Simulation Modelling Practice and Theory 47 (2014) 141151

Contents lists available at ScienceDirect

Simulation Modelling Practice and Theory

journal homepage: www.elsevier.com/locate/simpat

Analysis of freeway service patrol with discrete event-based

simulation
Weitiao Wu a, Luou Shen a,, Xiaojin (Jerry) Ji b, Wenzhou Jin a
a
Department of Civil and Transportation Engineering, South China University of Technology, 381 Wushan Road, Jiaotong Building, Room 522, Tianhe,
Guangzhou, Guangdong 510641, China
b
ACS Infrastructure Development Inc., One Alhambra Plaza, Suite 710, Miami, FL 33134, United States

a r t i c l e i n f o a b s t r a c t

Article history: As an efcient incident management program, freeway service patrol has gained wide
Received 19 August 2013 popularity. There is increasing need for developing systematic procedures of planning the
Received in revised form 16 March 2014 program to ensure efcient resource allocation and to achieve maximum benets. A dis-
Accepted 25 March 2014
crete event-based simulation model is developed to replicate the patrol process that
includes districting, incident generation, response vehicle dispatching and routing, and
incident clearance. Model validation proves its capability to accurately replicate the inci-
Keywords:
dent generation/distribution process. Later, the Interstate-95 freeway in south Florida is
Incident management
Freeway service patrol
used as a model application example to test alternatives, and the improved alternative
Discrete event-based simulation shows better performance compared with the previous even-length districting method
Patrol vehicle dispatching and the currently used even-activity districting approach. Besides traditional roving
Patrol districting scheme, the simulation model also has the capability to evaluate the prepositioning
Prepositioning scheme schemes. Furthermore, sensitivity analysis is performed to examine the impact of opera-
tional parameter settings, such as beat districting, incident detection rate, and incident fre-
quency. Although continuous efforts are needed to develop a more comprehensive
simulation model for the freeway service patrol program, the application example pre-
sented in this study reveals promising application results, and provides an insight into
the patrol process as well as a better understanding of the impact of parameters on
performance.
2014 Elsevier B.V. All rights reserved.

1. Introduction

Non-recurrent congestion on freeway is mainly caused by incident, and contributes to a large part of trafc delay.
Incidents cause excess fuel consumption as well as air pollution as reported by Lindley [1], also result in secondary accidents
by Zhan et al. [2]. Besides, Red Cross report [3] shows that 57% of all victims who are killed die in the rst minute after the
crash, before the arrival of the emergency services. Effective incident detection and verication, response, and clearance are
essential to improve safety and trafc conditions. The attention of the US government has been directed toward effective
freeway incident management strategies, such as the use of freeway service patrol (FSP), automatic incident detection,
and closed circuit television (CCTV) camera, to reduce the impacts of incidents. Kaass [4] review indicates that many eval-
uation studies of FSP have reported high benet-cost ratio.

Corresponding author. Tel.: +86 13760752636.

E-mail addresses: wu.weitiao@mail.scut.edu.cn (W. Wu), ctlshen@scut.edu.cn (L. Shen), xji@acsinfra.com (Xiaojin (Jerry) Ji), ctwzhjin@scut.edu.cn (W. Jin).

http://dx.doi.org/10.1016/j.simpat.2014.03.012
1569-190X/ 2014 Elsevier B.V. All rights reserved.
142 W. Wu et al. / Simulation Modelling Practice and Theory 47 (2014) 141151

The FSP is a complex process with random inputs, and the operational parameters of a patrol program, such as size of
patrol area, eet size, hours of operations, and response vehicle depot locations, inuence the required time of response vehi-
cles to reach the incident location. For the evaluation of a patrol program, it is necessary to develop simulation models that
can reproduce the patrol process as completely as possible. Given the infrequency feature of incidents under a daily time
frame, the developed model should be operated on a year basis in order to capture the random pattern of incident distribu-
tion, which brings another requirement of time efciency to the simulation model. This paper reports a phase output of an
ongoing effort toward developing a systematic procedure for FSP that is a discrete event-based simulation model with the
capability of replicating the patrol process including districting, incident generation, response vehicle dispatching and rout-
ing, and incident clearance.

