Fifty shades of black: Uncovering physical models from symbolic regressions for scalable building heat dynamics identification

Abstract

The rapid growth of machine learning (black-box) techniques and computing capacity has started to transform many research domains, including building performance analysis. However, physical interpretation of these models remains a challenge due to their opaque nature. This paper outlines an experiment to unveil analytical expressions from an open-source machine-learning-based algorithm, i.e., symbolic regression. From 241 residential buildings in the Netherlands, 50 unique analytical expressions were produced demonstrating overall better characterization accuracies than an XGBoost baseline, while providing a powerful mean of interpretability from model structures and coefficients. These insights present a starting point for further work towards highly scalable models yielding new characterizations of residential buildings.