Data-Driven Building Energy Modelling – Generalisation Potential of Energy Signatures Through Interpretable Machine Learning

Building energy modeling based on data-driven techniques has been demonstrated to be effective in a variety of situations. However, the question about its limits in terms of generalization is still open. The ability of a machine-learning model to adapt to previously unseen data and function satisfactorily is known as generalization. Apart from that, while machine-learning techniques are incredibly effective, interpretability is required for a "human-in-the-loop" approach to be successful. This study develops and tests a flexible regression-based approach applied to monitored energy data on a Passive House building. The formulation employs dummy (binary) variables as a piecewise linearization method, with the procedures for producing them explicitly stated to ensure interpretability. The results are described using statistical indicators and a graphic technique that allows for comparison across levels in the building systems. Finally, suggestions are provided for further steps toward generalization in data-driven techniques for energy in buildings.