A hybrid framework combining data-driven and expert-driven approaches in building energy modeling based on simulation

Buildings account for approximately 39% of global energy consumption and 38% of greenhouse gas emissions, underscoring the urgency of improved energy analysis. This study introduces a hybrid framework combining machine learning algorithms and multi-criteria decision-making to predict and prioritize factors influencing energy consumption in buildings. A real office building in Mehrabad, Tehran, was modeled and simulated using EnergyPlus, with 4,200 monthly data records generated via parametric modeling. After feature selection using Lasso regression, three ML algorithms (XGBoost, Random Forest, and SVM) were trained, with XGBoost achieving the best performance (RMSE = 215.62, R2 = 0.99). A dual sensitivity analysis was then conducted: Shapley Additive Explanations (data-driven) and Analytic Hierarchy Process (expert-driven) were applied to rank the eight key input features. The comparison revealed partial alignment but also substantial differences in feature importance, highlighting the strengths and limitations of each approach. The integration of SHAP and AHP enables a comprehensive, explainable, and balanced feature prioritization, rarely addressed together in previous studies. This hybrid methodology contributes a novel perspective to energy modeling by linking quantitative data with qualitative expert judgment, supporting more robust decision-making in sustainable building design and operation.