FSR‐SSL: A fault sample rebalancing framework based on semi‐supervised learning for PV fault diagnosis

Abstract

Abstract Photovoltaics face the threat of many potential faults in daily operation, which calls for accurate fault diagnosis to avoid huge economical losses. This paper investigates practical and troublesome scenarios, where the photovoltaics station has a large number of unlabeled samples and only a few are labelled. In the labelled set, the sample sizes of different types are unbalanced. To this end, a new fault sample rebalancing framework based on semi‐supervised learning (FSR‐SSL) is proposed. Specifically, a dual‐threshold selection mechanism is proposed to choose trusted pseudo‐label samples from unlabeled data to expand the training set. Moreover, a fault sample rebalancing strategy is designed to further filter the obtained trusted pseudo‐label samples, thereby flexibly adding different amounts of pseudo‐label data to various types. As the training rounds increase, the fault samples are gradually rebalanced and the model learning bias caused by type imbalance is well overcome. The extensive numerical experiments show that the proposed FSR‐SSL method reaches 99% accuracy. Compared with existing methods, the accuracy is increased by up to 33%.