You are on page 1of 8

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/329613662

A Dynamic Signal Control Approach for Integrated Ramp and Mainline


Metering*

Conference Paper · November 2018


DOI: 10.1109/ITSC.2018.8569567

CITATIONS READS
0 75

5 authors, including:

Amir Ghiasi David Kivilcim Hale


Leidos, Inc. Leidos, Inc.
17 PUBLICATIONS   138 CITATIONS    137 PUBLICATIONS   106 CITATIONS   

SEE PROFILE SEE PROFILE

Joe Bared Alexandra Kondyli


U.S. Department of Transportation University of Kansas
62 PUBLICATIONS   792 CITATIONS    60 PUBLICATIONS   344 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Cooperative Vehicle-Highway Systems to Improve Speed Harmonization View project

Fundamentals of Travel Time Reliability View project

All content following this page was uploaded by Amir Ghiasi on 01 April 2019.

The user has requested enhancement of the downloaded file.


A Dynamic Signal Control Approach for Integrated Ramp and
Mainline Metering*
Amir Ghiasi1 , David Hale1 , Joe Bared2 , Alexandra Kondyli3 , Jiaqi Ma4

Abstract—Freeway merges are one of the most significant traffic signals at freeway on-ramps. Numerous studies have
sources of traffic congestion. This study proposes a dynamic been conducted to design and implement ramp metering
signal control algorithm for an integrated ramp and mainline approaches to mitigate freeway bottlenecks (e.g., [1]–[3]).
metering control strategy to improve freeway operation and
reduce environmental impacts at freeway merging areas. The While these studies were successful in reducing the weaving
presented algorithm uses real-time traffic sensor information to frictions between the mainline and the ramp traffic to some
predict the future status of downstream queue and dynamically extent, some oversaturated traffic conditions on the mainline
control the signal. For safety reasons, the mainline signal approach may require a more effective approach to keep
control is activated only when a significant speed drop is the traffic from flowing into the existing downstream queue
detected at all lanes of the mainline and the traffic has been
already almost stopped. The results of the simulation analyses and allow the bottleneck traffic speed to recover. This idea
on a real-world calibrated network show a maximum of 15.7% is basically applied in variable speed limit (VSL) or speed
improvement in average speed, 20.9% reduction in average harmonization approaches that aim to reduce traffic speed
delay, and 13.7% decrease in CO emission. variations by limiting the mainstream traffic flow when it is
Keywords: Mainline Metering; Ramp Metering; Dynamic necessary. Numerous studies have developed VSL and speed
Signal Control; Traffic Control; Traffic Prediction
harmonization to improve traffic (e.g., [4]–[8]). A number
of studies have further improved traffic performance by
I. I NTRODUCTION integrating VSL techniques with ramp metering approaches
Freeway merge areas are one of the primary sources of (e.g., [9]–[11]). While these methods have been providing
congestion and traffic breakdown. These areas are typically notable successes in freeway operations all around the world,
associated with increased turbulence due to the increased they mainly suffer from users’ compliance that can limit
lane changing and slowing down to accommodate the ve- the control benefits. Other possible solutions are mainline
hicles that want to enter the freeway. This turbulence is metering strategies that control the amount of traffic entering
associated with the beginning of congestion (i.e., break- a freeway section using traffic lights. This approach can
down) at those segments. In most cases these breakdowns be considered as a special case of VSL, where the speed
initiate on the right side of the mainline, as incoming limit is set to zero for a limited time. Although simulation
ramp vehicles produce excessive weaving maneuvers, and experiments may show better improvements with a gen-
then spread to all lanes. Once a breakdown has begun, the eral VSL method than a special mainline metering one,
resulting shockwaves and queue spillback can eventually these experiments mainly assume that most users observe
stretch back for miles, causing area-wide traffic operational the posted speed limit. Therefore, with a proper mainline
and environmental problems. Given the prevalence of such metering technique, we can make sure that the majority of
problems, a low-cost solution to reduce the friction associ- users comply the rules provided by the control strategy. The
ated with vehicle maneuvers, or facilitate smoother merging freeway mainline metering strategy with traffic lights has
operations, could produce great savings. A potential cost- not received significant attention in the literature. Although
efficient solution to the freeway merge bottleneck problem some research findings appear to indicate that mainline
is to develop an effective signal control strategy to reduce metering can improve freeway operations, such a scheme
the frictions between the mainline and merge traffic. is only implemented in special situations such as tunnels
One of the well-known approaches in this category is and bridges [12], [13]. One of the first mainline metering
ramp metering that regulates the inflow ramp traffic using systems has been introduced by the Port Authority of New
York during the late 1950s and 1960s. They implemented
*This work was supported by the FHWA Saxton Transportation Opera- a traffic control system to increase the throughput of the
tions Laboratory.
1 Amir Ghiasi and David Hale are with Transportation Solutions Division, tunnels under the Hudson River [14]–[16]. Later, a signal
