DC microgrids are widely used in sensitive applications where fast and accurate protection is essential. Many existing studies focus on detecting short-circuit faults and identifying the faulty line, but in practice this is not enough. For effective repair and minimal downtime, the exact fault location along the line must also be determined. This paper presents a data-driven method for fault location in a bipolar DC microgrid. The faulty line is assumed to be known, and the method does not require the fault type. Voltage and current signals measured at each bus are used to extract time-domain and time-frequency features using wavelet transform. After extracting various features, a random forestbased feature selection method is used to identify the most useful features, and for each bus–line pair, a regression neural network is employed to estimate the exact fault location as a percentage of the faulted line length. To achieve the highest possible accuracy, the models are trained using 14 different training algorithms combined with three data normalization algorithms (no normalization, Min-Max scaling, and Z-score standardization), and the best-performing combination is selected for each scenario. Finally, a scoring-based consensus combines the individual estimates into a single, more accurate value. Simulation results across various fault locations and fault resistances show that the proposed scheme achieves low location error and improved robustness compared with single-device estimation, with an overall RMSE of about 1.38%.

Keywords