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

Area of Science:

• Building Science
• Machine Learning
• Energy Efficiency

Background:

• Early-stage architectural design significantly influences a building's energy consumption.
• Surrogate models trained on benchmarks require defined validity limits for reliable application.

Purpose of the Study:

• To present a risk-aware, reproducible, and explainable machine learning workflow for screening heating load (HL) and cooling load (CL) indicators.
• To evaluate the performance of different machine learning models using concept-stage geometric and envelope parameters.

Main Methods:

• Utilized the Energy Efficiency (ENB2012) dataset with 768 building configurations.
• Evaluated Ridge regression, Random Forest, and Extremely Randomised Trees (Extra Trees) models.
• Incorporated diagnostics for geometry-blocked validation, conformal prediction intervals, and TreeSHAP attribution.

Main Results:

• The Extra Trees model achieved low Root Mean Square Error (RMSE) for HL (0.60 kWh/m2) and CL (1.45 kWh/m2), with high R2 values (0.996 and 0.977).
• Diagnostic assessments indicated that model predictions are best suited for benchmark interpolation, not external generalization.
• Overall height, roof area, and glazing area were identified as key predictors, aligning with building physics principles.

Conclusions:

• The study provides an auditable workflow for transparent reporting, uncertainty communication, and conservative deployment of building energy models.
• The workflow emphasizes interpreting model predictions within defined limits, distinguishing interpolation from generalization.
• Results confirm the importance of geometric factors like height, roof, and glazing area in determining building energy loads.