Evaluation of the Operational Effects of Autonomous and Connected Vehicles through Microsimulation

COM API

The COM script has access to nearly all data inside VISSIM, and can be made visible in a list window. Certain simulation parameters such as acceleration and headway are only readable, whereas parameters such as desired speed and desired lane are both readable and writable. Thus, COM may be used to enable changes to driving behavior and vehicle movement parameters, but may not be utilized to explicitly control a specific vehicle’s immediate actions during runtime. The COM API does not depend on a certain programming language. COM objects can be developed in a wide range of programming and scripting languages.

External Driver Model (EDM)

The EDM is an external driver model compiled in C++ to substitute the VISSIM default driver model. VISSIM passes the current state of a vehicle and its surroundings to a user-defined dynamic link library file (DriverModel.dll), which then computes the “reaction” of the vehicle based on this information. Thus, EDM may be used to specifically control the movements of a given vehicle, unlike COM. VISSIM allows some or all vehicles to be modeled with the DriverModel.dll. The DriverModel.dll may be used to specify driving behaviors based on CAV logic.

DrivingSimulator.dll Interface

This interface allows testing of the interaction between externally controlled vehicles or pedestrians in VISSIM (also known as “human-in-the-loop” simulation). All decisions and calculations of externally controlled vehicles have to be specified, and are based on the reaction to “other” vehicles in the simulation. Each vehicle moves through the VISSIM network based on simulator instructions, that is, steering movements and acceleration. Other vehicles “react” to what the external vehicle is doing.

Data Fetching

COM API has real-time read access to most of the network elements, and write access to many elements. EDM provides data (acceleration, speed, position, etc.) related only to the subject vehicle and neighboring vehicles.

Longitudinal Movement Control

EDM allows users to control longitudinal movement by updating a vehicle’s “desired acceleration” parameter. Through the COM API a user may update a vehicle’s desired speed, which results in a lag in implementation, or update a vehicle location which ignores the acceleration and speed change process, resulting in non-realistic trajectories.

Lateral Movement Control

EDM allows users to completely model lateral movement, whereas the COM API only allows users to set the desired lane where the lane change process is carried out by VISSIM.

Car-Following Mode ( 7 )

The “car-following mode” is executed if the closest leading target with potential influence on behavior is a vehicle (rather than a signal). The acceleration calculation for a vehicle mainly considers two elements: safety constraints and vehicle movement.

Equation 1 calculates the distance the AV should keep from the leading vehicle if the leading vehicle decelerates at its maximum deceleration level:

Δ x n = ( x n − 1 − x n − l n − 1 ) + v n τ + v n − 1 2 2 a n − 1 decc

(1)

where

n and n-1 represent the AV and its leader, respectively;

x n is the location of vehicle n;

l n is the length of vehicle n;

v n is the speed of vehicle n;

τ is the reaction time of vehicle n;

a n decc is the maximum deceleration of vehicle n; and

v limit is the speed limit of the road.

In this study, τ  = 0.1 s, a n decc = − 6 m / s 2 , v limit = 20 mph

Equation 2 provides the safe space headway, considering the minimum of safe distance and sensor detection range (a distance of 300 m was used). It assumes that there is a vehicle at a complete stop outside of the sensors’ detection range, which cannot be detected by the sensors at the time of decision-making:

Δ x = min { Sensor Detection range , Δ x n }

(2)

Equation 3 calculates the maximum safe speed ( v max ) considering both the safe distance and maximum deceleration:

v max = min ( − 2 a n decc Δ x , v limit )

(3)

Equation 4 shows the movement model that considers the acceleration of the leading vehicle, the speed difference between the AV and the leading vehicle, and the difference between the real headway and reference headway S ref :

a n d ( t ) = k a a n − 1 ( t − τ ) + k v ( v n − 1 ( t − τ ) − v n ( t − τ ) ) + k d ( S n ( t − τ ) − S ref )

(4)

where

S n is the spacing;

S ref is the minimum of minimum distance ( S min ), following distance based on the reaction time ( S system ), or safe following distance ( S safe );

a n d is the acceleration of vehicle i; and

k a , k v , and k d are model parameters.

Based on the recommendations of Van Arem et al. ( 8 ), k a = 1 . 0 , k v = 0 . 58 , k d = 0 . 1 . In this study, the minimum distance ( S min ) is set at 2.0 m.

