Simulation-based analysis of variable speed limit for traffic performance and safety during incidents Open Access

CC0

standstill distance

CC1

desired time headway

CC2

additive part of safety distance

E

estimated volume

$Mexisting$

value of measure M under the existing scenario

$MVSLi$

value of measure M under the VSL$i$ strategy

n

number of time periods

V

observed volume

$yi$

observed mean speed on period i

$y^i$

estimated mean speed on period i

Traffic congestion in urban areas – especially when triggered by unexpected incidents such as accidents or vehicle breakdowns – poses serious environmental, economic and societal challenges. These non-recurring events reduce road capacity and lead to significant delays, unreliable travel times, increased emissions and operational inefficiencies. Among them, traffic accidents often cause the most disruption due to their severity and the behavioural responses they trigger (Mfinanga and Fungo, 2013). To mitigate such impacts, intelligent transportation systems (ITS) and active traffic management (ATM) strategies have been increasingly adopted. One prominent ATM approach is the variable speed limit (VSL) system, which adjusts speed limits dynamically in response to real-time traffic and roadway conditions. The VSL system has been shown to enhance both safety and mobility under certain conditions (Boyles et al., 2018).

Thailand’s motorway network frequently experiences high-speed traffic and a high rate of accidents. Between 2019 and 2024, over 11 000 accidents were recorded on Thai motorways, underscoring the urgent need for effective incident-management strategies (DOH, 2024). While prior research confirms the potential of VSL systems, their effectiveness varies with implementation context, and field deployment in developing countries remains limited. This study investigates a rule-based VSL approach using microsimulation with the Vissim–Matlab component object model (COM) interface. The objective is twofold: (i) to evaluate the effectiveness of VSL strategies on traffic performance and safety during incidents, and (ii) to evaluate the impact of varying VSL coverage distances (1 km versus 2 km) on system effectiveness.

The VSL system is reviewed, followed by recent VSL studies, rule-based algorithms, microsimulation modelling using PTV’s Vissim software for traffic incidents, and safety assessment using surrogate safety measures.

VSL systems consist of a network of speed limit signs and traffic detectors that monitor prevailing road conditions. The displayed speed limits are determined based on both local traffic data and conditions at adjacent upstream and downstream locations (Grumert et al., 2018). The concept of VSL was first introduced in the UK in the 1960s with the primary aim of enhancing road safety (Lu and Shladover, 2014). This was followed by Germany in the early 1970s and the Netherlands in the 1980s (Soriguera et al., 2017). In these early implementations, the VSL system was used mainly for incident detection and warning, as demonstrated by the motorway control systems in Sweden and the Netherlands (Grumert et al., 2018; van den Hoogen and Smulders, 1994).

Over the past decade, VSL systems have evolved significantly. Simulation-based approaches have enabled the development of more sophisticated algorithms focusing on both safety and mobility (Lu and Shladover, 2014). Moreover, the scope of VSL systems has expanded to address environmental concerns, such as reducing vehicle emissions by optimising traffic flow (Zegeye et al., 2011).

Research has identified three key benefits of VSL systems. First, the VSL enhances safety by minimising speed differences between lanes, synchronising driver behaviour and discouraging sudden lane changes. Second, the VSL mitigates traffic flow breakdowns near bottlenecks by reducing incoming vehicle speeds in advance (Hegyi et al., 2005). Third, the VSL contributes to environmental sustainability through reduced fuel consumption and emissions resulting from smoother traffic operations (Farrag et al., 2020; Khondaker and Kattan, 2015; Sadat and Celikoglu, 2017).

Grumert et al. (2018) classified VSL algorithms into four major categories, each with different levels of complexity and application contexts.

• Rule-based algorithms: these use predefined thresholds to adjust speed limits and are the most commonly implemented due to their simplicity, cost-effectiveness and ease of deployment in real-world scenarios.

• Fuzzy logic-based algorithms: these apply fuzzy inference to determine speed limits based on the degree to which input data matches predefined linguistic rules. This approach allows for more adaptive control, especially under uncertain traffic conditions.

• Analytical algorithms: these models use real-time traffic data to calculate quantitative traffic states. Their strength lies in producing precise assessments for informed speed control, especially in well-instrumented environments.

• Control theory-based algorithms: the most sophisticated type employs feedback control systems to continuously adjust speed limits based on real-time conditions. They are suitable for complex, real-time traffic networks and often require robust computational support.

In practice, the choice of VSL algorithm depends on the system’s goals, available infrastructure and technical feasibility.

