A physics-informed patch transformer with interpretable reasoning for early anomaly warning in wind turbines

Abstract Wind power has experienced rapid development in recent years. Under harsh environmental conditions, achieving accurate and stable anomaly warning of wind turbines is the crucial technical approach to improving operational economy and safety. To address the challenges of processing low-quality operational data, difficulties in modeling operating behavior, and the resulting poor accuracy and stability of anomaly early warning, this paper proposes a physics-informed patch Transformer model. By incorporating physical knowledge constraints to extract high-quality data, and applying a temporal patch attention mechanism to achieve deep fusion of multivariate variables, earlier and more stable anomaly warnings are realized. In addition, shapley additive explanations and the custom root cause priority index methods are employed to rapidly and accurately localize fault causes. Case studies on real-world wind turbine anomalies demonstrate that the proposed method can provide early warnings for specific faults up to 5 days in advance, with no false alarms observed during the 10 day monitoring period following fault resolution. Moreover, the proposed fault localization approach enables engineering personnel to quickly and accurately identify fault locations, thereby further reducing maintenance response times.