Investigating wintertime air pollution in Hangzhou, China

Hangzhou, one of the most prosperous cities in China, suffered from severe air quality degradation in wintertime, and the ambient atmospheric particulate matter (PM) has become the most public-concerning air pollutant. In this work, a case study in wintry Hangzhou is made, for analysis of air pollutants and prediction of PM2.5/PM10 using two machine learning models, recurrent neural network (RNN) and random forest. The results signify that statistic-based and inventory-free machine learning is competently alternative to the inventory-predicted atmospheric models. Variable importance (VI) indicates that CO was the predominant factor for both PM2.5 and PM10. Dew-point deficit played an essential role in shaping gaseous air pollutants. Water vapor pressure and hydrostatic energy had trivial impact on atmospheric pollutants. RNN and random forest both show high accuracy in predicting PM2.5 and PM10. The inter-annual consistence of PM's components is confirmed. A method to pinpoint whether the high episodes of PM were spawned by long-range transport or increase of gaseous pollutants (SO2, NO2, and CO) is proposed. Additionally, the possible chemical bond between CO and PM needs to be further investigated.