2. Previous research work

Freeway incident response, as a subarea of emergency response, has been a hot topic in the operation research commu-
nity. Most studies focused on the development of methodological framework to reduce incident duration by optimizing the
number of emergency vehicles, response policies, routing schemes, and relevant jurisdiction areas. In general, methodologies
used in previous work can be summarized into two groups: analytical method and simulation method.
Zografos et al. [5] rst proposed an analytical framework with linear programming technique that can minimize the free-
way incident delays through the optimum deployment of trafc ow restoration units. Pal and Sinha [6] constructed a mixed
integer programming model to determine optimal locations for response vehicles, given the frequencies of incidents at
potential sites in the network, which could minimize the annual cost, but the determination was subject to the eet size
constraint. An opportunity cost-based model proposed by Sherali and Subramanian [7] demonstrates that dispatching the
closest available vehicle to the site of the current accident is not always the optimal incident response strategy when con-
sidering the expected future demands. A simple analytical method was proposed by Khattak et al. [8] to determine the most
benecial locations for patrol deployment by combining crash rate, congestion level, estimation of incident-induced delay,
and estimated B/C ratio of FSP. However, this method did not tackle the specic patrol process. Yin [9] developed a minmax
bi-level programming model to decide the allocation of FSP. Later, he also [10] proposed a mixed-integer nonlinear program-
ming model to allocate FSP vehicles by minimizing the expected loss associated with incident occurrence.
Recognizing the highly stochastic nature of the incident management operations, researchers started to apply simulation
to solve the problem. Developed by Nathanail and Zografos [11], the rst tool applied Poisson distribution to the occurrence
of incidents, whereas uniform distribution is assumed for incident spatial distribution. Incident duration for lane-blockage
incidents (incident occurred on travel lane) was extracted from archived data, while the duration for shoulder incidents
(incident occurred on shoulder) was based on empirical judgments. Zografos et al. [12] presented a decision-support system
for incident management decision-making by integrating mathematical models, rules, and algorithms in a user-friendly
environment to minimize incident response time. Pal and Sinha [13] also developed a simulation tool for the selection of
FSP operational parameters. The trafc ow was modeled at a macroscopic level mainly based on queuing theory to calculate
incident delays and determine diversion assumptions. However, the prepositioning scheme was not modeled.
In a paper by Ozbay and Bartin [14], a simulation tool was developed with Arena simulation package. This study focused
on testing the impact of the incident detection methods. Results showed that in failing to detect the incident, it would take a
maximum of 20 min for the responsible agencies to be informed by other means (cell phones). Haghani et al. [15] developed
a simulation model to evaluate a real-time emergency medical service system, which dealt with general emergency response
problems. However, freeway service patrol has its special characteristics, such as different incident types, change of travel
duration with time, and work load balance problem. Recently, to assess the impact of the parameters on service patrol oper-
ations, Hadi et al. [16] developed a simulation tool that was proven to have the capability to accurately replicate the incident
generation/distribution, the travel duration change with time and the patrol process. Unfortunately, as the initial stage of
their work, the pre-positioning schemes and inuences of patrol conditions are not investigated.
As demonstrated in the literature review, for the evaluation of a patrol program, the latest studies tend to focus on devel-
oping simulation models that can reproduce the patrol process. However, very few studies have been conducted to develop
systematic planning or evaluation procedures of those programs. This paper is a continuation of the work by Hadi et al. [16]
and initiated by the Florida Department of Transportation (FDOT) District 4 through investigating many patrol policies
including the pre-positioning schemes and testing various patrol parameter settings to further understand the patrol process
and assess the program. The following part of this paper rst describes the need for a simulation model, the current FDOT
practice, and the data in the sample application. Afterwards, the simulation model is illustrated, and the sample application
of the model with the Interstate-95 incident data is summarized. At the end, some conclusions and future work are
presented.

3. Need for a simulation model

Freeway patrol vehicles respond to incidents on the freeway as quickly as possible to provide emergency aid and relieve
the impact on congestion. The traditional roving strategy of FSP program can be summarized as: response vehicles patrol
within the assigned beat and watch for any incidents; upon detection of an incident, the vehicle reaches the incident location
 W. Wu et al. / Simulation Modelling Practice and Theory 47 (2014) 141151 143

and provides assistance; after the clearance of the incident, the response vehicle resumes its normal patrol operations. If
incidents are detected by automated detection technologies, patrol vehicles are directed to the reported incident location
following predened routing schemes.
During the patrol process there are many operational parameters, such as the size of patrol areas, eet size, hours of oper-
ations, and the response vehicle depot location inuence how quickly the response vehicles can reach the incident location.
Altogether, the complexity of the incident response process mainly results from four aspects: the high uncertainty of inci-
dents occurrences which usually modeled with statistics methods; the complex interactions between incidents occurrences
and the response of patrol vehicles; the various parameters of the system settings; and the dynamics of trafc.
Analytical methods present difculty to capture such complexity of FSP program. As shown in the previous research work
section, researchers recently start looking into simulation method for solution. By simulating the patrol process the effective-
ness of the program can be evaluated with the outputs. To achieve practical and useful results, the simulation must capture
all the important interactions between the components of a system. Therefore, it is required that the simulation model need
to be able to explicitly replicate the real operation of patrol vehicles response process on prevailing freeway trafc condi-
tions. Meanwhile, given the infrequency nature of incidents under a 24-h time frame, ideally, such model should be applied
on a year basis in order to fully capture the incident distribution pattern, which requires the feature of time-efciency.

