Open AccessTL–LEDarcNet: Transfer Learning Method for Low-Energy Series DC Arc-Fault Detection in Photovoltaic Systems
Author(s) -
Yoondong Sung,
Gihwan Yoon,
Ji-Hoon Bae,
Suyong Chae
Publication year - 2022
Publication title -
ieee access
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 0.587
H-Index - 127
ISSN - 2169-3536
DOI - 10.1109/access.2022.3208115
Subject(s) - aerospace , bioengineering , communication, networking and broadcast technologies , components, circuits, devices and systems , computing and processing , engineered materials, dielectrics and plasmas , engineering profession , fields, waves and electromagnetics , general topics for engineers , geoscience , nuclear engineering , photonics and electrooptics , power, energy and industry applications , robotics and control systems , signal processing and analysis , transportation
The arc-fault phenomenon in photovoltaic (PV) systems has emerged as a major problem in recent years. Existing studies on arc-fault detection in conventional PV systems primarily focus on detecting typical stable arc-faults. Low-energy arc-faults are more challenging to detect than stable arc-faults because of their low current distortions, short durations, and nonlinear properties. These low-energy arc-faults, which are precursors to stable arc-faults, could even inflict serious damage on the system components. Here, a transfer learning-based low-energy arc-fault detection network (TL–LEDarcNet) using a two-stage training method is proposed to proactively detect series DC arc-faults by considering low-energy arc-faults. A one-layer long short-term memory network combined with a lightweight one-dimensional convolutional neural network was developed to detect low-energy arc-faults by only using the sensed current information. The results of offline and online experiments conducted with a commercial grid-connected PV inverter indicate that the proposed method can perform real-time operations on a single-board computer and detect low-energy arc-faults with an accuracy of 95.8%, which is higher than previous methods considered in this study.