PaperlayerView exact revision on arXiv

Source-linked AI summary

Predictive Rerouting of Connected and Automated Vehicles Using Traffic and Charging Demand Forecasts

Tony Kinchen, Ting Bai, Andreas A. Malikopoulos

arXiv:2610.10532v1cs.NIcs.SYeess.SY

TL;DR

The paper addresses predictive rerouting for CAVs in mixed traffic, where routing decisions must account for traffic and EV charging demand. It combines DCRNN forecasts with a policy for evaluating feasible routes, and reports improved average TTI versus no rerouting and the strongest congestion–rerouting tradeoff among active methods.

  • Problem

    In mixed traffic, only some vehicles can follow a routing policy, and predicted conditions must be evaluated against realized network performance after rerouting.

    Source: A. Motivation · p. 1, B. Related Work · p. 2

  • Method

    The framework combines graph-based traffic forecasting and charging-demand prediction with a routing procedure that ranks feasible alternatives and checks route connectivity and cost before rerouting.

    Source: C. Contributions of This Paper · p. 2

  • Results

    DCRNN improves average TTI by approximately 1.8% relative to no-reroute and has the lowest average TTI, highest average speed, and lowest total CO2 emissions among active routing methods.

    Source: A. Aggregate Method Comparison · p. 5, E. DCRNN Accepted-Reroute Benefits · p. 7

  • Takeaways & Limitations

    Among the evaluated methods, DCRNN achieves the strongest congestion–rerouting tradeoff, while the no-reroute reference retains lower aggregate distance and emissions.

    Source: E. DCRNN Accepted-Reroute Benefits · p. 7

    Future evaluation should include larger networks and repeated demand and disruption realizations, explicit charging queues and station availability, and matched-route comparisons that separate forecast from routing-policy effects.

    Source: §IV · p. 7

Abstract

from arXiv · show

In this paper, we consider the problem of predictive rerouting of connected and automated vehicles (CAVs) in mixed traffic with electric vehicle charging demand. We provide a framework that combines a diffusion convolutional recurrent neural network (DCRNN) with a routing policy that accounts for congestion, CO2 emissions, route length, and charging demand. The DCRNN uses historical network observations to forecast traffic conditions and charging demand. These forecasts are then used to evaluate feasible alternative routes for eligible CAVs. A route change is accepted when the alternative preserves connectivity to the original destination and improves the prescribed route cost. We evaluate the proposed framework in SUMO under controlled traffic disruptions at five CAV penetration levels, ranging from 5% to 45%. We compare its performance with K-shortest-path routing, predictive-density routing, V2X proactive routing, and a reference scenario without rerouting. In the considered scenarios, the proposed framework reduces the average travel-time index by approximately 1.8% relative to the reference scenario and achieves the lowest average travel-time index, highest average speed, and lowest aggregate CO2 emissions among the active routing methods. Across all five penetration levels, it accepts 41 route changes, compared with 146 for K-shortest-path routing and 153 for V2X proactive routing. The reference scenario retains lower aggregate emissions and distance traveled, illustrating the tradeoff between congestion reduction and the additional travel associated with rerouting.

A. Motivation

In mixed traffic, routing policies can affect only eligible vehicles, and a route that helps one vehicle may add congestion elsewhere. The paper motivates combining forecasts of traffic and charging demand with route evaluation that considers network-wide effects.

A. Motivation

  • Forecasted traffic and charging demand can inform route selection before a vehicle reaches an affected area, while route evaluation must account for emissions, distance, and subsequent traffic conditions.Charging trips add traffic on nearby roads, and congestion affects travel time, energy use, and access to charging infrastructure.Source: A. Motivation · p. 1

B. Related Work

  • Prior work covers connected-vehicle routing, multi-objective routing, traffic forecasting, and EV charging interactions, motivating their integration for CAV rerouting –, –.DCRNN combines diffusion convolution over a directed graph with recurrent temporal modeling; forecasting accuracy alone does not establish routing effectiveness because rerouting changes traffic conditions.Source: B. Related Work · pp. 1–2
  • A central question is how predicted conditions can guide route changes that improve traffic performance when only a subset of vehicles is controllable.The study emphasizes evaluating both the information used to select routes and realized network performance after routing decisions are applied.Source: B. Related Work · p. 2
  • In microscopic simulation, a candidate route’s quality depends on its interactions with surrounding traffic, not only its length or free-flow travel time.The paper compares active routing benchmarks with a no-reroute reference to distinguish routing-logic performance from the tradeoff introduced by intervention.Source: B. Related Work · p. 2

C. Contributions of This Paper

The paper integrates predictive traffic and charging-demand information into individual CAV rerouting in SUMO. It evaluates feasible route alternatives using a multi-factor cost and compares the approach with active routing methods and a no-reroute reference.

