Content uploaded by Xiao-Yun Lu
Author content
All content in this area was uploaded by Xiao-Yun Lu on Mar 08, 2015
Content may be subject to copyright.
A Novel Speed-Measurement Based Variable Speed Limit/Advisory
Algorithm for a Freeway Corridor with Multiple Bottlenecks
by
Xiao-Yun Lu*
California PATH, University of California, Berkeley
1357 S. 46th Street, Richmond Field Station, Bldg 452, Richmond, CA 94804-4648
Tel: 1-510-665 3644, Fax: 1-510-665 3691
Email: xiaoyun.lu@berkeley.edu
Steven E. Shladover
California PATH, University of California, Berkeley
1357 S. 46th Street, Richmond Field Station, Bldg 452, Richmond, CA 94804-4648
Tel: 1-510-665 3514, Fax: 1-510-665 3537
Email: sess@berkeley.edu
Iman Jawad
Transportation Solutions and Technology Applications Division
Leidos (formerly SAIC), FHWA
11251 Roger Bacon Drive, Reston, VA 20190
Tel: 202.277.1032 email: iman.s.jawad@leidos.com
Ram Jagannathan
Transportation Solutions and Technology Applications Division
Leidos, FHWA
11251 Roger Bacon Drive, Reston, VA 20190
mobile: 703-303-7134 email: ram.j@leidos.com
Thomas Phillips
Transportation Solutions and Technology Applications Division
Leidos, FHWA
11251 Roger Bacon Drive, Reston, VA 20190
Tel: 703 973-8711 email: thomas.h.phillips@leidos.com
For presentation and publication
at the 94th Annual Meeting
Transportation Research Board
Washington, D.C.
January 2015
# Words: 5499
Plus 8 Figures and 1 table x 250 (2000)
Total: 7499
*Corresponding author
Abstract
Most previous variable speed limit/advisory (VSL/VSA) algorithms for bottleneck flow
maximization were local and occupancy measurement-based. Some algorithms for freeway
networks with multiple bottlenecks were proposed based on the second order METANET model,
but these were rather complicated. This paper proposes a novel simple VSL/VSA control
algorithm for bottleneck flow improvement based on speed measurements, and a strategy to
expand to a freeway network with multiple bottlenecks using a distance-based interpolation. The
algorithm has been validated with a well-calibrated traffic network model for I-66 inside the
Capital Beltway. Simulation results showed that the overall system performance was improved
even with a 10% compliance rate, and the improvement was not sensitive to the compliance rate
if the latter is greater than 10%. The performance parameters include: total travel time, total
travel distance, total number of stops, average speed variation, total delays, and flow at the
bottlenecks. The VSA/VSL can be implemented using infrastructure-to-vehicle communication
to a display or to determine the set speed for an adaptive cruise control system or using variable
message signs on the roadside. The algorithm has been conceptually tested with three vehicles
that used the VSL as set speed for their infrastructure-to-vehicle (I2V) Cooperative ACC
systems.
3
INTRODUCTION
This paper presents a Variable Speed Limit/Advisory (VSL/VSA) strategy, including Speed
Harmonization (SH) as a special case for a freeway network with multiple bottlenecks. The
objective for the algorithm is to use feedback control to improve traffic flow at bottlenecks,
particularly the most downstream bottlenecks, and eventually to achieve system-wide
performance improvement. In contrast to previous work based on first-order or second-order
traffic models, the proposed approach is simply based on measured average speed in bottleneck
section(s).
The VSA is designed for in-vehicle implementation, and for display to all drivers using
roadside Changeable Message Signs (CMS). The objectives include not only speed
harmonization, but also throughput improvement and shockwave reduction. The minimum
requirement is to have one CMS at the I-66 bottleneck at N. Sycamore St. More CMS upstream
would further benefit overall system performance since advisory speeds are generated for each
section of the system.
For the overall system with geographic scope shown in Figure 1, there were two major
bottlenecks identified, i.e. the major upstream bottleneck at the merge section of I-66 and VA267
and the downstream bottleneck at the Sycamore Street onramp (red circles in Figure 1).
Figure 1. Geographic scope of the study area and typical traffic situation in PM peak hours
The main contributions of this paper include:
Propose a new bottleneck flow maximization algorithm using feedback control based on
measured speed within the bottleneck
4
Propose a strategy to extend this approach to freeway networks with multiple bottlenecks
Conduct simulations in Aimsun for different compliance levels from 5% to 100%.