In recent years, the development of VSL systems has increasingly been driven by artificial intelligence (AI), reinforcement learning (RL) and data-driven optimisation approaches. These emerging technologies aim to enhance both responsiveness and adaptability in real-time traffic management.

Feng et al. (2023) proposed a lane-specific VSL control model using deep neural networks, which can automatically learn complex relationships between traffic conditions and optimal speed limits. The model was optimised using a covariance matrix adaptation evolution strategy, a stochastic, population-based algorithm, suitable for solving difficult non-linear optimisation problems. By applying this approach, the system dynamically adjusted speed limits per lane, achieving up to 23% improvement in travel time and a notable reduction in vehicle emissions. Similarly, Yang et al. (2024) applied deep deterministic policy gradient, a model-free, off-policy actor–critic RL algorithm, for managing speed limits during emergency situations. The system learned optimal speed adjustment strategies by interacting with simulated traffic environments and receiving feedback in the form of safety and performance metrics. Their results demonstrated over 28% improvement in safety performance under various incident scenarios.

Zheng et al. (2023) developed a coordinated multi-agent RL framework designed to manage consecutive freeway bottlenecks. In this architecture, each VSL sign acts as an independent agent, yet they are trained in a coordinated fashion using a centralised training and decentralised execution paradigm. This design allows the agents to collaborate effectively while making local decisions. The study showed that this approach led to a 17–18% reduction in total system travel time. Fang et al. (2023) utilised multi-agent proximal policy optimisation, an actor–critic-based deep RL algorithm known for its stability and sample efficiency, to coordinate VSL control in areas where motorway lanes merge into urban roads. By treating each control point as an agent and training them collectively, the system reduced overall network waiting times by 15.8% and achieved a significant decrease in carbon dioxide emissions, reflecting both performance and environmental benefits.

Field-based studies have also contributed to VSL development. Isaenko et al. (2024) conducted a field-based study on an Italian motorway to evaluate the impact of a non-mandatory (advisory) VSL system. By applying statistical analysis and clustering techniques to trajectory and speed data, they were able to classify traffic patterns and identify behavioural responses to advisory speed limits. Their findings showed a modest 4% improvement in travel time and a 12–20% reduction in speed variability, suggesting that even non-enforced VSL strategies can improve traffic flow and consistency.

While advanced AI-based VSL models demonstrate high performance, they often rely on connected vehicle technologies and real-time data collection infrastructure, which may not yet be fully available in all regions. In the context of developing countries such as Thailand, where such systems are still in the early stages of deployment, rule-based VSL algorithms offer a more practical and scalable alternative. These models, though simpler, have shown considerable success in improving traffic flow and safety under a variety of conditions.

Rule-based algorithms remain one of the most widely implemented VSL control strategies due to their simplicity, interpretability and ease of deployment. These systems rely on fixed thresholds – such as traffic volume, occupancy and average speed – to determine appropriate speed limits in response to traffic conditions.

Allaby et al. (2007) developed a fundamental rule-based VSL system using the Paramics software program, deploying 13 VSL signs linked to loop detectors spaced every 500–600 m. Speed limits were adjusted every 20 s based on predefined conditions, with logic that considered real-time volume, speed and occupancy. The study found that while safety improved during congestion, travel time slightly increased during free-flow conditions. Further tuning of parameters led to safety improvements and minimised delay, demonstrating the value of tailored rule-based logic.

Sadat and Celikoglu (2017) built on this work by incorporating realistic driver compliance rates in the Vissim software program and calculating new desired speeds using the Matlab software program. The system accounted for variations in speed adherence and introduced constraints on speed changes to maintain traffic stability. They found that VSL systems could effectively reduce congestion levels and confirmed reductions in fuel consumption and emissions (carbon monoxide and nitrogen oxides), and highlighted the sensitivity of rule-based VSL systems to driver behaviour.

Boyles et al. (2018) applied a threshold-based rule algorithm on a freeway corridor in Texas, integrating it into Vissim through Matlab. By responding to data from downstream detectors, the system reduced total delay by 16.8% and improved average travel speed by 10%. The study reinforced the benefit of using simple but responsive VSL algorithms in real-time operational settings. Additionally, their non-recurring congestion modelling showed that even under incident conditions, VSLs enhanced mainline travel time and smoothed flow.

Overall, rule-based VSL systems have consistently demonstrated effectiveness in contexts with limited infrastructure and technological maturity. While less adaptive than AI-driven approaches, they offer a practical, cost-effective and evidence-based solution for enhancing traffic safety and performance during incident scenarios.

