Accurate identification of influential building parameters through an integration of global sensitivity and feature selection techniques

The development of building energy performance simulation models often requires significant time and effort to achieve an acceptable degree of prediction accuracy. As such, energy modelers introduce various simplifications and assumptions that require a high degree of modeling literacy to avoid any errors in energy predictions. Previous studies relate these simplifications to the identification of influential building parameters using engineering judgment techniques that are often subjective and differ based on experts' opinion. The proposed methodology accurately defines influential and non-influential building parameters to formulate a guideline minimum dataset in the context of residential building energy models. The methodology integrates two feature selection techniques (Bayesian Information Criteria and Least Absolute Shrinkage with Selection Operator) with parametric analysis to determine the set of influential parameters. The study uses Irish residential archetypes to compare and validate the subsets of influential parameters using sensitivity rankings and established validation metrics. The predicted annual energy use lies within 10% of measured data for both subsets of influential parameters. Thereby, energy modelers could significantly reduce the time and effort spent on model development while maintaining the desired accuracy. The formulated datasets represent only influential features and hence, could be used by urban planners and energy policymakers to estimate energy retrofit investment costs, emission reductions and energy savings.