Identification of multiple power quality disturbances in hybrid microgrid using deep stacked auto-encoder based bi-directional LSTM classifier

• A novel adaptively fast complementary ensemble local mean decomposition (AFCELMD) technique used for power quality signal de-noising. • *AFCELMD involves adaptive way of selecting the key operational parameter (upper frequency limit) while fixing the other two critical parameters (noise amplitude and ensemble trials). • A deep-stacked auto-encoder (dSAE) reduces the data dimensionality and extracts intrinsic features, a BiLSTM network is used to predict and classify the PQD classes. • Finally the combined AFCELMD-dSAE-Bi-LSTM is applied on different PQDs to classify different disturbances with greater accuracy, minimal cross-entropy loss, and robust anti-noise performance. • With fewer codified features from the last AE layer, dSAE discriminates all the PQD classes with minimum overlapping and results in least average error values. In recent years microgrid technology has created widespread interest for the integration of renewable energy sources into main utility grid to supply clean energy to the end users. However, the use of power electronic equipments, electronic controllers, and uncertain nature of the renewable energy sources, in the microgrid network power quality disturbances (PQD) are becoming quite complex and challenging task. Thus to design an effective PQD recognition system, this paper proposes a novel time-frequency analysis method based on adaptively fast complementary ensemble local mean decomposition (AFCELMD) technique that decomposes the multicomponent PQD signal into a series of demodulated product functions (PFs). Out of the several PFs the most sensitive one is selected adaptively and used for feature extraction and classification through a deep stacked auto-encoder (dSAE) hybridized with a time-recursive bi-directional long short term memory (BiLSTM) network classifier. The proposed BILSTM classifier captures the temporal features and their long term dependencies from the processed PF data samples and detects single and simultaneously occurring twenty complex power quality disturbances in the grid connected mode and five PQDs during uncertain PV insolence variation and load and capacitor switching during islanded mode of microgrid operation with significant accuracy of 99.90%.