PTV Vissim is a widely-used microscopic simulation tool known for its ability to model detailed vehicle behaviour in complex traffic environments. It provides realistic simulation outputs based on the Wiedemann car-following model and allows for integration with external tools such as Matlab through the COM interface (PTV, 2023). According to the Federal Highway Administration (FHWA), accurate calibration of simulation models with real-world data – such as speed, travel time and delay – is essential to ensure credible outputs (Dowling et al., 2004).

Due to the unpredictable nature and complexity of real-world traffic incidents, simulating them using analytical models is often insufficient. Microsimulation enables researchers to reproduce vehicle-level behaviours under various incident scenarios. Since Vissim lacks a built-in incident module, several researchers have proposed innovative methods to simulate incidents.

Xie et al. (2022) and Farrag et al. (2021) demonstrated the use of a ‘parking lot’ object placed on a traffic lane to simulate a road-blocking incident. When a vehicle occupies the parking lot, it mimics a stalled vehicle, effectively reducing lane capacity. Karaer et al. (2020) extended this by incorporating routing decisions to divert approaching vehicles using partial routes and connectors.

To evaluate the impact of VSL strategies on safety, this study adopted a two-stage process. First, Vissim generated vehicle trajectory data, capturing key indicators such as speed, acceleration, deceleration and time-to-collision. These data were then analysed using the Surrogate Safety Assessment Model (SSAM) developed by the FHWA (Kurker et al., 2014).

SSAM identifies potential traffic conflicts – classified as rear-end, lane-change or crossing conflicts – based on driver behaviour patterns. The safety impact in this study was quantified using the number of conflicts that occurred because a higher number of conflicts increases the likelihood of incidents.

This section outlines the methodology used to develop, simulate and evaluate a VSL strategy using microsimulation modelling. The study includes the selection of the study area, model development, incident simulation, VSL implementation, scenario design, performance measures, model calibration, and sensitivity analysis.

As shown in Figure 1, the selected site is the inbound direction of a 2.7 km segment of Motorway No. 7 (km 19 + 000 to km 21 + 700). This segment comprises five lanes at the beginning, which later merge into four lanes, with each lane measuring 3.6 m in width. Traffic composition includes 93% cars and 7% trucks. In this study, trucks are restricted to the two leftmost lanes. Traffic demand, vehicle speed and vehicle types were provided by the Inter–City Motorway Division (DOH, 2024). The data set covers working days from 15:30 to 18:00 across 12 days, divided into 5-min intervals. The first 30 min (15:30–16:00) is the warm-up period; the analysis begins at 16:00 and continues until 18:00. Note that in Thailand, vehicles are driven on the left side of the road.

The microsimulation model was developed in PTV Vissim, beginning with a link–node diagram representing the network structure. Link attributes, including lane configurations and lengths, were entered to replicate the geometric layout. Desired speeds were assigned based on empirical data, using the 15th and 95th percentiles as lower and upper bounds, respectively.

Traffic volumes and vehicle compositions were input to reflect observed data. Driver behaviour was modelled using Vissim’s Wiedemann 99 car-following model. Parameters CC0, CC1, and CC2 were adjusted for calibration purposes. In this study, the desired range of 2.0 with a desired 95% confidence has been chosen, so the minimum number of simulation repetitions is four (Dowling et al., 2004). The number of repetitions adopted in the study is eight.

Incident scenarios were simulated near the downstream end of the segment. Vissim lacks a native incident feature; therefore, a parking lot object was used to block one or more lanes, simulating real incident conditions. Start and end times of the incident were controlled by way of the parking lot configuration as shown in Figure 2.

Incident data are gathered by requesting information about accidents that have occurred on the motorway over the past five years (2017–2023) in the study area section. The data include details on the duration of each incident and the number of lanes that were closed during each incident. The percentages of lane closures from the leftmost to the rightmost lane are 60%, 6%, 8%, and 26%, respectively. The incident duration averages 41 min, with the 90th percentile reaching 87 min.

VSL control was implemented using the Vissim–Matlab COM interface. Detectors were placed upstream to collect vehicle count, average speed and occupancy every 60 s. These data were transmitted to Matlab, which calculated the new speed limits based on a rule-based VSL algorithm (Sadat and Celikoglu, 2017), then sent commands back to Vissim as shown in Figure 3.

