关键词: 因子关联矩阵, 多因子融合时空图神经网络, 图卷积网络, 图注意力网络, 门控循环单元

Abstract: To solve the problem of insufficient mining of spatio-temporal characteristics of road network level traffic parameter prediction, a multi-factorfusion spatio-temporal graph neural network (MF-STGNN) is proposed. The model further explores the spatial characteristics of the road network from three perspectives: road network geographic topology, road traffic data correlation and traffic capacity, and then extracts the temporal features through the temporal module (Gated Recurrent Unit, GRU). The multi-factor fusion graph neural network (MF-GNN)constructs three factor correlation matrices, in which the capacity correlation matrix C and the traffic data correlation matrix D are input into the frequency domain based graph convolution network (GCN), and the spatial nearest neighbor matrix A is input into the graph attention network (GAT) to extract three different spatial features. Then MF-GNN fuse the three features through the channel attention model. Finally, temporal features are extracted by gated recurrent unit. Using the floating car data set of some regions in Yubei District of Chongqing for experiment. The results show that MF-STGNN can better capture the spatiotemporal characteristics of the road network and improve the prediction accuracy.