Constructing the 3D spatial distribution of PM2.5 concentrations during the 2022 Beijing Winter Olympics using LiDAR vertical observation networks and machine learning models

Air pollution in China has shown significant improvement due to the strict implementation of emission control measures; however, pollution episodes caused by regional transport still occur frequently. This study investigates the three-dimensional distribution and transport mechanisms of PM_2.5 in the Beijing-Tianjin-Hebei (BTH) region during the 2022 Beijing Winter Olympics, focusing on heavy pollution episodes before, during, and after the event. We integrated 34 LiDAR stations and surface monitoring data with six machine learning models (XGBoost, Random Forest, LightGBM, RNN, CNN-RNN, and CNN-BiLSTM) to reconstruct spatiotemporal PM_2.5 dynamics. The CNN-BiLSTM model, with an R² of 0.924, RMSE of 6.805 μg/m³, and MAE of 4.640 μg/m³, outperformed the others, benefiting from its dual capability to capture spatial and temporal dependencies. Based on the model results, distinct PM_2.5 distribution patterns were identified across the three Olympic phases. During the pre-event period, high concentrations (∼180 μg/m³) were concentrated in the upper atmosphere (1 km above ground) south of BTH. In contrast, Beijing and Tianjin experienced near-surface pollution peaks during the event, likely driven by short-range transport. In the post-event phase, PM_2.5 concentrations decreased overall, with pollutants transported southward under persistent northerly winds. Transport flux intensity (TFI) analysis highlighted key pollution pathways, with pre-event TFI peaking at 3.2 × 10⁵ μg·m^-1·s^-1 in Beijing, event-phase TFI reaching 3.8 × 10⁵ μg·m^-1·s^-1, and post-event TFI peaking at 7.8 × 10⁵ μg·m^-1·s^-1. These findings underscore deep learning's role in PM_2.5 inversion and reveal significant transregional pollution transport dynamics, offering insights for regional air quality management.