Machine learning models for predicting interactions between air pollutants in Tehran Megacity, Iran

Abstract

Air pollution has significant detrimental impacts on the environmental compartments and human health. The present study investigates the interactions between selected air pollutants, including CO, O3, NO2, SO2, PM10, and PM2.5, in Tehran megacity, the capital of Iran, using statistical modeling and simulating approaches. Data on selected air pollutants have been extracted from 16 sensors across Tehran during 2013–2022. Using the development of three machine learning models, including XGBoost (XGB), LightGBM (LGBM) and Random Forest (RF), significant relationships between the air pollutants were observed. The comparison of the models demonstrated that the RF model has the highest level of optimality in forecasting the concentrations of CO, O3, NO2, and SO2 (R2CO = 0.63, R2O3 = 0.6, R2NO2 = 0.54, R2SO2 = 0.55). The LGBM model provided higher optimality for NO2 and PM2.5 (R2NO2 = 0.56, R2PM2.5 = 0.88), while the XGB model exhibited higher accuracy for the PM10 (R2PM10 = 0.88). Regarding the results obtained from the statistical models, it can be inferred that the RF model has superior performance in the forecasting of gaseous pollutants. Moreover, the XGB and LGBM models exhibited comparable performance and therefore may be regarded as appropriate options for PM prediction. The findings from the simulations indicated that a rise in an individual pollutant likely leads to an increase in other pollutants. Consequently, implementing air quality management strategies for a specific pollutant meaningfully, directly influences other pollutants, highlighting the significance of considering the chemical aspect of air pollutants interaction and enforcing air pollution management rules.

Keywords

• Air pollution
• Air quality
• Modeling
• Machine learning
• Simulation