Results showed overall system performance improvement in TTT (Total Travel Time),
TD (Total Delay), TTD (Total Travel Distance), speed variation, TNOS (Total Number
of Stops) and downstream bottleneck throughput. The network model represents I-66
inside the Beltway developed before [1].
Using VSL/VSA for traffic throughput improvement should have the following functions:
Maximize or improve bottleneck flow
Traffic in the section between bottlenecks should be smoothed (harmonized).
This is the way to achieve both throughput and safety improvements. Other benefits such as
emissions and fuel consumption reduction are natural by-products of the strategy since they are
consistent with overall system traffic smoothness and throughput improvements.
Although occupancy and flow are basic output measures from loop detectors, roadside or
gantry-mounted radar sensors such as those used on the highway section that was evaluated in
this project measure speed better. This algorithm is intended for use with those sensors.
Furthermore, speed estimation using single loop event data [1, 3] or with dual loop event data [4]
can be reasonably accurate if properly processed. Therefore, the algorithm can also be applied in
such situations.
The paper is organized as follows: Section 2 is a brief literature review of closely related
algorithms and the most recent studies in this area; Section 3 presents the measured-speed based
algorithm for bottleneck flow maximization, which is a proportional feedback control; Section 4
is devoted to performance evaluation and validation through simulation; Section 5 is for field test
and data analysis; the last section is concluding remarks.
LITERATURE REVIEW
An extensive literature review on Variable Speed Limit/Advisory (VSL/VSA) has been
reported in [5], which includes theory, algorithms, simulations and field implementations in
several countries. The following covers the most relevant and most recent studies in this area.
The previous work [6] proposed a strategy for throughput maximization of bottlenecks
caused by lane reductions (or virtual lane reductions due to weaving). The basic idea is to block
traffic upstream of the bottleneck, and provide feeding flow to the bottleneck close to its capacity
flow (highest flow possible). This can be achieved with Critical VSL (Variable Speed Limit) or
VSA (Variable Speed Advisory) in the section immediately upstream of the bottleneck [6].
Usually, if there is an onramp in the section downstream of the critical VSL, then the onramp
needs to be metered if its demand is high. Otherwise, potential heavy flow from the onramp
would destroy the function of Critical VSL/VSA control immediately upstream. However, in our
case, although the Sycamore St. onramp is right in the bottleneck section, its demand is
reasonably low, as reported before. Therefore, the proposed feedback control algorithm can still
function properly.
5
The paper [7] evaluated the combination of VSL and ramp metering. The VSL strategy was
the modified logic tree VSL algorithm presented in a companion paper [8]. The ramp metering
was local responsive ALINEA (Asservissement LINéaire d’Entrée Autoroutière) and
heuristically coordinated HERO (HEuristic Ramp metering coOrdination) implemented in
Australia. The combined approach was evaluated over a calibrated microscopic simulation in
Aimsun for a critical section of the Auckland Motorway in New Zealand. Simulation results
showed that the combined HERO and Modified Logic Tree VSL control scenario outperformed
other evaluated controls in terms of the total travel time (TTT) and equity measured by the Gini
coefficient.
The work in [8] reported Aimsun simulation results for a modified logic tree VSL algorithm
based on previous work with this approach. The logic tree algorithm was basically a Fuzzy Logic
feedback control based on measurements instead of models. The main modification in this paper
was to refine flow thresholds and to increase occupancy threshold. The VSL was used to
improve bottleneck throughput. A microscopic simulation model was calibrated with GEH flow
criteria for a critical section of the Auckland Motorway. Simulation results indicated some
positive effects in bottleneck flow improvement.
Van de Weg et al. [9] presented some recent improvement on the authors’ previous work on
the moving jam reduction using combined ramp metering and VSL, by introducing vehicle
cooperation with vehicle-to-infrastructure (V2I) communication. Thanks to the V2I, VSL control
could potentially be executed at a microscopic level to specific vehicles, regarding when and
where to adopt what speed. It shows through simulation that a moving jam near a metered on-
ramp can be resolved.
The paper by Hadiuzzaman et al [10] investigated VSL freeway improvements in both travel
time and link throughput during congestion periods using Model Predictive Control (MPC). The
model adopted is the 2nd order model with some extra constraints and without a fundamental
diagram assumption. The authors intended to improve performance through: proper selection of
objective function, predictive horizon, control horizon and its parameters.
