Arc Fault Detection and Line Selection in DC Multi-branch Power Supply Systems

In recent years, photovoltaic (PV) and energy storage power supply systems have developed rapidly in direct current (DC) microgrid, construction industry, and commerial fields. In these DC multi-branch scenarios, altered arc fault characteristics would challenge existing detection methods. In this paper, a feature-driven long short-term memory (LSTM) method is proposed for arc fault diagnosis in DC multi-branch situations.Firstly, an arc fault experimental platform is constructed for studying arc fault characteristics in DC multi-branch systems. Next, the characteristics of arc fault current signals are studied using different time-frequency transformation technologies. By setting the lifting ratio Z for bus-level DC arc faults and MZ for branch-level arc faults, it is found that the lifting ratio Z and MZ between the fault state and the normal state are relatively small for the short-time Fourier transform (STFT), which may cause a mistaken system state judgment. Therefore, the wavelet transform method is more suitable to detect arc faults and select their occurrence line. Then a DC arc fault detection and line selection method is proposed based on LSTM network. Compared with the data-driven mode, the feature-driven mode achieves higher accuracy. Finally, viewed from the verification of various experimental data, the LSTM-based algorithm displays 13.36% higher detection accuracy on average than a random forest (RF) diagnosis model.