Fine-grained prediction of PM2.5 concentration based on multisource data and deep learning

Abstract

With China's economic development, air pollution has become increasingly serious. Studying the concentration of PM2.5, as the main component of haze, is the key to control air quality. Extent research lacks spatiotemporal correlations and full regional coverage. Therefore, a fine-grained prediction method was developed for continuous prediction of station-specific and regional PM2.5 concentrations over 24 h that comprise two parts: a site prediction model (TSRT model) and a grid prediction model. The TSRT model includes a long short-term memory-based temporal predictor that uses temporally related features as inputs, a Long Short-Term Memory-based spatial predictor that requires spatially related features as inputs, and a regression model based on regression-tree algorithm to aggregate the predictions of temporal and spatial predictors. The grid concentration prediction model based on a fully connected back-propagation neural network with grid characteristics as inputs achieved a PM2.5 regional prediction with 1-km2 non-overlapping grids across the study area. The approach was evaluated with data from 101 monitoring stations in the Beijing-Tianjin-Hebei area. The results imply that: (1) the forecasting performance of TSRT models is better than other possible methods with average R2 = 0.9 and average MAE = 10.35 μg/m3, (2) R2 and MAE values of the grid prediction model are 0.77 and 21.87 μg/m3, supporting the advantages of our method over well-known interpolation models, such as kriging interpolation, and (3) the method provides appropriate PM2.5 concentration prediction in the next 24 h. Therefore, this novel method may improve PM2.5 prediction, benefiting people's daily lives and influencing policies.