Leidos, Inc., Reston, VA 20190, USA amir.ghiasi@leidos.com control system was installed on Oakland Bay Bridge in San
and david.k.hale@leidos.com Francisco, CA to increase the westbound throughput [17].
2 Joe Bared is with the Office of Operation Research and Devel-
opment, Federal Highway Administration, McLean, VA 22101, USA These traditional systems mainly implement a fixed control
joe.bared@dot.gov logic or a basic feedback-based control algorithm that might
3 Alexandra Kondyli is with the Department of Civil, Environmental
be suitable for limited applications, however, may not obtain
and Architectural Engineering, University of Kansas, Lawrence, KS 66045,
USA akondyli@ku.edu satisfactory outcomes for freeways specially nearby critical
4 Jiaqi Ma is with the Department of Civil and Architectural Engineering merging traffic. A possible solution to address freeway traffic
and Construction Management, University of Cincinnati, Cincinnati, OH congestion near merging areas that could ensure users’
45221, USA jiaqi.ma@uc.edu
compliance is the integration of mainline and ramp metering
approaches. An effective and efficient dynamic mainline and
ramp metering system for freeway congestion near merging
areas is yet to be discovered.
This study proposes a dynamic signal control algorithm
that could be applied in an integrated ramp and mainline me-
tering control strategy. This algorithm is activated only when Fig. 1: Problem setting.
the downstream traffic queue spills back to the merging point
of the freeway. For safety reasons, the mainline traffic signal
will not be turned into red unless a significant speed drop is where 0 and 1 states represent the red and green signal states
detected at all lanes of the mainline. Activating the mainline at time t, respectively. Let g main and g ramp denote the start
signal control when the traffic has been already almost time of the most recent mainline and ramp green intervals,
stopped shall address the safety concerns regarding the usage respectively. Similarly, rmain and rramp are defined as the start
of traffic lights on freeways. When the signal control is time of the most recent mainline and ramp red intervals,
activated, the future downstream queue status will be pre- respectively.
dicted with a reliable and efficient real-time algorithm that A number of traffic sensors (e.g., loop detectors) are
is further used to determine the mainline signal red interval. deployed right before and downstream of the traffic signal to
Simulation analyses with a real-world calibrated network are measure traffic flow and speed. We assume that these traffic
performed to evaluate the algorithm performance and test the sensors can send real-time information to a local traffic
effects of different traffic conditions and parameter settings. controller. These traffic sensors are indexed as s ∈ S :=
The findings reveal that the proposed dynamic signal control {1, 2, . . . , S} and ls , ∀s ∈ S denotes the location of traffic
approach significantly improves traffic speed and reduces sensor s. The first traffic sensor is set at location 0 at the
travel delay and total emission. mainline approach (i.e., l1 = 0) and the last one is located
This paper is organized as follows. Section II presents the somewhere downstream of the bottleneck (i.e., lS > L),
problem setting. Section III describes the algorithm design where traffic speed is relatively recovered. After every ∆T
and explains the necessary modules. The simulation analyses time period, these sensors aggregate the measured traffic and
and results are presented in Section IV, and finally, Section send the information to the traffic signal controller. Fig. 1
V concludes the paper with a summary of the findings and illustrates the problem setting.
a brief discussion of future research directions.
III. DYNAMIC S IGNAL C ONTROL A LGORITHM
II. P ROBLEM S ETTING This section presents the dynamic signal control algorithm
This study considers a merging area of a freeway that design. The information provided by the traffic sensors are
connects the freeway mainline and ramp traffic. A lon- used to monitor traffic status at the merge area. When
gitudinal coordinate system is defined along the freeway the downstream traffic queue is propagated all the way to
that increases downstream. The origin of the coordinate the merge point and causes a certain speed drop at this
system is set right at the merging location of the freeway point, the dynamic signal controller is activated to prevent
and the ramp. Let N denote the number of lanes of the the upstream mainline traffic from joining the downstream
freeway mainline upstream of the merge point (i.e., location queue until it dissipates. The proposed dynamic signal
0). Without much loss of generality, the ramp roadway control algorithm includes three main modules. In the first
is assumed to have one lane. An acceleration lane begins module, the traffic flow and speed are updated by each traffic
after merging location 0, and ends at a point downstream sensor s at current time t, which are denoted by qs (t) and
of the merge. Let L denote the length of the acceleration vs (t), respectively. This information is updated at every time
lane. Therefore, a lane drop occurs at L, and thus traffic point with interval of ∆T . Thus the proposed dynamic signal
congestion may occur at this point of the freeway, which control algorithm runs iteratively at the same discrete time
can be propagated backwards to the upstream sections. We points. Given that, the second module activates the dynamic
consider a stream of vehicles enter the merging section at signal controller providing that a certain traffic speed drop is
location 0. Let v̄ and ā denote the maximum traffic speed detected at all lanes of the merge point. This module aims
