IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 16, NO.

2, APRIL 2015 653

Traffic Flow Prediction for Road Transportation

Networks With Limited Traffic Data
Afshin Abadi, Tooraj Rajabioun, and Petros A. Ioannou, Fellow, IEEE

Abstract—Obtaining accurate information about current and in a traffic network simply because most links are not equipped
near-term future traffic flows of all links in a traffic network has with traffic sensors.
a wide range of applications, including traffic forecasting, vehicle Traffic flow disruptions that affect the estimates of link traffic
navigation devices, vehicle routing, and congestion management.
A major problem in getting traffic flow information in real time is flows can be categorized as predictable and unpredictable.
that the vast majority of links is not equipped with traffic sensors. Predictable disruptions include traffic signals, stop signs, public
Another problem is that factors affecting traffic flows, such as transit services, scheduled sport events, music concerts, road
accidents, public events, and road closures, are often unforeseen, constructions/repairs, etc. Unpredictable disruptions include
suggesting that traffic flow forecasting is a challenging task. In automobile accidents, breakdowns, and emergency road clo-
this paper, we first use a dynamic traffic simulator to generate
flows in all links using available traffic information, estimated sures. The impact of disruption on traffic flow depends on the
demand, and historical traffic data available from links equipped location, the duration of the disruption, and the demand during
with sensors. We implement an optimization methodology to ad- the time of the disruption. Studies regarding the impact of these
just the origin-to-destination matrices driving the simulator. We types of disruption on traffic flows include [1]–[3].
then use the real-time and estimated traffic data to predict the The problem that arises is whether we can predict traffic
traffic flows on each link up to 30 min ahead. The prediction
algorithm is based on an autoregressive model that adapts itself flow ahead of time given the historical traffic information,
to unpredictable events. As a case study, we predict the flows of a information about scheduled events, and real-time traffic data
traffic network in San Francisco, CA, USA, using a macroscopic where available. In principle, due to unpredictable disruptions,
traffic flow simulator. We use Monte Carlo simulations to evaluate long-term predictions may not be accurate enough for reliable
our methodology. Our simulations demonstrate the accuracy of practical use. However, short-term traffic prediction, if properly
the proposed approach. The traffic flow prediction errors vary
from an average of 2% for 5-min prediction windows to 12% for done, may reach an accuracy level that is useful for several
30-min windows even in the presence of unpredictable events. applications when compared with no prediction or inaccurate
prediction.
Index Terms—Historical time traffic flows, least squares
method, optimization, traffic flow prediction. There have been many studies in the literature regarding
short-term traffic flow prediction. Short-term forecasting mod-
els include nonlinear models such as neural network models
I. I NTRODUCTION [4]–[8] and linear models such as Kalman filters [9]–[13] and
autoregressive integrated moving average (ARIMA) models
T RAFFIC flow prediction is considered as a challenging
problem in transportation planning and car navigation
systems. Traffic flows in a network can be estimated using his-
[14]–[17]. ARIMA models are linear estimators based on the
past values of the modeled time series [18]. The nature of data
torical traffic flow data. However, traffic flow prediction cannot and the type of application determine the modeling method
solely rely on past traffic data due to the following reasons: used for traffic prediction. Schmitt and Jula [19] investigated
1) On-road traffic events such as accidents, road closure, etc., the limitations of linear models that are commonly used and
affect the traffic flows in the network, and their effect cannot be observed that “near-future travel times can be better predicted
predicted a priori; 2) off-road events can have a major impact by a combined predictor.” A combined predictor is a linear
on traffic flows and may not be included in the usual historical combination of a historical mean predictor and a current real-
traffic flow data; and 3) traffic data are not available for all links time predictor. Guo et al. [20] compared different modeling
approaches for short-term traffic prediction and concluded that
using a prediction error feedback approach improves the pre-
Manuscript received December 19, 2013; revised April 7, 2014 and June 26,
diction accuracy under normal and abnormal conditions. In
2014; accepted June 29, 2014. Date of publication August 14, 2014; date another study, Smith et al. [21] compared parametric (seasonal
of current version March 27, 2015. This work was supported in part by the ARIMA) and nonparametric (data-driven regression) models
United States Department of Homeland Security through the National Center
for Risk and Economic Analysis of Terrorism Events (CREATE) and in part
and showed that “traffic condition data are characteristically
by MEtropolitan TRANSportation (METRANS). The Associate Editor for this stochastic, as opposed to chaotic.” Moreover, it was argued
paper was S. Sun. that seasonal ARIMA models have better performance than
A. Abadi and T. Rajabioun are with the Department of Electrical Engineer-
ing, University of Southern California, Los Angeles, CA 90089 USA. nonparametric regression models. Furthermore, experimental
P. A. Ioannou is with the Center for Advanced Transportation Technologies, studies showed that ARIMA models outperform heuristic fore-
University of Southern California, Los Angeles, CA 90089 USA. cast benchmarks [15]. The performance of ARIMA models
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org. can be improved by considering temporal–spatial correlations.
Digital Object Identifier 10.1109/TITS.2014.2337238 Multivariate models are introduced to take into account these

