Application of machine learning algorithms to improve numerical simulation prediction of PM2.5 and chemical components

Abstract PM2.5 and its toxic chemical components are critical air pollutants that have been associated with respiratory diseases and human mortality. Therefore, the accurate prediction of PM2.5 and its chemical components is necessary to formulate emission reduction measures. Although chemical transport models can provide continuous temporal and spatial estimates of pollutants, the uncertainties in emission inventories, meteorological processes, and chemical mechanisms reduce the accuracy of modeling results. In particular, the discrepancy between simulated PM2.5 components and ground monitoring data remain large, with varying degrees of over and underestimation. To improve the accuracy of numerical simulation forecasts, we applied three typical machine learning algorithms to calibrate deviations. For this, we employed major meteorological parameters along with simulated and observed pollutant concentrations as inputs. The results showed that random forest and support vector regression (SVR) models presented much better predictive performance (R = 0.71–0.81 and 0.76–0.82, respectively) compared with multiple linear regression (MLR, R = 0.41–0.61). Moreover, their root mean square error and mean absolute error were 17–76% and 33–79% lower than those of MLR, respectively. The SVR model presented the most accurate prediction of PM2.5 components. The predicted proportions of PM2.5 components reflected the variations in pollution sources, which can be used to analyze the causes of pollution and thereby support the air quality management. More accurate prediction of PM2.5 components can promote exposure assessments and provide a basis for health studies.