Work [11] quantitatively investigated the impact of driver compliance rate on efficiency and
performance of an optimal VSL approach. The investigation used MPC based, on an improved
2nd order model without Fundamental Diagram assumption. The simulation evaluation showed
that both mobility and safety benefits of VSL are positively correlated with driver compliance
rate. Specifically, the travel time, throughput, and collision probability are improved in the range
of 5- 15%, 6-8%, and 50-60%. This result is consistent with expectations.
The work in [12] studied the transferability of a self-learning VSL strategy, which is an
interesting fundamental problem. The authors argued that VSL in urban areas should be for
mobility and VSL in rural areas should be for safety, which makes sense due to traffic load.
Traffic congestion may have different causes, so the chosen VSL strategy should take into
account those differences such as: high onramp traffic demand, road geometry (lane drop or
freeway merge/split/weaving), off-ramp-spillback, different vehicle types, and weather
conditions etc.
6
SPEED-BASED BOTTLENECK FLOW MAXIMIZATION
Instead of using occupancy based feedback control, we propose a speed-based feedback. It
can be stated as:
The speed in the bottleneck
m
vk
(assuming section m is a bottleneck, indexed from
most downstream to the most upstream as we did before in [1]) is measured with fixed
sensors and filtered with a low-pass filter to be suggested later (to smooth; the objective
is to reduce speed variation);
Then the VSL/VSA at the bottleneck section is determined by
1.1,1.5 ; default value: 1.3
mmm
mm
uk vk
(Eq. 1)
Then the VSL/VSA at the immediate upstream section relative to the bottleneck is
determined by
1
, if
, if
0.7,0.9 ; default value: 0.8
free m sw
mmm m sw
mm
VokO
uk vk ok O
(Eq. 2)
f
ree
V- free-flow speed;
m
ok
- measured occupancy in bottleneck section;
s
w
O- the switch
threshold of occupancy close to the capacity flow (suggested value for
s
w
O: 10.0~12.5%). The
algorithm is intended to delay traffic breakdown and for throughput improvement for heavy
traffic. The threshold for occupancy for field implementation will need to be tuned based on
sensor characteristics.
Although the algorithm looks very simple, its function realizes the control philosophy we
stated before and also in [5]. The idea is explained as follows:
This is a feedback control based on sensor-measured bottleneck speed;
(Eq. 2) says that: if speed in the bottleneck decreases, the VSL/VSA of its immediately
upstream section will be proportionally reduced, which is equivalent to saying that
feeding flow to the bottleneck will be reduced;
With this limit to its upstream section, the speed in the bottleneck is encouraged to
increase as by (Eq. 1) since the suggested speed is higher than measured speed from this
point and downstream; it is a proportional control since the desired speed is proportional
to the measured speed;
(Eq. 2) also implies that: if measured speed in the bottleneck increases, then the
corresponding VSL/VSA in the immediate upstream section will be proportionally
increased to continue to feed the bottleneck to its maximum capacity;
7
It is noted that this approach is measurement-based instead of model-based as previous works
using the METANET model [9]. Therefore, this approach cannot claim system optimization, but
emphasizes simplicity and practical field implementation.
Low Pass Filter
Traffic speed is measured in the bottleneck as
m
vk
(assuming section m is a bottleneck,
indexed from most downstream to the most upstream as indicated in Figure 2) with fixed
sensors and filtered with a low-pass filter (to smooth; the objective is to reduce speed variation).
In practical simulation, the following 3rd order elliptic filter has been used:
11 2 3
21 2 3
31 2 3
0.6688 ( -1) 0.0000 ( -1) 0.0000 ( -1) 0.0445 _ ;
0.0314 ( -1) 0.9993 ( -1) - 0.0378 ( -1) 0.0008 _ ;
0.0006 ( -1) 0.0378 ( -1) 0.9993 ( -1) 0.0000 _
x
kxkxkxkindata
x
k x k x k x k in data
x
k x k x k x k in data
12 3
;
_ 6.2117 - 0.0057 0.0007 0.1656 _ ;y out x k x k x k in data
(Eq. 3)
where in_data is the speed measurement from the sensor, and
123
,,
x
xx are filter states. This
low-pass filter has been designed in Matlab to serve two purposes: to filter the noise and to
smooth the signal. The filter frequency range was selected to be suitable for traffic detection
characteristics and to not cause excessive time delay in the output signal. Those factors are
traded off in the filter design.
