<p>Urban traffic congestion is exacerbated by rapidly increasing vehicle numbers, posing serious environmental and economic challenges. Accurate traffic flow prediction is essential for real-time management, yet it is hindered by the dynamic, non-Euclidean, and multimodal nature of traffic data. Existing deep learning methods often fail to simultaneously capture dynamic spatial correlations, long-term temporal dependencies, and properly align spatial and temporal feature distributions. To overcome these limitations, we propose the Optimal Transport based Spatial-Temporal graph attention model (OTST), which integrates three key components: 1) Graph Spatial Attention (GSA) to capture dynamic node correlations, 2) Graph Temporal Attention (GTA) to model both short- and long-term temporal dependencies, and 3) Optimal Transport to align spatial and temporal feature distributions effectively. Experiments on four PeMS datasets demonstrate that OTST achieves state-of-the-art performance, improving MAE by 9.4-13.2%, MAPE by 9.8-13.4%, and RMSE by 5.1-10.7% compared to existing methods, highlighting its potential for efficient traffic management.</p>