Adaptive graph-generating jump network for air quality prediction based on improved graph convolutional network

Abstract

Long-term exposure to PM_2.5 is harmful to human health, and it is important and necessary for accurate PM_2.5 forecasts. However, complex spatial correlations make air quality prediction challenging, and some studies are limited still by the priori knowledge and may lead to incomplete information transmission between different sites. To address this issue, this study proposes a Dynamic Adaptive Graph Generating Jump Network (DAGJN) to predict PM_2.5. Specifically, in terms of spatial modelling, this study is the first to treat the graph structure as a learnable part, which can continuously optimize the weights with the training to better captures the potential spatial correlations among sites. A jump graph convolutional network that uses channel attention to weight features for selection of graph signals at different depths to utilize spatial information and mitigate the over-smoothing problem. A multiple self-attention mechanism is used to capture the global temporal correlation in time series data. Lastly, a spatial-temporal fusion layer can dynamically fuse spatial-temporal information based on global and local features. Meanwhile, extensive experiments were conducted on air quality datasets from Beijing and Chongqing with R² of 0.514 (0.770), and RMSE of 62.284 μg/m³ (12.814 μg/m³) in 1–24 h prediction. The differences of PM_2.5 prediction is compared with seasonal scales and shows that the DAGJN model outperforms other models. This study contributes significantly to the field of PM_2.5 prediction, and these results illustrate the potential of the DAGJN model for PM_2.5 prediction.