Abstract
Abstract Urban traffic congestion poses a growing challenge to efficient mobility, and existing forecasting and routing methods often struggle to support timely decisions in dynamic road networks. To address this issue, this study proposes a hybrid framework that combines a spatiotemporal graph attention network with the Quantum Approximate Optimization Algorithm (QAOA) for joint congestion prediction and path optimization. In the prediction stage, a multi-scale spatiotemporal encoder is developed to capture short-term, intra-day, and multi-day traffic variation patterns, while an adaptive graph learning mechanism is introduced to model hidden spatial correlations among road segments. In the optimization stage, the routing problem is formulated as a QUBO model and solved by an improved deep QAOA with hierarchical parameter sharing, which helps stabilize training and improve solution quality. The predicted congestion probabilities are further incorporated into the routing objective, enabling the optimization module to avoid highly congested areas. Experiments on the PeMS-BAY and METR-LA datasets show that the proposed method reduces the RMSE of 15-min traffic prediction by 5.4% compared with the strongest baseline. For 50-node routing tasks, it achieves an optimality ratio of 0.952, outperforming standard QAOA. In congestion-aware routing, the average travel time is further reduced by 13.7%. These results indicate that the proposed framework is effective for integrated traffic prediction and dynamic path planning.