Building occupancy number prediction: A Transformer approach

Abstract

Buildings substantially impact global energy usage and emit significant carbon dioxide. Building occupancy is crucial to enabling energy conservation in buildings and achieving zero emissions by 2050. While existing occupancy prediction methods have made remarkable progress, their analysis is limited to complex practical scenes. In addition, the expectations of Transformers are high for predicting building occupancy. To address these problems, we introduce Transformer algorithms for building occupancy number prediction. We publicly provide an actual operating dataset for 2 weeks in 6 zones, including multi-sensor sensing information (e.g., temperature, occupancy, relative humidity, FCU control, FCU fan feedback, and FCU on/off feedback). To evaluate the performance, we provide an experimental analysis and comparison among our Transformer-based occupancy prediction algorithm and different machine learning methods (e.g., random forest, decision trees, XGBoost, and long short-term memory networks). Our Transformer-based occupancy prediction algorithm performs better on our dataset than other existing algorithms. The code, dataset, and demo are available at https://github.com/kailaisun/occprediction.