Traffic Flow Forecasting via Graph Transformers: A Deep Learning Approach

Accurate traffic flow prediction plays a key role in enhancing traffic management, reducing congestion, and improving road safety within intelligent transportation systems. This paper introduces a Graph Transformer-based model for traffic flow forecasting, which merges the spatial dependency learning capabilities of Graph Attention Networks (GAT) with the temporal pattern recognition strengths of Transformer architectures. This integrated approach allows the model to efficiently capture both the spatial relationships among traffic sensors and the temporal dynamics of traffic flow. The model is evaluated on the widely-used PeMS04 dataset, containing traffic flow data from 307 sensors across California, and the PeMS08 dataset. The training process begins with the Adam optimizer and is followed by fine-tuning with the SGD optimizer. Performance is assessed using key metrics, achieving a loss of 0.0865, MAE (Mean Absolute Error) of 0.0674, RMSE (Root Mean Squared Error) of 0.0928, and MAPE (Mean Absolute Percentage Error) of 0.1382. These results demonstrate a significant improvement over traditional methods, such as ARIMA and Support Vector Machines (SVM), which often struggle to fully capture the complex spatiotemporal dependencies in traffic data. By leveraging the Graph Transformer architecture, our model delivers enhanced accuracy and robustness, making it a promising approach for real-time traffic flow prediction in large-scale urban environments.