and maximum acceleration, respectively. v̄ parameter can be to stop the mainline traffic from entering the downstream
set equal or a little greater than the posted speed limit and queue until it is about to dissipate. Based on the information
ā can be set to an average of various possible acceleration provided by the traffic sensors, the third module, which is a
values from standstill acceleration to accelerations at higher sub-routine to the second module, predicts the downstream
traffic speeds. traffic queue dissipation time. This prediction is then used
A traffic signal is deployed at location 0 with two signal by the second module to let the upstream traffic smoothly
groups, one for the mainline approach and the other for the merge into the downstream queue.
ramp traffic. Let Itmain , Itramp ∈ {0, 1} , ∀t denote the signal The algorithm initiates at time t := 0. Both Itmain and Itramp
status of the mainline and ramp approaches, respectively, are initially set to 1 (i.e., I0main := 1, g main := 0, I0ramp := 1,
B. Signal Control
This module designs the signal timing plan according to
the outputs of the first module. Basically, as long as v1 (t− )
is high enough, the signals on both mainline and ramp ap-
proaches are kept green, i.e., Itmain = 1 and Itramp = 1. When
the downstream queue propagates backward and reaches
the merge point and v1 (t− ) drops below a certain speed
threshold, denoted by V threshold , this module sets Itmain := 0
to prevent the upstream traffic from further joining the
queue1 . Although from a traffic operations perspective it
may be too late to let the downstream queue propagate to
the upstream merge point, it may not be safe to turn the
freeway mainline signal into red at higher v1 (t− ) values.
Thus, we let the queue propagate that far only to ensure that
the upstream traffic is already stopped or it is safe enough
to be stopped. The mainline signal control is activated when
the Mainline Control Activation criteria are met, which are
defined at the end of this sub-section. During this time, the
ramp signal is kept green as this approach generally includes
less traffic demand, i.e., Itmain = 0 and Itramp = 1. Then,
using the third module, the time when the last vehicle of the
downstream queue (which is called the downstream queue
tail hereafter) passes the bottleneck is predicted. Let Ttend
denote the queue dissipation time that is predicted at current
time t. Further, let T reach denote the time that it takes for the
leading vehicle(s) of the upstream traffic to accelerate from
Fig. 2: Algorithm flowchart. a zero-speed state at location 0 and reaches location lS that
depends on lS , v̄ and ā. Therefore, if we release the mainline
red signal T reach seconds before Ttend , the upstream traffic
and g ramp := 0) until the dynamic signal controller algorithm can be expected to smoothly join the downstream queue tail.
is activated. Moreover, we set rmain := −∞ and rramp := Further, we require that the red intervals at both approaches
−∞. After activation, Itmain and Itramp are controlled by our to be no greater than pre-determined maximum red intervals.
dynamic signal controller algorithm. Fig. 2 illustrates the Let Rmain and Rramp denote the maximum red intervals for
algorithm flowchart. In this flowchart, δT represents a very the mainline and ramp approaches, respectively. Moreover,
small time interval that indicates the passage of time during to reduce the conflicting maneuvers, we set the ramp signal
the algorithm implementation. The algorithm modules are to red at the time when the mainline red signal is released.
described in detail in the following sub-sections. Finally, to ensure stability and efficiency of the signal timing
plan, we require that the red and green intervals at both
approaches be no less than a certain minimum time interval
that is denoted by T min . The mainline signal control is
A. Traffic Data Collection
deactivated when the Mainline Control Termination criteria
This module runs iteratively and collects the traffic sen- are met.
sor information at every ∆T time point. Since the traffic With this, the Mainline Control Activation and Termina-
sensors aggregate data with time interval of ∆T , these data tion are defined as follows.
should be represented in the middle of each time interval • Mainline Control Activation criteria:
( ( −) ) ( )
(t − ∆T, t]. Let t− := t − ∆T define the middle time v1 t ≤ V threshold ∧ t ≥ g main + T min .
point. Therefore, this module collects qs (t− ) and vs (t− )
at each discrete time point and for all s ∈ S. Note that • Mainline Control Termination criteria:
( ( ) )
the traffic information at each ls is aggregated among all t ≥ min max Ttend − T reach , rmain + T min , rmain + Rmain .
lanes of the freeway, except for v1 (t− ) that is set to the
maximum observed speed among all N lanes right before If the Mainline Control Termination criteria are met, then we
the traffic signal. The reason for this exception is that the set Itmain := 1 and Itramp := 0. The ramp signal turns back to
mainline control is not activated until the speed at all lanes green (i.e., Itramp := 1) when the Mainline Control Activation
of the merging point drops below a certain threshold at all 1 For real-world implications, a sufficient yellow interval may be required
lanes before the traffic signal. Finally, this module updates before the red intervals. While it is easy to determine appropriate yellow
the signal status Itmain , Itramp , g main , g ramp , rmain , and rramp . intervals for both directions, it is excluded in this model.
criteria are met or the ramp red interval reaches Rramp (i.e.,
t ≥ rramp +Rramp ). The Queue Dissipation Prediction module
is presented in the following sub-section.

