Multi-label fault detection in hybrid microgrid using a unified spatiotemporal deep learning approach

This paper presents a Temporal Convolutional Network–Long Short-Term Memory (TCN-LSTM) framework for fast and accurate fault detection and diagnosis in hybrid AC/DC Microgrid (MG). The model effectively captures high-frequency transients using dilated causal convolutions while learning long-term fault evolution patterns through gated memory mechanisms. The proposed method consistently identifies multiple fault types, including ABC-to-ground, three-phase, and phase-to-phase faults occurring between 0.03 s and 0.06 s. The system exhibits strong detection performance, with sharp transitions from 0 to 1 in all affected phases during fault intervals. For Phase A faults, the model achieves 99.74% accuracy, 99.97% precision, 99.25% recall, and an F1-score of 99.61%, based on 53,877 true negatives (TN), 22,822 true positives (TP), 6 false positives (FP), and 172 false negatives (FN). Combined multi-label evaluation shows that TN is 6.294 × 10⁴, TP is 1.368 × 10⁴, FP is 61.25, and FN is 195.8, demonstrating high precision with very low misclassification. The PR curve and macro-averaged indices further confirm consistently strong performance for all fault categories, indicating that the TCN-LSTM architecture offers a robust, real-time, and phase-selective solution to improve protection, reliability, and stability in renewable-rich MG.