Multiple dynamic graph based traffic speed prediction method

Traffic speed prediction is a crucial and challenging task for intelligent transportation systems. The pre-diction task can be accomplished via graph neural networks with structured data, but accurate traffic speed prediction is challenging due to the complexity of traffic systems and the constantly dynamic changing nature. To address these issues, a novel evolution temporal graph convolutional network (ETGCN) model is proposed in this paper. The ETGCN model first fuses multiple graph structures, and uti-lizes graph convolutional network (GCN) to model spatial correlation. Then, the spatial-temporal depen-dence and their dynamical changes are learned simultaneously to predict traffic speed on a road network graph. Especially, a similarity-based attention method is proposed to fuse multiple graph adjacency matrices. Then, the gated recurrent unit is combined with GCN to capture spatial-temporal correlations and their changing status, simultaneously. Extensive experiments on two large-scale datasets show that our methods provide more accurate prediction results than the existing state-of-the-art methods in every prediction horizon. (c) 2021 Elsevier B.V. All rights reserved.

Automated summary

This paper proposes a novel evolution temporal graph convolutional network (ETGCN) model, which fuses multiple graph structures and utilizes graph convolutional network (GCN) to model spatial correlation, while simultaneously learning spatial-temporal dependence and their dynamical changes for traffic speed prediction.