Existing traffic flow prediction methods face two critical challenges in dynamic traffic systems: Homogeneous graph-based models (e.g., STGCN) fail to capture time-varying node interactions caused by traffic incidents or peak-hour congestion; State-of-the-art heterogeneous models like H \(^2\) STGCN only consider static spatial proximity, ignoring the non-stationary propagation characteristics of traffic flows. To address these limitations, we propose QTA-H \(^2\) STGCN, a novel framework integrating quantum attention mechanism and Transformer architecture.The quantum attention layer models dynamic node relationships using quantum bit superposition, enabling probabilistic encoding of spatiotemporal correlations with 40% fewer parameters than traditional attention. The Transformer module captures long-range dependencies (e.g., inter-regional peak-hour interactions) through multi-head self-attention, while parallel processing accelerates training by 2.3 \(\times\) on large-scale datasets. These components are synergistically integrated with heterogeneous graph convolution to form a unified prediction framework.Extensive experiments on PEMS03/04/07/08 datasets demonstrate that QTA-H \(^2\) STGCN outperforms SOTA methods: MAE (0.48%-1.02%), RMSE (0.71%-2.40%), and MAPE (0.15%-0.82%). Particularly, compared to H \(^2\) STGCN, our model achieves 0.82% MAE reduction and 1.53% RMSE reduction, validated by statistical significance tests ( \(p<0.01\) ). This work provides a new paradigm for dynamic traffic system modeling with quantum-enhanced learning.
