Article

Transportation Research Record

1–10
Ó National Academy of Sciences:
Integrated Incident Decision-Support Transportation Research Board 2018
Reprints and permissions:
using Traffic Simulation and sagepub.com/journalsPermissions.nav
DOI: 10.1177/0361198118782270

Data-Driven Models journals.sagepub.com/home/trr

Tao Wen1, Adriana-Simona Mihăitxă1, Hoang Nguyen1, Chen Cai1,

and Fang Chen1

Abstract
This paper introduces the framework of an innovative incident management platform with the main objective of providing
decision-support and situation awareness for transport management purposes on a real-time basis. The logic of the platform
is to detect and then classify incidents into two types: recurrent and non-recurrent, based on their frequency and characteris-
tics. Under this logic, recurrent incidents trigger the data-driven machine learning module which can predict and analyze the
incident impact, in order to facilitate informed decisions for transport management operators. Non-recurrent incidents acti-
vate the simulation module, which then evaluates quantitatively the performance of candidate response plans in parallel. The
simulation output is used for choosing the most appropriate response plan for incident management. The current platform
uses a data processing module to integrate complementary data sets, for the purpose of improving modeling outputs. Two
real-world case studies are presented: 1) for recurrent incident management using a data-driven model, and 2) for non-
recurrent incident management using traffic simulation with parallel scenario evaluation. The case studies demonstrate the
viability of the proposed incident management framework, which provides an integrated approach for real-time incident
decision-support on large-scale networks.

An initiative of great value in the era of ‘‘Big Data’’ is to the incident detection and classification as an example,
effectively and consistently generate insights and action- which is further elaborated in this paper, several data
able outcomes from multiple data sources. This would sources would complement each other for improving the
typically require several data handling processes, includ- reliability and accuracy. Related data sources may
ing data forwarding, parsing, and integration, before include real-time traffic data, real-time incident monitor-
data becomes usable by dedicated applications, such as ing data, crowd-sourced data, and so forth. Without a
machine learning models. These data handling processes well-developed mathematical modeling framework that
were conventionally deployed to enterprise data infra- implements several fundamental principles, such as the
structure, such as the enterprise data warehouse (EDW). decision tree and support vector machines, it is unlikely
With the advent of cloud-based infrastructure being epi- the generic analytics would arrive at a deterministic and
tomized by Amazon Web Service (AWS) and Azure, data meaningful outcome.
handling, storage, and computing are increasingly being In this paper, we present a part of the Advanced Data
integrated into a cloud data platform. Such a platform Analytics in Transport (ADAIT) platform (1). ADAIT
offers agility, scalability, and serviceability with substan- is a cloud data platform hosted on AWS and dedicated
tial financial cost incentives, when compared with EDW to urban transport data analytics. In particular, we
solutions. Generic machine learning analytics are also address the innovative approach of integrating data ana-
being integrated with cloud data platforms, for the bene- lytics with traffic simulation for incident decision-sup-
fit of generating valuable insights on-demand. Examples port. This approach exploits the distinction between
of these are IBM Watson and Google Cloud Platform.
Of greater complexity from the generic cloud data 1
DATA61, CSIRO, New South Wales, Australia
platform are the domain platforms. The latter would
require a considerable degree of knowledge on a particu- Corresponding Author:
lar domain, for example, urban transportation. Using Address correspondence to Tao Wen: tao.wen@data61.csiro.au
2 Transportation Research Record 00(0)

recurrent and non-recurrent traffic incidents. With an the general flowchart of the incident detection frame-
adequate number of observations, information about a work, which comprises various types of interconnected
recurrent incident can be parsed into several prominent modules, from the raw data processing and cleaning to
features and represented by a feature space. Data-driven the machine learning predictive module and traffic simu-
models, that is machine learning models, could then be lation core.
built to map the feature space into observed impact. The The occurrence of incidents is one of the main factors
more observed repetitions of the recurrent incident, the that affect urban transport mobility (2). To reduce the
better the data-driven model could explain the variance congestion caused by incidents, an incident decision-
in the incident impact. On the other hand, for a non- support system should have a prompt response time to
recurrent incident, building a viable data-driven model help operators make timely decisions. Various models
may be prohibitive because of the lack of data. Instead, and methods have focused on signal priority (3), sub-
the detected incident information could be used to popu- network and tunnel segment (4) or incident management
late a simulation model. The latter can then run multiple system infrastructure (5). However, scalability remains
scenarios in parallel, faster than real-time, to provide the an issue. In addition, some simulation-based incident
expected impact of the incident. This paper discusses decision-support systems can provide a quantitative esti-
how the two modeling techniques could be integrated via mation of different traffic operations, but struggle to
the ADAIT platform to provide incident decision- provide automated situation awareness (6). The major
support capabilities. characteristic of the ADAIT platform is that it offers a
continuous situation awareness by real-time monitoring
of the traffic condition, and triggers the data-driven pat-
Incident Decision-Support Framework
tern reconstruction or traffic simulation modules only
The work presented in this paper focuses on incident when reported incidents appear in the network. If no
detection and impact analysis on the traffic network and incident occurs, the platform will continue monitoring
is a part of the ADAIT general framework, currently and reporting the traffic state in the network. All outputs
under development in collaboration with the Transport of each module are stored in a dedicated cloud database
Management Centre (TMC) of Sydney. Figure 1 presents and the results are provided for consumption and further