4. FDOT current practice and data acquisition

In 1995, FDOT District 4 started their FSP program that initially covered the I-95 70-mile long corridor and then expanded
to I-75 and I-595. Called road rangers (RR) by Florida residents, the patrol vehicles are operated by a private contractor
under agreement with FDOT District 4.
In the past, incident detection was mainly dependent on the information reported by notications through the call box or
by personal cell phones. This approach could not avoid false alarms and patrol vehicles needed to keep traveling along the
freeway to detect incident occurrence.
As a trend, incident detection and verication are becoming active by deploying CCTV cameras. Starting from 2004, FDOT
District 4 has deployed CCTV cameras along most of the interstate freeways and urban expressways. Based on the data of
2011, approximately 55% of the incidents could be quickly detected and veried through this system by operators working
in the Trafc Management Center (TMC).
In 2004, FDOT developed an incident management database to record information about incidents and incident manage-
ment activities, such as incident time stamps (detection, notication, responses, arrivals, and departures), incident ID,
responding agencies, event details, chronicle of the event, and environmental information for all incidents in the region.
In this study, incident records of 2011 were extracted to develop the simulation model. FDOT also maintains a trafc-mon-
itoring system that collects trafc counts continuously throughout the state highway network. The trafc count data are
published annually in the Florida trafc information (FTI) DVD.

5. Methodology

A simulation model has been developed for FSP planning and evaluation, which involves replication of temporal and spa-
tial incident occurrence, trafc ow variation by time of day, dispatching of response vehicles and their movements along
their beat areas according to the predened policies, incident clearance process, and effectiveness evaluation of patrol ser-
vices. Following the patrol process, four major modules in the simulation model are dened as follows: (1) incident gener-
ation, (2) trafc ow simulation, (3) incident response simulation, and (4) performance evaluation. The framework of the
developed simulation model is described in Fig. 1.
Some operational parameters of FSP program are typically determined during the planning phase as follows: (1) beat size,
(2) number of patrol vehicles in each beat, (3) operation time, (4) dispatching policy, and (5) routing schemes of response
vehicles.
Although the FSP program aims to reduce incident-related delay, the objective of applying the simulation model is to
minimize the incident response time and balance the work load of each beat. Incident-related delay can also be minimized
through deploying advanced and equipped response vehicles. However, this is out of the scope and capability of the simu-
lation process.
The Arena simulation software developed by Rockwell Automation [17] is adopted in this study as the modeling environ-
ment. Arena, which uses a discrete, event-based, owcharting methodology for dynamic process modeling, has been widely
applied in the engineering industry. Discrete event simulation is also widely used in transportation engineering, for example:
Jaoua et al. [18] developed a simulation model to be applied in internal transport systems, in container terminals and in sur-
face mines. In addition, Motraghi and Varhanov Marinov [19] developed an event-based simulation model that used Arena to
solidify the merits of moving urban freight by rail.
As a key modeling issue in trafc simulation, vehicle interaction is not modeled here because it is not the main point of
patrol process. Macroscopically modeling the trafc is sufcient for a planning tool, and it could also reduce the development
efforts for the model. A visual basic for application (VBA) script language is the standard extension in Arena, and this makes it
144 W. Wu et al. / Simulation Modelling Practice and Theory 47 (2014) 141151

Incident Data
Generation Preprocessing

Freeway Network
Districting
Plan

Dispatch Dispatching
Patrol Vehicle Policy

Patrol Vehicle Routing

Movement Scheme

Model Alternative
Evaluation Generation
Fig. 1. Simulation model framework.

possible for developers to write codes to manipulate the simulation process and formulate customized patrol response
strategies.
The Arena software is a dynamic system through the ow of entities. Entities travel along a serious of modules where
actions are dened and performance data is collected. Therefore, the explanation of the full model is broken down into sec-
tions: the model attributes are presented in Section 5.1, and details about the sub-modules are presented in Sections 5.2-5.5.

5.1. Model attributes

In this simulation system, the entities are patrol vehicles and incidents. Both of them are created by CREATE modules. This
module denes the arrival pattern which is usually set to distributions for incidents or schedule for patrol vehicles. Incidents
are created to STATION and waiting at QUEUE for patrol vehicles response. Along the way owing through the system patrol
vehicles use the resources. Here, the resources are patrol drivers. Once patrol vehicles reached incident scenes, patrol vehi-
cles REMOVE the incident and SEIZE the patrol driver to DELAY module at the same time. After incident clearance, patrol
vehicle RELEASE the patrol driver and DISPOSE the incident. Besides, other modules such as: DECIDE, ASSIGN, BRANCH,
ROUTE and VBA are used in order to accurately model the patrol process. Arena screenshot of model attributes are shown
in Fig. 2.

5.2. Incident generation sub-module

The recorded data include incident type, lane blockage, rst notication time, verication time, response time, clearance
time, location description, longitude, and latitude. The incident location is recorded as the geographic coordinates. FDOT has
discretized the freeway network to segments represented by some predened key longitude/latitude locations as shown in
Fig. 3.
Therefore, a GIS tool is applied to convert the longitude/latitude coordinates to mileposts. Because the incident locations
are recorded in a discrete format (geographic coordinates), it is natural to model the freeway segments as nodes. This
approximation simplies the highway network modeling and is considered appropriate because the simulation model is
designed to support patrol program planning and policy development rather than detailed trafc modeling.
As a result, I-95 is modeled with nodes and links in Arena. Incidents occur to nodes according to the temporal/spatial dis-
tribution, and patrol vehicles move from node to node through the links in between. Incident durations are calculated from
the time attributes recorded in the database. The incidents are classied into the following two groups: lane-blockage inci-
dents and shoulder incidents. Because the lane-blockage incidents have greater impact on congestion, usually a higher pri-
ority is assigned to these accidents to achieve faster response. In this study, clearance time is dened as the time from the
arrival of response vehicle at the incident location to the clearance of incident and the reopening of lanes to trafc. Clearance
duration models are t for the two groups of incidents, respectively.
 W. Wu et al. / Simulation Modelling Practice and Theory 47 (2014) 141151 145

Fig. 2. Screenshot of model attributes in Arena.