C. Contributions of This Paper

  • The framework combines DCRNN forecasts with a SUMO routing policy that evaluates feasible alternatives using congestion, CO2 emissions, route length, and charging-demand exposure.Eligible CAVs receive route changes only when the acceptance conditions are satisfied; other vehicles retain their assigned routes.Source: C. Contributions of This Paper · p. 2
  • The routing procedure ranks feasible alternatives using predicted conditions and checks connectivity, route-cost improvement, and route-change conditions before accepting a reroute.The contributions describe evaluation against K-shortest-path, predictive-density, and V2X proactive routing, as well as a no-reroute reference under matched demand and disruption schedules at five CAV penetration levels.Source: C. Contributions of This Paper · p. 2

II. Modeling Framework and Predictive Routing

  • The evaluation examines travel-time index, average speed, emissions, distance traveled, recorded charging energy, and accepted route changes.The modeling-framework section introduces the simulation setting, forecasting and routing methods, and performance measures.Source: C. Contributions of This Paper · p. 2, §II · p. 2

A. Transportation Network and Simulation Setup

The study evaluates routing policies on the NYC-6 SUMO network across five CAV penetration cases. Active methods share demand, disruption schedules, and charging-station configurations within each case, while the reference applies no CAV rerouting.

A. Transportation Network and Simulation Setup

  • The NYC-6 SUMO network is a directed graph with 638 edges and 254 junctions, including 368 non-internal road edges and 211 non-internal junctions.Each penetration case uses four charging stations on four station edges.Source: A. Transportation Network and… · pp. 2–3
  • The five CAV penetration cases use 589, 582, 573, 567, and 558 vehicle or trip entries, respectively.At each penetration level, active methods use the same demand, disruption schedule, and charging-station configuration.Source: A. Transportation Network and… · p. 2

B. Disruptions and Reference Conditions

  • Controlled intersection disruptions stop selected vehicles in junction areas, while rear-end disruptions place a vehicle behind a stopped lead vehicle.The no-reroute reference uses the same mixed traffic and disruptions but applies no CAV route intervention.Source: B. Disruptions and Reference Conditions · p. 3

C. Common Routing Procedure

For an eligible CAV, the common procedure forms feasible candidate routes from its current edge to its original destination, selects the lowest-cost candidate, and applies route-change checks.

C. Common Routing Procedure

  • At each decision time, candidate routes are evaluated for an eligible CAV from its current edge toward its original destination.The candidate generator can return at most K loop-free routes connected under SUMO’s network connectivity.Source: C. Common Routing Procedure · p. 3
  • The selected route is the feasible candidate with minimum route cost.Candidate routes are ranked over the method-specific feasible set.Source: C. Common Routing Procedure · p. 3
  • A route change requires positive route-cost improvement and passing cooldown, route-change-limit, duplicate-route, and feasibility checks.Ties follow the deterministic order returned by the candidate-search routine.Source: C. Common Routing Procedure · p. 3
  • All active methods use a one-second decision clock, but DCRNN uses a shorter cooldown and fewer permitted non-forced changes per CAV than the baselines.K-shortest-path generation uses K=3; DCRNN also applies route-cost, CO2-improvement, and length gates for accepted non-forced reroutes.Source: C. Common Routing Procedure · p. 3

D. Forecasting and Routing Policies

The section combines edge-level DCRNN forecasts with route-cost policies to assess feasible alternatives for eligible CAVs. It also defines benchmark policies and separates decision-time route scores from realized outcomes.

D. Forecasting and Routing Policies

  • The DCRNN uses filtered SUMO road edges as graph nodes and features, forecasts over a 120 s nominal horizon, and live history available up to each routing decision.The chronological 80/20 train-validation split and live-history restriction avoid exposing the policy to future evaluation observations.Source: 1) DCRNN-Based Predictive Routing · p. 3
  • DCRNN route costs combine forecast TTI and EV-risk terms with current or proxy CO2, normalized edge length, and dimensionless charging-demand risk.The weights are 0.95, 3.1, 0.90, and 2.5 for TTI, CO2, length, and charging risk, respectively.Source: 1) DCRNN-Based Predictive Routing · pp. 3–4
  • For accepted DCRNN reroutes, recorded old-route minus new-route score components explain decision-time choices but are not independent measurements of realized performance.The implementation keeps an audit trail of routes, score components, and vehicles; realized outcomes are computed from SUMO outputs after simulation.Source: 1) DCRNN-Based Predictive Routing · p. 4
  • EV battery-output values are diagnostic aggregates, not primary evidence for energy-efficiency conclusions, because cumulative fields summed over time may overstate delivered energy.The main performance claims instead use TTI, speed, CO2, distance, and rerouting volume.Source: 1) DCRNN-Based Predictive Routing · p. 4
  • K-shortest-path routing evaluates up to three loop-free routes, while predictive-density routing penalizes edges nearing critical density and V2X routing avoids scheduled warning edges.V2X uses a 180 s warning horizon and eight-edge lookahead; candidates must pass common acceptance checks, with a collision-window fallback.Source: 2) Benchmark Routing Policies · p. 4

