VOC transport in an occupied residence: Measurements and predictions via deep learning

Abstract

Monitoring and prediction of volatile organic compounds (VOCs) in realistic indoor settings are essential for source characterization, apportionment, and exposure assessment, while it has seldom been examined previously. In this study, we conducted a field campaign on ten typical VOCs in an occupied residence, and obtained the time-resolved VOC dynamics. Feature importance analysis illustrated that air change rate (ACR) has the greatest impact on the VOC concentration levels. We applied three multi-feature (temperature, relative humidity, ACR) deep learning models to predict the VOC concentrations over ten days in the residence, indicating that the long short-term memory (LSTM) model owns the best performance, with predictions the closest to the observed data, compared with the other two models, i.e., recurrent neural network (RNN) model and gated recurrent unit (GRU) model. We also found that human activities could significantly affect VOC emissions in some observed erupted peaks. Our study provides a promising pathway of estimating long-term transport characteristics and exposures of VOCs under varied conditions in realistic indoor environments via deep learning.

Keywords: Deep learning; Indoor air quality; Long short-term memory network (LSTM); Residence; Volatile organic compounds (VOCs).