Integration of Global Attention Mechanism and Spatio-Temporal Characterization of Air Quality Prediction Methods

To address the limitations of conventional air quality prediction methods—such as the inability of static spatial topology to adapt to meteorologically driven pollution dispersion, unidirectional temporal modeling that overlooks bidirectional dependencies, and fixed-weight fusion mechanisms that fail to emphasize critical features—this paper proposes a meteorologically driven spatiotemporal graph neural network model named MDSTF-Net. The model first dynamically constructs topological relationships between monitoring stations based on real-time meteorological data, enabling accurate characterization of pollutant transport and diffusion patterns governed by meteorological conditions. Next, a Bidirectional Long Short-Term Memory (BiLSTM) network is employed to extract temporal features from the Air Quality Index (AQI), simultaneously capturing historical accumulation effects and future evolution trends of pollutant concentrations. Furthermore, a Global Attention Mechanism (GAM) is introduced to adaptively assign feature weights to key pollutants in the channel dimension and dynamically focus on critical polluted regions in the spatial dimension, thereby achieving deep fusion of multi-dimensional spatiotemporal features. Using air quality and meteorological monitoring data from Beijing and Shanghai between 2022 and 2023, systematic experiments were conducted under both normal conditions and extreme weather scenarios such as sandstorms and typhoons. The results demonstrate that the proposed model outperforms existing mainstream baseline methods in both prediction accuracy and robustness, validating its effectiveness and applicability for air quality forecasting in complex meteorological environments.