PaperlayerView exact revision on arXiv

Source-linked AI summary

Traffic Congestion Awareness and On-Demand Distribution in Vehicular Delay-Tolerant Networks in California I-210 Freeway

Xiaofei Liu, Milena Radenkovic

arXiv:2609.36156v1cs.NI

TL;DR

Flooding-based V-DTN warnings can create redundant copies, while existing approaches give limited attention to congestion events themselves. This paper uses PeMS data to model congestion recurrence and duration and applies event-driven dynamic thresholds to on-demand dissemination. Experiments report reduced network resource use, higher delivery rates, and shorter delivery delays.

  • Problem

    Flooding-based V-DTN warning dissemination creates redundant copies, while congestion-event recurrence and duration receive limited attention in existing approaches.

    Source: §I · p. 1

  • Method

    Using PeMS I-210 data, the paper models congestion recurrence and duration and uses event-driven dynamic thresholds for on-demand V-DTN dissemination.

    Source: B. I-210 Freeway Traffic Congestion… · pp. 3–4, A. Conclusion · p. 7

  • Results

    As the dynamic threshold rises, delivery rate increases from approximately 0.80 to 0.86 and median delivery delay decreases by about 38%.

    Source: §V · p. 7

  • Takeaways & Limitations

    Dynamic-threshold dissemination reduced network resource consumption, while duration prediction provided a basis for setting thresholds in the I-210 V-DTN use case.

    Source: §V · p. 7, A. Conclusion · p. 7

    The congestion-prediction method relies heavily on historical free-flow speed data.

    Source: B. Reflections and Future Work · p. 9

Abstract

from arXiv · show

In vehicular networks under edge computing environments, vehicle-to-vehicle delay-tolerant networking (V-DTN) can disseminate congestion warnings to other vehicles via a store-carry-forward mechanism, helping them proactively choose suitable routes. However, most existing in-vehicle information dissemination methods rely on flooding or limited flooding strategies, broadcasting alerts across the entire network whenever congestion is detected. This leads to excessive redundant copies and consumes node cache space. To address this issue, this paper proposes a congestion-event- aware on-demand message dissemination mechanism. By considering the recurrence and duration of congestion, the mechanism suppresses broadcasts of short-lived, self-dissipating congestion events. Based on real-world PeMS data from California's I-210 corridor, we construct an I-210 Freeway Traffic Congestion Use Case. Experiments show that predicting congestion recurrence heavily relies on historical free-flow speed data of lanes. Without incorporating additional feature dimensions, the accuracy of modeling and comparing lane free-flow speeds across different day types outweighs the choice of machine learning models, and such patterns are difficult to reproduce in simulators. Meanwhile, congestion duration prediction serves effectively as the basis for dynamic-threshold on-demand dissemination in V-DTN, which demonstrate that machine learning-based congestion awareness combined with dynamic-threshold on-demand dissemination significantly reduces network resource consumption and network-layer congestion, enhances delivery reliability, and lowers delivery latency.

I. INTRODUCTION

V-DTNs can carry congestion warnings to approaching vehicles, but flooding-based dissemination creates redundant traffic and can worsen network congestion. The paper identifies a gap in addressing congestion events themselves and calls for online prediction that accommodates dynamic traffic.

I. INTRODUCTION

  • Flooding and limited flooding broadcast warnings network-wide, creating redundant copies that consume cache, bandwidth, and energy and can undermine reliable delivery.Short-term periodic alerts may contribute to buffer overflow, message loss, and self-induced network congestion.Source: §I · p. 1
  • The study explores using predicted congestion duration and periodicity as thresholds for selective, on-demand V-DTN warning dissemination.Source: §I · p. 1

A. Traffic Congestion Awareness Based on Machine Learning

  • Earlier work describes machine learning for nonlinear traffic data and incident-duration prediction, including classifiers and CART regression using sensor data.The review also notes simulation-assisted anomaly detection and edge computing's role in real-time processing.Source: A. Traffic Congestion Awareness Based… · p. 1
  • Prior congestion-awareness research emphasizes accident-induced events, while many models rely on offline historical analysis that does not accommodate dynamic urban traffic needs.The passage calls for online prediction combining edge perception and lightweight machine learning.Source: A. Traffic Congestion Awareness Based… · p. 2

B. Gradient Boosting and LightGBM Algorithm

The background introduces gradient boosting and LightGBM, reviews approaches to non-recurrent congestion, and motivates event-aware forwarding in vehicular delay-tolerant networks.

