Abstract
Accurate and timely traffic flow prediction plays a crucial role in improving road utilization, reducing congestion, and optimizing public transportation management. However, modern urban traffic faces challenges such as complex road network structures and the variation in traffic flow across different temporal and spatial scales. These issues lead to complex spatiotemporal correlations and heterogeneity, resulting in low prediction accuracy and poor real-time performance of existing models. In this work, we propose a novel traffic flow prediction model called TMDCN (Temporal Attention and Multi-Graph Adjacency Fusion Using DynamicChebNet), which integrates temporal attention and multi-graph adjacency matrix fusion. First, to address the difficulty of capturing dependencies across multiple time scales, we construct a Temporal Feature Extraction Block that combines attention mechanisms with multi-scale convolutional layers, enhancing the model's ability to handle complex traffic pattern changes and capture flow variations and temporal dependencies. Next, we leverage multi-graph adjacency matrix fusion and dynamic Chebyshev graph convolutional networks to capture the spatial dependencies of the traffic network. Experiments on the PeMS04 and PeMS08 datasets show that, compared to conventional methods, the proposed method reduces the Mean Absolute Error (MAE) of traffic flow prediction one hour ahead to 18.33 and 13.72, respectively. The source code for this paper is available at https://github.com/tyut-zjb/TMDCN .