TPDGCN: Transmissibility-Periodicity with Dynamic Graph Convolutional Network for Traffic Flow Forecasting

Forecasting traffic flow plays a vital role within Intelligent Transportation Systems and has drawn considerable focus. The data associated with traffic flow is multifaceted and characterized by multiple cycles, while the road network displays a nonlinear topological architecture. Identifying the spatio-temporal correlations between evolving traffic flow data and the network's topology is essential for accurate traffic flow predictions. Prevailing methodologies utilize diverse techniques to discern these spatio-temporal relationships in traffic data; nevertheless, many merely leverage periodicity information without recognizing that present periodic patterns emerge from prior ones-a phenomenon we refer to as 'the transitivity of periodicity'. To bridge this gap, we introduce a cutting-edge Dynamic Graph Convolutional Network grounded in Transmissibility-Periodicity, dubbed TPDGCN, for enhanced traffic flow forecasting. Within the TPDGCN framework, we have devised a Transmissibility-Periodicity Learning Module aimed at unraveling the propagation dynamics of multi-periodic data. Moreover, we have intelligently integrated an Attention-based Dynamic Spatio-Temporal Extraction Module to dynamically capture the fluid nature of spatio-temporal characteristics. Rigorous experimentation on authentic datasets has confirmed that our TPDGCN model surpasses existing benchmark models in terms of performance.