光伏电站故障检测与分类中符号表达式的综合研究

Abstract Large-scale photovoltaic (solar) farms play a crucial role in harnessing solar energy for electricity generation through photovoltaic (PV) technology. However, the control and management of such systems pose significant challenges, particularly in fault detection. This paper introduces the application of a genetic programming symbolic classifier (GPSC) to a publicly available dataset for fault detection in photovoltaic farms. Given the imbalanced nature of the original dataset, the study necessitated the application of oversampling techniques to achieve a balanced representation of class samples. Additionally, the impact of scaling and normalizing techniques on the performance of the GPSC was thoroughly investigated. The GPSC was systematically applied to each scaled or normalized balanced dataset variation, and its hyperparameters were fine-tuned using a random hyperparameter values search (RHVS) method. The algorithm underwent training, via a 5-fold cross-validation (5FCV) process, and the best symbolic expressions (SEs) were determined based on accuracy, area under the receiver operating characteristics curve, precision, recall, and F1-score. The research yielded many SEs, which were used to develop a threshold-based voting ensemble (TBVE). The TBVE for each class was tested on the initial dataset and the threshold was finely tuned to achieve even higher classification performance in photovoltaic detection/classification. Results demonstrated that this approach produced highly accurate TBVE for each class (accuracy in the majority of cases equal to 1.0), showcasing the effectiveness of the GPSC and TBVE in fault detection/classification for photovoltaic farms.