Daily PM2.5 Forecasting Using Graph Convolutional Networks Based on Human Migration

Most existing deep learning methods for air pollution concentration forecasting mainly focus on temporal characteristics of air pollutants. However, the spatial characteristics of air pollution concentration are closely related between nearby cities. In this study, we construct Target-city Graphs (TCG) to reveal the features of air pollutants between cities by using intercity migration networks. Then, we develop Graph Convolutional Neural Networks (GCN) and Graph Attention Networks (GAT) with Sum Aggregation (Sum-agg) and Mean Aggregation (Mean-agg) functions to forecast daily PM2.5 concentration time series based on TCG graph representing. The experimental results indicate that the GAT with Sum aggregation performs the best in forecasting PM2.5 while considering the intercity migration data.