Air pollutant dispersion in street canyons based on an outdoor scale model and machine learning

Air quality often deteriorates in street canyons owing to poor ventilation and increased emissions. In this study, the factors controlling air pollutant dispersion in street canyons were examined using outdoor scale models and machine learning. CO2 concentrations were measured at different heights for different ratios of building height (H) to street width (W) (H/W = 1, 2, 3). The results showed that when H/W increased from 1 to 2 and from 2 to 3, the mean CO2 concentration at a height of 0.25H (0.75H) in the street canyon increased by approximately 1 and 2 (0.5 and 1.5, respectively) times, respectively. An eXtreme Gradient Boosting (XGBoost) regression model for CO2 concentration was developed using machine learning based on different street canyon morphologies, monitoring locations, and meteorological conditions. Cross-validation demonstrated that the XGBoost model performed well on the test set, with an R2 value of 0.95. The SHapley Additive explanation (SHAP) values calculated for all samples showed that the five features that contributed most to the CO2 concentration were H/W, d (along-canyon position of the sensor), T (air temperature), W/2 (cross-canyon position of the sensor), and b_ws (atmospheric environmental background wind speed), with contributions of 34.1%, 19.1%, 12.1%, 10.5%, and 10.0%, respectively.