Enhancing PM2.5 Forecasting via the Integration of Lidar and Radiosonde Vertical Structures

Accurate forecasting of near-surface PM2.5 concentrations remains challenging due to the complex coupling between atmospheric vertical structure, thermodynamic stability, and pollutant accumulation processes. Most existing surface-based statistical and deep learning approaches struggle to represent the three-dimensional state of the atmosphere, which limits their robustness under complex meteorological conditions. In this study, we propose a multi-source spatiotemporal learning framework(MST-Net) to enhance PM2.5 forecasting accuracy by integrating vertically resolved atmospheric information from lidar and radiosonde observations. The proposed approach incorporates vertical profile features together with surface measurements to provide complementary information on atmospheric vertical structure and its temporal evolution. Experimental results demonstrate that MST-Net consistently outperforms conventional time-series models across multiple forecast horizons. Notably, at extended lead times (12–24 h), the proposed framework exhibits enhanced stability and slower error growth. For 24 h forecasts, MST-Net reduces RMSE by approximately 13% and MAE by about 19%. These results indicate that leveraging multi-source vertical atmospheric information can effectively improve the reliability of urban air quality forecasting.