B. Gradient Boosting and LightGBM Algorithm

  • Gradient boosting trains base models sequentially against loss gradients; LightGBM uses GOSS and EFB to improve training speed and efficiency.The passage notes LightGBM can overfit smaller datasets and that anomalies or insufficient data can degrade predictive performance.Source: B. Gradient Boosting and LightGBM… · p. 2

C. Non-Recurrent Traffic Congestion

  • Non-recurrent congestion research has progressed from statistical decomposition to tensor and deep-learning methods, but often relies on post-event detection and high-quality labeled data.The cited discussion identifies overlapping incidents in complex urban environments and calls for proactive forecasting of congestion evolution, diffusion, and duration.Source: C. Non-Recurrent Traffic Congestion · p. 2

D. Vehicular Delay-Tolerant Network

E. Simulation of Urban MObility (SUMO) and The ONE Simulator

The background describes SUMO and The ONE as tools for evaluating vehicular delay-tolerant networks, and PeMS as a source of California freeway measurements.

E. Simulation of Urban MObility (SUMO) and The ONE Simulator

  • SUMO converts real-world traffic into microscopic vehicle-mobility models, while its trajectories can support communication-protocol evaluation.Source: E. Simulation of Urban MObility (SUMO)… · p. 2
  • The ONE evaluates store-carry-forward DTN protocols; prior disaster simulations found unlimited replication could cause cache overflow, while limited-replica approaches achieved higher delivery rates.In one cited Nepal earthquake scenario, Spray-and-Wait maintained a 94.6% delivery rate under constrained resources.Source: E. Simulation of Urban MObility (SUMO)… · p. 2

A. Data Source: Performance Measurement System (PeMS)

B. I-210 Freeway Traffic Congestion Dataset Generative Pipeline

The pipeline uses three months of California PeMS data from the I-210 corridor to identify congestion events and generate recurrence-classification and duration-regression targets. It combines historical speed baselines with neighboring-station speeds to represent recurrence and spatial changes in congestion.

B. I-210 Freeway Traffic Congestion Dataset Generative Pipeline

  • The dataset uses September–November 2024 PeMS observations from I-210 and retains 49 mainline stations after quality screening.The figure caption identifies the original data as five-minute observations.Source: B. I-210 Freeway Traffic Congestion… · p. 3, Figure 2 · p. 3
  • Events below the same station, workday, and time-period 15th-percentile speed bound were classified as non-recurrent; other events were recurrent.The group median represented normal speed, while the 15th percentile represented the lower bound of the normal speed band.Source: B. I-210 Freeway Traffic Congestion… · p. 3, Figure 3 · p. 4
  • Speeds from adjacent upstream and downstream stations were added to capture spatial changes related to congestion duration.Downstream speeds support bottleneck identification, while falling upstream speeds indicate a tendency for congestion to spread downstream.Source: B. I-210 Freeway Traffic Congestion… · pp. 3–4
  • 5,693 congestion events were identified after applying speed thresholds and excluding events shorter than 15 minutes; features came from the first 10 minutes after each start.Congestion duration was the regression target, and whether an event was non-recurrent was the classification target.Source: B. I-210 Freeway Traffic Congestion… · p. 4, Figure 4 · p. 4

IV. MACHINE LEARING MODEL TRAINING

On the I-210 use case, limiting congestion-duration modeling to events of at most 240 minutes improved reported errors, while LightGBM performed best among the tested regression models. Upstream and downstream traffic features were especially informative.

A. LightGBM Regression: Predicting congestion duration

  • Upstream and downstream spatiotemporal features ranked first and second in regression feature importance, respectively.Downstream speeds indicate bottlenecks ahead, while upstream speeds indicate continued inflow and whether queuing may worsen.Source: A. LightGBM Regression: Predicting… · pp. 4–5, Figure 8 · p. 5
  • Limiting duration modeling to ≤240 minutes (about 97% of events) yielded 5-fold CV MAE 34.7 ± 1.1 and RMSE 52.6, versus full-sample MAE 64.3 and RMSE 113.1.Events exceeding 240 minutes account for about 3% of the data but contribute the vast majority of total error.Source: A. LightGBM Regression: Predicting… · p. 4, Table II · p. 5
  • LightGBM achieved the lowest MAE among five models, while tree-based ensembles outperformed linear and distance-based alternatives on this dataset.Random Forest ranked second; Linear Regression performed worst, and KNN was only slightly better than Linear Regression.Source: Figure 9 · p. 5, A. LightGBM Regression: Predicting… · p. 5

B. Classification: Non-Recurrent Congestion Determine

For identifying non-recurrent congestion, a direct free-flow-versus-congestion speed comparison performed comparably to machine-learning models. The models showed little performance separation and relied predominantly on a free-flow speed threshold; the study’s short data period limits its recurrence assessment.