Figure 1. Incident detection and impact analysis framework.

Wen et al. 3

use through public APIs. Amongst the modules of the 5. Response plan evaluation module, which chooses
platform, we detail the following: the best plan to mitigate the impact of the inci-
dent. The simulated outputs will be compared
1. Incident detection and classification module, which with the observed traffic data, which will enable
comprises a) real-time data fusion from various the transport operators to make informed deci-
sources for detecting recent incidents, b) incident sions by various scenario testing and evaluation.
ranking algorithms based on type and severity,
and c) incident duration classification based on Combining these various modules with different char-
available characteristics of the reported accident. acteristics and inputs/outputs, from data-driven incident
The main outcome of this module is to determine detection/classification/prediction to automatic traffic
if the reported incident follows a recurrent pat- simulation models, represents a unique and innovative
tern or not to support decision-making for the method to evaluate the impact of incidents in a highly
operator. It also estimates the incident duration affected traffic network. The advantages of the proposed
and severity to efficiently trigger the simulation platform in the aforementioned perspectives are:
module for impact analysis and response plan
evaluation. 1. Integrating and fusing multiple data sources to
2. Machine learning data-driven predictive module, improve the reliability and accuracy of results.
which is based on historical observations of both 2. Applying machine learning algorithms to learn
transport and incident log data, and can predict from previous incidents and make predictions.
the severity and duration of the detected incident. 3. Detecting and classifying incidents based on real-
3. Automatic simulation module, which is triggered time data. The innovation is to only trigger the
only if the incident pattern is identified as non- simulation model when the impact of a new inci-
recurrent. In this case, the incident feature infor- dent could not be determined.
mation and candidate response plans are passed 4. Automated simulation to provide quantitative
to the simulation module. The module then auto- evaluation of incident impact and response plan
matically selects the sub-network where the inci- performance.
dent took place based on the incident feature
information. It further generates a traversal The logic and mechanism behind the platform are to
demand matrix for the selected sub-network provide a viable and practical solution to incident man-
based on the path assignment obtained from pre- agement. Such an incident management system needs a
viously running regular macroscopic static traffic detailed decomposition and theoretical analysis in terms
assignment on the large-scale network. This will of processing flow, traffic modeling, predictive analytics,
facilitate the application of mesoscopic or micro- and response plan selection for mitigating the impact of
scopic traffic scenarios at the sub-network level traffic incidents on both normal and public transport
for a more detailed simulation of the condition in modes. In this paper, we will only focus on detailing the
the affected area while maintaining computation major modules for incident impact analysis and the way
efficiency. Each response plan is simulated in par- the information flow is propagated from one module to
allel and the corresponding outcome will be eval- another with a major focus on the incident detection,
uated according to user-defined performance duration prediction, and response plan evaluation.
metrics. Such parallel simulations can signifi-
cantly reduce computation time, while providing
Module Details
quantitative insights on a mesoscopic or micro-
scopic level. It is worth mentioning that this type In this section, we describe each module of the decision-
of simulation module requires periodic calibra- support platform, as well as the implications and the
tion and validation of the simulation outputs by afferent data sources needed for obtaining accurate
using existing data sources. insights.
4. Multimodal Origin-Destination (O-D) matrix esti-
mation module, which estimates the multimodal
O-D demand of large-scale network based on his-
Data Processing Module
torical transport data. The output of this module The data sets used in the platform have a wide variety of
is a fundamental input for subsequent analysis formats and specifications, are often sparse, and need
and is critical to the simulation module’s constant cleaning and monitoring. Amongst them, we
performance. cite:
4 Transportation Research Record 00(0)