VSL/VSA ALGORITHM FOR FREEWAY NETWORK
The VSL/VSA strategy for freeway network traffic improvement is stated as:
Define the freeway network to include coupled recurrent bottlenecks, which means that
traffic behaviors of those bottlenecks would affect each other significantly;
Traffic at the most upstream and most downstream ends of the freeway network are free
flow; therefore the VSL/VSA can be computed; it is clear that this can always be
achieved if the system scope is large enough;
VSL/VSA at the bottlenecks section and their immediately upstream can be determined
by (Eq.1 and Eq. 2);
VSL/VSA at other sections in between the bottlenecks and the most upstream and
downstream sections can be determined by distance-based interpolation.
The justifications for this approach are as follows:
(1) Bottleneck flow determines overall system performance. Since the algorithm is intended
to operate the bottleneck at its capacity flow, therefore, each bottleneck intends to push traffic
forward. The question is: is the downstream bottleneck able to receive it? This will be analyzed
as follows:
8
If
upstream bottleneck flow downstream bottleneck flow
ud
FF: this can only last
for a certain period of time since the section between the two bottlenecks will be filled up; in
this case, the downstream bottleneck can still be operated such that
downstream bottleneck capacity flow
dd
c
FF;
If ud
FF: this can only last for certain period of time, and then udd
c
F
FF
which is
what one can do; this situation is due to either upstream bottleneck demand flow is lower, or
(upstream bottleneck capacity flow)
ud
cc
FF;
If ud
FF: this is the ideal case for operation;
In summary: in all situations, application of the strategy for a freeway network could potentially
improve overall system performance; however, when a section between two bottlenecks used as
storage fills up, it might block an off-ramp which may become a disadvantage, further analysis
will be necessary to optimize and balance traffic storage to maximize all output flows.
(2) Distance based interpolation of VSA/VSL at the bottlenecks has the following
advantages:
Traffic will be smoothed in each section (speed harmonization)
VSA/VSL can be discretized for (constant in) each section for feedback to the driver
using Changeable Message Sign, or can be continuous to be used as set-speed for vehicle
with I2V communication and ACC, providing I2V CACC capabilities
(3) If the freeway has non-recurrent bottleneck(s), caused by incidents that can be detected,
the algorithm can also be reconfigured for such situation
The I-66 traffic network inside the Capital Beltway is used as an example to explain the
algorithm. The overall system in consideration is divided into 4 links as shown in Figure 2.
The stretch upstream of the merge on I-66
The stretch upstream of the merge on VA-267
The stretch between the merge and N. Sycamore St
The stretch between N. Sycamore St and George Mason Drive
Each link is divided into several cells based on road geometry, road length and sensor locations.
Links and cells are numbered from downstream to upstream since the advisory speed for each
cell is determined in this sequence.
9
Figure 2. Road partitioned into 4 links; each link is divided into several sections; with constant VSA value within each cell; and
sensor (trailer) locations indicated; red spots indicating two bottlenecks; green circles mean free-flow (boundary) sections of the
system
10
u[5] corresponds to the downstream bottleneck at Sycamore St. u[11]corresponds to the
upstream bottleneck at the I-66 & VA267 merge. Its upstream section VSL is critical, which
should be reduced. Meanwhile, there are two upstream sections to the bottleneck at the I-66 &
VA267 merge, i.e. u[12] and u[16]. Their speeds should be reduced if u[11] is congested;
Now the distance based interpolation is described as follows:
55
65
, if
1.3 , if
, if
0.8 , if
f
ree m c
mc
f
ree m c
mc
VokO
uk vk ok O
VokO
uk vk ok O
(Eq. 4)
where the feedback gains 1.3 (for speed increasing) and 0.8 (for speed decreasing) have been
selected empirically based on results of multiple simulations. Those gain values are related to
sensor detection characteristics, the real-time traffic state estimation algorithm, and estimation
accuracy and delay, and they will need to be tuned in field implementation.
Then
234
,,ukukuk are determined by distance interpolation between
5
uk and
1
uk
. The speed measurement point should be chosen at 15% of the total section length, from
upstream edge to downstream edge.
Similarly, one can determine
12
uk
and
16
uk
based on the speed measurement
11
vk
.
The VSL/VSA in all other sections can be determined by distance-based interpolation.
