DCRNN predictive rerouting cuts average travel-time index by about 1.8% versus no-reroute in SUMO while accepting only 41 route changes

The framework integrates directed-graph traffic forecasting and charging-demand prediction with CAV rerouting in microscopic simulation, scoring feasible alternatives by predicted congestion, CO2, route length, and charging-demand exposure. Prior work studied graph-based traffic forecasting, multi-objective eco-routing, and EV charging scheduling separately; this work feeds predicted edge-level conditions directly into per-vehicle reroute acceptance. Deterministic single runs in SUMO on the NYC-6 network (638 edges, 254 junctions) at five CAV penetration levels (5%, 15%, 25%, 35%, 45%) with matched demand and disruption schedules; DCRNN uses diffusion order 2, hidden dimension 64, and an 80/20 chronological split.

A route change is accepted only when the alternative preserves connectivity to the original destination and passes cooldown, route-change-limit, duplicate-route, and feasibility checks, making intervention selective rather than high-volume. Compared with K-shortest-path and V2X proactive routing, DCRNN accepts only 41 route changes across all penetration levels versus 146 and 153, respectively. Table II aggregates five penetration scenarios: DCRNN shows average TTI 3.520, average speed 27.618 km/h, and total CO2 770.928 kg, best among active methods; predictive-density routing changes only 5 routes but has the worst average TTI (3.591) and CO2 (781.528 kg).

Relative to the no-reroute reference, DCRNN reduces average TTI by about 1.8%, while the no-reroute reference retains lower aggregate CO2 emissions and total distance traveled. This separation shows a tradeoff between reducing congestion severity and adding travel distance and emissions, rather than dominance on every metric. The 1.8% is computed from scenario-averaged TTIs; the no-reroute reference has the lowest distance, V2X has the lowest distance among active methods, and DCRNN is close to V2X with far fewer route changes.

Internal records of accepted reroutes show consistency with the routing-score objectives: cumulative reductions of 403.675 in the TTI objective, 113275.495 in the CO2 objective, 10.219 in EV charging-risk exposure, and 1387.290 m in route length. These records explain why the policy accepted specific route changes rather than serving as independent counterfactual measurements of network state. Table III lists 41 reroutes by penetration; the 15%–45% cases each accept 8 reroutes with substantially larger objective reductions, while the 5% case accepts 9 because two vehicles receive repeated accepted updates but total objective reduction is smaller.

The result applies to urban road networks under controlled disruptions in mixed traffic, in settings with edge-level traffic and charging-demand observations and the ability to apply routing decisions to a subset of CAVs. It enables researchers to examine in microscopic simulation how forecast information can support a small number of high-value reroutes rather than converting control authority directly into reroute volume. The authors state that future work will extend to larger networks and repeated demand and disruption realizations, and will incorporate explicit charging queues and station availability to evaluate effects on charging service.

Results are single deterministic runs per case and method with no random-seed confidence intervals, so differences across penetration levels should be read cautiously. EV battery outputs are retained as diagnostic aggregates because some SUMO battery fields may be cumulative over time, so energy-efficiency conclusions do not rest on them. The score reductions in Table III are decision-time route-score evidence, not independent counterfactual measurements of network state. Prediction quality and routing-policy quality are not yet separated by replaying matched candidate route sets under alternative forecast sources, and charging queues and station availability are not explicitly modeled.