A Four-Phase Spatio-Temporal and Quantum-Enhanced Framework for Traffic Speed Forecasting With Calibrated Uncertainty

Short-term forecasting of traffic speeds is a foundational requirement for adaptive signal control, congestion mitigation, and emerging real-time digital twin infrastructures. Despite substantial progress, many deep learning approaches continue to experience difficulties in capturing long-range temporal behavior, handling non-stationary traffic conditions, and producing reliable measures of predictive uncertainty. To address these challenges, this work develops a four-stage hybrid spatio-temporal framework consisting of: (i) a graph-attention spatial encoder, (ii) a temporal Transformer module, (iii) a variational quantum fusion layer, and (iv) a conformal calibration component for uncertainty adjustment. In all our experiments we used the METR-LA benchmark which we did in order to maintain the like conditions for method and empiric study. In the put forth pipeline the Transformer which we introduced in Phase 2 does the best job in terms of point forecast accuracy, we see Mean Absolute Error (MAE) values of 3.46, 3.68, and 3.76 for the 5, 10 and 15 minute prediction horizons which is a 50% or greater improvement over the Spatial-Temporal Graph Attention Network (ST-GAT) base. In Phase 3 we looked at the quantumaugmented version which did see a small improvement at the short term horizon but as the time horizon increased we saw a slight drop in performance which we note is a common issue in present variational quantum circuits. Conformal calibration for uncertainty estimation produces prediction intervals with empirical coverage of 92.8%, significantly outperforming ensemble-based and Monte Carlo Dropout (MC-Dropout) methods. On an RTX-class Graphical Processing Unit (GPU), end-to-end inference takes about 150-250 ms, which is compatible with real-time operational requirements in intelligent transportation systems. A configuration with four attention heads, a one-hour historical input window, and a thirty-minute model update interval provides a workable trade-off between accuracy and computational cost, according to sensitivity analyses based on MAE and Root Mean Square Error (RMSE) combined with latency monitoring. Overall, the findings show that the suggested framework facilitates accurate traffic speed forecasting with calibrated uncertainty, allowing for more timely and informed decision-making in urban mobility applications.