MODIFICATION OF VSL/VSA FOR DRIVER ACCEPTANCE
The designed VSL/VSA value could be used for feedback to the driver through variable
message sign or in-vehicle display if wireless communication is available. It can also be used as
the set-speed of CACC vehicles with I2V data communications. It is noted that the designed
VSL/VSA values for each road section may not be multiples of 5 mph. This number could be
directly used as the set speed for ACC (Adaptive Cruise Control) and CACC (Cooperative
Adaptive Cruise Control) vehicles. However, for driver acceptance of roadside or in-vehicle
displays, this number will need to be modified. Considering the fact that most drivers intend to
drive at least 5 mph faster than the roadside posted speed limit, the following modification is
suggested. It is assumed that
m
uk
only has its integer part with the decimal part ignored. The
displayed VSL value
m
Uk
is the maximum of the multiple of 5 mph strictly smaller than
m
uk
. This can be expressed as:
11
5
0mod5
mm
mm
m
Uk uk
Uk uk
Uk
(Eq. 5)
PERFORMANCE EVALUATION THROUGH SIMULATION
System Performance Parameters in Aimsun
The following four system performance parameters have been used for evaluation of the
algorithm implemented in the simulation model:
Total Travel Time (TTT): this includes the queue time at entrance ramp;
Total Travel Distance (TTD): indicates the traffic accommodated by the network, which
should be encouraged to reduce/avoid traffic spill-back to arterial(s) and surface street(s),
particularly in peak hours; however, TTD and TTT may need to be traded off when
system demand is high;
Total Delay (TD): it is obtained by comparing the simulated traffic with mainline free-
flow assumption;
Speed Variation: it directly reflects the fluctuations of system-wide speed;
Total Number of Stops (TNOS): In Aimsun, the total number of vehicle stops is recorded
and used as a system performance parameter to indicate traffic smoothness. As is
generally recognized, stop & go traffic will significantly affect traffic throughput and
energy use as well as safety, since a significant portion of collisions happen due to
resulting shockwaves;
Flow (throughput) changes at bottlenecks: obviously, the most downstream bottleneck
flow improvement is the critical factor for overall system performance improvement –
since it means that the system dumps traffic faster (provided that upstream off-ramp
flows are not reduced significantly).
In Aimsun microscopic simulation, different random seeds will generate slightly different
demand flows from the entrance ramps and freeway mainline. It is clear that higher total demand
in the simulation time interval could result in longer TTT. To overcome this problem, the
following correction has been conducted in the estimation of the percentage time improvement
for TTT, TD, and TNOS.
12
% Change =
% Change =
% Change =
default vsa default vsa
default default
default vsa default vsa
default default
default vsa default vsa
default def
TTT TTT TTD TTD
TTT TTT TTD
TD TD TTD TTD
TD TD TTD
TNOS TNOS TTD TTD
TNOS TNOS TTD
ault
(Eq. 6)
It says that if the TTD is reduced (increased) by x% after the simulation run, then TTT and
TD should be increased (decreased) by x% as penalty accordingly. Here the subscripts indicate
whether the parameter is estimated from the default scenario or from the scenario with VSL/VSA
activated.
Simulation Results
The proposed algorithm has been simulated for three model days: Mar 12, Mar 13 and Mar 14,
2013 with compliance rates: 10%, 25% and 50%. Each model date has been simulated for 10
replications (random seeds) to get their mean. Performance parameter changes were obtained by
comparing controlled cases and the baseline case without control. The VSL/VSA for each
section was also plotted and compared with measured section speed and fixed location sensor
measured speed. Control gain and other parameter values selection in simulation were as follows:
51.3
, 60.8
, 11 1.3
, and 12 16 0.85
.
Figure 3 shows section-wise VSA/VSL compared with measured section speed and point
sensor speed, in an Aimsun model calibrated with March 12 data. Section speed is the average
speed (over space) of all vehicles in that section, and point sensor speed is the average speed
over time at the sensor location. Section indices are indicated in Figure 2. The trends are very
similar for the models calibrated using data from other dates..
13
Figure 3. Comparison of VSL/VSA and measured section speed and point sensor speed; Compliance level: 10%
14
Table 1 System Performance Parameter Change with VSL/VSA with 3 Compliance levels
Model
Date Comp.
rate TTT [%] TTD
[%] TD[%] Spd Var
[%] Ave # of
Stops [%]
Flow@
Syc.
