Fault Diagnosis of Photovoltaic Strings Based on SiamMN Networks and Panoramic I-V Features

Accurate fault diagnosis is crucial to ensuring the stable operation of photovoltaic (PV) systems. With the increasing availability of condition monitoring data such as current-voltage (I-V) curves, machine learning (ML)-based PV fault diagnosis methods have been widely investigated. However, existing methods heavily rely on large amounts of labeled data to develop ML models, which limits their practical implementation. To improve the effectiveness of ML-based PV fault diagnosis in scenarios with limited labeled data, this article proposes a similarity learning method using Siamese-MobileNet (SiamMN) networks and panoramic I-V features (PIVFs). Specifically, the designed SiamMN networks contain two parameter-sharing MobileNetv3 subnetworks, which take a sample pair formed by two PIVFs as input. PIVFs are derived from corrected and normalized I-V curves. By randomly pairing two PIVFs, the number of sample pairs used for training is greatly increased, thereby improving the generalization capability of the proposed method. Experimental validation and field trials are conducted in a 3.75 kW PV system and a utility-scale PV plant, respectively. The results show that the proposed method is able to diagnose PV faults under various conditions and severities with 100% accuracy, and the required labeled sample size (25–30 I-V curves per PV condition) is only 10%–13.89% of that in existing studies.