Integrated Anomaly Detection and Machine Learning Reconstruction for Offshore Wind Power Cleansing

Offshore wind energy, an efficient renewable source, has gained significant attention. However, its intermittency and data anomalies pose challenges for accurate power prediction. To address this, we propose an anomaly detection model based on power curve characteristics and domain knowledge. The model employs ensemble learning, integrating LOF, DBSCAN, and IF algorithms to overcome individual model limitations. We enhance raw multivariate features through feature engineering and use LightGBM for anomalous data reconstruction. Validated on a real-world dataset from a Fujian, China offshore wind farm, results show bin curve errors reduced by 11.8%-26.5% post-reconstruction. When applied to benchmark power prediction models, reconstructed data improves prediction accuracy by 4%-7%. Computational results confirm that our strategy accurately identifies and rectifies offshore wind power data anomalies, significantly enhancing prediction accuracy and ensuring efficient, stable wind farm operation.