Occupancy modeling on non-intrusive indoor environmental data through machine learning

The primary drivers of energy consumption within buildings are the occupants. Non-intrusive Internet of Things (IoT) technology can be utilized to detect occupancy and optimize energy performance while preserving the privacy of building occupants. This study explores the suitability of various indoor environmental data for occupancy detection in office rooms. Data was collected utilizing an IoT sensory device, recording CO2 concentration, air temperature, relative humidity, indoor air quality, sound pressure level, and illuminance in two double-occupied office rooms at a campus building in Munich, Germany. Machine learning models, namely Random Forest, XGBoost, and dense neural networks were trained on the dataset, and the models’ feature importance was analyzed through post-hoc explainability testing. The best features were then incorporated in a sliding window of 15 min to consider temporal dependencies in the models. The study also compared two hyperparameter optimization techniques: Random and Bayesian optimization. The results of this study indicate that with six days of ground truth data of sound pressure level, CO2, and illuminance, accuracy and F1-score for occupancy detection of above 0.95 and 0.93 can be respectively achieved. •Comparison of random and Bayesian hyperparameter optimization.•Bayesian optimization superior in terms of the number of trials and validation loss.•Sound pressure level, CO2 level, and ambient light as the highest-ranking features.•High accuracies with six days of ground truth on models with top-ranking features.