Memory-Augmented Spatio-Temporal Transformer for Robust Traffic Flow Forecasting

Accurate traffic flow prediction plays a critical role in intelligent transportation systems, supporting traffic management, congestion mitigation, and efficient utilization of road resources. Advances in neural network-based methods, particularly graph neural networks (GNNs) and attention-based models, have demonstrated strong capability in modeling spatio-temporal traffic dynamics. However, existing approaches still face notable challenges: GNN-based models often rely on static adjacency matrices, limiting their ability to capture dynamic and long-range spatial dependencies, while attention-based models usually involve complex architectures and heavy reliance on large-scale pre-training data. To address these limitations, this study proposes a novel traffic flow prediction model that integrates a learnable memory tensor into an attention-based framework. The introduced memory mechanism provides persistent global context for modeling long-term temporal dependencies in an end-to-end manner, enabling efficient and dynamic spatio-temporal representation learning with a lightweight architecture. Extensive experiments on multiple real-world traffic datasets demonstrate that the proposed model achieves superior prediction accuracy and robustness compared with existing baselines. The proposed approach offers a new perspective for memory-enhanced spatio-temporal modeling and provides valuable insights for traffic forecasting and related intelligent transportation applications.