Fault Diagnosis of DC and AC Microgrids Based on a TCN–BiLSTM–Attention Model

The large-scale integration of wind, photovoltaic (PV) and energy-storage units makes microgrid fault signals strongly nonlinear, time-varying and transient, so that threshold- or experience-based diagnosis methods can no longer reliably satisfy the requirements of high accuracy and robustness. This paper proposes a fault diagnosis method for microgrids based on a hybrid temporal convolutional network (TCN), bidirectional long short-term memory network (BiLSTM) and additive attention mechanism. The TCN extracts local, multi-scale transient features from multi-channel voltage and current waveforms through causal and dilated convolutions with residual connections; the BiLSTM models bidirectional temporal dependencies of the fault evolution; and the additive attention adaptively re-weights the key time steps to emphasize the most discriminative information. Two representative tasks are studied on simulation data: condition/fault-state identification for a wind–PV–storage DC microgrid and fault-phase identification for a wind–PV–diesel–storage AC microgrid. A sliding-window strategy is used to build and augment samples for the DC task. Experiments with twenty repeated trials show that, after sliding-window augmentation, the proposed model attains an average accuracy of 94.18% on the DC fault-diagnosis task and 96.75% on the AC fault-phase-identification task, with a favorable accuracy–stability trade-off compared with TCN, LSTM, BiLSTM and TCN–BiLSTM baselines. The results confirm that combining local feature extraction, bidirectional temporal modeling and adaptive attention effectively improves the accuracy and stability of microgrid fault diagnosis. The AC evaluation is a conditional fault-phase identification task performed after a fault has occurred; normal/fault detection is not addressed.