This study proposes an anomaly data cleaning method for wind power generation data, addressing issues caused by equipment failures, unstable weather conditions, and data collection module errors. To overcome the limitations of the traditional LOF algorithm—which primarily detects locally sparse outliers but struggles to distinguish anomalies in global datasets where there is no clear distinction between normal and anomalous data—an iForest-LOF weighted fusion algorithm is introduced. This method enhances the identification of outliers in the entire dataset. Moreover, Mahalanobis distance is employed to optimize the LOF algorithm, with the covariance matrix improving the consideration of feature correlations, thereby increasing the accuracy of anomaly detection. After anomaly detection with MLOF-iForest, the identified anomalous data is treated as a missing value imputation problem. The RF algorithm is then applied to explore the implicit relationships between weather features and wind power generation, thus addressing the limitations of traditional KNN algorithms, which struggle to reconstruct large continuous segments of anomalous data by only considering neighboring outliers. Case studies demonstrate that the proposed wind power data recognition and reconstruction method effectively cleans the data, providing a high-quality dataset for subsequent wind power forecasting.