End
I-95

SB

NB

I-595
I-95

Start

Fig. 3. Incident GIS locations.

The number of generated incidents varies with time according to the real temporal distribution. As presented in previous
studies, the incident occurrence can be assumed as Poisson distribution. It has been proven that if the number of incidents in
a given time interval [0, t] conforms to the Poisson distribution, with mean = kt (wherein k is the arrival rate), then the
lengths of the inter-arrival times (the time between two successive incidents) follow the exponential distribution, with a
mean of 1/k. Therefore, the exponential distribution is used for modeling of time sequential incident occurrence in Arena.
The generated incidents are then directed to each freeway node according to the spatial distribution.

5.3. Trafc ow simulation sub-module

Incident response operation is inuenced by trafc ows that vary with time. Varying trafc ows cause different con-
gestion levels, which are usually dened by volume/capacity (V/C) ratio. The V/C ratio can be directly used to estimate
the travel speed and travel time along the freeway segment during the patrol process modeling.
146 W. Wu et al. / Simulation Modelling Practice and Theory 47 (2014) 141151

Patrol vehicle Patrol vehicle

Current node Clear incident Current node Clear incident

Yes Yes

Incident at Yes Priority Incident at Yes Priority

current node? Incident? current node? Incident?

No No No No

Yes No No
Incident Priority incident Incident Yes Priority incident
elsewhere? elsewhere? elsewhere? elsewhere?

No Yes No Yes
Keep Roving Go to Incident Location Go to Dock Location Go to Incident Location

Fig. 4. Roving logic (left side) and prepositioning logic (right side).

In this paper, trafc ow is modeled with an HCM macroscopic approach, and the movements of the response vehicles are
also macroscopically tracked. Through aggregate trafc ow modeling, the inuence of trafc on the movement of response
vehicles could be sufciently captured, which saves computational time.
It is assumed that the RRs travel at the average speed of trafc with no incidents at the time of the incident response.
Incidents (particularly lane-blockage incidents) will result in speed reduction of the trafc stream. However, the RRs are
assumed not to be affected because they are allowed to travel on the shoulder when trafc is becoming congested.
The FDOT counts sites installed along I-95 record the hourly trafc, which is published in the FTI DVD. The hourly trafc
volumes obtained from the 2011 FTI DVD, combined with the number of lanes and speed limit information from a FDOT GIS
database are used to calculate the V/C ratio of different freeway segments. The methodology in the Highway Capacity Manual
2000 (HCM) [20] is adopted to estimate the travel speed and the travel time along the segments based on the V/C ratio.

5.4. Incident response simulation sub-module

Patrol districting determines the boundaries of each patrol beat, which is an area pre-assigned to specic patrol vehicles.
Patrol vehicles respond to incidents by following the predened dispatching policies and routing schemes inside their beat.
Dispatching policies determine whether the patrol vehicle will keep roving or stay at a dock location called prepositioning
[21], whether to provide distinctive responses to incidents with different priority levels, or to allow ignoring minor incidents
when responding to severe incidents. In this simulation model, the prepositioning scheme can be modeled in addition to the
traditional roving scheme. Under the prepositioning scheme, the patrol vehicle is always traveling back to the predened
dock location once the incident clearance is completed and no new incident is assigned. The ow chart in Fig. 4 describes
the logic of the two patrol vehicle response processes.
Routing schemes dene the route for patrol vehicles under different situations. Here, the shortest path is used for the
patrol vehicle response process. In this model, the patrol vehicles will keep moving along the freeway to reach the incident
location if the incident occurs at the same direction; or it will change to the other direction at the closest interchange if the
incident occurs at the opposite direction. If an incident is reported or detected on the back side of the current patrol vehicle
location, the vehicle needs to get off the freeway at the next off-ramp and then turn around to the other direction. Because
the freeway network in this study is mostly a straight line, the routing mechanism is simple, if the network is complicated,
more sophisticated algorithms should be adopted in this simulation model for searching the shortest path, such as Dijkstra s
label correcting algorithm [22] that is widely used in transportation modeling.
In the simulation model, incident detection and verication are simply implemented by adding the detected incidents
into the response queues according to the priority levels. Otherwise, the un-detected incidents will only be discovered by
patrol vehicles when they pass by the incident location. The incident clearance process is modeled by adding delay to the
patrol vehicle once it reaches the incident location. The delay corresponds to the incident clearance duration distribution.
The time between the incident generation and the patrol vehicle arrival is dened as the response time.