C. Queue Dissipation Prediction


The queue dissipation prediction module, which is a part
of the signal control module, predicts Ttend using the traffic
sensor information. This prediction algorithm is proposed
by [18] for a single-lane freeway section. This paper uses
this algorithm for a multi-lane freeway section to predict
the queue tail location at any future time t′ > t, denoted by
pt (t′ ) , ∀t′ > t. This algorithm uses vehicle indices as the
boundary conditions. Let cs (t− ) denote the estimated per-
wave ′ ′
lane average vehicle count at ls at time t− that is calculated Fig. 3: An illustration of yts (t ) and cwave
ts (t ) functions.
as:
( )
∆T qs (t− ) ∆T
cs := , ∀s ∈ S, traffic speed cannot ( exceed
) v̄, it is required to ensure that
2 2N the obtained (t̃ts , ỹts) points are feasible for all s ∈ StQ .
( ) ( ) qs (t− ) ∆T 3∆T Starting from t, ls′t in the space-time diagram, we shoot
cs t− := cs t− − ∆T + , ∀s ∈ S, t− ≥ . an upper bound trajectory at speed of v̄. Let t̂ ′ de-
N 2 tst
− note the intersection( )between this upper bound trajectory
With the estimated cs (t ) , ∀s ∈ S values, the wave
and yt(s ′ +1) t̃t(s′ +1) that is calculated as t̂t(s′t +1) =
pt (t′ ) , ∀t′ > t function is predicted as follows. Since there ( t t

