Abstract
Accurate prediction of PM2.5 concentrations is crucial for public health and environmental management. However, effectively capturing complex spatiotemporal dependencies across multiple time scales remains a persistent challenge for existing methods, particularly in regions with sparse monitoring stations. This study proposes a Multi-Scale Dynamic Graph Neural Network (MSDGNN) for PM2.5 forecasting in station clusters. The model incorporates multi-scale temporal modeling (hourly, daily, weekly) to capture both short- and long-term dependencies. A learnable mapping matrix dynamically groups stations to strengthen spatial correlation learning. Furthermore, MSDGNN employs multi-head attention and spatiotemporal graph attention mechanisms to construct dynamic graphs, utilizing adaptive adjacency matrices and Chebyshev graph convolutions for effective feature propagation. We evaluated MSDGNN on 22 air quality monitoring stations in the Chang-Zhu-Tan region. Results demonstrated that our model reduces MAE by 6.77% and RMSE by 8.67% compared to the best baseline, validating its capability to learn complex dependencies and deliver accurate predictions under diverse spatiotemporal conditions.