Machine learning-driven building energy optimization: integrating EnsembleXGBoost sensitivity analysis with meta-heuristic algorithms and carbon economics

While recent building optimization studies achieve significant energy reductions, they often overlook two critical gaps: (1) proceeding on fixed, non-optimized building geometries, and (2) using machine learning (ML) solely as computational surrogates rather than strategic optimization guides. This study introduces a novel four-stage framework that establishes formalized procedural linkages, with each stage systematically informing subsequent decisions. The methodology integrates Grey Wolf Optimization (GWO) for geometric baseline establishment, a robust EnsembleXGBoost-SHAP sensitivity analysis (SA), NSGA-II for targeted multi-objective parameter optimization, and ASHRAE standards implementation. The EnsembleXGBoost model demonstrated superior predictive accuracy (R 2 : 0.973–0.995), a finding statistically confirmed by paired t-tests (p < 0.002) and bootstrap stability analysis (CV < 0.12). Critically, validated SHAP insights—identifying key drivers (Infiltration, External Wall, Lighting, Heating Setpoint) accounting for ∼ 80% of performance variation—were used to formally define and constrain the NSGA-II decision variables, transforming SA from an interpretative endpoint to an optimization input. This integrated approach achieved a 59.4% reduction in total operational energy (rounded to 60.0%) across cumulative stages (GWO: 3.95%, NSGA-II: 49.7%, ASHRAE: 15.9%). The plausibility of this reduction is supported by validating the reference building’s baseline EUI (163.98 kWh/m2/year) against regional literature and the technical feasibility of the optimized parameters (U-value 0.09 W/m2K, Infiltration 0.2 ac/h) against empirical studies. Furthermore, the framework simultaneously improved occupant thermal comfort (average PMV improved from −0.75 to −0.15). A comparative economic scenario analysis revealed stark market-dependent viability: US market conditions supported positive NPV by the 16th year (WCP), while subsidized Iranian markets required carbon pricing mechanisms to achieve viability (36th year). The framework advances building optimization by bridging the disconnect between ML-based sensitivity analysis and multi-objective optimization while demonstrating that preliminary geometric optimization significantly improves parametric optimization baselines.