Sparse-Decomposition-Assisted CNN-BiLSTM Framework for Multi-Condition Fault Classification in Hybrid AC/DC Microgrids

Abstract

The reliable protection of converter-dominated microgrids is problematic since the fault signatures depend on the operating mode, the load level, the fault resistance, the converter current limiting and the grid connection. In this study, a multi-condition diagnosis framework based on voltage-current signal processing, sparse decomposition and CNN-BiLSTM classifier is developed. Three-phase voltage and current signals are gathered at the monitoring point, cleaned, normalized and segmented before a sparse representation is generated to preserve the prominent transient components while reducing the redundant material. A one-dimensional convolutional layer learns local waveform signatures and a bidirectional long short-term memory layer models the temporal evolution in both directions. Five operating classes are considered: normal operation, open circuit fault, line to ground fault, islanding and swing state. To a structured 56-case test matrix, four load levels, twelve open-circuit combinations, fifteen L-G combinations, five islanding load-mismatch circumstances and twenty swing combinations are derived. The assessment set we provide has 48 properly detected instances and eight misclassifications. The total accuracy is 85.71%. The weighted precision, recall and F1-score is 85.81%, 85.71% and 85.71% correspondingly. Comparative assessment is structured against impedance, overcurrent/overvoltage relay, RMS and FFT based detection. The findings indicate that the sparse signal representation with local and bidirectional temporal feature learning is effective for heterogeneous microgrid disruptions.