Dual-Channel Dynamic Gated Spatio-Temporal Graph for Traffic Flow Forecasting

Abstract

Traffic flow forecasting is a critical and essential technology in the field of Intelligent Transportation Systems (ITS), as it plays a pivotal role in optimizing traffic management, improving road safety, and enhancing the overall efficiency of transportation networks. However, current research neglects the relationships between the local and global traffic flow data. Additionally, the predefined static graph structure fails to adequately capture the dynamic spatial features of traffic flow. To address the these challenges, this paper proposes a Dual-Channel Dynamic Gated Spatio-Temporal graph network (DC-DGST) for traffic flow prediction. We consider hourly slices as the local feature and daily slices to be the global feature of traffic flow. The DC-DGST framework employs a dual-channel structure to capture spatiotemporal dependencies between global and local features. It transforms the predefined static graph into a dynamic graph, enabling the establishment of connections between input data and historical information. Furthermore, we design gated spatio-temporal blocks based on residual structures within the spatio-temporal module. Specifically, we utilize Graph Gated Neural Networks (GGNNs) to learn and integrate both static and dynamic graphs, while Transformer encoders are used to capture long-range dependencies in the temporal sequence. We conducted a series of experiments on four publicly available benchmark datasets: PEMS03, PEMS04, PEMS07, and PEMS08. The results demonstrate that our model significantly outperforms baseline models. Moreover, the dual-channel structure effectively captures the correlation between local and global traffic flow features, while the dynamic graph enhances the model's overall performance.