Graph-Based Semi-supervised Learning for Fault Detection and Classification in Solar Photovoltaic Arrays

• Published in: IEEE transactions on power electronics Vol. 30; no. 5; pp. 2848 - 2858
• Main Authors: Zhao, Ye, Ball, Roy, Mosesian, Jerry, de Palma, Jean-Francois, Lehman, Brad
• Format: Journal Article
• Language: English
• Published: IEEE 01.05.2015
• Subjects: Arrays; Circuit faults; Classification; Data models; Data visualization; Fault classification; Fault detection; Faults; Inverters; Learning; machine learning; Photovoltaic cells; Safety; semi-supervised learning; Semisupervised learning; Solar cells; solar photovoltaic (PV) system; Fault classification; solar photovoltaic (PV) system; semi-supervised learning; machine learning; fault detection
• ISSN: 0885-8993; 1941-0107
• DOI: 10.1109/TPEL.2014.2364203

Fault detection in solar photovoltaic (PV) arrays is an essential task for increasing reliability and safety in PV systems. Because of PV's nonlinear characteristics, a variety of faults may be difficult to detect by conventional protection devices, leading to safety issues and fire hazards in PV fields. To fill this protection gap, machine learning techniques have been proposed for fault detection based on measurements, such as PV array voltage, current, irradiance, and temperature. However, existing solutions usually use supervised learning models, which are trained by numerous labeled data (known as fault types) and therefore, have drawbacks: 1) the labeled PV data are difficult or expensive to obtain, 2) the trained model is not easy to update, and 3) the model is difficult to visualize. To solve these issues, this paper proposes a graph-based semi-supervised learning model only using a few labeled training data that are normalized for better visualization. The proposed model not only detects the fault, but also further identifies the possible fault type in order to expedite system recovery. Once the model is built, it can learn PV systems autonomously over time as weather changes. Both simulation and experimental results show the effective fault detection and classification of the proposed method.