Robust Universal Adversarial Perturbation Attacks on Renewable Energy Forecasting

Recent advances reveal that renewable energy forecasting (REF) models, particularly AI-driven approaches, may be vulnerable to adversarial attacks, potentially inducing substantial forecasting errors and disrupting power system operations. However, existing studies focused only on customized attack schemes tailored to specific REF models, single-time inputs, and predefined locations, which are computationally expensive and often suboptimal within practical dispatch intervals. To fill this gap, we first propose a universal adversarial perturbation attack method, formulated in a fully offline manner, which can degrade REF performance across different REF model architectures and spatiotemporal scenarios. To enhance attack robustness, we further develop a robust universal adversarial perturbation generation method tailored for black-box settings through ensemble proxy models. Our findings reveal the new vulnerability of advanced REF technologies to fixed yet small perturbations, which can significantly amplify forecasting errors and severely compromise prediction accuracy, emphasizing the critical need for further investigation.