Unsupervised spatiotemporal adaptive model transfer framework for nonintrusive occupancy detection using environmental data

초록

Occupancy detection is fundamental to the optimization of energy management and occupant comfort in smart buildings. However, existing methods raise privacy concerns and are poorly generalizable to different spaces. This study introduces an unsupervised spatiotemporal adaptive model transfer framework for nonintrusive occupancy detection. By leveraging existing environmental sensors in building automation systems, our approach ensures privacy, cost-efficiency, and deployability without requiring labeled data from target spaces. Using a long short-term memory-based spatiotemporal feature signature and spatiotemporal similarity matching, our frameworks dynamically selects the most effective model in the given context. It also mitigates the slow response of CO2 sensors with multimodal sensor fusion, which combines CO2 sensing with temperature, humidity, and illuminance sensing. Assuming binary occupancy detection and a 10-min aggregation window, our multi-modal fusion consistently outperform single-sensor setups across five offices and three experimental periods. Extensive benchmarking with 15 algorithms clarifies the spatiotemporal context dependency of the occupancy detection performance of models, aligning with the no-free-lunch theorem in occupancy detection. The accuracy of the proposed framework is 85%, surpassing those of the top four baselines (79% on average) and the simple statistical approach (81%) and effectively approaching the fully supervised upper bound (89%). Overall, this framework offers a scalable, label-free solution that performs near the supervised limits while remaining practically operable.

키워드