Figure 4 shows the speed limit decision framework. Speed thresholds were established based on the volume corresponding to level of service C capacity and a 30% occupancy rate as the threshold for flow breakdown from free-flow conditions. Separate speed limits were set for cars and trucks, with car speeds set 20 km/h higher than those for trucks. To avoid abrupt speed changes, the speed difference between adjacent VSL zones was limited to 20 km/h (Allaby et al., 2007).

Three main scenarios were simulated.

• Base case – existing condition without VSL.

• VSL with 1 km spacing (VSL1) – two VSL signs placed before the incident, as shown in Figure 5.

• VSL with 2 km spacing (VSL2) – one VSL sign placed 2 km upstream, as shown in Figure 6.

Table 1 illustrates the experimental design comprising 24 simulation scenarios. Each scenario is defined by a combination of VSL strategy (no VSL, VSL with 1 km spacing and VSL with 2 km spacing), number of closed lanes (one or two), closure patterns (indicating which lanes were blocked) and incident durations (41 or 87 min). These close patterns and incident durations are decided based on the data. Dark blocks labelled L1 to L4 represent the lanes 1 to 4 that were closed in each set-up. This design allows a systematic evaluation of how different VSL implementations perform under varying incident severities. Within these, combinations of lane closures, blockage patterns and incident durations resulted in 24 unique simulation scenarios. The incident variables are inspired by the work of Martin et al. (2011). Normally trucks are allowed to use only lanes 1 and 2 from the left, but to maintain traffic flow during certain incident scenarios, trucks were temporarily allowed to use lane 3 (normally restricted) when both lanes 1 and 2 were blocked.

Model outputs were divided into traffic performance and safety indicators. Traffic performance included average delay, average speed, total travel time and vehicle throughput. Safety was assessed using SSAM (Kurker et al., 2014), focusing on average number of stops, rear-end conflicts and lane-change conflicts (Yang et al., 2009; Farrag et al., 2020).

The model was calibrated using field data. The objective was to calibrate the simulation model so that the simulated capacity differed by no more than 5% from the estimated field capacity (Dowling et al., 2004). The motorway capacity was estimated based on the field study (DOH, 2019). The root mean square error (RMSE) for speed is determined as follows.

where $yi$ is the observed mean speed on period $i$⁠,$y^i$ is the estimated mean speed on period $i$ and n is the number of time periods.

The Geoffrey E. Havers (GEH) statistic for flow is determined from the following equation.

where E is the estimated volume and V is the observed volume.

For model calibration, the RMSE has to be less than 10 km/h (Hollander and Liu, 2008) and the GEH statistic for flow has to be less than or equal to 5 (Ahmed et al., 2023).

Calibration parameters (CC0, CC1, CC2) were varied across upstream, merging, and downstream segments (Srisurin, 2020). The final values used were

• Upstream: CC0 = 1.5 m, CC1 = 1.5 s, CC2 = 11 m

• Merging: CC0 = 2.2 m, CC1 = 0.9 s, CC2 = 13 m

• Downstream: CC0 = 1.5 m, CC1 = 1.5 s, CC2 = 12 m.

The final calibration results from plotting the speed flow graph showed a simulated capacity of 6414 vehicles/h, compared to the field-estimated capacity of 6697 vehicles/h, representing a deviation of only 4.2% – well within the acceptable range. Table 2 shows the estimated and observed volume and mean speed for 24 5-min periods from 16:00 to 18:00 on typical weekdays. The observed volume and mean speed were averaged based on 12 weekdays in December 2023, whereas the estimated volume and mean speed were averaged based on eight simulation runs. GEH statistics for traffic volume in the 24 periods were all below 5, and the RMSE for speed was 3.00 km/h, both meeting the specified accuracy targets.

Sensitivity analysis was conducted by varying demand levels to evaluate VSL robustness under different congestion intensities by varying flow at 0.5, 0.75, 1.00 and 1.25 times peak flow. This comprehensive methodology ensures accurate, reliable assessment of VSL strategies under realistic freeway incident conditions using microsimulation and rule-based control logic.

This section presents the simulation results for the 24 scenarios designed to evaluate the performance of VSL strategies under various incident conditions. Key traffic performance measures including average delay, average speed, total travel time and vehicle throughput were analysed alongside safety indicators including average number of stops, number of rear-end conflicts and number of lane-change conflicts. Table 3 shows the simulation results from the lane L1 closure scenarios; Table 4 the lane L4 closure scenarios; Table 5 the lanes L1 and L2 closure scenarios; and Table 6 the lanes L3 and L4 closure scenarios.

