This paper proposes a multi-scale patch-aware spatio-temporal dynamic graph network  for air quality prediction. The model operates through three core modules working in tandem: the PurePatchEmbedding module employs a non-overlapping patching strategy to extract local temporal features; the Decomposable Temporal Convolutional Network decomposes the sequence into trend, periodic, and residual components, enabling multi-scale temporal pattern learning; the Adaptive Spatio-Temporal Graph Network integrates static geographical constraints with dynamic attention mechanisms to accurately model spatio-temporal dependencies. This effectively addresses two issues: (1) traditional models struggle to effectively separate and model complex temporal patterns such as overlapping periodicity and trends in the data; (2) modeling approaches based on fixed spatial relationships cannot adapt to the dynamic characteristics of pollutant dispersion influenced by factors like weather and time. Experiments at three monitoring stations in Beijing showed that the model achieved average reductions of 40.13\%, 31.79\%, and 36.21\% in MSE, MAE, and MAPE metrics, respectively, significantly outperforming existing methods. Ablation experiments validated the effectiveness of each module, with DTCN and ASTG-Net contributing 16.9\% and 13.2\% performance improvements, respectively. The research findings provide new technical insights for air quality prediction and have important practical applications.