LSTM Deep Learning Models for Virtual Sensing of Indoor Air Pollutants: A Feasible Alternative to Physical Sensors

Abstract

Monitoring individual exposure to indoor air pollutants is crucial for human health and well-being. Due to the high spatiotemporal variations of indoor air pollutants, ubiquitous sensing is essential. However, the cost and maintenance associated with physical sensors make this currently infeasible. Consequently, this study investigates the feasibility of virtually sensing indoor air pollutants, such as particulate matter, volatile organic compounds (VOCs), and CO2, using a long short-term memory (LSTM) deep learning model. Several years of accumulated measurement data were employed to train the model, which predicts indoor air pollutant concentrations based on Building Management System (BMS) data (e.g., temperature, humidity, illumination, noise, motion, and window state) as well as meteorological and outdoor pollution data. A cross-validation scheme and hyperparameter optimization were utilized to determine the best model parameters and evaluate its performance using common evaluation metrics (R2, mean absolute error (MAE), root mean square error (RMSE)). The results demonstrate that the LSTM model can effectively replace physical indoor air pollutant sensors in the examined room, with evaluation metrics indicating a strong correlation in the testing set (MAE; CO2: 15.4 ppm, PM2.5: 0.3 μg/m3, VOC: 20.1 IAQI; R2; CO2: 0.47, PM2.5: 0.88, VOC:0.87). Additionally, the transferability of the model to other rooms was tested, with good results for CO2 and mixed results for VOC and particulate matter (MAE; CO2: 21.9 ppm, PM2.5: 0.3 μg/m3, VOC: 52.7 IAQI; R2; CO2: 0.45, PM2.5: 0.09, VOC:0.13). Despite these mixed results, they hint at the potential for a more broadly applicable approach to virtual sensing of indoor air pollutants, given the incorporation of more diverse datasets, thereby offering the potential for real-time occupant exposure monitoring and enhanced building operations.