Assessment of Occupancy Estimators for Smart Buildings

Dense distributed sensor systems deployed to monitor and control the built environment open up wide application areas for improving the quality of life through smart spaces. By implementing these systems in residential or commercial buildings, more efficient, pervasive, real-time and finally more robust decision support is achieved. As one of the key tasks, high accuracy occupancy detection and estimation within buildings offers the potential to improve HVAC utilization while reducing maintenance needs, enhancing energy savings for building operators as well as a more suitable thermal comfort for occupants. In this context the main contribution of the paper consists of an assessment of supervised machine learning techniques namely random forests and extreme machine learning neural networks for indirect occupancy profile estimation. The methods are applied on a benchmarking dataset of indoor carbon dioxide measurements and ventilation damper positions as collected from the building information system. The impact of data preprocessing through filtering and smoothing as well as hardware constraints and cloud distributed processing for algorithm deployment are also discussed. We highlight key challenges and discuss how the learned models can be integrated for online operation to improve energy efficiency.