STGMN: A gated multi-graph convolutional network framework for traffic flow prediction

Accurate traffic flow prediction is crucial for the development of intelligent transportation. It can not only effectively avoid traffic congestion and other traffic problems, but also provide a data basis for other complex tasks. The rapid development of social technology and the increasingly complex traffic environment lead to the emergence of massive traffic data. Traffic flow prediction as a spatial-temporal prediction problem has been widely concerned, but the traditional forecasting methods often ignore the spatial-temporal dependence, difficult to meet the prediction requirements. Therefore, this paper proposes a novel spatial-temporal model based on an attention one-dimension convolutional neural network (1D-CNN) and a gated interpretable framework, which models historical traffic data from the perspectives of time and space respectively. The core of the model proposed in this paper is to construct spatial-temporal blocks. First, a 1D-CNN based on channel attention mechanism and "inception" structure is proposed to extract temporal correlation. Then, considering the complexity of the actual traffic network, an interpretable multi-graph gated graph convolution framework is proposed to extract the spatial correlation. Finally, extensive experiments are carried out on real data sets, which prove the effectiveness of the proposed model, and it is very competitive compared with some state-of-the-art methods.