Collaborate Global and Local: An Efficient PV Compound Fault Diagnosis Scheme With Multilabel Learning and Model Fusion

Fast and accurate fault diagnosis of photovoltaic (PV) modules has been the focus of intelligent operation and maintenance for PV systems. Considering efficiency and cost, many machine learning (ML)-based methods are widely used in practice. Due to the increasing scale of PV stations, the series–parallel topology of the system becomes more complicated, and multiple faults may inevitably occur together. However, most existing methods only consider the case of a single fault occurrence or the most basic compound faults, which fails to assess the system state comprehensively. For this reason, a novel global–local dual-stream collaborative framework is proposed for multiclass PV compound fault diagnosis. In the proposed method, a local mining algorithm is designed for the characteristics of PV data to extract local details more conveniently and efficiently. Then, we construct a shared convolutional neural network (CNN) network to extract fault features from global and local streams to improve the modeling efficiency. Next, two model fusion mechanisms are proposed for different data conditions to collaboratively use the fault features of the global and local streams. Both simulations and experiments are conducted to verify the effectiveness of the proposed method.