AI-enhanced time-frequency domain reflectometry for robust series arc fault detection in DC grids

Abstract

The generation of series arc fault (SAF) within DC grids poses a significant hazard, potentially leading to electrical fires. However, developing a diagnostic method that is robust to the external environment and noise remains challenging. Conventional reflectometry methods are unable to discern minute impedance fluctuations at cable joint, as they analyze reflected signals generated at points of impedance mismatch to locate faults. This paper presents a new version of reflectometry, which merges artificial intelligence (AI) with the time-frequency domain reflectometry (TFDR) technique. In this study, we designed a structure that combines denoising autoencoders (DAE) with autoencoders (AE) to secure robust SAF detection performance against noise such as ambient environmental changes and weather fluctuations. Furthermore, the time-series generative adversarial network (TimeGAN) was utilized to generate synthetic data. Three experiments were conducted to verify the performance of the proposed method. (1) According to the UL1699B safety standard, an experiment was conducted under laboratory conditions to check whether a SAF could be detected within a maximum of 2.5s immediately after a SAF occurred. (2) To simulate external environmental conditions, a system connected to an inverter and a photovoltaic (PV) simulator was utilized to emulate the occurrence of SAFs. The manipulation of the PV simulator allowed for the simulation of solar irradiance changes due to weather variations and power fluctuations occurring immediately after the system's power was turned on. This created an experimental setup that closely resembled actual PV systems, where the performance of the proposed algorithm was validated. (3) Finally, a 1kW capacity PV system was constructed, and SAF was simulated within it. The system was tested using four SAF detection methods, and the experimental results were compared and validated.