)
is no further information available about the traffic status ls − l
( st
′ + v̄t )+ wt / (v̄ + w). Then, define ŷt(s′t +1) :=
wave
between the traffic sensors, we assume that the queue tail is y ts t̂ ′
t(st +1) ( , and choose the )lower(point (or the greater )

located at one of the sensor locations at a time. Let st ∈ S time) between t̃t(st +1) , ỹt(st +1) and t̂t(s′t +1) , ŷt(s′t +1) as
′ ′
wave ′
denote the sensor index at which the queue tail is located at the final predicted space-time point along yt(s′t +1) (t ). Next,
− wave ′
current time t, which is estimated using vs (t ) , ∀s ∈ S starting from this final point along yt(s′t +1) (t ), we shoot
information. Let ctail t denote the per-lane average vehicle another upper bound trajectory at speed of v̄ and determine
count that corresponds to the queue tail. Since Itmain = 1 and its intersection ( with the next ) shockwave, and then compare
ramp inflow traffic is relatively limited when this module is it with t̃t(s′t +2) , ỹt(s′t +2) to find the next final space-time
wave ′
activated, we assume that traffic density is zero between point along yt(s′t +2) (t ). With similar calculations, all space-
− ′
l1 and ls′t and thus ct = cs (t ) , ∀s = {1, 2, . . . , st }. time points along all shockwaves in StQ can be determined.
tail

Therefore, we only use the information provided by the We let (t̄ts , ȳts ) denote the final predicted space-time points
sensor set StQ := {s′t , s′t + 1, . . . , S} for pt (t′ ) , ∀t′ > t along yts wave ′
(t ) , ∀s ∈ StQ that are calculated as:
prediction. ( )
Starting from each traffic sensor in the time-space dia- t̄ts := max t̃ts , t̂ts ,
gram, we generate a backward shockwave at speed −w, ȳts := min (ỹts , ŷts ) , (3)
which is formulated as: ( −
)
(′ ) where t̂ts = ls − ȳt(s−1) + ( v̄ t̄)t(s−1) + wt / (v̄ + w) , ∀s ∈
wave ′ − Q ′ − ′
yts (t ) := ls − w t − t , ∀s ∈ St , t ≥ t . (1) St \ {st } and ŷts := yts t̂ts , ∀s ∈ St . Then, we simply
Q wave Q

connect (t̄ts , ȳts ) , ∀s ∈ StQ points to construct the piece-


Based on the kinematic wave theory proposed by [19],
wise linear pt (t′ ) , ∀t′ > t function. This function continues
vehicle index increases along these shockwaves with the rate
up to (t̄tS , ȳtS ) that is lower than lS . Since no other
of 1/p0 where p0 is the standstill jam spacing. With that, a
shockwave existed after this point, we simply let pt (t′ )
vehicle index function can be defined as:
( −) (′ ) function continue with vS (t− ) up to lS . With this, a
wave ′ − Q ′ −
cts (t ) := cs t + w t − t /p0 , ∀s ∈ St , t ≥ t . piece-wise function is constructed for pt (t′ ) , ∀t′ > t,
(2) which yields
Fig. 3 illustrates an example for shockwave yts wave ′
(t ) and Ttend := p−1 t (lS ) .
wave ′
vehicle index cts (t ) functions formations. In this figure,
Fig. 4 illustrates the derivation of (t̄ts , ȳts ) , ∀s ∈ StQ ,
t′i , i ∈ Z+ represent time points that are greater than t− ,
pt (t′ ) , ∀t′ > t, and Ttend .
where t′i − t′i−1 = p0 /w, ∀i = {2, 3, . . . }.
Then, this module predicts the future time-space points
along these shockwaves that correspond to the downstream IV. S IMULATION E XPERIMENT
tail
queue tail. Given ct and (1)-(2), we can find the points This study uses a merge section of I-35 located in Kansas
wave ′ wave ′ tail
along
( yts (t ), at which
( tail ) c ts (t ) =
( t)) c that are defined as City, KS to test the proposed algorithm. This section is
t̃ts := cwave−1
ts ct , ỹ ts := y wave
ts t̃ ts , ∀s ∈ S Q
t . Since approximately 2.2 kilometers long and is taken from a
Fig. 5: An example of signal timing plan.

