Source-linked AI summary
Optimal Scheduling of Road Maintenance Jobs Considering Impact on Traffic Flows
Charitha Nandepu, Lohitha Kalepu, Gabriele Ciavarella, SangWoo Park
TL;DR
Repeated equilibrium-flow evaluations make network-level road-maintenance scheduling computationally burdensome. The paper uses solver-generated equilibrium flows to train surrogate neural models, with an attention-based model achieving the highest predictive accuracy and most faithfully reproducing congestion patterns.
Problem
Repeatedly computing equilibrium traffic flows across maintenance configurations and demand conditions remains computationally burdensome for network-level maintenance planning.
Method
The study generates Beckmann-type equilibrium flows from StreetLight OD demand and trains four neural architectures to approximate arc flows directly from those inputs.
Results
The attention-based neural network achieved the highest predictive accuracy and most faithfully reproduced network congestion patterns, explaining approximately 98% of flow variance with the lowest overall RMSE.
Takeaways & Limitations
The framework provides a data-driven surrogate for representing endogenous network-level traffic responses in road-maintenance scheduling contexts.
Abstract
from arXiv · showhide
Network-level maintenance planning requires repeated evaluations of equilibrium traffic flows under road capacity reductions. While equilibrium traffic assignment models are well established, their repeated solution quickly becomes computationally prohibitive and challenging to embed within maintenance scheduling problems. This paper investigates data-driven surrogate models that approximate equilibrium arc flows directly from origin-destination demand, using optimization-based equilibrium solutions as ground truth. A real-world case study based on traffic data from the Newark, New Jersey area demonstrates the effectiveness of the proposed approach as a scalable building block for future maintenance scheduling frameworks.
1. Introduction and Background
Road maintenance is essential to infrastructure management, but poorly planned interventions can cause network-wide traffic disruption and economic losses. Even localized lane closures may produce broader congestion effects, while traditional scheduling models require traffic assignment within larger optimization frameworks.
- Motivation: Poorly planned road-maintenance interventions can trigger network-wide traffic disruptions and economic losses.The passage identifies road maintenance as a critical infrastructure-management activity whose impacts extend beyond the intervention site.
- Motivation: Localized lane closures can generate non-local congestion effects as drivers reroute through navigation tools such as Google Maps or Apple Maps.These rerouting responses amplify the systemic impact of congestion.
- Modeling challenge: Traditional maintenance-scheduling models often embed traffic assignment within larger optimization frameworks, requiring equilibrium-flow evaluations.This modeling structure contributes to the computational challenge motivating the study.
2. Problem Statement
Evaluating candidate road-maintenance schedules requires repeatedly computing equilibrium traffic flows under varying maintenance combinations and demand conditions. Although equilibrium traffic assignment models are well established and convexly formulable, repeated solutions across periods and configurations become computationally burdensome.
- Operational challenge: Maintenance-schedule evaluation requires repeated equilibrium-flow computations across different maintenance activities and demand conditions.These repeated evaluations are needed from an operational perspective.
- Model foundation: Equilibrium traffic assignment models are well established and can be formulated as convex optimization problems.
- Computational burden: Repeatedly solving equilibrium models across multiple time periods and maintenance configurations quickly becomes computationally burdensome.
3. Approach
The approach combines convex-optimization user-equilibrium traffic assignment with learning-based surrogates that map OD demand directly to equilibrium arc flows. A Newark, New Jersey case study uses StreetLight-derived demand, solver-generated equilibrium flows as ground truth, and four neural architectures for approximation.
- Equilibrium generation: For each day–hour scenario, Wardrop-based user equilibrium is computed through a Beckmann-type convex optimization formulation implemented in CVXPY and solved with MOSEK.The formulation uses destination-based flow decomposition, nonnegative decision variables, integrated travel costs, and destination-specific flow conservation.
- Case study: StreetLight-derived OD volumes define day–hour demand scenarios on a 49-road-segment Newark, New Jersey network.The OD volumes come from aggregated GPS-based trip samples scaled using StreetLight’s Network Origin-Destination methodology.
- Network representation: The transportation network is represented as a directed graph, with traffic-analysis zones as nodes, feasible links as arcs, and OD matrices specifying interzone demand with zero self-demand.Arc flows are the fundamental units of equilibrium computation, and each OD matrix corresponds to a day–hour scenario.
- Ground truth: The resulting equilibrium arc flows are converted to vehicles per hour, aggregated across scenarios, and used as ground truth for training and evaluating learning-based models.Equilibrium travel times are also evaluated using nonlinear arc cost functions.
- Learning models: The learning task approximates the nonlinear mapping from flattened OD matrices to globally indexed equilibrium arc-flow vectors using MLP, CNN, GNN, and attention-based neural architectures.The architectures differ in input representation, including flattened vectors, two-dimensional OD grids, and directed transportation graphs with node and edge features.
4. Results
The results compare equilibrium flow prediction models using RMSE and R2, with separate attention to high-flow arcs above 200 veh/hour. The attention-based model performs best overall, while the GNN is particularly strong on high-demand arcs.
- Performance is evaluated using RMSE and R2 for equilibrium flow prediction.
- 200 veh/hour isolates primary loading corridors from background flow and focuses evaluation on operationally significant arcs.
- The attention-based model achieves the strongest performance, explaining approximately 98% of flow variance with the lowest overall RMSE.
- The GNN also performs strongly, particularly on high-demand arcs.
5. Conclusion
The paper presents a network-level road maintenance scheduling framework that accounts for endogenous traffic-flow responses. It uses equilibrium flows computed from real-world demand as ground truth for learning-based models, with an attention-based neural network performing best among those evaluated.
- The framework optimizes road maintenance scheduling while accounting for endogenous traffic-flow responses at the network level.
- A convex Beckmann-type formulation computes equilibrium arc flows from real-world origin-destination demand derived from StreetLight data.These equilibrium flows serve as ground truth for training learning-based Deep Learning models.
- The attention-based neural network achieved the highest predictive accuracy among the evaluated models.The passage also states that it most faithfully reproduced the reported response, although the supplied text is truncated.