The details of each performance measure are explained individually in terms of traffic performance measures and traffic safety measures, as presented in the following sub-sections.

The estimated means and 95% confidence intervals of traffic performance measures – namely, average delay, total travel time, average speed and vehicle throughput – are evaluated. Figures 7, 8 and 9 show that when only one lane (either L1 or L4) is closed, the average delays, total travel times (TTTs), and average speeds of VSL with 1 km spacing (VSL1) and that of VSL with 2 km spacing (VSL2) are significantly better than the existing scenario. For the lane L1 closure scenarios, the estimated average delays, TTTs and average speeds from the VSL1 scenario are better than those of VSL2 across the two closure durations. When two lanes are closed (either L1 and L2 closure or L3 and L4 closure), the estimated average delays, TTTs and average speeds from the existing, VSL1 and VSL2 scenarios are not significantly different – a level of significance of 0.5 – as their 95% confidence interval estimates overlap across the two closure durations. Interestingly, for the two-lane closures, the average speeds of the existing scenario are slightly better than those of VSL1 and VSL2 for the 87-min closure duration. This may be a result of the queue spillback effects combined with the provision of information to drivers, which can further worsen traffic conditions for the 87-min duration of closure and the closure of two adjacent lanes.

In most cases, the vehicle throughputs of the existing, VSL1 and VSL2 scenarios are not significantly different – a significance level of 0.05 – as their 95% confidence interval estimates intersect as shown in Figure 10. However, for two adjacent lane closures and the 87-min closure duration, the vehicle throughput estimates of the existing scenario are significantly better than those of the VSL1 and VSL2 scenarios. This reiterates the detrimental impact of supplying information to drivers.

The estimated means and 95% confidence intervals of traffic safety measures – namely, average number of stops, number of rear-end conflicts and number of lane-change conflicts – are evaluated. Note that the average number of stops relates to the risk of accidents, particularly in scenarios involving sudden braking, merging and potential conflicts with other road users. The number of rear-end and lane-change conflicts relate to the collision risks. Rear-end collisions are a frequent type of traffic accident, particularly in urban areas. These conflicts arise when a lead vehicle slows or stops, and a following vehicle fails to react in time, often due to driver inattention, excessive speed or inadequate following distance. Lane changes disrupt traffic flow (Liu et al., 2023), and unsafe lane changes increase collision risk (Mahajan et al., 2022).

From Figures 11, 12 and 13, for single lane closures (either L1 or L4), the average number of stops, numbers of rear-end conflicts and number of lane-change conflicts of the VSL1 and VSL2 scenarios are significantly better than for the existing scenario. These traffic safety measures are better for VSL1 than for VSL2 for single lane closures across both closure durations. For two adjacent lane closures, the existing, VSL1 and VSL2 scenarios are not significantly different in terms of the traffic safety measures across the two closure durations as their 95% confidence intervals overlap.

The VSL1 strategy outperforms the existing scenario and VSL2 strategy in terms of traffic performance and traffic safety for single-lane closure scenarios across the short and long closure periods. However, the existing scenario, VSL1 strategy and VSL2 strategy are not different for two-lane closure scenarios across the short closure period in terms of traffic performance and traffic safety. Interestingly, under the long closure period, the existing scenario markedly outperforms both VSL1 and VSL2 in terms of vehicle throughput and slightly surpasses them in average speed. This outcome may be attributed to the combined effects of queue spillback and the provision of information to drivers, which can exacerbate congestion during the 87-min closure and the simultaneous shutdown of two adjacent lanes. This suggests that when a two-lane closure occurs and the expected incident-clearance time is prolonged, the VSL system should be deactivated.

To evaluate the VSL strategies, a sensitivity analysis was conducted by varying the traffic flow to 50%, 75% and 125% of the peak volume observed in the field (denoted by 0.5f, 0.75f, and 1.25f). The objective was to assess how deviations from peak flow conditions would impact the VSL1 and VSL2 strategies in terms of traffic performance measures and traffic safety measures during two incident durations and four lane-closure patterns. The percentage of change in a measure for a VSL strategy compared to the existing scenario is calculated as follows

where $MVSLi$ is the value of measure M under the VSL$i$ strategy and $Mexisting$ is the value of measure M under the existing scenario.

