Estimation of the number of occupants using carbon dioxide concentration and differential pressure data-based machine learning model

¹ Department of Architectural Engineering, Sejong University, Seoul 05006, Republic of Korea
² Department of Architectural Engineering, Sejong University, Seoul 05006, Republic of Korea

^* Corresponding author: This email address is being protected from spambots. You need JavaScript enabled to view it.
^** Corresponding author: This email address is being protected from spambots. You need JavaScript enabled to view it.

Abstract

Abstract. The building sector accounts for a significant 39% of total global energy consumption, with HVAC systems consuming the largest share at 40%. Precise indoor occupancy estimation is a critical factor for demand-controlled ventilation, which balances energy reduction with comfortable indoor air quality. While environmental data-based methods are privacy-friendly, they often face challenges such as time lags and unaccounted air leakage. This study aims to enhance estimation accuracy by utilizing differential pressure data and door/window opening status to reflect these leakage variables. Data was collected from an office laboratory using IoT sensors to monitor CO₂ concentration, ventilation system power, and differential pressure between the room and adjacent spaces. Three machine learning models—Multi-Layer Perceptron (MLP), Random Forest (RF), and Long Short-Term Memory (LSTM)—were evaluated across five experimental cases. A physics-informed approach was implemented by incorporating the CO₂ mass balance equation as a specific input variable to calculate airflow rates. The results indicated that the LSTM model outperformed MLP and RF models in all scenarios. Case 5, which integrated all variables including the occupancy equation results, achieved the highest accuracy of 0.612 and the lowest Root Mean Squared Error (RMSE) of 0.981. The inclusion of door and window opening status significantly reduced estimation errors compared to the baseline case. In conclusion, incorporating real-time leakage through differential pressure and opening status into a physics-based neural network framework substantially improves the reliability of occupancy estimation for smart building control.