Abstract
This paper addresses the critical link between predictive accuracy and optimization effectiveness in smart home energy management.Our key idea is that the superior accuracy of an advanced forecasting model is a direct enabler for achieving more significant and reliable energy savings.The novelty of this work lies in the rigorous, endto-end validation of a comprehensive framework that integrates a hybrid attention-based encoder-decoder neural model with a formal mixed-integer linear programming (MILP) optimizer.The proposed forecasting model significantly outperforms traditional baselines, achieving excellent accuracy (R² ≈ 0.91, MAPE ≈ 2.4%).This high-precision forecast serves as a critical input for the MILP-based Home Energy Management System (HEMS), which in turn facilitates a 22.5% reduction in energy cost and a 31.8%decrease in peak demand.The framework's robustness is confirmed through extensive validation, including multi-household testing and sensitivity analysis under forecast error.Interpretability is enhanced via SHAP (SHapley Additive exPlanations) and attention visualization, linking model behavior to meaningful temporal features.The results empirically validate our central thesis, demonstrating that a highfidelity forecasting component is fundamental to unlocking the full potential of formal optimization for demand response in intelligent residential environments.