Spatiotemporal Dynamic Graph Convolutional Networks for Air Route Network Congestion Prediction

Abstract

High-precision air traffic congestion prediction is the basis for implementing refined traffic flow management. However, current prediction approaches are limited to fixed route topologies, neglecting the dynamic spatial dependencies caused by congestion propagation between routes. Additionally, these models do not adequately consider the varying importance of different routes and time points within the time series. This paper proposes a spatiotemporal dynamic graph convolutional network to simultaneously capture the static and dynamic spatial dependencies among air traffic routes. We construct two types of adjacency matrices based on the connections between routes: a static adjacency matrix and a dynamic adjacency matrix. The static adjacency matrix reflects the basic connections between routes, while the dynamic adjacency matrix captures the congestion influence from upstream to downstream routes at each time interval. In the model, these adjacency matrices are first processed through graph convolutional network layers, where a spatial attention mechanism assigns weights to different routes. The processed data are then fed into gated recurrent unit layers, utilizing a temporal attention mechanism to assign varying weights to congestion data at different time. Experiments were conducted on the air route network in Guangdong–Hong Kong–Macao Greater Bay Area of China. The experimental results demonstrate that our proposed method outperforms several existing approaches, particularly those using only static graph convolution methods.