Traffic Flow Prediction Method Based on Deep Learning and Graph Neural Network

The continuous development of society is accelerating the process of urbanization and promoting the rapid construction of smart cities. As an important part of smart cities, many cities have begun to establish intelligent transportation systems (ITS). As the cornerstone of ITS, traffic flow prediction technology is crucial in trip prediction and urban traffic management. How to more effectively capture the spatiotemporal information present in traffic data is the fundamental problem of traffic flow prediction. However, existing methods have problems such as insufficient mining of spatiotemporal features and incomplete modeling of spatiotemporal relationships when performing traffic flow prediction tasks. Aiming at the problem that the existing study approaches lack dynamic modeling of traffic data, and RNN-based approaches have limitations in the capture of global information, this study proposes a self-attention-based spatial-temporal double graph convolutional networks for traffic flow forecasting (SASTDGCN). This method abstracts the influence between nodes in the time domain into a graph structure with relationships, and then uses relational graph convolutional neural networks (RGCN) for extraction of temporal features, so that road network nodes can directly obtain the traffic conditions of nodes at historical moments that are far apart, effectively addressing long-term forecasting. In addition, the model utilizes a module that is based on a dual-domain temporal and spatial self-concern mechanism for capturing the association between spatial and historical time steps of road nodes and generates a corresponding dynamic graph structure for the road network at each time step and the historical time steps of each node, successfully improving the model prediction accuracy by 13%.