Closer to Realtime: A Comparison of Machine Learning Methods to Reduce Artificial Switch-Off Times

Presence-based control of artificial lighting is an important method of reducing energy consumption for artificial lighting. Passive infrared sensors are the most commonly used detection technology due to their low investment costs and simple commissioning. However, these are typically tied to switch-off times to reduce nuisance false-off rates. If the switch-off times are set too long, this can impair energy efficiency. Probabilistic methods make it possible to reduce switch-off times based on real occupancy patterns. In order to automate this process and take seasonal effects into account, various machine learning-based modelling approaches are compared. The basis is a high-resolution data set from an open-plan office. The results show that the modelling approaches are highly accurate in predicting periods of inactivity and thus create a solid basis for deriving suitable time-variant switch-off times. This is reflected in a lower energy requirement.