Rethinking building envelope design: Machine learning approaches to evaluate its impact on indoor ozone exposures

Abstract

Ozone is widely recognized as an ambient air contaminant that causes acute and chronic health effects. However, there is a limited number of studies investigating indoor exposures to ozone in occupied houses and linking design variables to the predictive power of indoor ozone levels. This study focuses on typical envelope airflow paths used in residences in the Philadelphia area in the United States. The model development draws from the field data, including indoor and outdoor ozone concentration, environmental parameters, and building characteristics from four building envelopes. Five machine learning algorithms (i.e., support vector machine, lasso regression, random forest, Bayesian bridge regression, and gradient boosting) are employed, with indoor ozone concentration as the dependent variable, as it indicates how the hot and sunny weather that might lead to the possibility of indoor air quality (IAQ) alerts due to ozone. The results showed that gradient boosting model based on all field measurements had the highest R-squared value of 0.974 and low enough root mean square error (RMSE) and mean absolute error (MAE) which are 1.182 and 0.788, respectively. We conclude that indoor ozone forecasting model based on inputting environmental survey (ES) in addition to either design variables or indoor environment characteristics can effectively predict and can therefore be used at the building design phase to improve healthy living environments.