Publications

Wang, XP; Sun, WB (2019). Meteorological parameters and gaseous pollutant concentrations as predictors of daily continuous PM2.5 concentrations using deep neural network in Beijing-Tianjin-Hebei, China. ATMOSPHERIC ENVIRONMENT, 211, 128-137.

Abstract
The deep learning model can simulate the complex nonlinear relationship between PM2.5 and aerosol optical depth (AOD), and has great application potentiality in PM2.5 inversion. However, the underestimation of high PM2.5 concentrations problem is still exist in heavily polluted Beijing-Tianjin-Hebei (JingJinJi) region due to AOD cannot adequately represent the correlation between high PM2.5 concentrations and independent variables and neglected the effects of missing AOD. Thus, the long- and short-term PM2.5 exposure risk estimate was reduced. This work introduces gaseous pollutant data (NO2, SO2, CO, and O-3) related to primary emission and secondary transformation of pollutants as predictors into a deep neural network model to improve the underestimation of high PM2.5 concentrations based on AOD and meteorological factors. We predicted the PM2.5 concentration in the missing AOD areas, generated a daily continuous PM2.5 spatial distribution, and reduced estimated bias due to AOD deficiency. Grid-based 10-fold cross-validation (CV) was used to test the model performance. Results show that daily PM2.5 concentration CV R-2 is 0.87 and the root-mean-square prediction error (RMSE) is 27.11 mu g/m(3). The CV R-2 and RMSE are higher by 0.12 and lower by 9.72 mu g/m(3) than the model without gaseous pollutants (GAS(s)) as predictors. In including the missing AOD, the average concentration of PM2.5 CV R-2 is 0.86 and the RMSE is 16.95 mu g/m(3) in heavy polluted winter; the CV R-2 and RMSE are higher by 0.07 and lower by 3.95 mu g/m(3), respectively, than when the missing AOD was excluded. Prediction results of PM2.5 spatial distribution show that the model has high prediction accuracy and provides a complete and highly accurate spatiotemporal distribution characteristics for long- and short-term PM2.5 exposure studies, and reduces exposure misclassification of PM2.5 in heavily polluted areas.

DOI:
10.1016/j.atmosenv.2019.05.004

ISSN:
1352-2310