Predicting PM2.5 Concentration in the Yangtze River Delta Region Using Climate System Monitoring Indices and Machine Learning

Abstract

Changing meteorological conditions during autumn and winter have considerable impact on air quality in the Yangtze River Delta (YRD) region. External climatic factors, such as sea surface temperature and sea ice, together with the atmospheric circulation, directly affect meteorological conditions in the YRD region, thereby modulating the variation in atmospheric PM_2.5 concentration. This study used the evolutionary modeling machine learning technique to investigate the lag relationship between 144 climate system monitoring indices and autumn/winter PM_2.5 concentration over 0–12 months in the YRD region. After calculating the contribution ratios and lagged correlation coefficients of all indices over the previous 12 months, the top 36 indices were selected for model training. Then, the nine indices that contributed most to the PM_2.5 concentration in the YRD region, including the decadal oscillation index of the Atlantic Ocean and the consistent warm ocean temperature index of the entire tropical Indian Ocean, were selected for physical mechanism analysis. An evolutionary model was developed to forecast the average PM_2.5 concentration in major cities of the YRD in autumn and winter, with a correlation coefficient of 0.91. In model testing, the correlation coefficient between the predicted and observed PM_2.5 concentrations was in the range of 0.73–0.83 and the root-mean-square error was in the range of 9.5–11.6 µg m⁻³, indicating high predictive accuracy. The model performed exceptionally well in capturing abnormal changes in PM_2.5 concentration in the YRD region up to 50 days in advance.