1524-9050 © 2014 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html
Authorized licensed use limited to: Sathyabama Institute of Science and Technology. Downloaded on January 30,2025foratmore information.
08:03:33 UTC from IEEE Xplore. Restrictions apply.
 654 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 16, NO. 2, APRIL 2015

correlations. Kamarianakis and Prastacos [22] and Min and

Wynter [23] proposed the space–time autoregressive integrated
moving-average model to satisfy interrelations between links.
The given traffic prediction models become inaccurate under
partially missing data. Missing data indicates the unavailability
of traffic data for a certain period of time in part of a transporta-
tion network due to sensor malfunction or noise-contaminated
data. This problem frequently occurs in transportation networks
[24]–[26]. Many studies address the issue of traffic prediction
with partially missing traffic data. For instance, van Lint et al.
[4] presented a neural network for travel time prediction under
missing traffic data. Sun et al. [27] introduced a Bayesian
method to forecast traffic flows where a certain period of histor-
ical data is missing for some links of the transportation network.
The missing portion of historical traffic data is approximated
by using a Gaussian mixture model. Moreover, other statistical
and probabilistic methods are used to address the missing traffic
data problem [28]–[34]. These approaches are not applicable
to large transportation networks where historical traffic data
are not available for the majority of the arterial links simply
because of the lack of sensor measurements. Therefore, the
problem of short-term traffic flow prediction based on com-
pletely unavailable traffic data in some links due to lack of
sensors in the network is an open challenging problem. Fig. 1. Traffic flow completion methodology.
In this paper, we propose a methodology that predicts the
traffic flows for all the links in a transportation network over a
short time horizon. It consists of two steps: 1) traffic flow data
completion and 2) short-term traffic flow prediction. In the first
step, we use dynamic origin–destination (OD) matrix estima-
tion with the help of a macroscopic simulator to generate traffic Fig. 2. Three-stage model.
flow data at all links based on demand, historical data, and the
limited real-time data available using an online optimization Nodes are connected by links, and the links represent streets
methodology. In the second step, we use the traffic flow data or freeways. Zones are places that a considerable number of
at all links generated from the first step to recursively predict people visit such as schools, stadiums, commercial buildings,
future flows by adapting to changes in real-time data due to and so on. One zone is defined for each residential district.
unpredictable changes. The OD matrix determines the number of trips within zones
As a case study, we focus on a road network in the downtown in each time interval. Each OD matrix is assigned for one
region of San Francisco, California, which we refer to as transportation choice. We consider two types of transportation
the Downtown SF subnetwork. A Monte Carlo experimental choices, namely, personal cars and buses. There are many
method is used to evaluate the proposed algorithm under a wide studies regarding OD matrix estimation. In previous efforts,
range of uncertainties. the OD matrices are estimated using least squares techniques
This paper is organized as follows: In Section II, we present to minimize the difference between the measured link flows
the methodology for generating traffic flow data at links that and the estimated ones [35]–[37]. Some other studies focus on
have no sensors. In Section III, we present the short-term traffic dynamic OD matrix estimation [38]–[40].
flow prediction algorithm. We demonstrate our approach using In the proposed methodology, traffic flow data completion is
the Downtown SF subnetwork in Section IV. We present our performed in two steps: the initial OD matrix estimation and
conclusions in Section V. the dynamic OD matrix estimation. Let Vijn be the ith row and
the jth column element of an OD matrix. It determines the total
number of trips from zone i to zone j in the time interval n ∈
II. T RAFFIC F LOW DATA C OMPLETION {1, . . . , T }. Fig. 1 shows the block diagram of the proposed
Here, we consider a traffic network that contains links with methodology for traffic flow data completion. Dynamic OD
no available measurements. To address the issue of lacking matrix estimation consists of two parts: the dynamic traffic
traffic information for the majority of links in a transportation simulator and the optimization algorithm.
network, we define link-to-link dividing ratio (LLDR) as the
ratio of traffic flow that propagates from a specific link to
A. Initial OD Matrix Estimation
the adjacent links. These LLDRs vary for different links and
also depend on the time of the day. A transportation network The purpose of the initial OD matrix estimation is to estimate
consists of several elements such as links, nodes, zones, etc. the OD matrices based on estimated demand data from the
Authorized licensed use limited to: Sathyabama Institute of Science and Technology. Downloaded on January 30,2025 at 08:03:33 UTC from IEEE Xplore. Restrictions apply.
 ABADI et al.: TRAFFIC FLOW PREDICTION FOR ROAD TRANSPORTATION NETWORKS 655

one zone to another. The number of trips can be estimated as

follows:
n
Vijn = ϑnij .Pin .Anj .eψ dij
∀i, j ∈ Z, ∀n (3)

