MASTGN: Multi-Attention Spatio-Temporal Graph Networks for Air Pollution Prediction

In recent years, the importance of air pollution issues has been increasingly discussed by the public and the government. Air pollution prediction has become a crucial reference to rely on when the government needs to formulate environmental policies. Considering the sparsity of atmospheric monitoring stations in the spatial distribution, in this study, we consider the atmospheric data as a spatio-temporal structure graph for better utilization of spatio-temporal information. Accordingly, we propose using multi-attention spatio-temporal graph networks (MASTGN) to exploit the graph structure atmospheric data for air pollution prediction tasks. The MASTGN model allows better mining the high-level spatial, temporal, and physical features corresponding to the atmospheric data through the multi-attention mechanism of the spatial, temporal, and channel attention. We conduct experiments on two datasets gathered in Japan and China to predict the concentration of PM
_2.5
, O
₃
, and PM
₁₀
. The results indicate that the proposed MASTGN model outperforms the considered baseline approaches on prediction accuracy.