Air pollution poses significant global health challenges, underscoring the need for highly detailed spatial predictions to inform effective environmental policy and safeguard public health. This study introduces a novel Spatio-Temporal Attention-based Multi-Graph Convolutional Network (STAMGCN) designed for fine-grained air quality prediction. The framework constructs spatial, atmospheric pollution-pattern, and meteorological-pattern graphs to represent complex, non-Euclidean regional relationships. Through graph convolutional networks, the model aggregates contextual information from adjacent nodes, followed by a fine-grained attention mechanism that emphasizes interactions between the target and nearby monitoring stations. By leveraging gated recurrent units with temporal attention, STAMGCN effectively captures evolving air quality changes. Experiments conducted on the Beijing dataset demonstrate that the model improves prediction accuracy by 10.18% for immediate forecasts and 15.56% for 6-hour forecasts compared to baseline spatio-temporal models. These results highlight the model’s potential to support urban air-quality management and provide a robust scientific foundation for pollution-control strategies.