Spatial Prediction of Relative Building Energy efficiency in Slovenia Using Non-Energy Predictors from the National EPC Registry

Abstract

National registries of energy performance certificates have become a standard empirical basis for analysing building stocks, but most predictive studies built on such registries rely on direct energy indicators that closely mirror the certification output. This article develops a spatially aware machine learning framework for Slovenia that predicts relative building energy efficiency using only non energy predictors, isolating the contribution of administrative location to predictive performance. The analysis uses a public snapshot of the Slovenian EPC registry covering 58,310 records, 16 variables and 2,360 unique cadastral community codes between 2017 and 2025. The target is constructed as a year normalised binary indicator, equal to one when a building’s primary energy exceeds the national median primary energy in its issue year. After target construction all direct energy variables are dropped from the predictor set, leaving conditioned floor area, certificate type, EPBD category, issue year, building part number and the cadastral community code. Three logistic regression models are compared. Model A uses only the six non spatial predictors. Model B augments Model A with smoothed out of fold target encoding of the cadastral community code. Model C further adds a neighbourhood based spatial lag computed from cadastral adjacency. On a stratified 80/20 split, Model A reaches ROC AUC 0.730 and F1 0.699, Model B reaches 0.748 and 0.708, and Model C reaches 0.748 and 0.708. A temporal holdout that trains on 2017 to 2024 and tests on 2025 produces the same ordering with lower absolute scores (ROC AUC 0.640, 0.661 and 0.661). The paper contributes a year relative target formulation that is robust to the structural break in the primary energy unit observed between 2022 and 2023 in the public registry, an empirically grounded quantification of the predictive value of cadastral location when no direct energy variable is available, and a negative finding that a neighbourhood spatial lag adds no measurable value once smoothed target encoding has been applied to the same cadastral feature. The framework is implemented in Python and the full notebook, processed cadastral support files and per model output figures are released for reproducibility.