Efficient occupancy detection could be an important step towards improved performance of HVAC systems (heating, ventilation, and air conditioning), by modelling human presence and identifying residents' living patterns. This study aims to develop a method of predicting occupancy based on non-visual environmental sensor data, by means of creating an LSTM (long short-term memory) neural network. The model was trained on data provided by the KTH Live-In Lab, including CO2, relative humidity, and indoor temperature from one apartment, spanning from the summer of 2022 to the summer of 2023. Five distinct models were created, one for each season and one covering the entire year. Results showed that the best performance was achieved with the models for summer and winter, achieving an accuracy increase of 45% during winter compared to simply guessing that someone was home all of the time and summer achieving an increase of 24%. In contrast, models for autumn and spring only showed improvements of 5% and 8% respectively, and the full-year model achieved an overall increase of 24%.
Apéll et al. (Wed,) studied this question.