5.5. Performance evaluation sub-module

In Arena, a performance evaluation module is coded using the VBA script. The response time of each incident is recorded
and then aggregated by incident type and each beat. Work load is usually dened as the proportion of time the patrol vehicle
spends on incident clearance. Because the patrol vehicle will be delayed at the incident location until the incident is cleared,
 W. Wu et al. / Simulation Modelling Practice and Theory 47 (2014) 141151 147

it is natural to treat the work time as the patrol vehicle delay time that is also the incident clearance duration. The
performance evaluation module also records the incident clearance time by each patrol vehicle during the simulation, which
will be later aggregated to determine the average work load of each beat.
As discussed in previous section, the objectives of the simulation modeling are to reduce the response time and balance
the work load of patrol vehicles in each beat. Different parameter settings will be examined to nd the settings with better
performance.

6. Model application

The FDOT District 4 implements different patrol eet sizes during the weekday daytime, weekday nighttime, and week-
end. Preliminary analysis shows that more incidents occur during weekday daytime. Therefore, this study focuses on the
patrol process evaluation at the weekday daytime period.
In 2011, there were 26,338 shoulder incidents and 1583 lane-blockage incidents along the study I-95 corridor during a
total of 261 weekdays, representing an average of 100.91 shoulder incidents and 6.07 lane-blockage incidents per weekday.
The frequent occurrence of incidents along the I-95 segment conrms the importance of incident management. Table 1
shows the incident distribution by time of day and the expected inter-arrival time.
As discussed previously, in Arena, the random incident occurrences are scheduled in a time axis, and exponential distri-
butions are used for modeling the inter-arrival time between two incidents.
A review of the incident distribution by time of day shows that more incidents occur from 6:00 AM to 8:00 PM than other
time periods. A discussion with Road Ranger drivers also conrmed that there are still relatively more incidents occurring at
8:00 PM, and it is better to extend the patrol service daytime in weekdays from 7:00 PM to 8:00 PM. Therefore, the patrol
service period for the daytime shift from 6:00 AM to 8:00 PM, a total of 15 h, is determined for further simulation modeling.
The Gamma(k, h) distribution, a two-parameter distribution, is frequently a probability model for waiting time, with
shape parameter of k and scale parameter of h. The Gamma(k, h) has the mean of kh and variance of kh2, as well as the fol-
lowing form of probability density function:

ex=h
f x; k; h xk1 for x P 0 and h; k > 0 1
hk Ck
Based on the incident data, distributions for the incident clearance duration are tted with the input analyzer tool pro-
vided by Arena. The best tted distribution for shoulder incidents is a Gamma distribution of Gamma(25.9, 0.843) with a
mean of 21.8 min and variance of 18.4 min; lane-blockage incidents is also a Gamma distribution with an intercept
0.999 + Gamma(20.8, 2.31), mean of 48.0 min and variance of 111.0 min2. The data show that lane-blockage incidents have
longer clearance duration and higher variance.

Table 1
Incident distribution by time of day for 2011 weekdays.

Hour Shoulder incident Lane-blockage incidents

Total number of Number of incident/ Expected inter-arrival Total number of Number of incidents/ Expected inter-arrival
incidents weekday time (min) incidents weekday time (min)
0 480 1.8 32.63 38 0.15 412.11
1 325 1.2 48.18 20 0.08 783.00
2 250 1.0 62.64 16 0.06 978.75
3 197 0.8 79.49 22 0.08 711.82
4 248 1.0 63.15 26 0.10 602.31
5 243 0.9 64.44 24 0.09 652.50
6 1379 5.3 11.36 72 0.28 217.50
7 2020 7.7 7.75 116 0.44 135.00
8 1864 7.1 8.40 136 0.52 115.15
9 1710 6.6 9.16 81 0.31 193.33
10 1670 6.4 9.38 71 0.27 220.56
11 1346 5.2 11.63 59 0.23 265.42
12 1466 5.6 10.68 88 0.34 177.95
13 1268 4.9 12.35 74 0.28 211.62
14 1671 6.4 9.37 77 0.30 203.38
15 1555 6.0 10.07 84 0.32 186.43
16 1813 6.9 8.64 103 0.39 152.04
17 1700 6.5 9.21 140 0.54 111.86
18 1350 5.2 11.60 94 0.36 166.60
19 1005 3.9 15.58 64 0.25 244.69
20 828 3.2 18.91 44 0.17 355.91
21 531 2.0 29.49 55 0.21 284.73
22 666 2.6 23.51 44 0.17 355.91
23 753 2.9 20.80 35 0.13 447.43
148 W. Wu et al. / Simulation Modelling Practice and Theory 47 (2014) 141151