where Vijn is the number of trips from zone i to zone j in time

interval n. The correlation between two zones can be illustrated
in different formats such as exponential, linear, and so on [41].
In this paper, we choose the exponential format since it provides
Fig. 3. Dynamic traffic flow generation. the best fit for the selected region, which is an urban region [42].
dij is defined as the shortest distance between zones i and j.
region. For instance, students go to schools at 8 A . M . and leave Parameter ψ n is a negative constant, and ϑnij is a scaling factor
schools at 2 P. M . We can roughly estimate the demand for each to adjust the total number of trips from zone i to zone j in time
school in the morning based on the number of students who interval n to satisfy the following constraint equations [42]:
use the car and bus for commuting to school. There are two 
main methods in transportation forecasting [41]: 1—the gravity Vijn = Pin ∀i ∈ Z, ∀n (4)
1≤j≤Z
model; 2—the activity-based model. The activity-based model

is on the microscopic level and requires much more information Vijn = Anj ∀j ∈ Z, ∀n. (5)
than the gravity model. Since, in our case, the data are limited, 1≤i≤Z
we use the gravity model to estimate the initial OD matrices, as
explained below and shown in Fig. 2. The next step in Fig. 2, which is referred to as the trans-
The first stage in the model, which is referred to as trip portation choice stage, determines the proportion of Vijn that
generation, is the aggregated travel demand for each zone. Each use a specific transportation mode such as bicycles, single- or
zone has a certain production, i.e., the number of trips that begin multipassenger cars, or public transit vehicles. The proportion
at that zone, and a certain attraction, i.e., the number of trips of Vijn that uses single car and public transit are used to generate
that culminate at that zone. The information (i.e., production OD matrices of the trips at different intervals of time. For
and attraction) can be obtained through surveys and information instance, buses represent X% of the total number of vehicles,
from the region. For instance, in a big sport game, the attraction and their proportion is denoted by Vij,b n n
, where Vij,b = 0.01 ∗
includes thousands of cars and possibly hundreds of buses X ∗ Vij , and the rest (100-X)% are cars (Vij,c = Vijn − Vij,b
n n n
)
heading to the parking lots adjacent to the stadium before the for a pair of zones (i, j). The ratios of buses and vehicles
start of the game. The structure is the stadium, and one zone are variable depending on the location of zones. Buses have
is assigned for each parking lot. The production/attraction ratio a higher ratio in the downtown region than they have in a local
of each zone (parking lot) for this specific structure (stadium) residential area.
can be estimated based on the capacity of the parking lots.
Moreover, each residential district is considered as one structure
within one zone in the center of the district. Let us denote B. Dynamic OD Matrix Estimation
trip production of zone i and trip attraction of zone j in time The initial OD matrices that drive the dynamic traffic sim-
interval n, with Pin and Anj , respectively. Pin determines the ulator are adjusted based on an optimization procedure pre-
total number of trips originating from zone i, and Anj is the total sented below, which minimizes the error between measured
number of trips ending in zone j.The production and attraction flows (where available) and estimated ones as well as travel
of each zone can be generated as times. It is assumed that trips will follow routes that minimize
 travel time.
Pin = ηqn (i)Pqn ∀i ∈ Z, ∀n (1) 1) Dynamic Traffic Simulator: The physical network model
q∈Q consists of various objects such as links, nodes, zones, etc. The
 OD matrix represents the aggregated number of trips from one
Anj = ξqn (j)Anq ∀j ∈ Z, ∀n (2) zone to another. Traffic flows on each link can be obtained
q∈Q by feeding demand to a traffic flow simulator, while taking
into account physical network characteristics. Fig. 3 shows the
where Anq and Pqn are the total attraction and production diagram for dynamic traffic flow generation.
for structure q in time interval n. Structure q is a subset of The proposed methodology evaluates the propagation of
Q, which contains all the places in the region of study that flows in the network by calculating the LLDRs based on the
people commute in. The attraction/production of each struc- link flows generated by the dynamic traffic flow simulator and
ture is distributed to the adjacent zones. Each structure has a the optimization algorithm described below.
production/attraction ratio for each zone. ηqn (i) and ξqn (j) repre- 2) Optimization: Historical time traffic flows, derived from
sent production/attraction ratios for zones i and j, respectively. historical traffic flow data, are used as one of the inputs to
Parameter Z corresponds to the set of all zones in the network. the optimization algorithm. They are used in both the dynamic
The second stage of the model, which is referred to as trip OD matrix estimation and short-term traffic prediction mod-
distribution, determines the number of trips that traverse from els. Historical time traffic flows vary seasonally, weekday to
Authorized licensed use limited to: Sathyabama Institute of Science and Technology. Downloaded on January 30,2025 at 08:03:33 UTC from IEEE Xplore. Restrictions apply.
 656 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 16, NO. 2, APRIL 2015