[%]
Flow@
Merge
[%]
Mar 12 0.1 -8.94 0.98 -10.45 -10.76 -0.24 2.30 -0.60
Mar 13 0.1 -7.47 0.86 -8.30 -6.83 0.77 1.99 -1.27
Mar 14 0.1 -4.00 0.78 -3.14 -6.32 0.89 1.91 -1.09
Mean 0.1 -6.80 0.87 -7.3 -7.97 0.47 2.07 -1.0
Mar 12 0.25 -12.47 1.00 -17.25 -10.32 -0.85 2.21 -0.60
Mar 13 0.25 -8.65 0.99 -9.17 -7.92 -0.79 2.18 -1.01
Mar 14 0.25 -6.54 0.94 -5.58 -7.64 0.34 2.25 -0.68
Mean 0.25 -9.22 0.98 -10.67 -8.63 -0.43 2.21 -0.76
Mar 12 0.5 -13.73 1.24 -18.59 -9.55 -1.93 2.01 -0.81
Mar 13 0.5 -8.10 1.41 -8.75 -6.82 0.15 2.61 -0.55
Mar 14 0.5 -7.98 1.49 -7.70 -7.81 -1.55 2.91 -0.23
Mean 0.5
-9.94 1.38 -11.68 -8.06 -1.11 2.51 -0.53
It can be observed from Table 1 that most performance parameters improve modestly as the
compliance rate increases. In fact, as compliance rate increases further to 50% and 75%, there is
not much further improvement in performance. This means that the results are relatively
insensitive to the compliance rate, which can be explained as follows: the traffic in PM peak
hours is very congested in that corridor; therefore, making lane changes to take advantage of a
higher speed lane could be very difficult; as a consequence, if 10% of drivers comply with the
VSL/VSA, most other drivers have to follow.
It is noted from Table 1 that the flows at the upstream bottleneck at the I-66 and VA267
merge section were reduced for all three model dates. This is explained as follows: To operate
the downstream bottleneck at its capacity flow, more traffic has to be stored in its upstream,
which is the cause of the flow drop at the upstream bottleneck. The flow drop at the upstream
bottleneck is smaller than the flow increase at the downstream one. Besides, the overall system
performance is eventually determined by the downstream bottleneck throughput. This is why the
overall system performance still improved.
FIELD TEST WITH I2V CACC VEHICLES
A conceptual field test of the algorithm has been conducted. The objective was to use three
test vehicles driving in a congested traffic stream with the test vehicles’ set speeds updated
periodically based on real-time traffic conditions measured along the field experiment highway
segment. The set speed values changed continuously over distance along the route and speed
15
profiles are updated no more than once per 30 seconds. The influence of the test vehicles on the
vehicles immediately behind them was analyzed using probe vehicles in the trailing traffic
stream to find out how far upstream the traffic could be impacted by the test vehicles. This will
be useful to determine the minimum number of vehicles necessary to generate the desired effect
on overall traffic. This will lay the foundation for future larger scale testing with multiple
equipped vehicles.
This test was conducted to (a) initially set up the overall system; (b) test real-time
functionality of each component in the overall system; (c) determine if the VSA/VSL value
determined by the server is reasonable; (d) measure the difference between VSA/VSL and the
actual speed of the controlled vehicles; (e) determine how probe vehicles ahead of and behind
the test vehicles can be used for overall VSL/VSA design and refinement.
However, such small conceptual tests cannot be expected to generate measurable effects on
the overall system traffic.
System Setup and Integration
The three test vehicles with Infrastructure-to-Vehicle (I2V) Cooperative Adaptive Cruise
Control (CACC) capability can be driven fully manually or using the automatic speed control
from the ACC system, with set speed provided by the roadside server. The driver still provides
steering input and can cancel the I2V CACC functionality at any time with a simple button. The
I2V CACC mode of operation is expected to provide more accurate tracking of the set speeds
and easier driving by the test drivers, so they can focus their attention on observing local traffic
conditions.
System integration of the speed harmonization field experiment consists of four primary
components: connected mobile traffic sensing system trailers, Cadillac research vehicles, probe
vehicles, and the Saxton Lab server (Figure 4). There are four trailers being used in the
experiment, each equipped with RTMS cross-fire radar traffic detection systems.
16
Figure 4. System Integration Diagram
System Sensor Location and Test Procedures
Figure 5 (upper) shows the road geometry, sensor locations and freeway section range where
the I2V CACC was activated. Four RTMS sensors were used for roadside detection, which
provide lane-wise traffic speed, occupancy, flow, and density. The test vehicles were starting
from VA7, and exiting the freeway at Fairfax Drive.
