Machine learning-based white-box prediction and correlation analysis of air pollutants in proximity to industrial zones

The adverse health effects caused by long-term exposure to high pollution volumes from industries near urban areas are a growing concern. Determining accurate distribution models of pollutants is crucial for establishing safe distances between sectors and urban regions and continuously monitoring pollutant levels. This study was conducted in Siraf City, situated in the Pars special energy zone in southern Iran, to improve the accuracy of simulation results and identify the correlation between emission models and pollutant concentrations. To achieve this goal, concentrations of seven pollutants (CO, CO2, NO2, SO2, O3, PM2.5, PM10) were determined seasonally at 45 points within the study area using field sampling and numerical simulation with AERMOD software. Subsequently, the obtained results were seamlessly transferred into new domains with the primary objective of feature engineering. These engineered features were then fed into an XGBoost model for regression analysis to obtain coefficients, deriving seven equations to enhance pollutant concentration simulations' accuracy signifi-cantly. The developed equations improved the simulation accuracy for CO (12.54%), CO2 (12.91%), NO2 (0.94%), SO2 (6.7%), O3 (3.05%), PM2.5 (12.47%), and PM10 (4.62%). The findings demonstrate varying improved accuracy levels depending on the pollutant and simulation accuracy with well-known machine learning algorithms. The machine learning model effectively reveals the relationship between emission models and pollutant concentrations, offering valuable insights to enhance the accuracy of air pollutant emission predictions.