7 and 24.5 m, respectively. Thus we set p0 to the average


value of 7.61 m. The Wiedemann 99 car-following model
implemented in VISSIM simulation package is used for
the simulation analyses. Based on the calibrated simulation
network, v̄ = 100 kph, ā = 1.52 m/s2 , and w = 4.6
m/s. To evaluate the algorithm performance, three measures
(a) (t̄ts , ȳts ) , ∀s ∈ StQ of effectiveness are chosen: average traffic speed, average
traffic delay (per vehicle), and total CO emission. The
improvements to these measures are denoted as ∆E S , ∆E D ,
and ∆E E , respectively. These measures are network-wide
average values, which are averaged throughout the 2.2-
kilometer network. The simulation analysis is performed for
three 15-minutes time periods, which the first and the last
periods are assigned to warm-up and cool-down periods,
respectively.
Let Dmain and Dramp denote the traffic demands of the
mainline and ramp approaches. We first present the simula-
tion results for a default parameter setting. In this setting,
we set Dmain = 8000 vph, Dramp = 1500 vph, Rmain = 60
s, Rramp = 20 s, V threshold = 10 kph, T min = 10 s, and
(b) pt (t′ ) , ∀t′ > t, and Ttend ∆T = 2 s. Further, based on the merge geometry and
the calibrated traffic behavior, we obtain T reach = 19.2
Fig. 4: An illustration of queue dissipation prediction. s. To consider the stochasticity effects of the simulation
experiments, each simulation scenario is run 10 times with
different random seeds. The obtained average mainline and
simulation network that is calibrated with field measure- ramp green intervals among all random seeds are 151 and
ments. For the model calibration, a day without incidents 170 s, respectively, where on average 72% of the green
throughout the corridor or adverse weather was selected. intervals overlap. In this algorithm, the ramp red interval
The selected calibration period is the afternoon peak-hour lasts until another mainline control activation happens or the
of April 22, 2016. The westbound side of the network is red interval reaches Rramp . In the default case scenario, the
chosen for the analysis. The algorithm is tested on a merging time between every consecutive mainline control activation
area, which is located around 1.2 kilometers downstream of is greater than 20 seconds, and thus the ramp red interval
the input traffic. At the merge point, an on-ramp roadway is always equal to Rramp = 20 s. Moreover, the average
merges the three-lane mainline freeway (N = 3). Right after mainline red interval is 40 s. Fig. 5 shows an example of
the merge point, an acceleration lane with the length of signal timing plan for the first two mainline and ramp cycles
281 meters begins (L = 281 m). Therefore, at the end of resulted from the first simulation instance. Further, Fig. 6
this acceleration lane, a lane drop occurs, and thus traffic compares the simulation results of the signal control algo-
congestion may occur at this point of the freeway, which rithm with the benchmark case and without any control with
can be propagated backwards to the upstream sections. box-plots. The results indicate that significant improvements
In this network, six sets of loop detectors are deployed, are obtained for all these measures.
and each set contains a number of loop detectors embedded We perform a sensitivity analysis on the Dmain and Dramp
next to one another to cover all lanes at a specific longi- values. Then, the effects of Rramp are investigated through
tudinal location. Thus each loop detector set represents a simulation experiments. For these simulation analyses, dif-
traffic sensor in our analyses (i.e., S = 6). These traffic ferent scenarios are considered for the Dmain and Dramp
sensors are evenly distributed along the merge section with values, while other parameters are kept as their default
the distance of 69 meters from one to another. The locations values. The results of the sensitivity analysis on Dmain and
of these traffic sensors are set as: l1 = 0, l2 = 69, l3 = 138, Dramp values are shown in Fig. 7, 8, and 9 for ∆E S , ∆E D ,
l4 = 207, l5 = 276, and l6 = 345 m. Traffic includes 3.5 and ∆E E , respectively. These figures show the 3D box-
percent of trucks and p0 for regular vehicles and trucks are plots among all 10 random instances. From these figures
Benchmark
32 10
200 Benchmark
Control
190 9.5

Average Delay (sec/vehicle)


30
Average Speed (kph)

180

CO Emission (kg)
9
28
170
8.5
160
26
8
150

24
140 7.5

Benchmark 130 Control Control


22 7

Fig. 6: Simulation results for the default setting.


Fig. 8: Sensitivity analysis on Dmain and Dramp : ∆E D (%).

