CO2-Based Ventilation Control Strategy Via LSTM-Based Long-Term CO2 Prediction Model

Indoor air quality (IAQ) is essential to the health, well-being, and productivity of occupants. CO2-based demand ventilation control is considered an effective strategy for maintaining satisfactory IAQ when the occupancy levels vary greatly. However, few studies have established and validated the relationship between CO2 concentration and dynamic occupancy level. Additionally, the demand ventilation control is at the cost of reduced ventilation, which is not recommended during the COVID-19 period, so it is necessary to control the minimum fresh air requirement. CO2 concentration prediction can be used as the feedback of ventilation control, so prediction accuracy is vital for the effectiveness of ventilation control. In order to fully exploit the temporal features of the input variables and achieve better performance in long-term prediction, a deep learning prediction model based on long and short-term memory (LSTM) is developed to learn the dynamic patterns of CO2 concentrations over a variety of time horizons. The model integrates optimal time lags and periodicity features to improve the prediction performance. The results show that the proposed control strategy based on the dynamic CO2 level can reduce indoor CO2 levels and achieve a 4.7% energy saving. Adding periodic features as auxiliary data can improve prediction accuracy by up to 7.73% in the long-term prediction task of CO2 level. The performance of the direct strategy in long-term CO2 prediction is better than that of the recursive strategy in terms of prediction accuracy and computation time.