Cross-domain disaggregation of electricity for heating in all-electric school buildings – learning from school buildings with district heating

• Most buildings with electric heating lack sub-meters for heating loads. • Disaggregation of electricity for heating from hourly AMS meters in buildings. • Cross-domain training on energy measurements from buildings with district heating. • Data from 74 school buildings including novel dataset. • CatBoost achieves 0.91 R 2 , 2.6 % NMAE and 8 % peak load error for disaggregation. Electric heating is widespread in Norwegian buildings and significantly contributes to peak loads in the electricity grid. Non-residential buildings are typically heated either by district heating or a combination of electrical heating appliances. Despite its widespread use, most buildings lack sub-meters for electric heating. As a result, the true potential for energy efficiency and load flexibility from heating appliances in buildings remains unknown. Non-intrusive load monitoring and disaggregation techniques offer alternatives to sub-metering by using data-driven methods to extract electricity use for appliances from time-series data. However, little research has been conducted on disaggregating electrical heating loads from low-resolution data, partly due to the scarcity of sub-metered training datasets. Unlike all-electric buildings (EHBs), district heating buildings (DHBs) typically have separate, hourly heating energy meters. This paper examines feature extraction and multiple machine learning algorithms for disaggregation of electricity for heating from AMS-meter data in EHBs, and how cross-domain training from DHBs can contribute to this task. We use sub-metered data from 74 school buildings (54 DHBs and 20 EHBs) in Norway with over 3.8 million hours of recorded measurements, where parts of the dataset are published in a novel public dataset. Results show that CatBoost achieves high performance in disaggregating electricity for heating in EHBs when trained on data from DHBs, with an R 2 value of 0.91, NMAE of 2.6%, and a peak load estimation error of 8%, which is an improvement compared to training on EHBs. The study also shows that feature engineering can improve the disaggregation performance in some, but not all EHBs.