Indoor air parameters (CO₂, temperature, and humidity) as indicators of room occupancy

For citation: Rogal I.O., Nikitin A.A., Muraveinikov S.S., Kropis Yu.N., Nikitina V.A. Indoor air parameters (CO₂, temperature, and humidity) as indicators of room occupancy. Journal of International Academy of Refrigeration. 2026. No 2. p. 3-11. DOI: 10.17586/1606-4313-2026-25-2-3-11 (in Russian)

Abstract
This paper addresses the problem of estimating the number of occupants in indoor spaces based on indoor air parameters such as CO₂ concentration, temperature, and relative humidity. Particular attention is given to the use of a multilayer perceptron (MLP) model as one of the possible neural network–based solutions. The relevance of the study is driven by the need to improve the energy efficiency of ventilation systems in modern buildings by adapting their operation to the actual occupancy of spaces. In the study, a comparative analysis of existing occupancy estimation approaches is carried out, including direct methods (motion sensors and video surveillance systems), algorithmic methods (signal filtering models and empirical relationships), and machine learning techniques, including neural networks. It is shown that traditional algorithmic approaches exhibit limited accuracy under conditions of environmental variability and incomplete data, whereas neural network architectures are more robust to noise and capable of modeling complex nonlinear relationships. An investigation of the application of a multilayer perceptron neural network for room occupancy estimation based on measured indoor air parameters – temperature, CO₂ concentration, and relative humidity – is conducted. Sensor data with the specified features, obtained from an open dataset, are used to train the model, while the target variable is a binary occupancy indicator (occupied/unoccupied). The experimental results confirm the feasibility of the proposed approach: the MLP model achieves a high classification accuracy (>85%). In future, more advanced deep learning architectures, such as convolutional neural networks (CNNs), are planned to use for further improving the model performance.