The following equations calculate S system and S safe :

S safe = v n − 1 2 2 ( 1 a n decc − 1 a n − 1 decc )

(5)

S system = v n τ

(6)

S ref = min ( S safe , S system , S min )

(7)

a n ( t ) = min ( a n d ( t ) , k ( v max − v n ( t ) , a comf )

(8)

where

k is a model parameter; and

a comf is the comfortable acceleration level.

In this study, k = 1 . 0 , and a comf is assumed to be 2 . 5 m / s 2 .

Deceleration Mode

The deceleration mode is executed when the closest object to a vehicle is a signal head with a red or yellow indication, and the vehicle is approaching it.

Equation 9 calculates the distance to stop S d , that is, the distance required for the vehicle to come to a complete stop if it decelerates at a comfortable deceleration a comf decc :

S d = 1 2 V n ( t ) 2 a comf decc

(9)

where

S d is the acceleration distance;

a comf decc is the comfortable deceleration rate assumed to be − 3 . 5 m / s 2 . a n ( t ) if the distance to signal is less than the distance to stop S d .

VISSIM Control Mode

When neither of the two aforementioned situations occur (i.e., the vehicle is influenced neither by a leading vehicle nor by a signal) and the vehicle is making a right or left turn, “gap acceptance models” are required. The authors did not develop a gap acceptance model for this paper; therefore, the AV driver behavior control is temporarily given back to VISSIM to navigate the turn. In this case, the vehicle uses the VISSIM suggested acceleration until it exits the connector.

Free Driving Mode

When the vehicle is not affected by a leading vehicle, a leading signal, or turning, then the “free driving” mode is implemented. In this case the vehicle will aim to maintain or reach the desired speed. Acceleration is estimated as:

a n ( t ) = k ( v desired − v n ( t ) )

(10)

where k is the desired change rate, assumed to be 1 in this study.

This acceleration value is subject to VISSIM’s vehicle dynamics model and its constraints.

Acceleration Implementation through EDM

At each time step, the VISSIM suggested acceleration is stored in the variable “desired_acceleration,” through this command:

A function is written to check whether a vehicle ID is contained in the first column of the acceleration “.txt” file. If yes, then the vehicle is an AV, so it will use the acceleration from the.txt file. If no, then it is a CNV and it will use the VISSIM suggested acceleration (Figure 2 left part).

Travel Time Comparative Analysis

Before presenting CO₂ emissions, travel time results are briefly discussed as these provide for interesting comparison with the emissions. The travel time results are aggregated over five replicate runs. Figure 7, a and b, reflect the EB travel time for v/c’s of 0.7 and 0.9. Both plots show a trend of decreasing mean, median, and 75th percentile travel time with increasing AV, CV, and CAV penetration rates. These trends are more prominent in the AV and CAV results than the CV results. In addition, the travel time interquartile range (IQR), depicted by the vertical length of the boxes in the plots, generally decreases with increasing penetration rates. These trends are somewhat more pronounced under the higher demand level (i.e., v/c = 0.9). Therefore, for all AV/CV/CAV penetration levels, improved travel times and reduced variability are seen at higher penetration rates and volume levels. These results agree well with the previously presented speed and delay findings (Figure 6).

CO₂ Emission (MOVES Estimate) Comparative Analysis

CO₂ results are next presented in a similar format to that of the travel time. To allow for a comparison between routes and scenarios, CO₂ emissions are expressed as emissions per unit distance (in grams per foot) for each individual vehicle.

As seen in Figure 8, the mean and median emissions show a slight reduction with increasing AV penetration. However, mean and median emissions show an increase with increasing CV penetration, while for the CAV cases an increasing penetration rate results in increasing average but decreasing median CO₂ emissions. The emissions IQR also increases with CV and CAV penetration. AV demonstrated lower emissions than CV and CAV at all penetration rates. These trends are consistent at both demand levels.

Based on the speed, delay, and travel time results, it would appear that vehicles should have more efficient route traversal with increasing AV, CV, and CAV penetration. At first glance this would intuitively imply lower emissions with increasing penetration rate, assuming the improved KPIs are a result of fewer stops and acceleration/deceleration events. However, the anticipated emissions improvements were not seen. The next section will further explore these findings.

