Building Energy Surrogate Modelling – a Feature Selection Methodology

A mathematical regression model, referred to as a surrogate model as it was trained on a set of computer-simulated results, was developed to permit the rapid modelling of large commercial office buildings within a single climate zone. The model was developed using a large number of building features and their EnergyPlus simulated results. In previous building energy surrogate modelling, a research gap in selecting building features using statistical approaches was identified. This thesis investigates a feature selection method, including forward stepwise selection and least absolute shrinkage and selection operator (LASSO), to identify building features that, together, have the most significant impact on annual building energy use. The final model, with 23 features selected through this methodology, predicts annual building energy use at 11.3% error, on average.