17
Figure 5. Field Conceptual Test: Sensor locations (4 RTMS trailers), and section division
Relative positions of the six vehicles used in the test are shown in Figure 6. The distance
between each CACC vehicle and the probe vehicle behind was about 50~100 m. The reasons for
using the probe vehicles in the front and rear of the CACC vehicles are:
Since the RTMS sensors are sparsely located, dynamic traffic characteristics between
sensors cannot be detected; therefore, it is expected that the front probe vehicle could
provide more detailed and accurate traffic information (basically speed) to the central
sever, which could be used for refining the VSL/VSA calculation; obviously, availability
of such information will be proportional to the market penetration of V2I systems in the
long term;
Probe vehicles behind the test vehicles are mainly for understanding how the I2V CACC
would affect the traffic behind, which will be used to determine the minimum equipped
vehicle density that could adequately influence the overall traffic;
The three test vehicles put side-by-side may not be the best option if the market
penetration of CACC vehicles is low since it will slow all lanes down to the minimum
speed based on the most congested lane up ahead, even if that speed is lower than the
VSL selected speed, which may hinder traffic flow in practice;
18
Figure 6. Relative positions of the test vehicles
Preliminary Test Data Observations
Test runs were conducted on the following dates: 6/26, 7/1, 7/8, 7/22, 7/24, and 7/30 of 2014,
with one run per day during the evening peak period while congestion was building. Figure 7
(upper) is from the data on vehicle C3 on 7/24, which is used as a representative of all six runs
during the time when traffic flow is just starting to break down. '
i
Ts
in the figure are time points
used to explain the speed changes with seven vertical lines. The legend of those curves is on the
right. From the figure, we have the following observations:
The VSL/VSA target speed value, represented by the blue dots labeled as “C3
Recommended Speed”, starts from the measured speed at Trailer 3 (green, upstream),
then Trailer 4, and then gradually decreases toward the downstream speed at Trailer 5
(pink), which is the expected trend for VSL/VSA to reduce shockwaves at the moving
end of lower-speed traffic downstream;
The VSL/VSA target speed closely followed the speed measured at Trailer 6, the
downstream bottleneck speed, which is the expected trend for the algorithm;
19
Figure 7. Test Run 1 for ACC vehicle C3 on 7/24/14; Upper: speed trajectories of C3 and RTMS;
Lower: locations of all 5 test vehicles – Mile-Post vs time
The CACC vehicle followed a complicated speed trajectory, which can be explained as
follows in Figure 7:
o The Cadillac ACC design did not require the vehicle speed to be strictly limited
by the set-speed; if the front vehicle speed was higher or there was an adequate
clear space in front of the vehicle, the ACC vehicle speed could be somewhat
higher (in the time period
34
, TT) as the vehicle coasted without actively
braking;
o If the vehicle speed is 15 mph lower than the set-speed, the ACC generates a
significant acceleration, leading to over-shoot (as in the first part of time period
12
, TT);
o If the vehicle speed is 15 mph higher than the set-speed, braking is activated,
leading to a deceleration which could produce an under-shoot (as in the latter part
of
12
, TT; also in
45
, TT
);
20
o The speed overshoots and undershoots appear to be associated with problems in
this first implementation of I2V CACC, which will need improving before the
next large-scale test.
It can also be observed from the lower part of Figure 7 that the two probe vehicles (P1
and P2) had very similar speeds to their predecessors C1, C2 and C3; this indicated that
impact distance on the traffic upstream of the three controlled vehicles would be at least
the distance that was adopted in the test.
It is noted that the downstream bottleneck location used for the test was the exit of US29,
one section upstream of the N. Sycamore St. exit used in the simulations. For future large scale
I2V CACC vehicle tests, or for tests with set speed feedback to the drivers via roadside CMS (in
which impacts on the overall traffic flow should be measurable), the downstream bottleneck
selection is important, and needs to be the one with certain level of traffic flow drop.
This conceptual test also found some problems that need to be resolved before large scale
testing:
The V2I communication between probe vehicles and the server at the STOL Laboratory
was not reliable; therefore, the probe vehicle information could not be used for refining
the VSL/VSA to account for some local traffic fluctuations;
The ACC controller on the test vehicles needs to be replaced with one that can take the
set-speed as a limit while avoiding the over-shoot and under-shoot in speed control seen
in these initial test results; this is critical for I2V CACC application;
More traffic sensors need to be added to the system, or else probe vehicle data will need
to be used.
