Sliding sampling and successive variational mode decomposition CNN-BiLSTM-attention based fault detection and early warning method for DC microgrid.

A fault detection method that integrates sliding sampling, successive variational mode decomposition (SVMD), and CNN-BiLSTM-Attention model is proposed to address the problem of insufficient sensitivity and discrimination ability in fault signal diagnosis of DC microgrids. Firstly, sliding sampling is used to capture transient fault information, avoiding the loss of information in traditional fixed windows; Secondly, by decomposing the fault signal through SVMD, the penalty factor is optimized with the maximum mutual information coefficient (MIC), and the effective modal components (IMF) are selected by combining the correlation coefficient, spectral entropy, and Teager Kaiser energy ratio to achieve noise reduction and signal reconstruction; Finally, a CNN-BiLSTM-Attention classification model is constructed, using CNN to extract local time-frequency features, BiLSTM to capture sequence context relationships, and adaptive weighting of key fault features through attention mechanism to suppress noise interference. The experimental results show that the proposed method has an average classification accuracy of 92.35% in islanding mode and 91.13% in grid connected mode, which is significantly better than the compared methods, especially with an accuracy rate of over 95% in single-phase grounding faults; The accuracy exceeds 91% in different scenarios (radial/mesh topology), verifying its robustness and adaptability.