route r ∈ θ for OD pair s, which is denoted as cs,r,t , depends

on the traffic flow of the route. We define cost as the travel
time of loaded route in the network. We also denote v̄l,n as
the historical time traffic flow on link l, and vl,n represents the
estimated flow of link l in time interval n, where n is an integer,
and n ∈ {1, . . . , T } and t ∈ [0, T ], where t indicates time, and
T is the length of the total interval. Since traffic sensors are only
available for a portion of links in the network, we use the results
from the three-stage model that covers all links in the network
as the initial solution for the optimization algorithm. Let vs,r,t
be the generated flow (number of vehicles per time slot) of route
r for the OD pair s through the three-stage model. Therefore,
the flow of link l is calculated as
 ∀l ∈ L
vl,n = ρl,t
s,r,n vs,r,t (8)
Fig. 4. Overall, block diagram. n ∈ {1, . . . , T }
s∈S r∈θ t∈T
weekend, and during a specific day. For instance, major events
that occur adjacent to a road network and draw hundreds or where ρl,t
s,r,n is a decision variable equal to 1 if link l is on
possibly thousands of attendees can have huge impacts on route r connecting OD pair s during time t in time slot n and 0
network flows. otherwise. Note that cs,r,t is the function of ρl,t
s,r,n vs,r,t [41].
In this paper, traffic flows and information about major events Moreover, vl,n must satisfy the flow conservation law; thus,
are taken into account for M consecutive months. Given the we have
major event data such as date and capacity of the venue, histor-  ∀l ∈ L
ical time traffic flows can be also derived for each specific event. vl,n = Rl,k,n vk,n (9)
Days are classified into three general categories: 1) regular day; n ∈ {1, . . . , T }
k∈A(l)
2) holiday; and 3) major events. For regular days, the average
hourly traffic flow for link l (AHTFl ) is expressed as where vk,n represents the flow of links feeding link l, and Rl,k,n
is defined as the LLDR of link k to link. We define ϕn (R) as
 ∀h ∈ H a set of vl,n and the corresponding vs,r,t as well as Rl,n that
HTFl (h, d, w, m) d∈D
AHTFl (h, d, m) = w∈W
 (6) satisfies (8) and (9).
w∈W 1
m∈M Note that A(l) = {j ∈ L|head(j) = tail(l)}. This contains
l∈L all the links that are feeding link l. We refer to the optimization
where HTFl is the hourly traffic flow for link l, h indicates the algorithm formulated below as the master problem. Thus
specific hour of the day, d is the day of the week, w represents
the week, and m is the month of the year. For the second and min g(V )
third categories, (6) is also used only by taking into account w.r.t.
historical traffic flows corresponding to the specific event or     2
  v̄l,n − s∈S r∈θ t∈T ρl,t
s,r,n vs,r,t
holiday. Due to the important role of recent data, a weighting vs,r,t = α1
factor is introduced to have a better historical time traffic flow v̄l,n
n∈{1,...,T } l∈L
evaluation. The weighted-average hourly traffic flow for link l, 
(WAHTFl ) is calculated as + α2 cs,r,t vs,r,t (10)
s∈S r∈θ t∈T
WAHTFl (h, d)
subject to vs,r,t ≥ 0. (11)
|M |
 (|M |−m+1) 2
∀h ∈ H
= |M | e−ζm AHTFl (h, d, m) The objective function (10) minimizes the normalized varia-
m=1 k=1 (|M |−k+1)
2 d∈D
tion between the historical time link flows and the simulated
(7) ones as well as the total cost (travel time) of the network.
Constraint (11) imposes a nonnegative value for link flows.
where ζ is a constant positive coefficient, and the number of Coefficients 0 ≤ α1 , α2 ≤ 1 weigh the relative importance of
members in a set is denoted by | · |. For the sake of simplicity, average historical count data and total travel time. The opti-
in the rest of this paper, WAHTFl is denoted by v̄l,n , where n is mization (10) and (11) aim to minimize the total travel time in
the time interval. addition to minimizing the error between the measured flows
A physical transportation network is defined on a set of links (where available) and estimated ones by taking into account the
and nodes, i.e., g = (N, L), where N and L represent nodes and fact that, in general, drivers select traffic routes that minimize
links in the network, respectively. Let S denote all OD pairs, travel time. The objective is to guide the solution using new ef-
and θ is the set of all routes connecting OD pair s ∈ S. Each ficient routes with the minimum total network cost. The column
route r ∈ θ consists of one or more links from L. The cost of generation algorithm is one of the most common algorithms to
Authorized licensed use limited to: Sathyabama Institute of Science and Technology. Downloaded on January 30,2025 at 08:03:33 UTC from IEEE Xplore. Restrictions apply.
 ABADI et al.: TRAFFIC FLOW PREDICTION FOR ROAD TRANSPORTATION NETWORKS 657

find the efficient routes [43]. It starts with the subset of solution
(θ1 ⊂ θ) of (10), which can be presented as

min g(V )
w.r.t.
    2
  v̄l,n− s∈S r∈θ t∈T ρl,t
s,r,n vs,r,t
1
vs,r,t = α1
v̄l,n
n∈{1,...,T } l∈L

+ α2 cs,r,t vs,r,t (12)
s∈S r∈θ1 t∈T

subject to vs,r,t ≥ 0. (13)

