14-Day PM2.5 and O3 Forecasting over Eastern China Using a Hybrid CNN-Transformer Architecture Driven by Multilevel Meteorological Fields

Abstract Air pollution is modulated by multiscale meteorological processes spanning from local to regional domains and extending from the near-surface layer to the upper troposphere. However, existing data-driven models often neglect these complex spatiotemporal dependencies, limiting medium- to long-term forecast accuracy. Therefore, a hybrid convolutional neural network (CNN)-Transformer framework integrating multilevel regional meteorological fields and emission inventories is designed for 14-day-ahead forecasting of O3 and PM2.5 across central-eastern China. Combining information on large-scale meteorological evolution with the Transformer’s long-range attention improves predictive performance across diverse spatial scales and extended lead times, yielding relative R2 gains of 40% for O3 (0.58 to 0.81) and 18% for PM2.5 (0.44 to 0.52) over baseline models. Notably, the strategic incorporation of 850 hPa meteorological fields for O3 and emission inventories for PM2.5 yields additional accuracy gains of 5% and 9.8%. Overall, the model demonstrates superior predictive skill for O3 compared to PM2.5, particularly in heavily polluted regions, where it successfully captures 66–83% of O3 pollution episodes in North China. Furthermore, forecasts driven by GFS meteorological inputs exhibit higher O3 prediction skill, suggesting that reducing uncertainty in meteorological inputs is critical for improving operational O3 forecasting and enhancing early warning systems for regional pollution events.