifying solar energy reduction during hurricanes with deep learning

Solar energy is central to the clean energy transition, accounting for a large share of newly installed renewable capacity in recent years. Extreme weather events, particularly hurricanes with extensive cloud systems, can cause substantial reductions in solar photovoltaic generation. Traditional statistical approaches and physics-based simulations struggle to forecast these hurricane-induced reductions both efficiently and accurately. To address this gap, we introduce ResistNet, a machine learning framework based on high-resolution solar irradiance data and comprehensive storm characteristics across the Atlantic basin over two decades. Built on an encoder-decoder transformer architecture with multi-head self-attention and cross-attention mechanisms, ResistNet integrates time-evolving hurricane dynamics with station-specific contextual features to predict solar energy reduction at fine spatial and temporal scales. Our experiments using an event-based five-fold cross-validation scheme demonstrate improved predictive accuracy over traditional statistical and machine learning baselines. ResistNet achieves the lowest overall MSE and MAE among the compared models and consistently improves prediction accuracy across cross-validation folds, highlighting its robustness and generalizability. ResistNet provides an efficient tool for enhancing solar energy forecasting, planning, and resilient grid operations under intensifying climate extremes.