https://hal.univ-lorraine.fr/hal-02064976Calderon-Mendoza, EdwinEdwinCalderon-MendozaIJL - Institut Jean Lamour - INC - Institut de Chimie du CNRS - UL - Université de Lorraine - CNRS - Centre National de la Recherche ScientifiqueSchweitzer, PatrickPatrickSchweitzerIJL - Institut Jean Lamour - INC - Institut de Chimie du CNRS - UL - Université de Lorraine - CNRS - Centre National de la Recherche ScientifiqueWeber, SergeSergeWeberIJL - Institut Jean Lamour - INC - Institut de Chimie du CNRS - UL - Université de Lorraine - CNRS - Centre National de la Recherche ScientifiqueA double ended AC series arc fault location algorithm for a low-voltage indoor power line using impedance parameters and a neural networkHAL CCSD2018[SPI.NRJ] Engineering Sciences [physics]/Electric powerWeber, Serge2019-06-25 16:04:132022-01-04 03:58:302019-06-25 16:04:13enJournal articles10.1016/j.epsr.2018.08.0081This paper presents a novel method for the distance estimation of a series arc fault in a low-voltage indoor power line. This method is based on the SIMULINK modeling of an electrical line by using its RLCG parameters. Rather than using an arc fault model, arc faults are inserted at different points across the line, using measured data. Faults using carbonized paths and opening contacts between two copper electrodes are considered. The algorithm estimates the arc fault distances by employing both the voltages and currents at two ends of the line, calculating their DFFT (Discrete Fast Fourier Transform) and then inserting these magnitudes, into Kirchhoff equations which take into account impedance parameters of the simplified approach line model. The arc fault is generated at an unknown distance. As the impedance parameters depend on the fault location, sets of supposed fault distances (varying by one-meter steps) are inserted in these equations. Thus, the extraction of currents peak values at the fundamental frequency (50 Hz) generates a signature vector. A neural network is trained using signature vectors as inputs (for faults generated at different distances). Finally, the algorithm thus developed is validated in two steps. First, the fault distances are estimated using distances not considered in the training process. Secondly, the fault distances are estimated using other series arc faults data not considered in the learning process. The results obtained in this work show that fault location can be successfully estimated both for faults generated by opening contacts and for carbonized paths. Additionally, the algorithm has successfully tested on different line lengths and considering also changes on the impedance parameters of the line.