🤖 AI

HLSR: Hybrid Live Forecast Selective Dynamic Vehicle Rerouting for Real-Time Congestion Avoidance

This paper proposes HLSR, a selective hybrid live-forecast framework that integrates real-time speeds with short-horizon predictions and driver-tailored metrics to dynamically reroute a limited subset of vehicles for effective real-time congestion avoidance without requiring network-wide replanning.

Original authors: Xiao Wang, Shun Ren Yang, Hui Nien Hung

Published 2026-08-19

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Xiao Wang, Shun Ren Yang, Hui Nien Hung

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Cities are growing denser, and the roads within them are struggling to keep up. As more people move from the countryside into urban centers, the pressure on traffic networks intensifies. Simply building more roads is often impossible due to space and cost, so cities must find smarter ways to manage the flow of existing vehicles. Current navigation apps, like those found on smartphones, are good at finding the fastest route based on traffic conditions at the moment a driver starts their trip. However, they struggle when a driver is already on the road. If a traffic jam forms ahead, these apps often fail to re-route drivers who are already committed to a path, or worse, they might send too many drivers onto a new route, creating a new bottleneck elsewhere. The challenge for traffic engineers is to intervene just enough to smooth out congestion without causing chaos by changing the plans of every single car on the road.

To solve this, researchers have developed a new system called HLSR, which stands for Hybrid Live–Forecast Selective Dynamic Vehicle Rerouting. This approach was tested in a detailed computer simulation of the city of Tainan, Taiwan, using a microscopic traffic model that tracks individual vehicles. The core idea behind HLSR is to be selective rather than overwhelming. Instead of trying to re-plan the route for every vehicle in the city, the system identifies specific bottlenecks and only intervenes with the vehicles that are likely to get stuck or are approaching the trouble spot. It acts like a traffic manager who watches the flow, spots a developing jam, and gently guides the cars approaching it onto a different path before they get trapped, while leaving the rest of the traffic to flow naturally.

The system works in three main stages. First, it detects congestion by looking at two specific signs: how full the road is and how fast the cars are moving. It does not rely on just one of these factors; instead, it requires both the road to be crowded and the speed to be dropping significantly before it decides a jam is real. This prevents the system from reacting to minor slowdowns that would clear up on their own. Once a jam is confirmed, the system selects which vehicles to re-route. It looks at the cars currently on the road segments leading up to the jam, as well as those that are just about to enter those segments. This ensures that the intervention is timely, catching drivers before they are fully stuck in the queue.

The most significant innovation in HLSR is how it calculates the best new route. Traditional systems usually look only at the traffic conditions right now. HLSR, however, blends what is happening on the road at this exact moment with a short-term prediction of what will happen in the next few minutes. For the part of the route a driver will take immediately, the system trusts the live data from road sensors. For the parts of the route further down the line, where the driver will arrive later, the system relies on a forecast of future traffic speeds. This hybrid approach allows the system to avoid routes that look clear now but are predicted to become congested by the time the driver arrives. It also personalizes these predictions by accounting for how different drivers behave, such as whether they tend to drive faster or slower than the average.

The results of the simulation were striking. In a scenario with 8,000 vehicles, the new system reduced the average travel time to 380.6 seconds. This was significantly faster than a system that only used live data to re-route the same number of cars, which resulted in an average time of 438.1 seconds. It also outperformed a system that tried to re-route every single car in the city using only live data, which took 408.1 seconds on average. The researchers found that the hybrid live-and-forecast method was the key driver of this success; without the predictive element, the system's performance dropped noticeably. Even when the traffic volume was increased to 16,000 and then 20,000 vehicles, the HLSR system remained the most efficient, keeping travel times lower than all other tested methods.

The study also explored how far back the system should look when selecting cars to re-route. They found that looking too far back caused unnecessary changes for drivers who were far away from the jam, while looking too close missed drivers who needed help. The optimal setting was to look back nine road segments, which balanced the need to intervene early without disrupting too many people. The researchers also tested different ways of predicting future traffic. They found that a system designed to specifically rank the best routes, rather than just predicting speed accurately for every single road segment, worked better for the overall goal of reducing travel time.