Emission Analysis: VSP Operating Mode Bin Histogram

Figure 9, a–d, shows the VSP-bin density histograms for CV penetration rates of 0%, 20%, 60%, and 100%, respectively, for v/c = 0.9, simulation random seed = 1, on the EB route. Since the vehicle speed for the network never exceeded 25 mph, no data points were observed beyond VSP-bin 16 (Table A.1 in Xu et al. [ 16 ]). By comparing the plots, it becomes evident that with increasing CV penetration, more data points appear in bin 12 (cruising/acceleration) and fewer data points fall in bins 0 (deceleration), 1 (idle), and 11 (coasting). The CV algorithm has achieved its stated goal; that is, maximizing arrival-on-green, which is seen in the reduction in idle time. However, there is an increase in data points in bin 12, which has a higher emission value than bins 0, 1, and 11, resulting in more overall emissions. It is hypothesized that in the implemented CV logic, the vehicles (CV and non-CV following, as seen in Figure 4) are undergoing more cycles of acceleration and deceleration by undertaking frequent adjustments to their speed to maximize the likelihood of arrival-on-green. This increasing number of changes in acceleration results in higher emissions. However, in the real world there are certain limitations on how fast an acceleration change can be propagated through the drivetrain as well as limitations on the frequency and magnitude of acceleration reversals that are acceptable for driver comfort. It is therefore expected that a successful CV algorithm would smooth out the acceleration profile.

OPEN IN VIEWER

Figure 9.

Variations in VSP-bin distribution for eastbound (EB) through vehicles with increasing penetration in CV for v/c = 0.9, simulation random seed = 1: (a) 0% penetration (top left), (b) 20% penetration (top right), (c) 60% penetration (bottom left), and (d) 100% penetration (bottom right).

Conventional Vehicle (CNV) Following Behavior

When the lead vehicle is a CV and the follower a CNV, the CNV does not have the same information as the CV. Therefore, instead of slowing down to a lower cruising speed, as the CV does (see Figure 4b), the CNV aims to achieve its desired speed and enters a sharp oscillation pattern around the lower cruising speed of the leader-CV. These oscillations probably contribute to increased emissions, although it is likely this is an artifact of the low speed following behavior rather than expected real-world behavior. That is, it is unlikely that a human-driven vehicle would initiate this rate of acceleration and deceleration to assiduously maintain its desired speed, with most drivers tending to a smoother speed profile. However, this supposition should be further explored with field data.

VISSIM’s CV Movement Logic

When either the leader or the follower is a CV which enters the communication range when the signal is red, VISSIM’s CV logic utilizes a parameter called “SpeedMaxForGreenStart,” which is the maximum speed to arrive at the next green start (details of these calculations are provided in the model implementation section). The calculation of this parameter depends on the green time remaining. While our simulation runs at 0.1 s frequency, the COM API updates this parameter (internally) every 1 s by default (regardless of simulation frequency set). This leads to an approximation of “green time remaining” to the nearest second. This results in “spikes” every half a second (because of rounding) and results in the oscillation pattern shown in Figure 10a. This occurs only when the CV arrives during the red signal. The oscillation can be alleviated by calculating the “SpeedMaxForGreenStart” parameter manually according to the chosen simulation frequency as shown in Figure 10b. Using the VISSIM CV code an acceleration and deceleration profile is again seen that is unlikely in a human-driven vehicle or an AV that has human passengers where such an oscillation could result in significant discomfort. However, for the purposes of demonstration we chose to use VISSIM’s CV code as is, and this may be one of the contributing factors to increased emissions with CVs.

OPEN IN VIEWER

Figure 10.

Optimal speed provided by VISSIM and manually estimated: (a) acceleration of leader-CV and follower-CV with 30 s red remaining (VISSIM’s CV logic) and (b) comparison of VISSIM CV logic and manually estimated speed.

Artifact of Binning

Finally, it is possible that a portion of the results may be an artifact of the binning process used in the MOVES approach. The emissions may be being overestimated; if the vehicle operations are occurring at the boundaries of bins 1, 11, and 12, for example, a minor change in vehicle operations is shifting the emissions calculation from bin 1 to bin 12. Future efforts will explore this potential impact.