For the traffic volumes of 0.5f and 0.75f, VSL1, VSL2 and the existing scenarios are not different in terms of vehicle throughput on all lane-closure patterns and durations. This is because the traffic volumes of 0.5f and 0.75f represent low, uncongested conditions, under which different VSL strategies and the baseline scenarios produce nearly identical vehicle throughputs. VSL1 and VSL2 are not different in terms of the other measures for single-lane and double-lane closures across the two lane-closure durations. The VSL1 and VSL2 strategies perform significantly better than the existing scenarios in all cases, except for a couple instances where their performance is comparable. Specifically, these are (i) VSL2 and the existing scenario in terms of rear-end conflicts when lanes L1 and L2 are closed for a 41-min duration and (ii) VSL1 and the existing scenario in terms of lane-change conflicts when lanes L3 and L4 are closed for an 87-min duration. Across all cases, the maximum VSL benefits observed were up to 48% reduction in average delay, 18% increase in average speed, 15% reduction in TTT, 88% reduction in the number of stops, 88% reduction in rear-end conflicts and 80% reduction in lane-change conflicts. Notably, the magnitude of improvement increased as traffic demand decreased, reinforcing the trends observed under peak flow conditions.

When traffic flow increased by 25%, VSL strategies were no longer significantly better than the existing condition in most cases, and several cases showed worse outcomes. Specifically, at the traffic volume of 1.25f, during both the 41-min and 87-min closure periods, the VSL1 and VSL2 strategies are either outperformed by or perform comparably to the existing scenarios across all traffic performance and safety measures. This suggests that when a two-lane closure occurs in congested traffic conditions, the VSL system should remain inactive regardless of the anticipated duration of incident clearance.

This study simulated the VSL strategy using a variable message sign and detector system that responds to flow, average speed and percentage occupancy in real time through Vissim and Matlab by way of the COM interface. The VSL system was tested on Motorway No. 7, from KM 19 + 000 to KM 21 + 700, during incident conditions. The VSL system aims to improve traffic performance and safety by assigning an appropriate speed limit based on prevailing traffic conditions, allowing vehicles to reduce speed before entering a non-recurring congestion area, thereby decreasing the likelihood of traffic breakdowns and secondary accidents. The study evaluated the effectiveness of the VSL system by comparing its performance to the existing condition, and also between two configurations of VSL coverage: 1 km (VSL1) and 2 km (VSL2).

Under peak volume conditions, the VSL1 strategy outperforms both the existing scenario and the VSL2 strategy in terms of traffic performance – measured by average delay, TTT, average speed and vehicle throughput – as well as traffic safety, reflected in the average number of stops, rear-end conflicts and lane-change conflicts. This advantage holds for single-lane closure scenarios across both short and long closure periods. However, under peak volume conditions, for two-lane closure scenarios during the short closure period, there is no significant difference among the existing scenario, VSL1 and VSL2 in either traffic performance or safety outcomes. Interestingly, during long closure periods, the existing scenario markedly outperforms both VSL1 and VSL2 in vehicle throughput and shows a slight advantage in average speed. This result is likely due to the combined effects of queue spillback and driver information provision, which can intensify congestion when two adjacent lanes are closed for an extended period. These findings suggest that under peak volume conditions, in the case of two-lane closures with prolonged incident-clearance times, the VSL system should be temporarily deactivated to avoid worsening congestion.

The sensitivity analysis results show that for mild to moderate traffic volumes, VSL1 and VSL2 are not different in terms of traffic performance indicators (average speed, TTT and average delay) and traffic safety indicators (average number of stops, number of lane-change conflicts and number of rear-end conflicts) for single-lane and double-lane closures across the two lane-closure durations. The VSL1 and VSL2 strategies perform significantly better than the existing scenarios in almost all cases. The maximum benefits of the VSL implementation included up to a 48% reduction in average delay, an 18% increase in average speed, a 15% decrease in TTT, an 88% reduction in the number of stops, an 88% reduction in rear-end conflicts and an 80% reduction in lane-change conflicts. Notably, the extent of these improvements increased as traffic demand decreased, further reinforcing the patterns observed under peak flow conditions. When traffic demand increased by 25% relative to the peak volume condition, the VSL strategies no longer demonstrated significant advantages over the existing condition in most cases, and in several instances performance even deteriorated. This finding indicates that under congested conditions involving a two-lane closure, the VSL system should remain inactive, regardless of the expected incident-clearance duration.

Future research should consider alternative intelligent transportation strategies or adjust variable message sign placement to address VSL limitations during heavy congestion. Studies should also incorporate various incident durations and locations for broader applicability, and explore environmental and economic impacts to support comprehensive VSL evaluation and implementation.