Machine Learning Models for Indoor PM2.5 Concentrations in Residential Architecture in Taiwan

Abstract

People typically spend 80-90% of their time indoors. Therefore, establishing prediction models estimate particulate matter (PM2.5) concentration in indoor environments is of great importance, especially in residential households, in order to allow for accurate assessments of exposure in epidemiological studies. However, installing monitoring instruments to collect indoor PM2.5 data is both labor and budget-intensive. Therefore, indoor PM2.5 concentration prediction models have become critical issues. This study aimed to develop a predictive model for hourly household PM2.5 concentration based on the artificial neural network (ANN) method. From January 2019 to April 2020, PM2.5 concentration and related parameters (e.g., occupants’ behavior information and ventilation settings) were collected in a total of 62 houses and apartments in Tainan, Taiwan (tropical and subtropical region). Overall, 2136 pairs of data and 9 possible variables were used to establish the model. Meteorological data were primarily used to establish the model. Meanwhile, occupants’ behavior and building characteristics were generalized as effective opening areas to describe the importance of ventilation in subtropical areas. We performed five-fold cross-validation to assess prediction model performance. The prediction model achieved promising predictive accuracy, with a coefficient of determination (R2) value of 0.88 and a root mean square error value of 3.35 (μg/m3), respectively. Outdoor PM2.5 concentrations were the most important predictor variable, followed in descending order by temperature, outdoor carbon dioxide concentration, outdoor relative humidity, and opening effective areas. In summary, we developed a prediction model of hourly indoor PM2.5 concentrations and suggest that outdoor meteorological data, building characteristics, and human behavior can be powerful predictors. The results also confirm that the model can be used to predict indoor PM2.5 concentrations across seasons.