Abstract
Electric heating in buildings is the dimensioning factor of the Norwegian power grid and is expected to become increasingly important globally with electrification of the building sector. Due to the lack of sub-meters, the full demand side flexibility potential of buildings remains unknown. While several studies have focused on disaggregating the total electricity consumption in single buildings, little research has been done on doing this for the aggregated electricity consumption of multiple buildings combined. This study demonstrates how data-driven decomposition methods can be applied not only to individual buildings, but also to aggregated electricity consumption of multiple buildings to extract electricity use for heating from total hourly electricity use data using real-world electricity consumption data from 30 school buildings. Results show that a CatBoost model performs well for the decomposition when the number of aggregated buildings in the training and test sets is the same or similar, typically with R2> 0.85, NMAE < 0.2, and peak demand error < 0.15. In contrast, there is low transferability between models trained on low aggregation levels (e.g., 1–5 buildings) and high aggregation levels. Overall, while the study is for the time being limited to school buildings, it offers a promising proof of concept for applying data-driven decomposition methods to aggregated smart meter data.