Traffic prediction is fundamental for urban traffic management and guidance. Nowadays, Graph Convolution networks have been extensively applied to this topic due to its ability to restore road network topology, process high-dimensional data and extract complex spatiotemporal dependencies among regions. Existing methods mainly generate graph according to Euclidean distance between adjacent intersections While we observe that the non-Euclidean relationship among possibly distant intersections are also critical for accurate forecasting. Moreover, previous research lacks the consideration of intersection control scheme. In this work, we propose a multi-graph convolutional neural network model to predict the traffic flow. Firstly, the urban road network system is transformed into multiple graphs, where the node is the intersection, and the edge is the relationship between the intersections. We construct three representative graphs: a) Distance Graph, using the reciprocal of the distance to mark the weight between two intersections; b) Path Graph, calculating the number of paths between two intersections as the weight because intersections on an OD path tend to affect each other; c)Similarity Graph, considering intersection factors such as signal control system, saturation degree and the average stop delay, we perform a cluster analysis on intersections so that similar stations will be linked with higher weights. Secondly, we fuse the multiple graphs and then apply the convolutional layers on the fused graph to predict Traffic flow. Using the loop-detector and GPS date, the superiority of the proposed method is evaluated by a real-world experiment in Beijing, which reducing 13.1% prediction error compared with normal Convolution Neural Network. Key words: traffic flow prediction, multi-graph convolution network, graph weight,dynamic OD path,cluster analysis.

CICTP