Fig. 7: Sensitivity analysis on Dmain and Dramp : ∆E S (%). Fig. 9: Sensitivity analysis on Dmain and Dramp : ∆E E (%).

we see that all of these measures mainly improve at various the appropriate parameter settings.
demand scenarios. A maximum of 15.7% improvement in
average speed, 20.6% reduction in average delay, and 13.7% V. C ONCLUSION
decrease in CO emission is obtained. The maximum benefits This paper develops a dynamic signal control algorithm
are achieved at the mainline demand values of 7000 to that can be applied for an integrated ramp and mainline
7500 vph (or 2333 to 2500 vph/ln) and the ramp demand metering strategy. When the traffic queue, caused by a lane-
values of 1000 to 1500 vph, and the minimum improvements drop downstream of a freeway merge, spills all the way back
are resulted for relatively low demand values as traffic to the merge point, the dynamic signal control algorithm
is undersaturated in such demand scenarios. Overall, this is activated and turns the mainline signal to red to keep
experiment reveals that the proposed algorithm can achieve the mainline traffic from entering the downstream queue,
significant improvements when high traffic saturation exists. which otherwise would exacerbate the traffic congestion.
Next, we perform sensitivity analysis on impacts of Rramp For safety reasons, the proposed algorithm does not activate
parameter on the algorithm results. The results are shown as the signal control unless the traffic speed at all mainline
box-plots in Fig. 10, in which the green, yellow, and red box- lanes drops below a safe speed threshold (that is set to 10
plots correspond to reasonably low, moderate, and high time kilometers/hour in our experiments). After activation, this
intervals for Rramp . The results indicate that all measures algorithm predicts the future downstream queue status using
mainly decrease with Rramp . This is because greater Rramp the information provided by the deployed traffic sensors.
values would worsen the ramp traffic performance that This prediction is updated at every decision time point and
impacts the network average speed, delay and total emission. as a result, the prediction is modified using the new received
The maximum obtained benefits are: 15.7% improvement in information. The mainline traffic signal turns into green a
average speed at Rramp = 20 s, 20.9% reduction in average few seconds in advance of the predicted downstream queue
delay at Rramp = 10 s, and 13.4% decrease in CO emission dissipation time to let the mainline traffic smoothly join the
at Rramp = 20 s. Overall, these simulation analyses provide downstream traffic. At the same time, the ramp signal turns
insightful information regarding the algorithm benefits and into red to remove frictions between the merging and the
CAV-based VSL strategy to further improve the freeway
40 40
operational performance.
30 30