a) Survey data such as household travel survey data, the survey data are used as prior estimates of O-D matrix
census survey data, and so forth, which are very and are calibrated based on traffic counts (11). However,
time-consuming and become available every few the integration of the multimodal O-D estimation into
years; they can provide insights on the O-D the automated incident management platform is a novel
demand but may also be obsolete. On this plat- approach that offers valuable input for the simulation
form, survey data is processed and provided to module.
the multimodal O-D matrix estimation module as
prior estimates of O-D matrix.
b) Traffic counts are provided by the Sydney Incident Detection and Classification Module
Coordinated Adaptive Traffic System (SCATS) Incident Detection. Incident detection is considered an
system. The integration of SCATS data into the important component of many modern intelligent trans-
platform is challenging because of a high com- port systems. Multiple data sources may provide comple-
plexity and duplication of streams, which needs mentary data, and data fusion can produce a better
supplementary processing, outlier detection, error understanding of the observed situation by decreasing
elimination, and so forth. the uncertainty related to the individual (9). Traffic con-
c) Smart transit card data provides tap-on and tap- trol operators can set a threshold to trigger and clear an
off information of each user traveling in the city incident alert using the alert score. The method to esti-
by public transport modes. mate the alert score is explained below in Equation 1,
d) Public transit monitoring data includes informa- while the data used in this paper are explained later.
tion such as fixed-route schedules, routes, and bus By analyzing the historical data sources along with
stop data. The General Transit Feed Specification confirmed incident logs, the important factor of the indi-
provides such information on the Sydney network vidual source is evaluated using attribute ranking algo-
and is used for the validation of simulation rithms (e.g., information gain, principal component
results. analyses) (12).
e) Crowd-sourced data such as Twitter and Waze Besides the reliability of the individual source, the
data are used on the platform as they provide tex- alert score is dependent on spatial and temporal aspects.
tual information on incidents, and are processed Furthermore, a recent detection (e.g., within 10 minutes)
directly by the incident detection and classifica- should receive more attention than a past report (e.g.,
tion module. reported over 30 minutes ago). As a consequence, the
f) Real-time incident monitoring data provided by the spatial-temporal adjusted weight should be considered.
TMC, which reports incidents and actions taken The total alert score for an incident over the time is cal-
on a real-time basis. This data is used by the inci- culated as:
dent detection and classification module.
g) Travel time data can be provided by many ITS X
n
applications; on the current platform Google score(t) = Ri Wij S ðrÞT (t) ð1Þ
Travel Time is used for the validation of the simu- i=1

lation outputs. where, i represents data source i, Ri is the reliability score

for source ith, Wij is the weight for subtype of source i,
Multimodal O-D Matrix Estimation Module S(r) is the spatial adjustment function which depends on
road type r, and T(t) is the temporal adjustment which
The origin-destination demand matrix, which represents
depends on time t.
the number of trips from one urban centroid to another,
is a fundamental element in many transport models.
Having a reliable and accurate O-D matrix can signifi- Incident Classification. When an incident is detected and
cantly enhance the prediction quality (7). In this paper, confirmed by the system, a classification process is
road traffic demand and public transport demand are applied based on the incident description to divide it into
estimated independently, and data fusion techniques are different categories including accident, breakdown,
applied in both estimations. Multimodal O-D matrix delay, and so forth. Historical incident data are also used
estimation algorithms accounting for modal choice to train the machine learning incident classification
require additional information on behavioral analysis model to further classify incidents based on duration and
and choice modeling (8) and is not adopted here because severity.
of computation complexity and data availability. For incident duration classification, our system
Researchers have proposed various methods for data applies the method proposed by Taib et al. (13) to predict
fusion in public transport (9, 10), even the traditional the duration of an incident that has just happened by
road traffic O-D estimation is an example of data fusion; using the available characteristics known at the onset, for
Wen et al. 5

