Spatial-Temporal Traffic Forecasting Based on Optimized Dynamic Graph Convolutional Network

Traffic forecasting stands as a pivotal element within intelligent transportation systems. Achieving precise traffic forecasting presents a formidable challenge, primarily due to the latent correlations inherent in traffic network and the non-Euclidean characteristics of traffic data. Despite the proposition of numerous spatio-temporal modeling methodologies, they continue to demonstrate constraints in extracting intricate spatio-temporal features from traffic data and capturing their underlying spatio-temporal correlations. Furthermore, the extraction of spatial features is significantly contingent upon the configuration of graph representation structures. A preponderance of methods relies on either predefined or adaptively generated graphs, yet none of them can accurately mirror the authentic relationships among nodes. To tackle these intricacies, this paper introduces the Optimized Dynamic Graph Convolutional Network (ODGCN) for spatio-temporal traffic forecasting. The ODGCN primarily comprises a dynamic graph construction methods and optimized graph convolutional layers. The innovative model uses a novel dynamic graph structure generation method to construct the sensor adjacency matrix as a basis for extracting dynamic spatial features in the transportation network. The optimized graph convolutional layers proficiently capture intricate dependencies among nodes while preserving graph structural information through the utilization of stacked 1D convolutional components. An in-depth analysis conducted on two traffic datasets collected in the real world underscores the superiority of ODGCN in comparison to the most advanced baseline methodologies.