Cross-city PM2.5 predictions with recurrent neural network

Abstract

PM2.5 is inhalable particulate with a diameter less than 2.5 μM that easily enters the lungs and causes diseases and non-accidental death. The generation and dissipation of PM2.5 are strongly affected by a variety of environmental factors, thus the concentration of PM2.5 is presumably predictable with the observations of environmental conditions. This paper used multi-year meteorological and PM2.5 concentration data across multiple megacities in China (Beijing, Chengdu, and Shenyang) and sought for a universal predictive model. Our results showed that data-driven machine-learning model was able to not only capture PM2.5 dynamics at the city where the model was trained but also could be generalized to predict PM2.5 concentrations over other cities. Therefore, the modeling results indicated a universally existing predictive relationship between PM2.5 source-sink dynamics and the environmental drivers.