Using information from the region as well as sensor data, we

would like to estimate link flows for the specified time window Fig. 5. Case study traffic network.
on a specific link. Using an observation interval (such as a six-
month period), we are able to calculate the average flow for a The routes with the negative value of gm,p,w are eligible
particular time of the day using (7). routes and are added to the previous routes. As a result, the
Solution Methodology: We start with a subset of solution problem of finding a new eligible route to add to the previous
θ1 ⊂ θ (restricted master problem) using a column generation routes has changed to the dynamic shortest route problem with
algorithm. The restricted master problem is a way to expedite new link costs. Finding the minimum cost flows in this case
the process of finding the lowest cost route. The restricted is NP-complete. In the static case, the computation time for
master problem is defined as a subset of the master problem finding the shortest path in a graph with N nodes is O(n); how-
(θ1 ). The initial solution of the restricted master problem is ever, in the dynamic case, 2N −1 routes must be evaluated from
determined by the three-stage model, as explained in detail in node 1 to N . Thus, the computational time exponentially grows
the previous section. Initially, the column generation algorithm [44]. The LLDRs can be calculated based on the estimated link
begins with the restricted master problem with only a small flows using (9).
subset of routes (initial solution). Then, it adds new eligible
routes to the previous routes. The restricted master problem is
III. T RAFFIC F LOW P REDICTION
defined as
Here, the estimated LLDRs as well as current and historical
min g(V ) traffic flows are used by the traffic prediction algorithm to make
w.r.t. predictions, as explained in this section. Fig. 4 shows the overall
    2 structure of the prediction approach.
  v̄l,n− s∈S r∈θ t∈T ρl,t
s,r,n vs,r,t
vs,r,t = α1 1 The approach to be presented in detail below is summarized
v̄l,n as follows: First, an autoregressive model of the link flows is
n∈{1,...,T } l∈L
 introduced that takes into account the uncertain nature of traffic
+ α2 cs,r,t vs,r,t (14) and historical traffic data including the most recent ones as well
s∈S r∈θ1 t∈T as historical time traffic flows in its coefficients. Due to the
subject to vs,r,t ≥ 0 (15) nonstationary nature of traffic flows, we detrend the traffic flow
data by subtracting the corresponding long-term historical flow
where θ1 ⊂ θ. The initial solution θ1 is derived by feeding means ῡl,t , which are calculated in Section II, from the real-
the initial OD matrices into a dynamic traffic simulator. The time flows vl,t to obtain a stationary process. Then, we train the
column generation algorithm operates as follows. autoregressive model using the most recent data to calculate the
Step 1: Apply the initial solution. model parameters. Finally, we apply the least squares method
Step 2: Calculate the total cost. with the LLDR constraints to expand the short-term prediction
Step 3: Add new eligible routes. results to the entire network.
Step 4: Calculate new total cost, go to step 3. Let υl,t represent the flow for the specific link l at time t
We take a derivative with respect to vs,r,t to find new eligible that is assumed to be generated by the autoregressive model
routes, i.e., [15], i.e.,

∂g(V ) 
M
∀l ∈ L
gm,p,w = υl,t = β0 + βm υl,t−m + εl,t (17)
∂vm,p,w m=1
∀t
    
  v̄l,n− s∈S r∈θ t∈T ρl,t
s,r,n vs,r,t where the coefficients βs, s = 0, . . . , M are the parameters of
= 2α1
v̄l,n the autoregressive model, and εl,t forms a white noise process
t∈[0,T ] l∈L
that is the innovation process uncorrelated with υl,t . The order
× ρl,w
m,p,n + α2 cm,p,w . (16) M of the model is found by applying the Akaike information
Authorized licensed use limited to: Sathyabama Institute of Science and Technology. Downloaded on January 30,2025 at 08:03:33 UTC from IEEE Xplore. Restrictions apply.
 658 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 16, NO. 2, APRIL 2015

criterion [45]. Historical time traffic flow means, which are In general, the predictions need to be made for more than one
calculated in previous section, are denoted by ῡl,t = E{vl,t }. time step ahead in the future. Since the real-time information
Using the autoregressive model (17), ῡl,t is represented by is available for up to time t, êl,t+1|t , êl,t+2|t , . . . , êl,t+p|t are
−1
informationally equivalent, and

M
∀l ∈ L
ῡl,t+1 = β0 + βm ῡl,t−m (18) êl,t+p|t = E{el,t+p |el,t }
m=0
∀t.

Then, these historical means are subtracted from real-time ∀l ∈ L

= E{el,t+p |êl,t+1|t } = · · ·
traffic flows. We define the error process el,t as ∀t
∀l ∈ L = E{el,t+p |êl,t+p−1|t } (25)
el,t = υl,t − ῡl,t (19)
∀t.
∀l ∈ L
Then, from (17) and (18), we have êl,t+p|t = Apl el,t (26)
∀t.

