Unraveling Street-Level Air Pollution upon a Pivotal City of Yangtze River Delta, China

We use two machine learning models—random forest (RF) and recurrent neural network (RNN)—to analyze and predict air pollutants in a pivotal city of YRD, China. We quantitatively show the determinants for the atmospheric pollutants, providing insights for air pollution control policies. We propose, test and verify a five-step avenue to forecast the main atmospheric pollutants (SO2, NO2, CO, O3, PM2.5 and PM10) in the future 24 h. Step one, WRF is used to generate the meteorological conditions in the next 24 h. Step two, SO2 and CO are predicted by RNN using WRF-simulated meteorological conditions. Step three, NO2 is predicted by RNN using WRF-simulated meteorological conditions and RNN-simulated CO. Step four, O3 is predicted by RNN using WRF-simulated meteorological conditions and RNN-simulated CO and NO2. Step five, PM2.5 and PM10 are predicted by RNN using WRF-simulated meteorological conditions and RNN-simulated SO2, CO and NO2. The significant role that dew-point deficit plays in shaping SO2, NO2 and O3 is recognized. CO was strongly positively linked with NO2 and PM2.5. Decrease of CO may trigger the crescendo of ground-level O3. Stratospheric downward transport played paltry role in shaping tropospheric O3 at Hangzhou. We also identify that some illegal factories were surreptitiously emitting trichlorofluoromethane (CFCl3), one of the strongest stratospheric ozone-depleter that should have been forbidden since 2010.