PreSTAEFormer: A Hybrid GNN-Transformer Model with Pre-Attention Mechanism for Enhanced Urban Traffic Prediction

This study presents a hybrid GNN and Informer model for urban traffic prediction. By combining GNN’s spatial analysis with the Informer’s temporal processing, the model captures complex spatio-temporal dynamics in traffic systems. Key innovations include a pre-attention mechanism and a Probabilistic Attention Layer, enhancing the model’s efficiency and accuracy. Evaluated on five diverse traffic datasets—METR-LA, PEMS-BAY, PEMS03, PEMS04, PEMS07 and PEMS08—the model demonstrated robust and consistent performance across different traffic conditions and regions. The results highlight its adaptability to real-world traffic systems, capturing both short-term and long-term dependencies. These findings underscore the potential of advanced neural networks in developing intelligent urban traffic management systems, promoting sustainability and improved air quality.