M
∀l ∈ L The value êl,t+p|t is estimated after calculating β1 , . . . , βM .
el,t = βm el,t−m + εl,t (20)
∀t. The truncated error process el,k , k ∈ {t − T + 1, t − T, . . . , t}
m=1
is used to obtain the sample autocovariance functions
The uncertainty in the system is modeled by a white noise
1 
t−j
process; therefore, el,t is a stationary mean-zero process for all ∀l ∈ L
Cl (j) = el,i+j el,i (27)
the links in the network. Now, we represent this model in the T ∀j ∈ {0, 1, . . . , T }
i=t−T
matrix form as
where T is the length of the short-term historical data used for
∀l ∈ L
el,t = Al el,t−1 + l,t (21) fitting the model. The least squares method is used to estimate
∀t
the β  s parameters.
 −1 Considering the expanded predicted value
where el,t , A, and l,t are defined as υ̂l,t+1 = M β υ̃
m=0 m l,t−m , the mean square residuals are de-
⎡ ⎤ fined as [47]
εl,t
⎢ 0 ⎥ 1
T
el,t = υ l,t − ῡ l,t , l,t =⎢
⎣ ... ⎦
⎥
Rl (β) = Wt (el,t − êl,t )2
T t=1
0
⎡ ⎤
··· ··· 2
1 
β1 β2 βM T M
⎢ 1 0 ··· ··· 0 ⎥ = Wt el,t − βm el,t−m ∀l ∈ L
⎢ ⎥ ∀l ∈ L T t=1
Al = ⎢ 0
⎢ .
1 0 ··· 0 ⎥ m=1
⎣ .. .. .. .. .. ⎥ ∀t
. . . . ⎦ (28)
0 0 ··· 1 0
⎡vl,t ⎤ ⎡ v̄l,t ⎤ where W  s are the positive normalizing weights. The objec-
⎢ vl,t−1 ⎥ ⎢ v̄l,t−1 ⎥ tive is to minimize the residuals with respect to {βm ; m = 1,
v l,t =⎢
⎣ .. ⎥
⎦ v̄ l,t =⎢
⎣ .. ⎥.
⎦ (22) 2, . . . , M } by assuming el,0 = el,−1 = · · · = el,−M +1 = 0.
. . The coefficients (βm ) are calculated by setting the gradient of
vl,t−M +1 v̄l,t−M +1 the residual function to zero. The model contains M parame-
ters; hence, there are M equations as
Using the matrix form, it enables us to expand the prediction
algorithm to more than one time step in the future. Let êl,t be ∂Rl T M
the estimated value of el,t . The following estimator minimizes = −2 Wt el,t−i el,t − βm el,t−m = 0 ∀l ∈ L.
∂βi
the estimated value of el,t+1 given the flow data at times t, t=1 m=1
t − 1, t − 2, . . . [46]: (29)
The coefficients (β̂m ) are obtained by solving the linear
êl,t+1|t = E{el,t+1 |υ l,t } ∀l ∈ L, ∀t equation
= Al E{el,t |υ l,t } + E{l,t } ⎡ C (0) ⎡ ⎤
· · · Cl (M − 1) ⎤ β̂1
(23)
l Cl (1)
⎢ Cl (1) Cl (0) · · · Cl (M − 2) ⎥ ⎢ β̂2 ⎥
where êl,t+1|t is the estimate of el,t+1 given the flow data at ⎢ ⎥⎢ ⎥
times t, t − 1, t − 2, . . .. Since el,t+1 is calculated from (19), ⎣ . . . . ⎦ ⎣ .. ⎥
⎢
.. .. .. .. . ⎦
we have E{el,t |υ l,t } = E{el,t |el,t }. Considering the fact that Cl (M − 1) Cl (M − 1) ··· Cl (0) β̂M
the real-time el,t is known at time t and t has zero mean and is
uncorrelated with υl,t , the predicted value of traffic flow error ⎡ C (1) ⎤
l
becomes ⎢ Cl (2) ⎥
=⎢
⎣ ..
⎥
⎦ ∀l ∈ L. (30)
∀l ∈ L .
êl,t+1|t = Al el,t (24)
∀t. Cl (M )
Authorized licensed use limited to: Sathyabama Institute of Science and Technology. Downloaded on January 30,2025 at 08:03:33 UTC from IEEE Xplore. Restrictions apply.
 ABADI et al.: TRAFFIC FLOW PREDICTION FOR ROAD TRANSPORTATION NETWORKS 659

After the calculation of êl,t+p|t , v l,t+p|t , the prediction of

link flows at time t + p is obtained by
∀l ∈ L
v l,t+p|t = v̄ l,t+p + êl,t+p|t (31)
∀t.
The short-term traffic flow prediction model is recalibrated
once the new traffic data are received. However, the results
from the short-term traffic prediction model include the critical
coverage gaps where the real-time data are not available, which
can be a large portion of an urban transportation network.
Therefore, we take into account both historical data and pre-
diction results to minimize travel times with respect to the
predicted traffic flows (v̂l,t+p|t ) as

  vl,t+p|t − v̂l,t+p|t 2
minimize γ1
vl,t+p|t Fig. 6. Time-series analysis (scenario 1).
l∈L1

  v̄l,t+p − v̂l,t+p|t 2
+ γ2 (32)
v̄l,t+p
l∈L

s.t. v̂l,t+p ∈ ϕt+p (R) (33)

