A data-driven interpretable framework for multi-objective optimization: Decoupling thermodynamic and economic constraints in cold-region rural dwelling retrofits

Enhancing the energy efficiency of rural dwellings is crucial for achieving carbon neutrality. Existing optimization methods are often hindered by the “black-box” nature of machine learning models and the early inclusion of economic objectives, which can mask the true potential of passive design. This study presents a transparent, data-driven framework that integrates explainable AI with a decoupled two-stage optimization process. First, high-fidelity surrogate models developed with XGBoost replace slow simulations. SHAP analysis then reveals how design parameters nonlinearly affect energy use, thermal comfort, and daylight, identifying distinct seasonal drivers and critical interaction thresholds—such as a “risk-reversal” point for windows. These insights guide a two-stage NSGA-II optimization. The initial stage maps the maximum physical performance achievable by the passive envelope alone. The second stage evaluates this high-performance solution space against economic criteria. The resulting optimal retrofit strategy, which combines double-glazed Low-E windows with moderate insulation, reduces energy demand by 45.3% with a payback period of 1.37 years, while significantly improving indoor comfort. This work provides a clear and actionable pathway for deep, cost-effective retrofits of rural dwellings in cold climates.