Understanding end-use energy patterns in schools: Disaggregated EUI analysis with surrogate features and explainable machine learning

초록

Buildings account for a significant portion of total energy consumption in developed countries. Although detailed information on energy consumption and building characteristics is essential to identify the drivers of building energy patterns, publicly available datasets often lack the level of detail required for in-depth analysis. To enable end-use- and system-aware analysis under such limited public data, this study proposes a data-enrichment approach that augments building-stock datasets with inferred end-use EUIs and surrogate indicators of mechanical-system configuration derived from utility data. Specifically, this study disaggregated whole-building energy use intensity (EUI) into base, cooling, and heating EUIs using a simplified energy disaggregation method (SED) and developed proxy variables-specifically the gas consumption ratio and district heating ratio-to represent the types of mechanical systems used in school buildings. For 609 schools located in metropolitan areas of Korea, various potential predictors were collected from public data sources. Random forest (RF) regression models were developed to estimate the major end-use EUIs-base and heating-since cooling energy use accounted for only a minor portion of total consumption. The SHAP method was then applied to interpret the relationships between factors and end-use EUIs. The models achieved R2 values of 0.51 for base EUI and 0.33 for heating EUI. For base EUI, the most influential predictors were gas consumption ratio, number of students and teachers per area, and register date. For heating EUI, gas consumption ratio, register date, and district heating ratio were identified as key drivers. Overall, the proposed data-enrichment strategy demonstrates how actionable, end-use-specific insights can be extracted from widely accessible public datasets, supporting targeted energy-saving strategies for school buildings.

키워드