Article’s

The global transition toward carbon neutrality and the integration of intermittent renewable energy sources have necessitated unprecedented sub-hourly forecasting precision within Smart Grid environments. Given that building energy consumption accounts for nearly 40per of global total energy use and 33per of greenhouse gas emissions, predicting volatile load patterns is fundamental for demand-side management and grid stability. This paper presents a rigorous, multi-layered investigation into hybrid Long Short-Term Memory (LSTM) networks for multivariate time-series modeling. Unlike traditional stochastic models such as ARIMA, which assume stationarity and linearity, our proposed framework captures the non-linear, multi- seasonal nature of electricity demand by integrating Discrete Wavelet Decomposition for multiresolution feature extraction. To address the ”vanishing gradient” problem and improve long- term sequence stabilization, we implement a ”Teacher Forcing” training strategy, utilizing ground-truth outputs to prevent error accumulation. Furthermore, we propose a hybrid architecture that leverages Support Vector Regression (SVR) for residual re- finement and Deep Extreme Machine Learning (DELM) for rapid sequence learning. All model hyperparameters were tuned using the Developed Henry Gas Solubility Optimization (DHGSO) algo- rithm to ensure structural robustness. Experimental validation was conducted using the London Smart Meter dataset and a two-year multi-campus university dataset. Results highlight a (15–20)per improvement in Root Mean Square Error (RMSE) and significant reductions in Mean Absolute Error (MAE) compared to standalone LSTM and MLP models. Finally, we employ Layer-wise Relevance Propagation (LRP) to enhance model interpretability, identifying critical 24-hour and 168-hour temporal lags. This study provides a robust solution for utility providers seeking to minimize operational risks through precise, high-fidelity demand prediction.