example, location, time, type of incident, lanes affected, set, the time interval was five minutes and segment
operator in charge, and so forth. Understanding the esti- lengths were predefined by the data supplier.
mated incident duration is useful to traffic operators To build the DBN traffic network, the road segments
when choosing an appropriate response plan. Estimated are presented by a set of Nh random variables,
incident duration is also one of the inputs for the traffic O(i)
t 2 ½0, 1, where i represents a congested segment at
simulation module. Although several machine learning time t. Snapshot t is a storage of the traffic condition
detection methods (14, 15) have been applied to detect an from all segments in the network at time t.
incident, few of them combined advanced machine learn- In a DBN, the transition (denoted as B!) and obser-
ing, active learning, and outlier detection techniques and vation model P(Zt jZt1 ) is then defined as a product of
achieved approximately 90% accuracy in predicting inci- the conditional probability distribution (CPD) in the
dent severity (16). This severity classification approach is 2TBN:
integrated into our proposed system.
Finally, the detailed incident impact on the road net- Y
N
ð iÞ ðiÞ
work is then predicted by machine learning approaches P(Zt jZt1 ) = P(Zt jPa(Zt )) ð2Þ
i=1
(for recurrent incidents/congestions) or simulation mod-  
ules (for unseen incidents). ðiÞ ð iÞ
where Zt is the ith node in snapshot t and Pa Zt are
ðiÞ
the parents of Zt . The unconditional initial state distri-
bution P Z11:N is presented by a standard Bayesian
Machine Learning Data-Driven Predictive Module Network, namely B1. Together, B1 and B! define the
When incidents occur in an urban traffic network, they DBN.
are likely to affect the traffic flows of surrounding areas, Suppose we have a simple traffic network which com-
especially the traffic leading to the congested roads. prises three segments: EB and GB are connected to BA.
Some of the causal congestions follow the same patterns As EB and GB both lead to BA, when BA is congested,
or sequences over the time. Therefore, it is useful to dis- it becomes the potential cause for congestions at EB and
cover frequent patterns (if any) of congestion propaga- GB in the next time frame. The corresponding DBN net-
tions by reviewing historical data in the traffic networks. work is presented in Figure 2.
This section reviews an algorithm that finds a conges- When the propagation pattern is generated, the joint
tion propagation pattern by looking at the relationships distribution for a known-structure tree which includes T
of congestions from the earliest data record through the consecutive snapshots (slices) can be obtained by ‘‘unrol-
latest one (17). The main insight is that congestion C1 is a ling’’ the network until we have T slices, and then multi-
parent of congestion C2 if C1 occurred before C2 in time plying together all of the CPDs:
and they are spatially connected.
A frequent tree represents an expected congestion pro-   Y
N
(1:N ) ðiÞ ðiÞ
pagation pattern when an incident happens on a root seg- P Z1:T = PB1 (Z1 jPa(Zt )
i=1
ment. A root of a congestion tree is defined as a segment ð3Þ
Y
T Y
N
where traffic from other segments is flowing into it and 3
ð iÞ ðiÞ
PB! (Z1 jPa(Zt )
causing congestion. t=2 i=1
A Dynamic Bayesian Network (DBN) approach is
used to model the spatial–temporal characteristics of a
recurrent congestion propagation. Using this method,
the probability of a recurrent incident’s impact can be
estimated. A DBN is usually referred to as a Two-
Timeslice Bayesian Network (2TBN) because at any
given time T, the value of a variable, is computed from
the internal regressors and the immediate prior value
(time T-1) (18). Therefore, DBN is reasonably close to
the real-world phenomenon of traffic congestion in
which the status of a segment at a specific time can be
determined by its previous condition and previous condi-
tions of connecting segments. However, this assumption
depends on the length of the time interval and traffic seg-
ment. If either the time interval or traffic segment is too
long or too short, the dependency may not be applicable. Figure 2. Modeling congestion propagation of a three-segment
In our experiment, because of the availability of a data traffic network by DBN.
6 Transportation Research Record 00(0)

