Data-driven building energy modelling – An analysis of the potential for generalisation through interpretable machine learning

Data-driven building energy modelling techniques have proven to be effective in multiple applications. However, the debate around the possibility of generalisation is open. Generalisation involves the ability of a machine-learning model to adapt to previously unseen data and perform in a satisfactory way. Besides that, while machine-learning techniques are extremely powerful, interpretability, i.e. the ability for humans to predict how the model output will change in response to a change in input data or algorithmic parameters, is essential to attain a "human-in-the-loop" approach and creating feedback loops aimed at continuous improvement of efficiency measures in buildings. A flexible regression-based approach is developed and tested on a Passive House building in this study. The formulation employs dummy (binary) variables as a piecewise linearization method, and the rules for creating them are explicitly stated to ensure interpretability. Furthermore, the possibility of automating the model selection process using statistical indicators is described, including specific indicators used in Measurement and Verification (M&V) for the acceptance of calibrated energy models. The valuable insights that can be found using data-driven methods are reported and discussed, emphasising limitations and constraints, as well as the potential for future research focused on systems of (interpretable data-driven) models that can exploit the techniques' spatial and temporal scalability. Finally, the physical interpretation of model coefficients and the analytical formulations for energy model decomposition can be used to supplement the scalability of data-driven techniques and create more sophisticated systems of interconnected models.