Neural Network-based Occupancy Detection on the Edge

Building occupancy information is essential to improve building performance and occupant comfort. Various non-intrusive detection methods using environmental sensors have already been explored, and ML methods have been shown to perform well in identifying occupancy patterns, but the scalability of the proposed approaches is still an open research question. Transfer learning offers a promising solution by adapting models to different targets and using shared knowledge for tasks with limited data. This study uses neural networks to detect indoor occupancy with sensor data in a real-world office setting, evaluating the performance and transferability of the models. We collected data with LoRa sensors in an office for three months and used them to train and test the Feed Forward Neural Network, Convolutional Neural Network, Long Short-Term Memory, and XGBoost models. We also explored model deployment on edge devices with TensorFlow Lite, achieving an average F1-score of 0.77 for occupancy detection and 0.62 in the transfer scenario. Our analysis identified critical features for occupancy detection.