Ultimately, this research demonstrates that a targeted, smart approach to traffic management is more effective than a blanket one. By combining real-time observations with short-term forecasts and intervening only where necessary, cities can reduce travel times and emissions without the massive disruption of re-routing every vehicle. The simulation suggests that this method is robust enough to handle heavy traffic loads, offering a practical path forward for intelligent transportation systems that aim to keep cities moving smoothly.

Technical Summary: HLSR – Hybrid Live–Forecast Selective Dynamic Vehicle Rerouting

1. Problem Statement

Urban traffic congestion significantly reduces productivity and increases travel costs and emissions. While network-wide live shortest-path rerouting (e.g., re-planning every vehicle at every decision interval) shows promise in simulations, it relies on the unrealistic assumption that every on-road vehicle can be replanned continuously. Furthermore, existing selective rerouting methods suffer from three primary limitations:

Congestion Detection: Single-indicator or locally scoped detectors often miss heterogeneous congestion patterns and network-wide coupling.
Vehicle Selection: Fixed spatial windows or greedy selection strategies tend to either over-reroute distant vehicles or under-react to upstream traffic approaching a bottleneck.
Route Allocation: Policies based solely on current live traffic ignore short-horizon forecast information required for segments that will be traversed in future decision periods, leading to sub-optimal routing as conditions evolve.

The paper addresses the need for a framework that intervenes selectively on congestion-affected vehicles while fusing real-time observations with short-term forecasts to optimize route allocation under limited intervention scope.

2. Methodology: The HLSR Framework

The authors propose HLSR (Hybrid Live–Forecast Selective Rerouting), a three-phase framework that integrates dual-threshold detection, calibrated vehicle selection, and hybrid live–forecast route costing.

A. System Model and Sensing

The system operates on a directed urban road network partitioned into zones. It utilizes:

Live Sensing: Real-time segment occupancy ($RO$) and velocity ratios ($RV$) derived from loop detectors and vehicle reports.
Decision Windows: Discrete aggregation windows (default $T=300$ s) where statistics are accumulated and control loops are invoked.
Speed Values: The framework distinguishes between Live Speed ( $V^{live}$ ) from current sensors and Predicted Speed ( $V^{pred}$ ) generated by a spatio-temporal forecaster for future horizons.

B. Phase 1: Dual-Threshold Congestion Detection

Congestion is declared only when both normalized occupancy and velocity metrics exceed specific thresholds ( $RO > \delta_O$ and $1-RV > \delta_V$ ). This dual-threshold approach prevents false positives where occupancy and speed diverge. A composite severity metric ($RS$) is calculated to weight the intervention.

C. Phase 2: Calibrated Vehicle Selection

Instead of fixed spatial windows, HLSR employs a calibrated upstream hop depth ( $\theta_{ur}$ ) combined with an approaching-vehicle expansion mechanism:

Upstream Selection: Identifies vehicles currently on segments within $\theta_{ur}$ hops upstream of the bottleneck.
Approaching Expansion: Supplements the set with vehicles outside this window that are on a planned route to the bottleneck and have a remaining hop count of $\le 2$ .
This ensures the intervention scope is bounded to vehicles likely to impact the bottleneck, avoiding unnecessary rerouting of distant traffic.

D. Phase 3: Hybrid Live–Forecast Route Allocation

This is the core innovation. For each selected vehicle, the system generates $N_{cr}$ candidate paths (using Yen's K-shortest path algorithm) and allocates the optimal route based on a hybrid cost function:

Horizon-Adaptive Speed Fusion: The cost of a path segment is calculated using a fused speed $\tilde{V}$ , which blends live and predicted speeds based on the prediction horizon ( $h$ ):
$\tilde{V}(s, \tau, h) = \alpha(h) V^{live}(s, \tau) + (1 - \alpha(h)) V^{pred}(s, \tau, h)$
where $\alpha(h)$ decays linearly with the horizon. Near-term segments rely on live data; distant segments rely on forecasts.
Driver-Tailored Prediction: The prediction module uses LSTAN_GERPE, a network-level spatio-temporal forecaster. Crucially, it incorporates a driver-behavior parameter ( $B_d$ ) to personalize predicted speeds for specific vehicles, rather than using a generic network average.
Ranking-Aware Fine-Tuning: The forecaster is optionally fine-tuned with a k-shortest-path ranking loss. This ensures the model preserves the relative ordering of candidate routes, not just edge-level accuracy.
Multi-Objective Scoring: The final route is selected by minimizing a weighted sum of normalized travel time, path length, route similarity, and network occupancy balance.

3. Key Contributions

The paper claims four primary contributions:

Proactive Selective Rerouting: A bounded intervention framework using dual-threshold detection and calibrated upstream selection that limits rerouting to congestion-related vehicles while capturing approaching traffic.
Hybrid Live–Forecast Costing: A novel route costing mechanism that fuses live and predicted speeds with horizon-dependent weights, coupled with travel-time-weighted K-shortest-path generation.
Driver-Aware Prediction: A network-level forecaster (LSTAN_GERPE) enhanced with driver-behavior personalization and ranking-aware fine-tuning, serving as a supporting module for routing rather than a standalone end-to-end solver.
Fair-Scope Evaluation: A comprehensive evaluation on a reproduced Tainan SUMO scenario (seed 42) comparing HLSR against classical selective baselines, reimplemented state-of-the-art methods, and network-wide live Dijkstra, across varying traffic demands (8,000 to 20,000 vehicles).

4. Experimental Results

Experiments were conducted on a Tainan road network (204 intersections, 561 segments) using SUMO with seed 42.

Performance at 8,000 Vehicles:
- HLSR achieved a mean travel time of 380.6 s.
- This outperformed HLSR-LIVE (selective rerouting with live speeds only) by 57.5 s (438.1 s), demonstrating the value of hybrid forecasting.
- It outperformed CAIE-TT-Scoped (live Dijkstra on the same selective set) by 11.0 s (391.6 s).
- It remained competitive with CAIE-TT (network-wide live Dijkstra, 408.1 s) while intervening on far fewer vehicles.
- It significantly outperformed other selective baselines like Du-GAQ (425.5 s) and NRR (525.5 s).
Scalability (16,000 and 20,000 Vehicles):
- HLSR maintained the best performance at higher loads: 895.7 s (16k) and 971.7 s (20k).
- The performance gap widened against same-scope live-only controls (e.g., HLSR-LIVE degraded to 1281.3 s at 16k vehicles), indicating that hybrid forecasting is critical under heavy congestion to prevent over-rerouting.
Ablation Studies:
- Removing hybrid costing (HLSR-LIVE) caused the largest performance degradation (+57.5 s).
- Removing driver personalization or using a non-ranking-aware forecaster resulted in smaller but consistent penalties.
- The upstream hop depth $\theta_{ur}=9$ was identified as optimal, balancing intervention scope and travel time.

5. Significance and Claims

The paper asserts that HLSR demonstrates that hybrid live–forecast costing is the primary driver of performance improvement in selective rerouting, rather than the specific choice of the prediction backbone or the sheer volume of rerouted vehicles.

Key claims include:

Efficiency: HLSR achieves lower mean travel times than network-wide live rerouting schemes while intervening on a significantly smaller subset of vehicles, thereby reducing computational overhead and driver annoyance.
Robustness: The framework remains effective under high-demand scenarios where live-only selective methods fail due to over-rerouting.
Integration: The work successfully integrates dual-threshold detection, calibrated selection, and horizon-adaptive costing into a unified pipeline, addressing the recurring limitations of existing three-phase rerouting methods.
Role of Forecasting: The forecasting module is positioned as a supporting component for route costing, not a standalone end-to-end rerouter, ensuring that real-time observations remain the authoritative reference for near-horizon decisions.

The authors conclude that by fusing real-time sensing with short-horizon forecasts under a bounded intervention scope, HLSR offers a practical and effective solution for real-time urban congestion mitigation.