Predicting and optimizing energy efficiency in buildings: a multi-output machine learning framework with explainability and constrained optimization

Purpose Nine multi-output regression models were compared for simultaneous HL and CL prediction, supplemented by domain-informed feature engineering, SHapley Additive exPlanations (SHAP explainability), robustness testing and a calibrated physics baseline comparison. The best-performing model, Gradient Boosting (Root Mean Square Error (RMSE) = 0.95 kWh, R2 = 0.990), was used as a surrogate within a Non-dominated Sorting Genetic Algorithm II (NSGA-II) multi-objective optimization framework to identify Pareto-optimal building designs minimizing total energy and peak load. Design/methodology/approach This study presents an integrated machine learning framework for predicting, interpreting and optimizing the heating load (HL) and cooling load (CL) of residential buildings. Findings SHAP analysis revealed that overall height and relative compactness dominate both loads, while orientation has negligible influence. Partial dependence analysis uncovered nonlinear response patterns, and counterfactual scenario testing confirmed that the energy gap between single-story and two-story buildings remains constant regardless of glazing level, establishing height as the primary design lever. The NSGA-II optimization identified designs achieving total energy as low as 17.06 kWh. Research limitations/implications The framework bridges prediction and actionable design guidance for energy-efficient buildings. Originality/value This work is a novel optimization of energy in a building using nine models.