Enhancing interpretability and automation in data-driven energy modelling: an analytical approach to change-point regression models

Reliable, automated Measurement & Verification (M&V) and portfolio-scale energy analytics need models that are both accurate and interpretable. Current practice often relies on change-point regressions whose balance points are found via grid search or optimisation. As an alternative, an analytical formulation for simplified and automated identification of three-parameter heating (3PH), three-parameter cooling (3PC), and five-parameter (5P) models as defined in ASHRAE Guideline 14:2023 is proposed in this paper, with the goal of preserving interpretability also in more sophisticated workflows at the state of the art, which can use this novel formulation at different temporal scales (monthly, daily, and hourly). Standardised test datasets (39 in total) for 3PH, 3PC, and 5P models' testing and the Inverse Modelling Toolkit (IMT) have been used, showing comparable results in the large majority of cases and minor discrepancies in the others. The total batch runtime has been markedly reduced compared to the original implementation. Moreover, datasets from prior studies have been employed to evaluate applicability in real-world scenarios, demonstrating analogous results in this instance as well. While the current formulation is tested with monthly and daily interval data, its incorporation in hourly and sub-hourly resolution modelling workflows can promote further research developments in the area of interpretable data-driven analytics towards the “digital twins” paradigm, where interpretability of machine learning techniques and physical interpretation of underlying parameters is relevant to deliver effective and trusted solutions. Open-source code and datasets are made available to encourage further research on robust, transparent, and scalable data-driven energy modelling methodologies based on M&V principles. In this regard, additional efforts may be pursued to expand the concepts presented for the analytical formulation's development to encompass various automated processes with different objective functions (e.g., lasso, elastic net regression, etc.), model formulations, and constraints (e.g., physics-based interpretation of slopes and change points).