Hybrid BiLSTM-Ensemble Learning for Fine-Grained Electricity Forecasting: Interpretability and Uncertainty-Aware Modeling

Forecasting household electricity consumption at minute resolution is increasingly vital for smart-grid reliability and cost control, yet behavior-driven volatility and nonlinearity hinder traditional approaches. This study presents a hybrid deep learning framework that couples a bidirectional LSTM with a tree-ensemble residual corrector for fine-grained residential load forecasting. Our novelty lies in a principled two-stage design that (i) additively corrects systematic underfit of the sequence model, (ii) yields uncertainty-aware outputs via calibrated prediction intervals with coverage diagnostics, and (iii) provides model-side interpretability through feature attributions quantifying short-/long-lag and rolling-stat effects. Using over two million IHEPC observations, the model attains RMSE 0.64 and MAE 0.35, outperforming strong base-lines. Diagnostic results show low bias, fast residual decorrelation, and approximately Gaussian errors; a demand-response illustration indicates peak reduction. Limitations include using only internal lag/time features (excluding weather/occupancy), validation restricted to IHEPC, and a simplified peak-capping DR heuristic. Deployment requires stream-latency control, periodic recalibration to track drift, and continuous interval monitoring, positioning the framework as an interpretable, uncertainty-aware basis for grid-edge automation and risk-informed energy management.