Occupancy-based energy consumption modelling using machine learning algorithms for institutional buildings

Abstract Occupancy information is one of the crucial variables in modelling and predicting the energy use in buildings. However, the presence of occupants is often stochastic in nature. In the presented case of eight space typologies derived from three institutional building blocks, a substantial variation in the correlation between occupancy and energy consumption is found for different space types during the semester and semester breaks for various resolutions of day and time. Further, it has been identified that a weak correlation between occupants and energy use is due to the use of common plug and lighting loads such as office printers, projectors, lab instruments, and fluorescent lamps. However, in the spaces studied, such as offices and computer rooms, the control of plug and lighting loads can be at individual occupancy levels for avoiding energy wastages. To characterize occupancy and energy consumption patterns by different space types and occupant types, this study develops and presents an integrated, data-driven modelling framework and results for different space types (like classrooms, studios, computer rooms, office spaces and laboratories and time resolution (hourly to semester-long intervals) for the case study building. It is found that the Deep Neural Network (DNN) model exhibits a slightly better prediction accuracy than conventional machine learning/regression models such as gradient boosting, support-vector network, and feed-forward neural network. However, in terms of computation time, the gradient boosting model is found to be faster than the DNN model for comparable outcomes. The developed integrated model will better equip the building facility management for an occupant-oriented control of building systems for energy efficiency.