Because the simulation model is designed for planning purposes, replication of annual operations is required to capture
the incident occurrence patterns. As a result, annual patrol operation including 261 weekdays is conducted with the simu-
lation model. Aggregated simulation results of these 261 weekdays are calculated for model evaluation.
The I-95 section in Broward County covering 24.629 miles starts from the Ives Dairy Road and ends at W. Palmetto Park
Road with 18 interchanges in between. There are a total of 23 on-ramps and 24 off-ramps along the Northbound and 26 on-
ramps and 20 off-ramps along the Southbound.
There are no special incident spatial distribution patterns along the entire corridor except the interchange with interstate
freeway I-595. This location has more incidents compared with others because of the high trafc volume and complex trafc
movements.
The even-length and even-activity districting methods are replicated with the simulation model. Even-length districting
separates the beats with approximately even length and this was used by FDOT District 4 before 2011, while even-activity
districting that separates each beat by approximately even number of incidents, is used by FDOT District 4 after 2011.
To nd a better beat districting method, different scenarios are carried out by adjusting the beat boundaries. After many
tries, an improved districting method with the current roving scheme, which has less response time and well-balanced work
load, is determined. Table 2 presents the locations and length of beat start and end, as well as percentage of incidents and
performances of the three districting methods written in shortened forms of even length, even activities, and improved listed
in Table 3.
As shown in Table 2, the even-length districting method divides the four beats in an average length of approximately 6
miles, whereas the percentage of incidents varies from 19.62% to 33.46%; the even-activity districting method splits the four
beats in an average of 25% incidents, whereas the beat length varies from 4.178 to 8.374 miles. The improved districting
method sets up each beat boundary to minimize the response time and balance the work load; therefore, the even length
or even activities are not held. The length varies from 5.135 to 7.340 miles, and the percentage of incidents varies from
19.62% to 38.16%.
In Table 3, the average response time for shoulder incident is about 14.5 min and for lane-blockage incident is about
8 min. The improved districting performs better than both even-activity districting and even-length districting, while
even-activity districting performs slightly better than even-length districting. Particularly for lane-blockage incident, the
improved districting reduces the response time by about 10% (1 min) compared with even-length districting. In addition,
the work load for each beat is well-balanced in the improved districting within a range of 4.94% compared with 14.14%
for even-activity districting and 5.24% for even-length districting.
In this study, the prepositioning location is assumed at the middle of each patrol beat. Three scenarios are designed for
the prepositioning patrol scheme to assess the performance: two prepositioning vehicles with 55% TMC incident detection

Table 2
Beat districting results for the three scenarios.

Beat Even-length Even-activity Improved

Beat 1 Start SR 858 (Hallandale Beach Blvd.) SR 858 (Hallandale Beach Blvd.) SR 858 (Hallandale Beach Blvd.)
End SR 818 (Grifn Rd.) SR 818 (Grifn Rd.) SR 848 (Stirling Rd.)
Length 6.148 miles 6.148 miles 5.135 miles
Incidents 25.40% 25.40% 20.70%
Beat 2 Start I-595 East/West I-595 East/West SR 818 (Grifn Rd.)
End SR 816 (Oakland Park Blvd.) SR 838 (Sunrise Blvd.) SR 816 (Oakland Park Blvd.)
Length 6.327 miles 4.178 miles 7.340 miles
Incidents 33.46% 25.16% 38.16%
Beat 3 Start SR 870 (Commercial Blvd.) SR 816 (Oakland Park Blvd.) SR 870 (Commercial Blvd.)
End SR 814 (Atlantic Blvd.) SR 840 (Cypress Creek Rd.) SR 814 (Atlantic Blvd.)
Length 5.863 miles 5.929 miles 5.863 miles
Incidents 19.62% 20.78% 19.62%
Beat 4 Start Copans Rd. SR 814 (Atlantic Blvd.) Copans Rd.
End SR 810 (Hillsboro Blvd.) SR 810 (Hillsboro Blvd.) SR 810 (Hillsboro Blvd.)
Length 6.291 miles 8.374 miles 6.291 miles
Incidents 21.52% 28.66% 21.52%

Table 3
Performance measurements for each districting scenario.

Scenario Work load (%) Shoulder incident Lane-blockage incident

Beat 1 Beat 2 Beat 3 Beat 4 Response time (min) Change (%) Response time (min) Change (%)
Even-length 25.61 21.68 20.89 19.37 14.78 8.62
Even-activity 28.57 14.43 20.52 27.73 14.69 0.61 8.28 3.94
Improved 24.10 22.82 23.3 19.16 14.54 1.62 7.82 9.28
 W. Wu et al. / Simulation Modelling Practice and Theory 47 (2014) 141151 149

Table 4
Prepositioning routing scheme performances.

Scenario Work load (%) Shoulder incident Lane-blockage incident

Beat 1 Beat 2 Beat 3 Beat 4 Response time (min) Change (%) Response time (min) Change (%)
Improved (Basis) 55% detection rate pre-positioning 24.10 22.82 23.3 19.16 14.54 7.82
two vehicles
55% Detection rate pre-positioning two vehicles 22.45 24.71 19.12 17.28 39.20 169.60 40.47 417.52
100% Detection rate roving one vehicle and 23.56 25.60 20.46 18.76 17.91 23.18 13.2 68.80
pre-positioning one vehicle
55% Detection rate 22.99 24.66 20.09 18.29 17.82 22.56 10.75 37.47

rate, two prepositioning vehicles with 100% TMC incident detection rate, and one roving vehicle and one prepositioning vehi-
cle with 55% TMC incident detection rate. The model performance outputs for these three scenarios are presented in Table 4.
It can be concluded from Table 4 that the prepositioning scheme results have longer response time for both incident types
compared with the traditional roving scheme. The response time increases dramatically if both patrol vehicles are preposi-
tioned with TMC detection rate similar to the base condition. This is due to the lack of capability to detect incidents. How-
ever, even with 100% incident detection rate, the response time is still longer than that under the roving scheme. The last
scenario with one patrol vehicle roving and the other vehicle prepositioned shows better results than the other two prepo-
sitioning scenarios, but still has longer response time compared with the roving scheme.
Therefore, it is concluded that prepositioning could not improve the response time, but the patrol vehicle running time
will decrease because the vehicle is at idle status when it is not assigned to an incident. In future work, converting the run-
ning time savings to monetary value could help quantify the benets. Further analysis concerning both the quantied ben-
ets and the incident delay costs due to the response time increase needs to be performed to validate the applicability of the
prepositioning scheme.

