Using neural networks for the classification of activities in an office: An application for a better prediction of CO2 concentration in tertiary buildings

Events such as the opening and closing of doors and windows and the number of people entering and leaving an office, can’t be predicted and play a major role in CO2 (carbon dioxide) concentration of an occupied zone. In this paper we propose a method for detecting the changes these events cause in the time series using neural network models. The outcome allows to improve the accuracy of CO2 concentrations predictions in a numerical model simulation.

A Recurrent Neural Network (RNN) model was used. Training and test sets were generated to train the model and test its performance. Data on the presence/absence of office occupants and the opening/closing of doors and windows were registered manually at a 30-minutes time step. The occupants’ performance register along with the concentration of CO2 measured with an installed air quality sensor were introduced as inputs to the models to achieve the training process. As an output, the RNN model predicts the status of the openings and the number of occupants. Different time-series steps and data sizes were tested.

Results were encouraging. Presence was globally well identified, while the detection of door openings and closings was more difficult, due to a delay between the action and its impact on the CO2 concentration. The transient profile of the CO2 curves makes it difficult to identify if a given pattern is due to an opening or a change in the presence of occupants.

Finally, the RNN model is generic and can be used on other case studies where there are no event sensors. In perspective, several approaches are being tested to reduce the delay in detecting opening/closing events such as reducing the time step of data registration or finding an optimal position of the CO2 sensor like placing it closer to the opening.