Probabilistic fault localization via artificial neural networks in MV distribution grids

Medium voltage grids, even when operated radially, are often designed with open ring or meshed structures to enable reconfiguration after a fault by closing one or more switches, with the aim of reducing unsupplied load. To enable rapid grid reconfiguration and resupply, it is important to localize faults as quickly and precisely as possible. However, accurate fault location is often not possible due to the limited number of available fault measurement devices in medium voltage substations and switchyards. This work aims to demonstrate an innovative fault localization method based on an artificial neural network that can be easily adapted to process heterogeneous data provided by protection devices and fault recorders in the grid. The conceived neural network derives probabilistic information about the fault location results: namely, it determines the set of lines suspected to be faulty together with the associated level of probability. The performance of the fault localization algorithm is demonstrated on a sample medium voltage grid with a meshed structure, considering the impact of distributed generation and taking into account both three-phase and single-phase to ground faults.