7. Sensitivity analysis

Three categories of sensitivity analysis are performed: incident frequency change, different districting and different num-
ber of patrol vehicles in a beat, and different TMC detection rates.
The incident frequency change scenarios with the improved districting scenario as the basis are designed to test the
model performance when the number of incidents increases or decreases. The results shown in Table 5 indicate that reduc-
ing the incident frequency by 20% will lead to a shorter response time for shoulder incidents but not for lane-blockage inci-
dents. This may be due to the rarely occurring lane-blockage incidents, and the response time is not sensitive to the incident
frequency. In case the number of incidents increases by 20%, the response time for both types of incident signicantly
increases, particularly for lane-blockage incidents.
More extremely, a scenario of one beat-8 vehicles per beat is created, and adverse results are presented with more than
100% response time increase for both types of incidents.
A sensitivity analysis is performed to assess the impact of the TMC detection rate. Recently, automatic incident detection
techniques, such as the CCTV system, the incident detection algorithm, as a trend of the incident management, have been
adopted by more and more states. Currently, the Smart Sunguide TMC CCTV system could detect 55% of incidents. In this
analysis, the TMC detection rates vary from 0% to 100%, and results are shown in Table 6.
As shown in Table 6, the variation of response time for shoulder incidents is not signicant, but for lane-blockage inci-
dents, it varies from 11.01 to 7.36 min for 0% to 100% TMC detection rate. This is reasonable because the number of lane-
blockage incidents is far smaller than that of shoulder incidents. Therefore, the increase of incident detection rate will more
signicantly improve the response of lane-blockage incidents than the shoulder incidents.

Table 5
Sensitivity analyses results.

Scenario Work load (%) Shoulder incident Lane-blockage incident

Beat 1 Beat 2 Beat 3 Beat 4 Response time (min) Change Response time (min) Change (%)
Improved (basis) 24.10 22.82 23.3 19.16 14.54 7.82
Improved 20% 19.88 20.96 16.93 16.17 12.16 16.37 8.24 5.37
Improved + 20% 29.15 26.78 25.81 23.54 17.64 21.32 10.88 39.13
Four beats one vehicle per beat 44.47 45.28 38.17 36.81 45.19 210.80 23.48 200.26
Four beats three vehicles per beat 15.75 16.37 13.52 12.52 9.70 33.29 7.02 10.23
Two beats four vehicles per beat 22.28 22.82 17.82 22.56 12.11 54.86
Three beats three vehicles per beat 18.44 20.71 18.23 14.30 1.65 11.33 44.88
One beat eight vehicles per beat 19.28 31.93 119.60 17.07 118.29
150 W. Wu et al. / Simulation Modelling Practice and Theory 47 (2014) 141151

Table 6
TMC detection rate sensitivity analyses.

Scenario (%) Work load (%) Shoulder incident Lane-blockage incident

response time (min) response time (min)
Beat 1 Beat 2 Beat 3 Beat 4
0 24.16 11.29 20.66 18.45 12.31 11.01
25 23.43 15.71 20.59 18.47 14.13 9.62
45 23.49 21.32 19.87 19.37 14.78 10.24
55 (Basis) 24.10 22.82 23.3 19.16 14.54 7.82
65 24.46 26.50 20.36 18.77 15.58 8.71
75 23.37 31.71 19.58 18.64 14.45 8.83
85 24.38 34.11 20.18 18.34 14.06 8.34
100 24.15 34.30 20.52 18.74 12.92 7.36

16

14

12

10

0
0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%

Fig. 5. Response time change with TMC detection rate change.

Compared with the current 55% detection rate, the 100% percent detection rate only improves by 0.46 min, which is not
signicant as expected. This may be due to the fact that 55% is already a high detection rate, and any reduction of response
time could become more difcult through increasing the detection rate.
The response time change pattern due to the change of the TMC detection rate is difcult to be discovered based on the
data listed in Table 6. Therefore, those data are plotted in Fig. 5. With the plot as basis, the response time for lane-blockage
incidents is observed to show approximately consistent decreasing trend with the increase of detection rate, whereas the
response time for shoulder incidents increases at the beginning then decreases as TMC detection rate increases, which
reveals the complexity of the patrol response process. The possible explanation is that combining the impacts of high inci-
dent rate and high detection rate, the vehicles will always be directed to new discovered incidents and ignore the incidents
close to their current location, which affects the response efciency.

