A hybrid spatio-temporal deep learning framework for urban travel time prediction and intelligent routing: Integrating recurrent and graph neural networks

Accurate short-term travel-time forecasting is important for intelligent transportation systems, but reported gains can be misleading when temporal leakage, ambiguous graph definitions, normalized error units, weak baselines, or untested routing claims are not controlled. This study evaluates an integrated forecasting and Web-GIS framework through two explicitly separated evidence streams. Dataset A contains 6,592 empirical platform-derived fixed-route observations from Tehran and is evaluated using within-day eight-step sequences, a 10-min prediction horizon, chronological day-level splitting, training-only scaling, and physical-unit errors. XGBoost is the strongest Dataset-A benchmark (test R²=0.9197; RMSE=1.9229 min), while the validation-selected Tuned LSTM–RNN is the best neural temporal configuration (test R²=0.8998; RMSE=2.1484 min). Event-stratified analysis shows a clear failure regime during accidents (N=33; RMSE=7.2936 min), so robustness under rare disruptions is not claimed. Dataset B is a controlled simulation-augmented multi-route experiment with 10,000 synchronized records on a deterministic 10-node route-interaction graph. Across three seeds, RNN–GCN–Skip achieves the best graph-neural within-network result (R²=0.9633; RMSE=2.885±0.023 min), only modestly improving on persistence. Leakage-free expanding-window stacking reduces RMSE to 2.7443 min (R²=0.9668), although the graph-neural base learner receives negligible meta-weight. Leave-one-route-out evaluation exposes a spatial-generalization gap: persistence is most stable (macro RMSE=2.706±1.239 min), whereas graph-neural models deteriorate on unseen routes. Finally, an executed standard-Dijkstra benchmark on the simulated corridor graph shows that stacking reduces mean realized path travel time from 77.988 to 76.729 min, saving 1.259 min relative to shortest-distance routing and matching the oracle path in 88.6% of test decisions. Dataset-A targets remain navigation-platform estimates without independent GPS ground truth, and Dataset-B/routing results are controlled simulation evidence. The contribution is therefore a reproducible, leakage-controlled, deployment-oriented evaluation framework rather than a claim of universal graph-model superiority.