A deep learning traffic flow prediction framework based on multi-channel graph convolution

• Published in: Transportation planning and technology Vol. 44; no. 8; pp. 887 - 900
• Main Authors: Zhao, Yuanmeng, Cao, Jie, Zhang, Hong, Liu, Zongli
• Format: Journal Article
• Language: English
• Published: Abingdon Routledge 17.11.2021; Taylor & Francis Ltd
• Subjects: Convolution; Data analysis; Deep learning; graph convolution (GCN); Long Short-Term Memory (LSTM); spatial-temporal features; Traffic congestion; Traffic flow
• ISSN: 0308-1060; 1029-0354
• DOI: 10.1080/03081060.2021.1992180

Accurate and timely traffic flow prediction is a critical part of the steps to alleviate traffic congestion. Fully considering the spatial-temporal dependencies of traffic flow is the key to accurately predicting traffic flow. Addressing the problem that traditional methods are difficult to capture the complex spatial-temporal dependence of urban traffic flow, and therefore cannot meet the accuracy requirements for medium and long-term prediction tasks, this paper uses Graph Convolution (GCN) and Long Short-Term Memory (LSTM) methods to capture time and space dependence through data analysis, and proposes a new type of deep learning model MCGC-LSTM. GCN is utilized to learn spatial dependence by analyzing the topological structure of an urban road traffic network, while LSTM is utilized to learn temporal dependence by analyzing the dynamic changes of traffic flow. The experimental results based on a real data set show that this method can achieve better prediction accuracy.