B. Classification: Non-Recurrent Congestion Determine

  • Classification models had no significant overall performance differences and relied chiefly on the lower bound of the 15th percentile of free-flow speed.The analysis reports that models did not fully capture more complex spatiotemporal traffic evolution patterns.Source: B. Classification: Non-Recurrent… · p. 5, Figure 11 · p. 6, Figure 10 · p. 6
  • Directly comparing free-flow speed with speed during a congestion event achieved performance comparable to machine-learning models.The authors report that, under the current data conditions, machine-learning advantages were not fully realized.Source: B. Classification: Non-Recurrent… · p. 5
  • The recurrence assessment is limited because the study uses only September–November 2024 data, insufficient to capture traffic across months, seasons, and holidays.The authors recommend longer historical records and finer grouping by day type and other factors.Source: B. Classification: Non-Recurrent… · p. 6

V. ON-DEMAND DISTRIBUTION EVALUATION BASE ON MACHINE LEANING MODEL

The evaluation uses SUMO-generated I-210 congestion durations to set dissemination thresholds in The ONE. Increasing the threshold reduces message generation and forwarding while improving delivery rate and latency.

V. ON-DEMAND DISTRIBUTION EVALUATION BASE ON MACHINE LEANING MODEL

  • SUMO traffic data fed to a machine-learning pipeline yielded 174 congestion events; approximately one-third lasted under 30 minutes and another third lasted 30–60 minutes.The predicted durations form the basis for dynamic thresholds in on-demand dissemination.Source: §V · p. 6, Figure 15 · p. 7
  • SUMO produced synthetic five-minute lane-level speed, flow, and occupancy series from an I-210 freeway-and-ramps network built from OpenStreetMap data.Congestion was simulated both through excessive traffic volume and through reduced lane capacity from speed-limit changes and lane closures.Source: §V · p. 6, Figure 12 · p. 6, Figure 13 · p. 7
  • The ONE evaluation compared broadcasting warnings for every event against 30- and 60-minute thresholds that suppress events predicted to last less than the cutoff.The setup used six congestion points and 90 free-flowing vehicles as relays and receivers.Source: §V · pp. 6–7
  • As the dynamic threshold rises, message creation falls from 8,870 to 5,043 (about 43%), forwarding drops about 17%, delivery rate rises from approximately 0.80 to 0.86, and median delivery delay falls about 38%.These results indicate reduced network resource consumption and improved delivery reliability and latency.Source: Figure 17 · p. 7, §V · p. 7

A. Conclusion

The conclusion reports the I-210 use case and its machine-learning models for congestion duration and recurrence. LightGBM performed best for duration prediction, while recurrence classification depended more on historical lane free-flow speeds than on model choice.

A. Conclusion

  • The study combines PeMS I-210 detector data, machine-learning congestion perception, and event-driven dynamic-threshold dissemination in V-DTN.Regression and classification models target congestion duration and recurrence patterns, respectively.Source: A. Conclusion · pp. 7–8
  • For duration prediction, LightGBM performed best among compared models, with 5-fold cross-validation MAE 34.7 ± 1.1 and RMSE 52.6 over a 240-minute window, improving to MAE 26.7 ± 0.1 and RMSE 38.1 over 180 minutes.The conclusion attributes underfitting beyond 240 minutes to too few long-duration samples and identifies sample imbalance as a key factor.Source: A. Conclusion · p. 8
  • Recurrence models achieved nearly identical F1-scores of approximately 0.85, comparable to rule-based methods; historical lane free-flow speeds were the most decisive feature.Without additional feature dimensions, comparing free-flow speeds across day types mattered more than model choice.Source: A. Conclusion · p. 8
  • Raising the dissemination threshold reduced message generation from 8,870 to 5,043 (about 43%) and forwarding by approximately 17%, while delivery rate rose from around 0.80 to 0.86 and median delay fell about 38%.The conclusion reports reduced network resource consumption and network-layer congestion alongside improved delivery reliability and latency.Source: A. Conclusion · p. 8

B. Reflections and Future Work

The authors identify incomplete simulator coupling, reliance on historical free-flow speeds, and limited protocol and geographic coverage as directions for future work.

B. Reflections and Future Work

  • Full SUMO vehicle-trajectory transfer to The ONE was incomplete because of simulator and computing-resource constraints, so the evaluation used an abstracted, decoupled scheme.The authors suggest focusing on specific intersections or using other simulator-coupling approaches to reduce hardware burden and improve experiment reliability.Source: B. Reflections and Future Work · pp. 8–9
  • Congestion prediction currently relies heavily on historical free-flow speeds; the authors propose deep-learning extensions and additional traffic or social-media features.Suggested signals include upstream-downstream speed, queue length, and social-media data for identifying non-recurrent congestion.Source: B. Reflections and Future Work · p. 9
  • The approach still needs testing with additional DTN routing protocols and energy-saving considerations, and broader urban scenarios beyond the I-210 freeway.The authors identify broader detector-data scenarios as a route toward improving transferability and applicability.Source: B. Reflections and Future Work · p. 9

APPENDIX A SOURCE CODE

The appendix provides a GitHub repository for the project source code.

APPENDIX A SOURCE CODE

  • The project source code is available on GitHub, with file descriptions in the README.md.The passage supplies the repository link.Source: Appendix A · p. 9
Loading 2609.36156v1…