Estimating concentrations for particle and gases in a mechanically ventilated building in Hong Kong: multivariate method and machine learning

Lack of characterization of indoor pollutant concentrations has been identified as a key barrier for exposure and health estimates. In this study, a field campaign was conducted to measure indoor concentrations of PM1, PM2.5, PM10, CO, and NO2 in a mechanically ventilated building. Statistical method using multivariate linear regression (MLR) and machine learning using random forest (RF) were used and compared to quantify variations in observed concentrations and were then used to predict indoor concentrations for selected pollutants. The two methods were consistent in identifying major predictors for each pollutant. Outdoor particles were the single largest predictors found for PM1 and PM2.5, while indoor environment and occupant-related variables were dominant predictors for PM10, CO, and NO2 in the selected mall. Based on MLR models, outdoor PM accounted for 91%, 64%, and 25% of variations in indoor PM1, PM2.5, and PM10 during opening hours. More than 30% of indoor CO variations were related to time-dependent activities. Nearly 50% of the indoor NO2 variations were explained by temperature and relative humidity. Both models are useful in predicting indoor concentrations. In the tenfold cross validation, RF models showed high prediction capability for PM1 (R2 > 0.9) and moderate (R2: 0.5 ~ 0.7) for the other four pollutants in both periods except for PM10 during non-opening hours (R2 = 0.3). MLR models exhibited comparable prediction power for PM1 and PM2.5, but generally lower for PM10 and gases. Availability of parameter information in modern cities facilitates the application of such models on large scale, based on proper validation, for better characterizing of indoor air quality.