Improving PM2.5 and PM10 predictions in China from WRF_Chem through a deep learning method: multiscale depth-separable UNet

Abstract

Accurate predictions of atmospheric particulate matter can be applied in providing services for air pollution prevention and control. However, the forecasting accuracy of traditional air quality models is limited owing to model uncertainties. In this study, we developed a deep learning model, named multiscale depth-separable UNet (MDS-UNet), to improve PM_2.5 and PM₁₀ concentration forecasts from WRF_Chem over China. Results showed that MDS-UNet was able to capture the complex nonlinear errors between model predictions and observations, which was helpful in correcting the biases and spatiotemporal distribution patterns of PM_2.5 and PM₁₀ concentrations predicted by WRF_Chem. MDS-UNet made a better performance in the improvement of both PM_2.5 and PM₁₀ prediction accuracy than UNet and CNN during the 0-24 forecasts. Using MDS-UNet, the reductions in the root-mean-square error (RMSE) of the regionally averaged PM_2.5 and PM₁₀ concentration forecasts were 35.08% and 17.74%, respectively. During the 0–24-h forecast period, MDS-UNet performed well in terms of PM_2.5 and PM₁₀ over six key urban agglomerations in China. Taking a pollution process as a case study, results demonstrated that, compared with WRF_Chem, MDS-UNet was able to make the best improvement in YRD, the Sichuan Basin, and central China, with reductions in the RMSE of the PM_2.5 forecasts of 55.22%, 55.53%, and 52.17%, respectively; and for PM₁₀ forecasts these reductions were 44.90%, 40.97%, and 46.79%, respectively. Through this analysis, it was apparent that MDS-UNet demonstrated a better effect in terms of improving both PM_2.5 and PM₁₀ predictions in these key urban agglomerations during an important pollution process.