In case the detected incident belongs to the root of the This is because transport operators can have multiple
congestion propagation tree and the probability to form candidate response plans, and the simulation module
a propagation pattern is higher than a predefined thresh- can enable them to choose the most suitable plan. On
old, the traffic controller may decide to rely on this this platform, the mesoscopic simulation has been
impact pattern to control the traffic rather than execut- adopted because it is computationally more efficient
ing simulation, which is more time-consuming. If there is than microscopic simulation, and less data-demanding
no propagation that exceeded the predefined threshold, while capturing the essentials of the traffic dynamics.
the simulation technique is applied to test the impact of Note that all the aforementioned simulation processes
an incident on the road network. When the system is are automated from the beginning, which is also critical
implemented, a real-world threshold will be suggested by to reducing the computation time to satisfy the strict
TMC to decide when it is reliable to use the predicted requirement on response time for incident management
patterns. purposes.
The reliability of this decision-making process relies
on a simulation model that represents the system’s
Simulation Module behavior closely enough (26), and therefore the simu-
In the real world, the location, type, and severity of an lated outputs are validated with several data sources:
incident may vary significantly and therefore several inci- travel time data (e.g., probe vehicle data, taxi data, or
dents may not be accurately and reliably predicted by Google Travel Time data), traffic counts (e.g., loop
the machine learning data-driven predictive module. In detector data), and public transport data (e.g., smart
this case, the simulation module needs to be activated to card data and public transport monitoring data). If the
evaluate the impact of the incident quantitatively. results do not closely approximate the observations,
However, simulation of large-scale networks in a reason- further calibration is conducted to fine-tune the para-
able time can hardly be viable because of the computa- meters such as sub-network traversal demand, meso-
tional complexity. Some real-time traffic simulation scopic reaction time, and so forth, until the result
models have opted for simulation at a corridor/motor- quality is satisfactory.
way level (19–21), but few offer automatic sub-network
selection and real-time incident simulation based on
duration prediction. Response Plan Evaluation Module
Therefore, we have proposed an automated parallel
As each city or traffic management center has its own
simulation module to address these issues. In this mod-
unique characteristics and preferences, it is important
ule, a large-scale traffic simulation model is constructed
that the best response is chosen based on bespoke
for the city of Sydney, Australia, and a macroscopic
metrics, so that the ramification of an incident is miti-
assignment process is implemented every day based on
gated in a user-defined way. For example, the average
the new incoming data received from the data processing
travel time per kilometer can be used to evaluate
module. Macroscopic assignment models have been well
performance:
studied by many researchers and many existing models
are available (22); in the case study, the algorithm by PNnetwork
Florian is used (23). i = 1 TTi
TTnetwork = ð4Þ
To achieve a reasonable computation time, the large- Nnetwork
scale network is subdivided into small sub-networks. For
where, TTi is the average travel time per kilometre of
example, the Bureau of Transport Statistics in New
vehicle i, Nnetwork is the number of vehicles in the network.
South Wales, Australia, applies a geometrical zoning
When travel time reliability or other factors are consid-
configuration which is open to the public (24). In addi-
ered important, they may also be used to choose the most
tion, some research papers also provide methods for
effective response plan. Such a user-defined metric embo-
automatic zoning when zoning information is not avail-
dies the platform’s flexibility and customizability.
able (25). Based on the incident information (location,
severity, lanes affected), the module automatically selects
the afferent sub-network, where further analysis and pre-
Case Study
diction are applied. Then, the traversal demand matrix is
generated for the selected sub-network. Here, the traver- Here, we present a case study of the Sydney large-scale
sal demand matrix represents the travelers that travel network consisting of more than 70,000 links and 2,000
through, in, and out of the sub-network. To further centroids. This case study focuses mainly on the incident
reduce the computation time, each response plan is simu- detection and classification, recurrent incident impact
lated simultaneously by the parallel simulation module. prediction, and non-recurrent incident simulation.
Wen et al. 7

became congested. Until 8:45 a.m., which is 15 minutes

after the incident happened, similar congestion propaga-
tion patterns were detected on all five segments. The case
study shows the efficiency and capability of the machine
learning module in decision-support and impact predic-
tion for recurrent incidents.

