Graph Attention Convolutional Network: Spatiotemporal Modeling for Urban Traffic Prediction

Abstract

Multi-nodes traffic flow prediction in road network has been a topical issue in ITS (Intelligent Transportation Systems) and V2X (Vehicle to Everything). As an extension of single node prediction, `multi-nodes' leads to spatial correlations among nodes which complicated spatiotemporal modeling. Many studies introduced graph neural network, e.g., graph convolution network (GCN) and graph attention network (GAT) to model spatial features. But it is deficient to model traffic exchanging among neighbor nodes only by road network or by value attribute of single node. In this study, GAC-Net (graph attention convolutional network) was proposed to tackle this problem. GAC-Net introduces learnable parameters to combine GCN and Directed GAT for spatial feature extraction. For temporal correlation on sequential series, temporal convolutional network is introduced twice with different structures. Novel temporal graph convolutional network and spatiotemporal self-attention network are proposed to capture traffic status features and spatiotemporal feature from input sequence. Experiments were conducted based on two public traffic speed benchmark datasets METRLA and PEMS-BAY. Results show that GAC-Net achieves improvement of performance over traditional methods and state-of-the-art baselines.