Estimating the number of occupants using machine learning based on carbon dioxide concentration data from IoT sensors in two different spaces

Measuring the real-time number of occupants in a building is crucial for optimizing HVAC control, estimating infection risks, and predicting pollutant levels. There are two main approaches to estimate the number of occupants: using environmental data and using image data. The environmental data-based method typically estimates the number of occupants using carbon dioxide concentration as the primary input, along with other environmental variables like temperature, humidity, and ventilation rate. The image data-based method uses CCTV footage to estimate the number of occupants. However, when using environmental data, applying a classification model has limitations as it is only applicable to the specific space where the data was collected. If data from a specific space used for training a classification model is used to validate the model with data from a different space, significant errors occur. To reduce these errors, retraining the model with new data from the different space is necessary, making immediate application challenging. Furthermore, existing research has not considered changes in carbon dioxide concentration due to leakage. This study measures the pressure difference between indoor and surrounding spaces in two Living Labs A & B, preprocesses this data, and uses it along with environmental data as input variables for machine learning models. Artificial Neural Network (ANN) and Random Forest models are trained. The data learned from Living Lab A is used to validate both Living Lab A and B, and the model trained from Living Lab B is used to validate both Living Lab A and B. Additionally, a model trained using combined data from Living Labs A and B is also validated. The ANN model trained on Living Lab A showed a minimum Root Mean Squared Error (RMSE) of 2.41 when validated on Lab A data, but the RMSE increased to 2.85 when validated using Lab B data. This is significantly higher than the RMSE of 1.78 obtained when Lab B data was used to validate Lab B. The model trained using both Living Lab A & B data had an RMSE of 2.67 in Lab A and 2.19 in Lab B, an increase in error over the model trained in one space. This research compares models that estimate the number of occupants using environmental data acquired from different spaces and analyzes the accuracy when leakage is considered. It is anticipated that utilizing indoor-outdoor pressure differences in the future can lead to the development of more generalized models