Predicting occupancy counts using physical and statistical Co^sub 2^-based modeling methodologies

Energy consumption and indoor environment quality (IEQ) of buildings have been linked to human occupants. Predicting the number of occupants in a space is essential for the effective management of various building operation functions as well as improve energy efficiency. This study is the first to compare the performance of physical and statistical models in predicting occupant counts in a high volume lecture theatre (Occ = 200) using CO2 sensors. CO2 measurements and actual occupant numbers were obtained for 4 months to provide robust data comparison of the methodologies. It was found that that the dynamic physical models and Support Vector Machines (SVM) and Artificial Neural Networks (ANN) models utilizing a combination of average and first order differential CO2 concentrations performed the best in terms of predicting occupancy counts with the ANN and SVM models showing higher predictive performance. RMSE values for the corresponding models were 12.8, 12.6 and 12.1 respectively and correlation coefficients were all greater than 0.95. The relatively good agreement between dynamic physical model predictions and ground truth shows that the dynamic mass balanced model is adequate for predicting occupancy counts provided that the air exchange rates measured are accurate. Model average accuracies across all tolerance was between 70 and 76% demonstrating good performance for a large number of occupants. A discussion on the merits and limitations of each model types was presented to provide guidance on the adoption of various models.