An unsupervised hourly weather status pattern recognition and blending fitting model for PV system fault detection

Detecting PV system faults in a timely fashion is important to ensure the safe operation of equipment and reduce their impact on the economy of the PV systems. It is necessary to further improve the time-sensitive performance evaluation of the system. However, the hourly weather scenario segmentations are seldom considered during the hour-level online monitoring process. Therefore, a hybrid method based on unsupervised hourly weather status pattern recognition and blending fitting model is proposed for hourly fault detection to improve the performance evaluation of PV systems. The proposed solution includes three parts, firstly, in the data preprocessing stage, the measured power with the errors and noise under normal operation situation caused by the environment changes is corrected by monthly linear fitting. Secondly, an unsupervised hourly weather status pattern recognition method is constructed using the measured radiation data, including unsupervised clustering and the Multiclass-GBDT-LR classification process. Finally, after eliminating the anomalies and errors, the blending fitting model of the hourly sub-weather status is established. Through the analysis of power plants in Australia and China, the proposed solutions are validated and evaluated to be superior to existing data-driven solutions in terms of fitting accuracy, detection validity, and response time. Numerical results of case studies indicate that the developed methodology under sub-weather has improved the detection accuracy up to 97.71% and 99.29% compared to benchmark models.

Automated summary

In this paper, a hybrid method based on unsupervised clustering and blending fitting model is proposed for PV system fault detection. By recognizing the hourly weather status patterns, the performance evaluation of PV systems is improved in terms of accuracy and response time. Experimental results show that the proposed method outperforms existing data-driven solutions in terms of fitting accuracy and detection validity.