Abstract: The output power of photovoltaic power generation systems is highly complex. Current short-term anomaly detection suffers from increased result errors due to single feature extraction and lack of targeting. An anomaly detection method for photovoltaic power generation output power based on light gradient boosting machine (LGBM) is proposed. According to current detection requirements, this method extracts and merges abnormal features of output power to comprehensively capture abnormal operation modes of the system; then establishes an output power anomaly location identification method based on LGBM to enhance detection targeting, and combines dynamic continuous detection to achieve anomaly handling. Taking K Photovoltaic Power Station as the test object, traditional XGBoost and GRUbased anomaly detection methods for photovoltaic output power, as well as traditional sub-band processing and correlation coefficient-based methods, are set as comparison methods, and anomaly detection on the power output is conducted respectively with the proposed method.Test results show that on the premise of not excessively increasing complexity, the proposed method has smaller absolute error of anomaly detection, stronger detection targeting, can avoid interference in complex backgrounds, and improve detection accuracy and efficiency.

    Key words: light gradient boosting machine; photovoltaic power generation; output power; anomaly detection