STAMGCN: A Spatio-Temporal Attention-Based Multi-Graph Convolution Model for Fine-Grained Air Quality Analysis

Air pollution is one of the major challenges faced globally. Implementing spatially fine-grained predictions of urban pollutants is crucial for establishing effective environmental policies, allowing for the protection of public health. This paper proposes a novel spatio-temporal attention-based multi-graph convolutional network (STAMGCN) for predicting air pollution. The model first constructs spatial graphs, atmospheric pollution pattern graphs, and meteorological pattern graphs to model various relationships between non-Euclidean regions. Graph convolutional networks are then employed to learn and integrate information from adjacent nodes of the corresponding graphs. Afterwards, the convolved graphs are processed, and a fine-grained attention mechanism is introduced to further focus on the potential relationships between the target station and nearby stations. Finally, by leveraging the time-series processing capacities of gated recurrent units, the model's ability to perceive dynamic changes in air quality is increased through the incorporation of a temporal attention mechanism. Air quality data, meteorological data, geographic locations, and regional types from Beijing were collected, and experimental results from extensive experiments show that, for unmonitored target areas, the proposed model improved the prediction accuracy of the air quality index by 10.18% for the current time and by 15.56% for the next 6 hours.