Traffic Flow Prediction Based on Multi-Spatiotemporal Attention Gated Graph Convolution Network

• Published in: Journal of advanced transportation Vol. 2022; pp. 1 - 9
• Main Authors: Ge, Yun, Zhai, Jian F., Su, Pei C.
• Format: Journal Article
• Language: English
• Published: London Hindawi 09.09.2022; Hindawi Limited; Wiley
• Subjects: Analysis; Artificial neural networks; Computer applications; Convolution; Deep learning; Feature extraction; Machine learning; Neighborhoods; Neural networks; Predictions; Recurrent neural networks; Research methods; Social networks; Time series; Traffic accidents & safety; Traffic congestion; Traffic control; Traffic flow; Transportation; Unstructured data
• ISSN: 0197-6729; 2042-3195
• DOI: 10.1155/2022/2723101

Accurate prediction of traffic flow plays an important role in ensuring public traffic safety and solving traffic congestion. Because graph convolutional neural network (GCN) can perform effective feature calculation for unstructured data, doing research based on GCN model has become the main way for traffic flow prediction research. However, most of the existing research methods solving this problem are based on combining the graph convolutional neural network and recurrent neural network for traffic prediction. Such research routines have high computational cost and few attentions on impaction of different time and nodes. In order to improve the accuracy of traffic flow prediction, a gated attention graph convolution model based on multiple spatiotemporal channels was proposed in this paper. This model takes multiple time period data as input and extracts the features of each channel by superimposing multiple gated temporal and spatial attention modules. The final feature vector is obtained by means of weighted linear superposition. Experimental results on two sets of data show that the proposed method has good performance in precision and interpretability.