CONCLUDING REMARKS
Based on the simulation results, it is concluded that the speed based variable speed advisory
algorithm has the following advantages:
It is simple to understand and easy to implement
System performance improvements have been achieved in several aspects: TTT, TTD,
TD, Total Number of Stops, speed variation and downstream bottleneck throughput
Posted VSL/VSA values for the speed based approach are close to the measured sensor
speeds at fixed locations for most sections
A simple conceptual test has been conducted using this algorithm. Results showed that as
long as the recurrent bottleneck location selection is correct, the algorithm provided reasonable
VSA/VSL values, which are intuitively reasonable for speed harmonization over the freeway
traffic network. Several problems were also found during the tests, including V2I communication
reliability, need to refine VSL/VSA calculations based on probe vehicle data because of the low
density of detector data, the need for a CACC controller with tighter speed control functionality
21
for I2V and V2V CACC; need for improved sensor density and location selections. The lessons
learned and experience gained will be very valuable to future larger scale tests.
Acknowledgement
This paper is based upon work supported by the FHWA's Operations Research and
Development Contract out of the Saxton Transportation Operations Laboratory (Task 2); we
offer our appreciation to Daniel Dailey (FHWA), Taylor Lochrane (FHWA), Joe Bared (FHWA),
Hesham Rakha (VTTI), Joyoung Lee (New Jersey Institute of Tech.) and Peng 'Patrick' Su
(Leidos) for their valuable and constructive insights.
References
[1] Lu, X. Y., J. Y. Lee, D. J. Chen, J. Bared, D. Dailey, and S. Shladover, Freeway Micro-
simulation Calibration: Case Study Using Aimsun and VISSIM with Detailed Field Data,
93rd TRB Annual Meeting, Washington D. C., Jan.12-16, 2014
[2] Coifman, B., Improved velocity estimation using single loop detectors, Transportation
Research A, 35, 863-880, 2001
[3] Lu, X. Y., P. Varaiya, R. Horowitz, Z. Guo, and J. Plalen, Traffic speed estimation using
single inductive loop event data, Transportation Research Record, #2308, 2012, pp. 157–166
[4] X. Y Lu, A. Skabardonis, Using Sub-second Dual Loop Station Data for Congestion Onset
Detection, 2nd Int. Symp. on Freeway and Tollway Ops, Hyatt Regency Waikiki, Honolulu,
Hawaii, June 21-24, 2009
[5] X. Y. Lu and S. Shladover, Review of Variable Speed Limits/Advisories – Theory,
Algorithms and Practice, Transportation Research Record, TRB, # 2423, 2014, pp. 15–23.
[6] Lu, X. Y., P. Varaiya, R. Horowitz, D. Su and S. E. Shladover, A novel freeway traffic
control with variable speed limit and coordinated ramp metering, Transportation Research
Record #2229, TRB, 2011, pp 55-65
[7] Li, D., P. Ranjitkar, A. Ceder, An Integrated approach combining ramp metering and variable
speed limits to improve motorway performance, 93rd TRB Annual Meeting, Jan. 12-16, 2014,
Washington D. C.
22
[8] Li, D., P. Ranjitkar, A. Ceder, A logic tree based algorithm for variable speed limit 4
controllers to manage recurrently congested bottlenecks, 93rd TRB Annual Meeting, Jan. 12-
16, 2014, Washington D. C.
[9] van de Weg, G., A. Hegyi, H. Hellendoorn, S. Shladover, Cooperative systems based control
for integrating ramp metering and variable speed limits, 93rd TRB Annual Meeting, Jan. 12-
16, 2014, Washington D. C.
[10] Hadiuzzaman, M., J. Fang, C. Lan, T. Z. Qiu, Impact of Mainline Demand Levels and
Control Parameters on Multi-objective Optimization Involving Proactive Optimal Variable
Speed Limit Control, 93rd TRB Annual Meeting, Jan. 12-16, 2014, Washington D. C.
[11] Hadiuzzaman, M., J. Fang, M. Karim, Y. Luo, T. Z. Qiu, Modeling Driver Compliance with
Variable Speed Limits and Quantifying Impacts of Compliance Levels and Control Strategy
in Mobility and Safety, 93rd TRB Annual Meeting, Jan. 12-16, 2014, Washington D. C.
[12] Sabawat, V. and R. K. Young, Transferability of control strategies across rural variable
speed limit corridors, 93rd TRB Annual Meeting, Jan. 12-16, 2014, Washington D. C.
