Fault Detection and Classification in Microgrid Using Long Short-Term Memory

Abstract

The protection of a microgrid against faults is one of the challenging tasks due to its operation in different modes with penetration of renewable energy sources. To address the challenges, in this paper, the Long Short-Term Memory (LSTM) method-based model is implemented for microgrid fault detection and classification. Most of the previously suggested fault detection and classification techniques extracted features using signal processing methods from the fault waveforms while the proposed method only requires instantaneous values. A microgrid is modeled in MATLAB/Simscape using an IEEE 13-node test feeder connected with renewable energy sources and its operation is simulated both in grid-connected and islanded modes for ten faults at different distances on a distribution line to collect the datasets needed for training and testing. This deep learning-based method can detect and classify microgrid faults for both modes of operation though fault currents are significantly lower in the islanded mode than the grid-connected mode. The method can successfully identify faults in 1/8 cycle data window of instantaneous voltages and currents. The results show that the proposed method gives impressive performance for both grid-connected and islanded modes of operation with an average accuracy of 99.93% and 99.92%, for respective modes.