3DVar sectoral emission inversion based on source apportionment and machine learning.

Air quality models are increasingly important in air pollution forecasting and control. Sectoral emissions significantly impact the accuracy of air quality models and source apportionment. This paper studied the 3DVar (three-dimensional variational) emission inversion method, which is based on machine learning, and then expanded it to sectoral emission inversion combined with source apportionment. Two machine learning conversion matrices were established to implement this method: a matrix that converts the total pollutant concentration to sectoral source apportionment results and a matrix that converts the sectoral source apportionment results to corresponding emissions. Combined with the O3 (ozone) concentration contributed by VOCs (volatile organic compounds) and NOx (nitrogen oxides) precursors in source apportionment, the inversion ability for O3-NOx-VOCs nonlinear processes was improved. Taking the BTH (Beijing‒Tianjin-Hebei) region from January 15 to 30, 2019, as an example, the results revealed that the regional errors of PM2.5 and O3 in the inversion experiment were reduced by 47% and 45%, respectively, and the temporal errors were reduced by 44% and 16%, respectively.