Spatial-Temporal Dynamic Fusion Graph Convolutional Network for Traffic Forecasting

Traffic forecasting remains a topic of active research in the domain of Intelligent Transportation System (ITS) and continues to garner significant attention. Given the intricate spatial relationships and temporal dependencies involved, it is a challenging task to forecast future traffic situations accurately. Most of the current methods developed complex neural networks by utilizing predefined graphs to capture spatial correlations and achieved superior performance. However, node attributes which capture significant data regarding traffic conditions have not been fully utilized to demonstrate spatial relationships. Furthermore, spatial correlations of the road networks are in perpetual change. To tackle the abovementioned matters, we propose a new spatial-temporal dynamic fusion graph convolutional network (STDFGCN). STDFGCN combines spatial-temporal convolutional networks with a dynamic fusion graph block that dynamically generate the optimal graph Laplacian matrices from two aspects: short-term dynamic fluctuation graph and global long-term static stable graph. Different from traditional graph, our model introduces hypergraph to capture high-order spatial correlations of edges. Then long-term associations in temporal and spatial features are captured by employing spatial-temporal convolutional blocks, comprising graph-based convolutional neural network and gated temporal convolutional network. Experiments validate the effectiveness of our STDFGCN model in capturing extensive spatial-temporal correlations comprehensively through modeling traffic networks at multiple scales. It achieves advantages over several baseline models on various datasets collected in real-world road traffic scenarios.