基于机器学习的风电机组雷击快速识别系统

Lightning is a leading cause of wind turbine blade failures in the United States and globally, resulting in significant financial losses for the industry. Rapid detection of turbine strikes is essential to reduce these costs. While nearby lightning strikes can be detected with high accuracy, confirming whether a specific turbine was struck remains challenging. Current confirmation relies on manual inspections with drones, which may take hours to years if damage develops slowly during operation. This work presents a scalable, three-step framework for lightning strike confirmation that integrates lightning measurements, turbine alarms, and supervisory control and data acquisition (SCADA) based machine-learning anomaly detection. The first step analyzes the magnitude (kA) and proximity of nearby lightning strikes. The second step evaluates historical alarm patterns associated with lightning-induced damage. The third step applies machine learningbased anomaly detection to post-event SCADA signals, focusing on rotor speed, wind speed, and pitch angle behavior. The framework was evaluated using 26 U.S. wind turbines with confirmed lightning strikes between 2021 and 2024, together with 1650 turbines that experienced nearby strikes without direct impact. This timealigned dataset enables robust model training and validation. The proposed approach is designed for fleet-wide deployment and demonstrates strong scalability. At the highest confidence level, recall and precision were 96% and 86% at the next level, 100% and 81%. Deployment across Vestas' U.S. fleet could conservatively save over ${\$}$ 16 million annually in avoided blade repair costs, excluding additional benefits from reduced turbine downtime, thereby contributing to lower wind energy costs.