A Novel Long Short-Term Memory-Based Approach for Microgrid Fault Detection and Classification Using the Wavelet Scattering Transform

Protecting microgrids poses significant challenges owing to their diverse operational modes, multiple power infeed sources, variable fault current levels, heterogeneous control strategies, and the high penetration of intermittent renewable energy sources (RESs). These characteristics undermine the effectiveness of conventional protection schemes, particularly those based on overcurrent relays (OCRs) that are designed for systems with predictable and high fault currents. During islanded operation, a common mode in microgrids, fault currents are often reduced, making fault detection and isolation even more difficult. These limitations underscore the urgent need for intelligent, adaptive, and fast-responding fault detection and classification algorithms tailored specifically to the nature of microgrids. To address this gap, this study proposes a novel fault detection and classification approach that combines the Wavelet Scattering Transform (WST) for robust feature extraction with a Long Short-Term Memory (LSTM) network for accurate temporal pattern recognition and classification. Although WST has demonstrated remarkable performance in audio and image classification, its use in power system analysis remains largely unexplored. In this study, WST was used for the first time to process fault current signals in microgrids, resulting in significant feature matrices for training and testing. WST is especially excellent for capturing hidden properties in current waveforms, making it ideal for classification applications. The proposed methodology was validated using a Consortium for Electric Reliability Technology Solutions (CERTS) microgrid testbed. All simulations were performed in MATLAB, which served as the platform for modeling the CERTS microgrid, configuring the WST, and implementing the LSTM network. A comparative analysis with conventional techniques highlighted the superior classification accuracy and robustness of the proposed WST-LSTM framework.