E. Performance Measures

Performance measures are computed from retained traversal records, with averages and sums defined separately. EV battery output remains diagnostic rather than an independent energy-efficiency measure.

E. Performance Measures

  • The retained records exclude internal movements, staging-vehicle records, and unmatched reference records before performance metrics are calculated.Each record includes duration, distance, CO2 mass, and reference time.Source: E. Performance Measures · p. 4
  • Average TTI and speed are arithmetic means over retained traversals, while total distance and CO2 are summed; speed is in km/h, distance in km, and CO2 in kg.SUMO CO2 values are converted from milligrams by 10^6.Source: E. Performance Measures · p. 4
  • The reported 1.8% TTI improvement uses scenario-averaged TTIs, while EV kWh is retained only as a diagnostic aggregate because sampled battery fields may be cumulative or monotonic.Reroutes, CO2, distance, and diagnostic EV kWh are summed across five scenarios; average reroutes, TTI, speed, and safety rows are arithmetic averages.Source: E. Performance Measures · p. 4

III. Results

The results section compares four active routing methods across five CAV penetration levels under the same disrupted traffic conditions as the no-reroute reference. The supplied figure caption identifies average TTI across penetration levels.

III. Results

  • The comparison covers K-shortest-path, predictive-density, V2X proactive, and DCRNN routing at CAV penetration levels of 5%, 15%, 25%, 35%, and 45%.The no-reroute reference uses the same disruptions without CAV route intervention.Source: §III · p. 4
  • Each method and case is a single deterministic run, with no random-seed confidence intervals reported; collision outputs serve only as safety diagnostics.Figure 1 presents average TTI across CAV penetration levels with the no-reroute reference.Source: §III · p. 4, Figure 1 · p. 5

A. Aggregate Method Comparison

Across the aggregate comparison, DCRNN has the strongest average TTI, speed, and CO2 results among active methods while using fewer route changes than K-shortest-path and V2X.

A. Aggregate Method Comparison

  • DCRNN achieves the lowest average TTI, highest average speed, and lowest total CO2 emissions among the four active routing methods.Table II aggregates the method comparison across all CAV penetration levels.Source: A. Aggregate Method Comparison · p. 5, Table II · p. 5
  • DCRNN accepts 41 route changes across all penetration levels, compared with 146 for K-shortest-path and 153 for V2X.The results characterize this as improving the congestion–rerouting tradeoff without high-volume route intervention.Source: A. Aggregate Method Comparison · p. 5
  • Predictive-density routing is the most conservative method, with five reroutes and the lowest diagnostic EV kWh, but it has the highest average TTI and CO2 among active methods.V2X has the lowest active-method distance but requires the largest rerouting volume.Source: A. Aggregate Method Comparison · p. 5

B. No-Reroute Reference Comparison

DCRNN reduces average TTI relative to the no-reroute reference, while its emissions and distance comparisons depend on whether it is compared with active routing methods or the reference. Its TTI improvements are clearer at higher CAV penetration, despite maintaining a small, stable number of accepted reroutes.

B. No-Reroute Reference Comparison

  • Approximately 1.8% lower average TTI than the no-reroute reference is reported for DCRNN across the five penetration levels.The result is described under the tested disruptions.Source: B. No-Reroute Reference Comparison · p. 5
  • DCRNN maintains a small, stable number of accepted reroutes as K-shortest-path and V2X interventions rise with penetration.The no-reroute reference has zero route changes by definition.Source: B. No-Reroute Reference Comparison · p. 5, Figure 2 · p. 5
  • DCRNN has the lowest aggregate CO2 emissions among active routing methods, but the no-reroute reference has lower aggregate emissions.The supported emissions advantage is limited to comparisons among active methods.Source: Figure 3 · p. 5, B. No-Reroute Reference Comparison · p. 5
  • The no-reroute reference has the lowest aggregate distance; among active methods, V2X has the lowest total distance, with DCRNN close and using many fewer route changes.DCRNN balances route length with predicted congestion, CO2, and charging-risk terms, so an accepted route can be longer when its score indicates a congestion benefit.Source: B. No-Reroute Reference Comparison · p. 5, Figure 4 · p. 6
  • DCRNN shows clearer positive TTI reductions at higher CAV penetration, while rerouting volume alone does not ensure the best congestion or emissions outcomes.K-shortest-path and V2X reroute more vehicles without matching DCRNN’s average TTI or speed; predictive-density routing changes very few routes but does not achieve the best congestion or emissions outcomes.Source: B. No-Reroute Reference Comparison · pp. 5–6, Figure 5 · p. 6

