Urban Traffic Speed Prediction with Multi-Graph Convolutional Networks: A Short-Term Prediction Approach

In order to improve the accuracy of short-term traffic speed prediction for urban road network, this paper defines a factor called intersection influence factor in the network based on intersection delay time. In the time dimension, a time convolution block composed of 1D convolution with GLU activation function is constructed to extract the temporal features of road network speed. In the spatial dimension, a spatial convolution block consisting of a graph convolution block that extracts the distance relationship between the location of the road segment and a graph convolution block that extracts the intersection influence factor is constructed to extract the spatial characteristics of road network speed. The Distance-Intersection Graph Convolutional Network (DI-GCN) model, a multi-graph convolution network model considering distance and intersection influence, is composed of a time convolution block and a spatial convolution block. Depending on the arrangement of the spatial convolution block, the DI-GCN model can be divided into DI-GCN (parallel) and DI-GCN (stacked) models. Finally, it is proved that the DI-GCN (stacked) model is more suitable for short-term speed prediction of urban road network with speed data from a small area. When predicting the speed for the next 15 min, compared with the STGCN model, the DI-GCN (stacked) model reduces the MAE, RMSE, and MAPE errors by 7.688%, 8.599%, and 6.437%, respectively.