Identification of essential variables for the estimation of energy demand in buildings at European scale

Accurate EU-scale building energy modelling remains challenging due to fragmented data, yet it is essential for assessing policy scenarios. This study proposes a data-driven methodology to identify a minimal set of essential variables for a reduced-form model capable of accurately predicting energy demand across European buildings. Using Random Forest machine learning applied to two complementary datasets: (1) a cross-country database of building typologies in EU28, and (2) the French Energy Performance Certificate (EPC) database—the analysis identifies five key variables that reduce mean absolute percentage error by 10 percentage points compared to using all 18 original available attributes. These variables are: living floor area, height, construction year, climatic conditions, and building functionality, while construction materials and other architectural features show limited predictive relevance. The study fills a gap in the literature by rigorously establishing essential variables required for EU-scale energy modelling, complementing previous work that either lacks cross-country generalization or focuses on single-country assessments. The main novelty lies in demonstrating that competitive predictive accuracy can be still achieved using standardized, widely available features, challenging the assumption that detailed physical data are required for robust models. This approach prioritizes scalability and cost-effectiveness, enabling wider development across Europe towards the creation of a detailed EU Digital Building Stock Model for energy-related purposes. As key variables can be derived from Earth Observation, cadastral records, and machine-learning-based inference, the findings support harmonized, building-level energy estimation and facilitate applications such as identifying worst-performing buildings, prioritizing renovation actions, and informing decarbonization strategies. • Insights are given for the creation of detailed EU-scale building energy modelling. • Open data sources are analyzed to identify essential model energy variables. • Machine learning models are fitted on a set of explanatory variables. • Floor area, height, age, climate and functionality selected as essential variables • Competitive predictive accuracy achieved with selected essential variables.