Based Matrix Fusion Spatial-Temporal Graph Neural Network for Traffic Flow Prediction

Nowadays, the traffic flow of motor vehicles in cities is increasing. Therefore, timely and accurate traffic flow prediction is very important for urban traffic control and guidance. The existing traffic flow prediction model can effectively extract the temporal and spatial characteristics of traffic flow, but it ignores the correlation between important traffic nodes and the indirect impact of external factors on traffic flow. Therefore, this paper proposes a traffic flow prediction method based on matrix fusion spatial-temporal graph neural network (MFSTGNN). Specifically, the graph neural network includes not only the adjacency matrix, but also the correlation matrix of nodes and the travel intention index generated according to external factors. In addition, these elements are fully combined through matrix fusion and self-learning weights to improve the modeling ability of spatial-temporal graph neural networks. Then, the temporal feature extraction module and spatial feature extraction module are designed. In the temporal feature extraction module, we use attention mechanism and gate recurrent unit to better capture the impact of long-term and short-term factors on traffic flow. For the spatial feature extraction module, we add the fusion matrix proposed in this paper to the graph convolution network, and combine the residual module. Finally, experiments on real data sets show that our model (MFSTGNN) is more effective than several existing models for traffic flow prediction.