Traffic Flow Prediction with Strong Dynamic Spatiotemporal Dependence Using a Hybrid Model of Transformer and TCN

The accurate prediction of traffic flow is a fundamental component of intelligent transportation systems and smart city planning. Conventional methodologies frequently encounter difficulties in capturing the intricate and evolving spatial-temporal interdependencies intrinsic to traffic data. Recent advances have employed Graph Neural Networks (GNNs) and attention mechanisms to address these challenges. However, existing models typically address spatial and temporal dependencies in isolation and may not fully leverage multi-modal interactions within the data. This paper proposes a novel framework, the Multi Modal Traffic Flow Encoder (MMTFE), which integrates temporal attention, spatial attention, and Temporal Convolutional Networks (TCN) for the joint modeling of the complex spatial-temporal patterns observed in traffic flows. By combining these components in a unified architecture, our model effectively captures dynamic dependencies and improves prediction accuracy. The superiority of the proposed approach is substantiated by comprehensive experimental investigations on actual traffic data sets, which reveal that it outperforms existing cutting-edge techniques.