Non-Recurrent Incident: Parallel Simulation and

Performance Evaluation
Although the machine learning data-driven predictive
module shows its efficiency in predicting incident impact,
several incidents may not be predicted correctly when lit-
tle historical information is available. The simulation
module needs to be triggered to assist transport operators
when such non-recurrent incidents are detected. As previ-
ously explained, simulation of the whole Sydney network
on a mesoscopic or microscopic level is extremely time-
consuming, which cannot satisfy the computational time
requirements for transport management purposes.
Therefore, the Sydney network has been subdivided into
many sub-networks beforehand using the Statistical Area
definition in (24), so that the simulation module can
select a sub-network promptly for simulation. The trans-
port simulation model of the Sydney network is imple-
mented in Aimsun, which is regularly calibrated using
Figure 3. The frequent congestion propagation pattern in the periodically aggregated SCATS traffic counts and smart
Sydney CBD. card data. The Aimsun network simulation model uses
the multimodal O-D matrix previously estimated and
Recurrent Incident: Congestion Propagation and runs a macroscopic multimodal traffic assignment for the
Impact Prediction Sydney network. The output will be saved for later use in
When an incident is detected, the incident detection and generating a traversal demand for the chosen sub-
classification module will also analyze and determine network where the incident happens.
whether the incident is recurrent or not. Here, an exam- The non-recurrent incident presented here is reported
ple of a recurrent incident is presented, to demonstrate by the incident detection and classification module with
that the machine learning data-driven predictive module the following information:
can predict the congestion propagation and incident
impact for transport operators to make informed a. Location (including x and y coordinates):
decisions. Pyrmont Bridge Road, Pyrmont.
As illustrated in Figure 3, at 8:30 a.m. on a weekday, b. Estimated duration: 30 minutes.
an incident was detected at George Street near Campbell c. Severity: major accident affecting all lanes in both
Street in the Sydney central business district (CBD). directions.
Given that the incident first happened at the root seg- d. Start time: 07:15 a.m.
ment BA, using a congestion discovery algorithm and e. Incident pattern: non-recurrent.
DBN, a five-segment congestion propagation pattern
was detected with a joint distribution probability esti- The non-recurrent incident pattern triggers the simula-
mated at 74%. This probability is higher than the prede- tion module. To demonstrate the network state and per-
fined threshold, therefore this predicted impact can be formance before and after the incident, a two-hour
used by the operator to manage the incident. simulation period is chosen, from 7:00 to 9:00 a.m.
The detected congestion pattern was then validated Using the incident location, a sub-network has been
using the real-time traffic data collected from SCATS. automatically selected from the list of available sub-
The congestion was initially detected at segment BA. networks in the city, which is identified as Pyrmont.
Five minutes later, both segments DB and CB also Pyrmont is a suburb adjacent to Sydney CBD and the
8 Transportation Research Record 00(0)

Figure 4), and redirect all traffic in intersection 1

toward surrounding intersections.
4) RP4: Activate the VMS to redirect all off-ramp
flow from the bridge toward Little Mount Street.

RP1 is intended to keep monitoring the network but

take no action, in order to evaluate the true impact of
the incident if no action is taken. RP2 redirects all traffic
in intersections 1 and 2 toward adjacent intersections to
prevent vehicles from queuing and eventually blocking
the intersection. RP3 has the role of activating the VMS,
which will inform drivers to make a left turn before
reaching intersection 1. Also, traffic will be redirected at
intersection 1 to prevent queuing. RP3 aims to let the
major traffic from the bridge (Western distributor)
bypass intersection 1. RP4 simplifies RP3 by keeping
only the VMS activation action. Note that the para-
meters for driver behavior in the microsimulation (such
as acceleration and reaction time) will remain unchanged
because of the difficulty in collecting sufficient data.
Figure 4. The selected sub-network (Pyrmont, New South
Wales). These response plans are then simulated in parallel on
a microscopic level by the microsimulation engine in
Aimsun. Figure 5c presents the average travel time per
majority of its area is zoned for commercial purposes kilometer (including buses and private vehicles) obtained
(see Figure 4). after applying each of the four response plans. RP1 is the
After automatically selecting the sub-network area in baseline and demonstrates that the average travel time
which the incident has occurred, the traversal demand reaches a high point at 8:00 a.m., when the incident has
matrix for Pyrmont is generated and calibrated for the already ended. This is due to the accumulation of queue
morning peak hour (7:00 to 9:00 a.m.). The model is also and the increase in traffic and public transport demand.
validated by comparing the average travel time obtained RP4 shows a very marginal improvement over RP1; the
from simulation (STT) on each road section with the average travel time over the two hours is also quite close
average travel time from Google (GTT) or the average to RP1. Although RP2 mitigates the congestion between
travel time obtained from the SCATS data in Pyrmont 7:45 a.m. and 8:15 a.m., the travel time increases gradu-
(SCATSTT). Figure 5a presents an example of compari- ally after 8:15 a.m., making the eventual average travel
son between the average STT and GTT on the road sec- time over the two hours very similar to RP1. Overall,
tion 2839_2840 from Pyrmont, on a Wednesday morning RP3 performs the best; it smooths the travel time after
from 7:00 to 9:00 a.m. The plot of travel time every 15 the incident happens, while having a 7% reduction in
minutes indicates that the simulation provides good average travel time over the two-hour simulation period.
results of the TT on this section as it falls between the The finding indicates that the best response plan for
5th and 95th percentile of the GTT. This finding is vali- mitigating congestion produced by a non-recurrent inci-
dated once more on a different section (5_2839), where dent is actually a combination of various actions that
the STT is compared with GTT SCATSTT, which is complement each other and help to reduce the incident
available for computation (Figure 5b). clearance time. Therefore, a possible extension of this
Based on the received incident information, assume work is to automatically recommend the best combina-
the operators choose the following incident response tion of response plans to apply for efficiently easing
plans (RPs) for evaluation: congestion.