C. Average Multi-Objective Radar Comparison

The radar comparison summarizes tradeoffs among active methods, while the accepted-reroute analysis and aggregate results show DCRNN’s performance relative to both active routing methods and the no-reroute reference.

C. Average Multi-Objective Radar Comparison

  • Each radar spoke is normalized across active methods, with 1 marking the best method and 0 the weakest; the plot is relative rather than a substitute for raw values.Lower raw values are preferred for reroutes, TTI, CO2, distance, and diagnostic EV battery output, while higher speed is preferred.Source: C. Average Multi-Objective Radar… · p. 6
  • Fig. 6 shows DCRNN strongest among active methods on normalized TTI, speed, and CO2, with better rerouting selectivity than K-shortest and V2X.Predictive-density leads selectivity and diagnostic EV-kWh because it reroutes very few vehicles; V2X leads distance but has the highest rerouting volume.Source: Figure 6 · p. 6

E. DCRNN Accepted-Reroute Benefits

  • Across penetration levels, DCRNN accepted 41 reroutes associated with cumulative route-score reductions of 403.675 in TTI, 113275.495 in CO2, 10.219 in EV charging-risk exposure, and 1387.290 m in route length.These are candidate-route objective records, not independent counterfactual measurements of network state.Source: E. DCRNN Accepted-Reroute Benefits · p. 6, Table III · p. 6
  • Higher-penetration cases had the largest objective reductions, while the 5% case had more accepted changes because two vehicles received repeated updates but a smaller total reduction.The passage emphasizes that reroute count alone does not measure route-score change.Source: E. DCRNN Accepted-Reroute Benefits · p. 7
  • DCRNN occupies an intermediate intervention regime: it reroutes more than predictive-density but far fewer vehicles than K-shortest-path and V2X, while achieving the strongest aggregate TTI and speed among active methods.The passage interprets this pattern as consistent with selective rather than unrestricted acceptance.Source: E. DCRNN Accepted-Reroute Benefits · p. 7
  • No-reroute retains lower total distance and aggregate emissions, so the reported congestion benefit does not mean rerouting improves every aggregate measure.The reported average TTI improvement relative to no-reroute is approximately 1.8%.Source: E. DCRNN Accepted-Reroute Benefits · p. 7
  • Safety outputs check whether rerouting placed CAVs in constructed incident windows; they are not used to estimate a network-wide crash rate.The scripted disruptions create controlled congestion and conflict conditions, while mobility and emissions metrics evaluate network performance.Source: E. DCRNN Accepted-Reroute Benefits · p. 7
  • DCRNN improves average TTI by approximately 1.8% relative to no-reroute and leads active methods in average TTI, speed, and aggregate CO2, with fewer changes than K-shortest-path and V2X.The no-reroute reference retains lower aggregate CO2 and total distance.Source: E. DCRNN Accepted-Reroute Benefits · p. 7

IV. Conclusion

The conclusion presents a predictive routing framework for CAVs in mixed traffic with EV charging demand and reports its aggregate performance under controlled disruptions. It also identifies evaluation extensions and analyses needed to clarify the findings’ scope.

IV. Conclusion

  • In SUMO disruption scenarios across five CAV penetration levels, DCRNN achieved the lowest average TTI, highest average speed, and lowest aggregate CO2 among active methods, with fewer route changes than K-shortest-path and V2X.The no-reroute reference had lower aggregate emissions and distance, while DCRNN improved aggregate TTI relative to that reference.Source: §IV · p. 7
  • The framework combines DCRNN forecasts with feasible-route evaluation using congestion, CO2 emissions, route length, and charging-demand terms.The conclusion describes predictive routing for CAVs in mixed traffic with EV charging demand.Source: §IV · p. 7
  • Future work calls for larger networks and repeated demand and disruption realizations, controlled comparisons of forecasting and charging-demand contributions, and explicit charging queues and station availability.It also proposes matched candidate-route replay under alternative forecasts and a verified incremental charging-energy measure.Source: §IV · p. 7
Loading 2610.10532v1…