Dynamic graph convolutional networks based on spatiotemporal data embedding for traffic flow forecasting

Abstract

Traffic flow forecasting has always been a challenge owing to its complicated spatiotemporal dependencies. Few of previous works can exploit the implicit interactions among traffic flows, leading to the superficial extraction of spatiotemporal features. Though graph convolutional networks have showed exciting performance in traffic flow forecasting, existing works either ignore the dynamic characteristics of the correlations among sensors or fail to extract the hidden fine-grained correlations among sensors, which makes it difficult to model the spatial dependency deeply. Therefore, a new deep learning model is proposed in this study to overcome these drawbacks and to achieve accurate traffic flow forecasting. First, a new spatiotemporal data embedding method is proposed to convert the original traffic flows into traffic flow vectors, so that the implicit correlations among traffic flows can be quantified and measured. Then, to sufficiently extract the non-linear global temporal features, a new temporal vector convolutional neural network is proposed to deal with the traffic flow vectors. Finally, a new dynamic correlation graph construction method is proposed to exploit the dynamic characteristics of correlations among sensors and explore the hidden fine-grained correlations among sensors, which is conducive to learning deep non-Euclidean spatial features. Experiments on five traffic datasets demonstrate that the proposed model is superior to state-of-the-art baseline models.