1) RP1: Do nothing.
Conclusion
2) RP2: Redirect all traffic in intersections 1 and 2
(marked as red rectangular in Figure 4) towards In this paper, we introduced the general framework of
adjacent intersections. the ADAIT platform and explained the main function of
3) RP3: Combined actions: activate the variable each module. The platform can detect and then classify
message sign (VMS) to redirect all off-ramp flow incidents into a recurrent and non-recurrent pattern; the
from the bridge toward Little Mount Street (see former triggers the machine learning data-driven
Wen et al. 9

be predicted by the machine learning module, and the

simulation module can help choose the best response
plan to mitigate the negative incident impact. In short,
data-driven incident detection/classification, machine
learning analytics for incident prediction, and automatic
traffic simulation models are integrated into the cloud-
based platform, which represents a unique and innova-
tive method to evaluate the impact of incidents in real-
time for large-scale networks.
There are various possible opportunities to further
extend the platform: the response plans can be gener-
ated automatically by advanced machine learning
techniques based on the information of a detected inci-
dent (such as location, duration, severity, etc.), and
therefore save time on manually input response plans.
Also, the platform’s modularity allows integration of
advanced transport algorithms in each module, which
can enhance the platform’s applicability. Because the
DBN model depends on the length of the time interval
and traffic segment, in our future work, the time inter-
val between two continuous snapshots will be consid-
ered as an additional parameter. Furthermore, the
algorithm will be tested on different traffic network
settings with different average segment lengths to eval-
uate the effect.

Acknowledgments
The work presented in this paper is partially funded by the
New South Wales Premier’s Innovation Initiative. The authors
of this work are grateful for the work and support of the
Traffic Management Centre from Transport for New South
Wales Australia. Data61 is funded by the Australian Federal
Government through Commonwealth Scientific and Industrial
Research Organisation.

Author Contributions
The authors confirm contribution to the paper as follows: study
conception and design: Tao Wen; data collection: Tao Wen,
Hoang Nguyen, Adriana-Simona Mihăitxă; analysis and inter-
pretation of results Tao Wen, Hoang Nguyen, Adriana-Simona
Mihăitxă; draft manuscript preparation: all authors. All authors
Figure 5. Average travel time comparison between (a) reviewed the results and approved the final version of the
simulation and Google Travel Time, (b) simulation, Google TT, and manuscript.
SCATS, and (c) each response plan (RP1, RP2, RP3, and RP4).

References
predictive module, which predicts the incident duration
1. ADAIT. Advanced Data Analytics in Transportation.
and impact, so transport management operators can
Data61, CSIRO. www.adait.io.
decide if the simulation module needs to be activated. A
2. Steenbruggen, J., P. Nijkamp, J. M. Smits, and M. Grothe.
non-recurrent incident is directly passed to the simula- Traffic Incident Management: A Common Operational Pic-
tion module; the performance of candidate response ture to Support Situational Awareness of Sustainable Mobi-
plans is evaluated quantitatively, and then operators can lity. 2011.
opt for the best plan to mitigate the negative impact of 3. Ahmed, F., and Y. Hawas. An Integrated Real-time Traf-
an incident. Case studies demonstrate that the impact of fic Signal System for Transit Signal Priority, Incident
a recurrent incident, such as congestion propagation, can Detection and Congestion Management. Transportation
10 Transportation Research Record 00(0)

