Data-driven building energy models for design and control of community energy systems

Abstract

Building energy models are used to forecast building energy use to design and control efficient building energy systems. Building energy use can generally be decomposed into heating, ventilation and air conditioning, refrigeration, appliance and lighting loads. These loads will depend on multiple factors such as outdoor weather conditions, occupants, building type, controls and scheduling. Data-driven models have become more popular with the increase in smart meter data available that can be used to train and fit the models. Additionally, buildings with high refrigeration loads have greater heat harvesting potential, however, few data-driven models have been developed for buildings such as supermarkets and ice rinks. In this work, linear regression models are used to predict the disaggregated space cooling, heating, baseload and refrigeration components of building energy use. In most cases, measured aggregate electricity use is available, however individual appliances or component loads require submetering equipment which can be expensive. Therefore the proposed models use time-based and weather features to separate the thermal and baseload portions of the electrical load. A generalized approach is also used to predict new buildings with data from existing buildings. Furthermore, a simplified model is used to predict hourly space heating from monthly natural gas measurements and hourly weather measurements. The models were evaluated on real data from buildings in Ontario and the disaggregated loads were verified with synthetic data. The results found that aggregate use was predicted reasonably well using linear regression methods, with most building types having a median normalized root mean squared error between 0.2 and 0.3, depending on the forecasting period. The model is flexible as it does not require detailed information related to the building such as lighting or setpoint schedules, however, it can be adapted in the future to include additional information and improve predictive capability.