8. Conclusion section

8.1. Conclusion

A discrete event-based simulation model in Arena is developed to model the freeway patrol service process. The trafc
ow is modeled through HCM approach, and the patrol vehicle movements are modeled macroscopically. The model could
be used for new freeway service patrol program planning and existing program evaluation or improvement.
The application of the simulation model with the I-95 segment in Broward County data shows encouraging results. The
new districting method after adjustment of the beat boundary could reduce 9.28% incident clearance time, which can be
implemented in the eld. Furthermore, the prepositioning scheme is investigated for the very rst time. Although no signif-
icant improvements are identied, it provides instructions for future work on quantitative cost-benet analysis methods. At
the end, detailed sensitivity analysis is conducted, which provides insight understanding and valuable ndings about the
freeway service patrol with different inputs and parameter settings.

8.2. Future work

Continuous efforts are needed to improve the simulation model which includes incorporating more patrol schemes, dif-
ferentiating the severity of lane-blockage incidents, and providing higher priorities for incidents with more lanes blocked.
It would also be benecial to research embedding automotive alternative searching algorithms or evolution computation
into the simulation model and developing routing schemes for network form freeway system.
 W. Wu et al. / Simulation Modelling Practice and Theory 47 (2014) 141151 151

References

[1] J. Lindley, Urban freeway congestion: quantication of the problem and effectiveness of potential solutions, ITE J. 57 (1987) 2732.
[2] C. Zhan, L. Shen, M. Hadi, A. Gan, Understanding the characteristics of secondary crashes on freeways, in: Presented at 87th Annual Meeting of the
Transportation Research Board, Washington, DC, 2008.
[3] Road Safety is No Accident, Red Cross, Red Crescent, Switzerland, 2005. <http://www.redcross.int/EN/mag/magazine2005_2/4-9.html> (accessed
29.12.13).
[4] K. Kaas, Benet of Trafc Incident Management, National Incident Management Coalition (NTIMC), 2006. <http://www.transportation.org/sites/ntimc/
docs/Benets11-07-06.pdf> (accessed 29.12.13).
[5] K.G. Zografos, T. Nathanail, P. Michalopoulos, Analytical framework for minimizing freeway-incident response time, J. Transport. Eng., ASCE 119 (1993)
535549.
[6] R. Pal, K. Sinha, A Framework for Locating Highway Incident Response Vehicles in Urban Areas, INFORMS National Meeting, San Diego, California, 1997.
[7] H.D. Sherali, S. Subramanian, Opportunity cost-based models for trafc incident response problem, J. Transport. Eng., ASCE 125 (3) (1999) 176185.
[8] A. Khattak, N. Rouphail, K. Monast, J. Havel, Method for priority ranking and expanding freeway service patrols, Transport. Res. Rec., J. Transport. Res.
Board 1867 (2004) 110.
[9] Y. Yin, Optimal eet allocation of freeway service patrols, J. Netw. Spatial Econ. 6 (2006) 221234.
[10] Y. Yin, Scenario-based model for eet allocation of freeway service patrols, J. Netw. Spatial Econ. 8 (2008) 407417.
[11] T.G. Nathanail, K.G. Zografos, Simulation tool for evaluating effectiveness of freeway incident response operations, Transport. Res. Rec., J. Transport.
Res. Board 1485 (1995) 105111.
[12] K.G. Zografos, K.N. Androutsopoulos, G.M. Vasilakis, A real-time decision support system for roadway network incident response logistics, Transport.
Res. Part C 10 (2002) 118.
[13] R. Pal, K. Sinha, Simulation model for evaluating and improving effectiveness of freeway service patrol programs, J. Transport. Eng., ASCE 128 (4) (2002)
355365.
[14] K. Ozbay, B. Bartin, Incident management simulation, Simulation 79 (2) (2003) 6982.
[15] A. Haghani, Q. Tian, H. Hu, Simulation model for real-time emergency vehicle dispatching and routing, Transport. Res. Rec., J. Transport. Res. Board
2004 (1882) 176183.
[16] M. Hadi, L. Shen, J. Li, Y. Xiao, D. Chen, Simulation tool to assess the impacts of the parameters of service patrol operations, Transport. Res. Rec., J.
Transport. Res. Board 2178 (2010) 111118.
[17] W. Kelton, R. Sadowski, D. Sturrok, Simulation with Arena, fourth ed., McGraw-Hill, New York, NY, 2007.
[18] A. Jaoua, D. Riopel, M. Gamache, A simulation framework for real-time eet management in internal transport systems, Simulat. Model. Pract. Theory
21 (2012) 7890.
[19] A. Motraghi, M. Varhanov Marinov, Analysis of urban freight by rail using event based simulation, Simulat. Model. Pract. Theory 25 (2012) 7389.
[20] Highway Capacity Manual 2000, Transportation Research Board, National Research Council, Washington, DC, 2000.
[21] A.J. Ballard, Incident Management for Managed Lanes. FHWA/TX-05/0-4160-17. Texas Transportation Institute, the Texas A&M University System,
2004. <http://tti.tamu.edu/documents/0-4160-17.pdf> (accessed 29.12.13).
[22] E.W. Dijkstra, A note on two problems in connexion with graphs, Numer. Math. 1 (1959) 269271.