CO₂-Based Occupancy Detection in Apartments: A Longitudinal Study for Enhanced Indoor Comfort and Energy Efficiency

This paper presents a sensor-based framework for multi-class indoor occupancy detection using environmental data in a residential setting. Over a 50-day period, continuous measurements of CO<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</inf> concentration, temperature, and relative humidity were collected from a residential bedroom. These data were used to train a long short-term memory (LSTM) model to estimate occupancy levels ranging from 0 to 3 individuals. Ground-truth annotations were established using camera footage solely during the data collection phase, ensuring that the deployed system relies exclusively on non- intrusive ambient sensors. The LSTM model achieved a test accuracy of 97.1%, with balanced precision, recall, and F1-scores across all occupancy classes. Detailed analysis of the temporal patterns revealed distinct correlations between occupancy states and fluctuations in environmental parameters, validating the effectiveness of the selected features. The proposed framework supports real-time occupancy inference and offers significant potential for informing adaptive control strategies in heating, ventilation, and air conditioning (HVAC) systems. By leveraging ambient sensor data, this approach provides a privacy-preserving, data-driven solution that enhances energy management and improves indoor comfort in residential environments.