An accurate and efficient forecast framework for fine PM2.5 maps using spatiotemporal recurrent neural networks

Abstract

Commonly used numerical prediction models for PM_2.5 maps suffer from low accuracy and high computation cost, which cannot meet the requirements for fine-scale air pollution control. In this study, we propose a framework based on the spatiotemporal recurrent neural network (PredRNN) to efficiently generate accurate 3-h and 6-km PM_2.5 maps with a lead time of 5 days. In this framework, two PredRNN networks are initially utilized to forecast PM_2.5 concentration at ground monitoring sites and the spatial distribution of aerosol optical depth (AOD) by assimilating the output of numerical prediction model. Subsequently, the 3-h and 6-km PM_2.5 forecasted maps with a lead time of 5 days can be inferred by establishing the regression links between the forecasted results of PM_2.5 concentration at ground sites and AOD maps. We evaluate the proposed framework in the Beijing-Tianjin-Hebei urban agglomeration region during 2017–2020. Compared with the numerical prediction products of the Copernicus Atmosphere Monitoring Service, the proposed framework achieves higher accuracy, with R² of 0.83 at the forecast base time and 0.70 at the fifth day. The spatial information richness is also enhanced by approximately 15.67% according to the information entropy metrics. Notably, the proposed framework only requires 1 min for forecasting 5-days PM_2.5 maps. These results demonstrate that our framework can efficiently generate accurate and fine PM_2.5 maps with a lead time of 5 days.