Spatio-temporal prediction of air pollutants in Beijing based on GNN-CNN

Abstract To accurately predict the variation trends of atmospheric pollutant concentrations, this study analyzed the spatiotemporal evolution characteristics of major air pollutants in Beijing from 2021 to 2025 using hourly monitoring data from 35 ambient air quality automatic monitoring stations between January 1, 2021, and August 31, 2025. The analysis incorporated spatial visualization techniques and temporal conversion processing. Building on existing prior knowledge, a hybrid Graph Neural Network-Convolutional Neural Network (GNN-CNN) model was developed to estimate time series concentrations of major air pollutants in Beijing. Additionally, the CNN model and GNN-CNN model were employed to predict the spatiotemporal distribution of major air pollutant concentrations in the study area for 2025. The results indicate the following: (1) Beijing’s air quality showed gradual improvement from 2021 to 2024. The monthly variations of major air pollutants PM 2.5 , PM 10 , NO 2 , and SO 2 exhibited a ‘U-shaped’ trend, with significantly lower concentrations in summer and autumn compared to winter and spring. O 3 pollutant concentrations were notably higher from late spring to early autumn than during other periods, while CO pollutant concentrations showed a relatively stable yet fluctuating monthly distribution. (2) Monitoring data indicates that the distinct pollution areas in Beijing are located in the urban districts. The spatial distribution of major pollutants PM 2.5 , PM 10 , NO 2 , and CO displayed a ‘higher in the south, lower in the north’ pattern across Beijing. (3) The constructed GNN-CNN model demonstrated superior prediction accuracy for major air pollutants compared to the GNN model, with a significant improvement in the R 2 metric. For the six pollutants, the correlation coefficient between predicted and observed values exceeded 0.9 at more than 90% of the sites, except for PM 10 .