Risk-aware explainable multi-output workflow for early-stage screening of building heating and cooling load indicators

Early-stage design decisions largely determine a building's operational energy demand, but benchmark-trained surrogate models must be used with explicit validity limits. This paper presents a risk-aware, reproducible and explainable multi-output machine-learning workflow for screening simulated heating load (HL) and cooling load (CL) indicators from concept-stage geometric and envelope parameters. Using the Energy Efficiency (ENB2012) dataset (768 simulated building configurations), Ridge regression, Random Forest and Extremely Randomised Trees (Extra Trees) are evaluated under five-fold cross-validation and a held-out test set, with additional revision diagnostics for geometry-blocked validation, split-conformal prediction intervals, compact TreeSHAP attribution, multi-output ablation and fixed-setting XGBoost, LightGBM and MLP comparators. The Extra Trees model achieves test-set RMSE of 0.60 kWh/m2 for HL and 1.45 kWh/m2 for CL, with R2 of 0.996 and 0.977, respectively. The added diagnostics show that these values should be interpreted as benchmark interpolation rather than external generalisation or HVAC plant-sizing evidence. The contribution is therefore not a new learning algorithm, but an auditable workflow combining transparent reporting, residual checks, feature-dependence-aware interpretation, scenario screening, uncertainty communication and conservative deployment gates. Results indicate that overall height, roof area and glazing area dominate the fitted model's predictions, consistent with building-physics intuition and published studies.