Applicability and accuracy of physical CO₂ mass balance in multi-zone mechanically ventilated office buildings to estimate real-time occupancy

Accurate real-time occupancy estimation in office buildings is crucial for efficient system control and security. While AI-based methods have recently gained prominence, this study revisits the physical CO₂ mass balance approach, traditionally applied to simpler environments, by evaluating its performance in a complex multi-zone mechanically ventilated office with high occupancy. A detailed sensitivity analysis identified the change in indoor CO₂ concentration as the most influential variable, guiding the application of smoothing filters. A vision-based ground truth dataset over 25 days enabled rigorous model validation using various error indicators. The best-performing model, using an Exponential Moving Average filter with a 0.9 smoothing factor and 10-minute data averaging, achieved a Mean Absolute Error of 1.56 persons and Root Mean Square Error of 2.98 persons—under 10 % of the maximum recorded occupancy. However, it tended to overestimate during unoccupied hours due to sensitivity to minor CO₂ fluctuations. With a five-person tolerance, the model achieved over 90 % accuracy, demonstrating practical applicability for system control. Model variations, including rounding and enhanced variable monitoring, further improved results. The study confirms that indoor CO₂ measurement quality is critical, particularly sensor placement. During working hours, the model reliably detects presence but struggles with exact occupancy. Compared to AI models, this approach is simpler, does not require training, and is broadly applicable where relevant variables are monitored. Future work will explore sensor layout optimisation and dynamic variables like outdoor CO₂. Overall, the CO₂ mass balance remains a viable, low-complexity alternative for occupancy estimation in multi-zone offices.