A Graph Neural Networks approach predicted spatiotemporal changes of Ozone Concentrations in the Yangtze River Delta (China)

Abstract With global warming and increasing emissions of air pollutants, urban ozone (O3) pollution events frequently occurred, which pose severe threats to human health. The formation of O3 is influenced by complex interactions among air pollutant emissions and meteorological conditions, leading to generally low accuracy in current O3 concentration forecasts. By using observational data from environmental quality monitoring stations, meteorological reanalysis datasets, and emission inventories in typical cities across the Yangtze River Delta (YRD) region from 2016-2023, we developed a spatiotemporal graph convolutional model for air forecast (STCGNFA) that integrates Graph Convolutional Networks and Gated Recurrent Units. This model was employed to simulate the variability in O₃ concentrations and to quantify the contributions of different variables. The results demonstrate that the STGCNFA model outperforms conventional statistical models and machine learning approaches, substantially reducing Mean Absolute Error (MAE) by 29.4% and 17.8% respectively, and Root Mean Squared Error (RMSE) by 31.9% and 17.1%, respectively. From 2016 to 2023, STGCNFA model simulated annual mean O3 concentrations exhibited a significant increasing trend (+0.32 ppvb·year⁻¹, p < 0.05), which is in line with the observed decline in O₃ concentrations. Spatially, high-concentration areas expanded toward the northern Hangzhou Bay region, while coastal areas exhibited a weakening trend. This pattern aligns well with the observed characteristics of O₃ concentrations. Shapely additive explanations (SHAP) analysis further reveals that 2 meters air temperature are the primary contributors to O3 variability, with stronger correlations with O3 concentration than emission factors.