A Semisupervised Learning Diagnosis Method for Series Arc Faults in PV Arrays Under Shading Conditions

ABSTRACT This paper proposes an intelligent wavelet decomposition and reconstruction‐frequency enhanced decomposition transformer‐semisupervised learning (WDR‐FEDformer‐SSL) method for detecting series arc faults in photovoltaic (PV) arrays under shading conditions, using limited labeled on‐site data. Fault arc patterns are classified by analyzing the interaction of I‐V curves between the PV array under shading conditions and the fault arc. Transformer neural networks are first introduced and designed to detect and identify DC fault arcs by using the FEDformer network with the multihead attention mechanism to deeply explore the fault features. Additionally, wavelet packet decomposition combined with dynamic soft thresholding denoising algorithms is employed to effectively suppress environmental noise interference and to obtain time‐scale signals for the transformer neural network. Aiming at the lack of labeled data on‐site and to enhance training accuracy, semisupervised learning (SSL) methods are proposed by leveraging a large amount of unlabeled sample data from on‐site with a pseudo‐label confidence propagation mechanism, and the model of the PV system with fault arc and shading condition is used to increase the labeled samples under shading conditions. Experimental results show that the proposed method achieved an average recognition accuracy of 98.6% under shading conditions.