Research Part C: Emerging Technologies, Vol. 60, 2015, 16. Nguyen, H., C. Cai, and F. Chen. Automatic Classification
pp. 52–76. of Traffic Incident’s Severity Using Machine Learning
4. Mitrovich, S., G. Valenti, and M. Mancini. A Decision Sup- Approaches. IET Intelligent Transport Systems, Institution
port System (DSS) for Traffic Incident Management in of Engineering and Technology, 2017.
Roadway Tunnel Infrastructure. Proc., 34th European Trans- 17. Nguyen, H., W. Liu, and F. Chen. Discovering Congestion
port Conference (ETC), Strasbourg, France, 2006, p. 240. Propagation Patterns in Spatio-temporal Traffic Data.
5. Li, H., H. Lu, and I. A. Qureshi. Intelligent Decision Sup- IEEE Transactions on Big Data, Vol. 3, No. 2, 2017, pp.
port System for Traffic Incident Management. 9th Interna- 169–180.
tional Conference on Applications of Advanced Technology 18. Ghahramani, Z. Learning Dynamic Bayesian Networks.
in Transportation (AATT), Chicago, IL, 2006, pp. 292–297. Adaptive Processing of Sequences and Data Structures,
6. Yurshevich, E., and I. Yatskiv. Decision Support System 1998, pp. 168–197.
for Transport System Management and Microscopic Traf- 19. Transport Simulation Systems. Aimsun in San Diego Inte-
fic Modeling. Proc., 7th International Scientific Conference grated Corridor Management Project, San Diego, Califor-
TRANSBALTICA, 2011, pp. 74–81. nia, 2012. https://www.aimsun.com/integrated-corridor-
7. Bierlaire, M., and P. L. Toint. Meuse: An Origin-destina- management-project-in-san-diego/.
tion Matrix Estimator that Exploits Structure. Transporta- 20. Dia, H., and N. Cottman. Evaluation of Arterial Incident
tion Research Part B: Methodological, Vol. 29, No. 1, Management Impacts Using Traffic Simulation. IEEE Pro-
1995, pp. 47–60. ceedings-Intelligent Transport Systems, No. 153, 2006. pp.
8. Pattanamekar, P., and D. Park. Estimating Multimodal 242–252.
OD Matrix from Mode Specific OD matrices. Presented at 21. Hidas, P. Modelling Vehicle Interactions in Microscopic
88th Annual Meeting of the Transportation Research Simulation of Merging and Weaving. Transportation
Board, Washington, D.C., 2009. Research Part C: Emerging Technologies, Vol. 13, No. 1,
9. El Faouzi, N.-E., H. Leung, and A. Kurian. Data Fusion 2005, pp. 37–62.
in Intelligent Transportation Systems: Progress and Chal- 22. Wen, T., C. Cai, L. Gardner, V. Dixit, S. T. Waller, and F.
lenges–A Survey. Information Fusion, Vol. 12, No. 1, 2011, Chen. A Strategic User Equilibrium for Independently Dis-
pp. 4–10. tributed Origin-Destination Demands. Computer-Aided
10. Kusakabe, T., and Y. Asakura. Behavioural Data Mining Civil and Infrastructure Engineering, Vol. 33, No. 4, 2018,
of Transit Smart Card Data: A Data Fusion Approach. pp. 316–332.
Transportation Research Part C: Emerging Technologies, 23. Florian, M., and D. Hearn. Network Equilibrium Models
Vol. 46, 2014, pp. 179–191. and Algorithms. Handbooks in Operations Research and
11. Wen, T., L. Gardner, V. Dixit, S. T. Waller, C. Cai, and F. Management Science, Vol. 8, 1995, pp. 485–550.
Chen. Two Methods to Calibrate the Total Travel Demand 24. Australian Bureau of Statistics. Main Structure and Greater
and Variability for a Regional Traffic Network. Computer- Capital City Statistical Areas, 2011. Australian Statistical
Aided Civil and Infrastructure Engineering, Vol. 33, No. 4, Geography Standard (ASGS): Volume 1 Institution of
2018, pp. 288–299. Engineering and Technology, 2017.
12. Guyon, I., and A. Elisseeff. An Introduction to Variable 25. Menon, A. K., C. Cai, W. Wang, T. Wen, and F. Chen.
and Feature Selection. Journal of Machine Learning Fine-grained OD Estimation with Automated Zoning and
Research, Vol. 3, No. March, 2003, pp. 1157–1182. Sparsity Regularisation. Transportation Research Part B:
13. Taib, R., D. Yee, F. Chen, and W. Liu. Use of Machine Methodological, Vol. 80, 2015, pp. 150–172.
Learning for Improved Incident Management Outcomes. 26. Barcelo, J., and J. Casas. Methodological Notes on the
Proc., 13th ITS Asia Pacific Forum 2014, Auckland, New- Calibration and Validation of Microscopic Traffic Simula-
Zealand, 2014, p. 10.
tion Models. Presented at 83rd Annual Meeting of the
14. Chong, M., A. Abraham, and M. Paprzycki. Traffic acci-
Transportation Research Board, Washington, D.C., 2004.
dent analysis using machine learning paradigms. Informa-
tica, Vol. 29, No. 1, 2005.
The Standing Committee on Urban Transportation Data and
15. Sohn, S. Y., and H. Shin. Pattern Recognition for Road
Information Systems (ABJ30) peer-reviewed this paper (18-
Traffic Accident Severity in Korea. Ergonomics, Vol. 44,
06575).
No. 1, 2001, pp. 107–117.