where 0 < γ2 < γ1 < 1, vl,t+p|t is the result from the short-
term traffic prediction model. v̄l,t+p is the historical time traffic
flow at time t + p for the link l, and v̄l,t+p|t is the final traffic
flow prediction for link flows based on the historical data as
well as the prediction model. It is worth noting that we take
γ1 > γ2 because of the relative importance of the role of the
results from the prediction model in computing link flows. L1
is defined as the number of links with available real-time data
and, L denotes the number of all the links in the transportation
network. The objective is to find v̄l,t+p|t for all the links in the
Fig. 7. Prediction scatterplot (scenario 2).
network including links where data are unavailable. Constraint
(33) yields link flows satisfying the LLDRs derived from the
traffic flow completion model. The output of the given model from the big venues (e.g., AT&T Park stadium), which provides
is the set of link flows and updated LLDRs. The LLDRs are a benchmark for the traffic flow completion model. Traffic data
calculated from link flows using (8). are received every 1 min for those 16% links, and no data are
available for the rest of the links.
Three scenarios are used to demonstrate the accuracy of the
IV. C OMPUTATIONAL E XPERIMENT
proposed model. The first one presents a time-series scenario.
The San Francisco, California downtown region is chosen At each step, traffic flows are predicted for the next 5 min.
as the case study to perform experiments and demonstrate the For instance, traffic flows for time t + 5 are predicted based
effectiveness of the developed algorithm. Traffic flow and event on the traffic data up to time t, and for time t + 6 up to time
data as well as incident data are provided by UC Berkeley PATH t + 1. The second scenario is defined as the comparison of
[48]. Clearly, the quality and quantity of data have an impact on the predicted and measured volumes for the 30-min horizon
the quality of the final solution. We use a macroscopic traffic prediction. Finally, the third scenario indicates the prediction
simulator based on the commercial software VISUM for the of flows on the targeted links for the next 5-, 10-, 15-, 20-,
traffic assignment step [49]. The inputs of VISUM are the OD 25-, and 30-min time horizon. Note that the targeted links
matrices, and the outputs are the estimated link flows. The OD are chosen to be in the category of nonavailable data links.
matrices are adjusted based on the available traffic volumes Therefore, the assumption is that no historical and real-time
and the initial OD matrices using the optimization formulation traffic data are available; however, in fact, both are available to
introduced in Section II. Fig. 5 shows the traffic network in the verify the proposed model results. Fig. 6 demonstrates the time-
VISUM format. series scenario for the next 5 min. Fig. 7 shows the scatterplot
The traffic network under study contains more than 20 000 for the second scenario, and Table I presents the root-mean-
links, and real-time traffic data are only available for about 16% square percentage error (RMSPE) for the third scenario, which
of the links in the network. The initial traffic flow estimation is the prediction of targeted links every 5 min for the next
model is used to generate traffic flows based on the information half hour.
Authorized licensed use limited to: Sathyabama Institute of Science and Technology. Downloaded on January 30,2025 at 08:03:33 UTC from IEEE Xplore. Restrictions apply.
 660 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 16, NO. 2, APRIL 2015

Fig. 8. Monte Carlo experiments results for 5-min-ahead predictions over a Fig. 9. Monte Carlo experiment results for 15-min-ahead predictions over a
30-min period. 30-min period.

TABLE I
RMSPE (S CENARIO 3) 4) The new LLDR set is fed to the predictive algorithm, and
the results are recorded.
These four steps are repeated 1000 times for each data
sample. The coefficient Γ is assumed to be a Gaussian ran-
dom variable Γ ∼ N (1, 0.1). The variance σ 2 = 0.1 is selected
based on the observations from Downtown SF traffic data (σ =
0.316). In the cases that the new LLDR is bigger than 1, the
fixed value of 1 is chosen for the specific LLDR. Figs. 8 and 9
show the error for a specific data sample for 5-min and 15-min
time horizon prediction. In this sample, the standard deviation
of the Monte Carlo results is σ = 0.09, which indicates that
the sensitivity of the predictive algorithm is about 28% with
respect to the LLDR variations. The sensitivity of traffic flow
The first column represents the prediction timing, i.e., 5, 10,
predictions to the error in LLDR calculations suggests that the
15, . . . min from present time. Columns 2–4 demonstrate the
prediction/simulation model is robust, and the noise does not
RMSPE for the targeted links at 11 A . M ., 3 P. M ., and 8 P. M . on
escalate through different steps of the model.
weekdays, respectively, defined as

