Large-Scale Floating Offshore Wind Turbine Anomaly Detection and Diagnosis System Based on Spatio-Temporal Transformation Features Combining Informer and XGBoost

This article presents a novel method for detecting and diagnosing anomalies in large-scale offshore floating wind turbines based on time-space transformation features combining Informer and extreme gradient boosting tree (XGBoost).To prevent the model from drifting due to the influence of faulty data during model establishment, and to ensure the integrity of key information in the supervisory control and data acquisition data, particularly the spatio–temporal transformation characteristics, this article have designed a data preprocessing method that combines the 3$\sigma$ rule, Kalman optimal estimation, and k-nearest neighbors algorithm. In addition, a features fusion construction method is introduced for spatio–temporal transformed features, which describe the dynamic changes of environmental conditions, the inertia information of wind turbine output under complex conditions, and the switching relationship of states in the motion space. This article applies the informer model to diagnose faults in offshore wind power systems and introduces the relative position coding algorithm to encode spatio–temporal transformation characteristics. This enables the model to better learn these characteristics. Anomaly detection is performed by setting the threshold based on the 99.74% confidence interval of the kernel density estimation fitted distribution curve, calculated using sliding window-kullback-leibler (KL) divergence calculation. Anomalies are detected using sliding window-based technology. XGBoost is then trained by utilizing expanded fault data from wind turbines of identical kind within the corresponding wind farm to attain precise abnormal location and diagnosis. Experimental results in real wind farms demonstrate the excellent performance of this method.