20 20 R EFERENCES
10 10 [1] M. Papageorgiou, H. Hadj-Salem, and J.-M. Blosseville, “Alinea: A
local feedback control law for on-ramp metering,” Transportation
0
0 50 100
0
0 50 100
Research Record, vol. 1320, no. 1, pp. 58–67, 1991.
[2] M. J. Cassidy and J. Rudjanakanoknad, “Increasing the capacity of
(a) ∆E S (%) (b) ∆E D (%) an isolated merge by metering its on-ramp,” Transportation Research
Part B: Methodological, vol. 39, no. 10, pp. 896–913, 2005.
[3] G. Gomes and R. Horowitz, “Optimal freeway ramp metering using
25
the asymmetric cell transmission model,” Transportation Research
20 Part C: Emerging Technologies, vol. 14, no. 4, pp. 244–262, 2006.
15
[4] R. Bertini, S. Boice, and K. Bogenberger, “Dynamics of variable
speed limit system surrounding bottleneck on german autobahn,”
10
Transportation Research Record: Journal of the Transportation Re-
5 search Board, no. 1978, pp. 149–159, 2006.
0
[5] R. C. Carlson, I. Papamichail, and M. Papageorgiou, “Local feedback-
0 50 100 based mainstream traffic flow control on motorways using variable
speed limits,” IEEE Transactions on Intelligent Transportation Sys-
(c) ∆E E (%) tems, vol. 12, no. 4, pp. 1261–1276, 2011.
[6] A. Tympakianaki, A. Spiliopoulou, A. Kouvelas, I. Papamichail,
Fig. 10: Sensitivity analysis on Rramp . M. Papageorgiou, and Y. Wang, “Real-time merging traffic control for
throughput maximization at motorway work zones,” Transportation
Research Part C: Emerging Technologies, vol. 44, pp. 242–252, 2014.
[7] J. Ma, X. Li, S. Shladover, H. A. Rakha, X.-Y. Lu, R. Jagannathan,
mainline traffic. This red interval is released when another and D. J. Dailey, “Freeway speed harmonization,” IEEE Transactions
speed drop is detected at the mainline approach (and thus on Intelligent Vehicles, vol. 1, no. 1, pp. 78–89, 2016.
the mainline signal turns into red) or it reaches a maximum [8] M. Tajalli and A. Hajbabaie, “Dynamic speed harmonization in
connected urban street networks,” Computer-Aided Civil and Infras-
ramp red interval. tructure Engineering, vol. 33, no. 6, pp. 510–523, 2018.
A number of simulation analyses is conducted using [9] A. Hegyi, “Model predictive control for integrating traffic control
a calibrated VISSIM network to evaluate the algorithm measures,” Ph.D. dissertation, Delft University of Technology, 2004.
[10] A. Hegyi, B. De Schutter, and H. Hellendoorn, “Model predictive
performance under different traffic conditions and parameter control for optimal coordination of ramp metering and variable speed
settings. The implemented case study is a section of I- limits,” Transportation Research Part C: Emerging Technologies,
35 freeway located in Kansas City, KS. In this study, vol. 13, no. 3, pp. 185–209, 2005.
[11] X.-Y. Lu, P. Varaiya, R. Horowitz, D. Su, and S. Shladover, “Novel
thee performance measures are considered: network-wise freeway traffic control with variable speed limit and coordinated
average traffic speed, delay, and total CO emissions. The ramp metering,” Transportation Research Record: Journal of the
results of the simulation analyses show a maximum of Transportation Research Board, no. 2229, pp. 55–65, 2011.
[12] E. L. Jacobson and J. Landsman, Case studies of US freeway-to-
15.7% improvement in average speed, 20.9% reduction in freeway ramp and mainline metering and suggested policies for
average delay, and 13.7% decrease in CO emission. Note Washington State, 1994, no. 1446.
that although these benefits are limited to a local control [13] K. A. Haboian, “A case for freeway mainline metering,” Transporta-
tion research record, no. 1494, 1995.
zone, the proposed control approach can be implemented at [14] K. Crowley and H. Greenberg, “Holland tunnel study aids efficient
the other parts of the networks. Therefore, further benefits increase of tube’s use,” Traffic Engineering, vol. 35, no. 6, pp. 20–22,
can be achieved for larger areas, and thus could justify the 1965.
[15] R. S. Foote and K. W. Crowley, “Developing density controls for
system deployment costs. improved traffic operations,” Highway Res. Board Rec, vol. 154, no. 1,
This study can be extended in a number of directions. pp. 24–37, 1967.
While the speed at which the control algorithm activates [16] R. S. Foote, “Instruments for road transportation,” Port Authority of
New York and New Jersey, Tech. Rep. Tbr-2-68A, 1968.
(i.e., V threshold ) is supposed to be set to a reasonably safe [17] M. MacCalden et al., “Traffic management system for the san
value, further safety analyses can be conducted to investigate francisco-oakland bay bridge,” ITE Journal, vol. 54, no. HS-037 200,
the potential safety concerns of this freeway signal control. 1984.
[18] A. Ghiasi, J. Ma, F. Zhou, and X. Li, “Speed harmonization algorithm
The level of users’ acceptance and compliance might be an- using connected autonomous vehicles,” in 96th Annual Meeting of the
other issue to explore. Further, this study can be represented Transportation Research Board, Washington, DC, 2017.
as a baseline for highway mainline metering techniques. [19] G. F. Newell, “A simplified car-following theory: a lower order
model,” Transportation Research Part B: Methodological, vol. 36,
With some future improvements in the prediction algorithm, no. 3, pp. 195–205, 2002.
we may be able to achieve better outcomes with fewer traffic
sensors to obtain more cost-effective solutions. Finally,
with the advent of connected and automated technology
(CAV), we may be able to feed the controllers with more
accurate and higher resolution data and to enforce better
compliance of mainline vehicle stopping, even with higher
speed thresholds. Moreover, the proposed integrated ramp
and mainline control algorithm can be combined with a

View publication stats

You might also like