<![CDATA[The rapid proliferation of clean energy microgrids—integrating solar photovoltaic (PV) generation, bidirectional power converters, and battery energy storage systems (BESS)—requires robust fault detection and isolation (FDI) to guarantee continuous operational stability. However, non-Gaussian residual profiles caused by severe irradiance intermittency and dynamic switching transients compromise traditional fixed-threshold strategies, resulting in elevated false alarm and missed detection rates. To address this challenge, this study presents a hybrid FDI framework combining physics-guided deep sequential forecasting with non-parametric adaptive thresholding. A Temporal Convolutional Network integrated with Long Short-Term Memory (TCN-LSTM), constrained by equivalent circuit and DC bus power balance equations, is developed to forecast multi-modal nominal trajectories and extract reliable diagnostic residuals. Raw electrical streams sampled at 10 kHz are down sampled via moving-window averaging to 10-second intervals to accommodate prognostic forecasting horizons. Non-parametric Kernel Density Estimation (KDE) is subsequently implemented to dynamically compute adaptive threshold boundaries from empirical non-Gaussian residual distributions. Simulation experiments under stochastic irradiance profiles and dynamic load cycles confirm that the proposed TCN-LSTM architecture achieves nominal forecasting RMSE values of 0.007 V for cell voltage and 0.306 °C for temperature, reducing the missed detection rate of PV shading mismatches and critical sensor biases compared to static thresholding. Furthermore, integration with a Local Outlier Factor (LOF) anomaly detector provides early warning margins of 83.5 minutes for micro-short circuit voltage dips and 24.6 minutes for thermal runaway precursors prior to hard-limit BMS alarms.]]>