 N  2
1  v̂l − vl V. C ONCLUSION
RMSPE =  (34)
N vl In this paper, we have estimated traffic flows in all links in
l=1
a traffic network where traffic data are unavailable and used
where v̄l is the predicted traffic flow, and vl indicates the mea- the information to predict short-term traffic flow for the entire
sured one for link l. Results in Figs. 6 and 7 and Table I present transportation network. A large network in the San Francisco
the accuracy of the proposed methodology to predict traffic area was used to demonstrate the efficiency and accuracy of the
flows in the urban region with the complex characteristics. methodology. Monte Carlo simulations were used to account
Fig. 7 also suggests that the proposed predictor is unbiased. for random effects and uncertainties. The results demonstrate
In addition to the given experiments, which were performed accurate predictions of traffic flow rates up to 30 min ahead
under normal conditions, Monte Carlo experiments are carried of time under normal operations. In the case of events, the
out to evaluate the effects of random uncertainties due to the prediction algorithm adapts to the changes and modifies its
stochastic nature of traffic data. The most possible source of prediction outputs with good accuracy.
error in the proposed predictive algorithm is that the estimated One of the limitations of this paper is the lack of adequate
LLDRs from the historical traffic data might not be valid for number of data during normal and incident traffic conditions
some links of the network at the current time due to accidents to perform additional tests. Future work will involve the col-
or social events. To measure the sensitivity of the model with lection of additional real-time data as they become available
respect to these variations, Monte Carlo experiments are per- due to emerging traffic sensor technologies, the use of vehicles
formed. The method is as follows. as probes, and vehicle-to-infrastructure communications. Such
1) At each experiment, the LLDRs are calculated. data may be used to improve the accuracy of our approach as
2) 50% of the LLDRs are randomly selected. well as validate it under different traffic scenarios and different
3) Selected LLDRs are multiplied by a random variable Γ. networks.
Authorized licensed use limited to: Sathyabama Institute of Science and Technology. Downloaded on January 30,2025 at 08:03:33 UTC from IEEE Xplore. Restrictions apply.
 ABADI et al.: TRAFFIC FLOW PREDICTION FOR ROAD TRANSPORTATION NETWORKS 661

ACKNOWLEDGMENT [23] W. Min and L. Wynter, “Real-time road traffic prediction with spatio-
temporal correlations,” Transp. Res. Part C Emerging Technol., vol. 19,
Any opinions, findings, and conclusions or recommendations no. 4, pp. 606–616, Aug. 2011.
in this paper are those of the authors’ and do not necessarily [24] L. Q. L. Qu, J. H. J. Hu, L. L. L. Li, and Y. Z. Y. Zhang, “PPCA-
based missing data imputation for traffic flow volume: A systematical
reflect the views of the United States Department of Homeland approach,” IEEE Trans. Intell. Transp. Syst., vol. 10, no. 3, pp. 512–522,
Security and METRANS. Sep. 2009.
[25] C. Chen, Y. Wang, L. Li, J. Hu, and Z. Zhang, “The retrieval of intra-day
trend and its influence on traffic prediction,” Transp. Res. Part C Emerging
Technol., vol. 22, pp. 103–118, 2012.
[26] B. Smith, W. Scherer, and J. Conklin, “Exploring imputation techniques
R EFERENCES for missing data in transportation management systems,” Transp. Res.
[1] RAND, Moving Los Angeles: Short-Term Transportation Policy Options Rec., vol. 1836, pp. 132–142, 2003.
for Improving Transportation, RAND Corporation, 2008. [27] S. Sun, C. Zhang, G. Yu, N. Lu, and F. Xiao, “Bayesian network methods
[2] “Traffic choices study: Summary report,” Seattle, WA, USA, 2008. for traffic flow forecasting with incomplete data,” in Proc. Mach. Learn.,
[3] T. P. Mamuneas and M. I. Nadri, “Contribution of highway capital to in- ECML, 2004, pp. 419–428.
dustry and national productivity growth,” Transportation Research Board, [28] L. Li, Y. Li, and Z. Li, “Efficient missing data imputing for traffic flow
Washington, DC, USA, 1996. by considering temporal and spatial dependence,” Transp. Res. Part C
[4] J. W. C. van Lint, S. P. Hoogendoorn, and H. J. van Zuylen, “Accurate Emerging Technol., vol. 34, pp. 108–120, 2013.
freeway travel time prediction with state-space neural networks under [29] J. Haworth and T. Cheng, “Non-parametric regression for space–time
missing data,” Transp. Res. Part C Emerg. Technol., vol. 13, no. 5/6, forecasting under missing data,” Comput. Environ. Urban Syst., vol. 36,
pp. 347–369, Oct.–Dec. 2005. no. 6, pp. 538–550, Nov. 2012.
[5] J. Yu, G.-L. Chang, H. W. Ho, and Y. Liu, “Variation based online travel [30] H. Tan et al., “A tensor-based method for missing traffic data completion,”
time prediction using clustered neural networks,” in Proc. 11th Int. IEEE Transp. Res. Part C Emerging Technol., vol. 28, pp. 15–27, Mar. 2013.
Conf. Intell. Transp. Syst., Oct. 2008, pp. 85–90. [31] A. Khosravi et al., “Prediction intervals to account for uncertainties in
[6] K. Y. Chan, T. S. Dillon, J. Singh, and E. Chang, “Neural-network-based travel time prediction,” IEEE Trans. Intell. Transp. Syst., vol. 12, no. 2,
models for short-term traffic flow forecasting using a hybrid exponen- pp. 537–547, Jun. 2011.
tial smoothing and Levenberg–Marquardt algorithm,” IEEE Trans. Intell. [32] S. S. S. Sun and C. Z. C. Zhang, “The selective random subspace predictor
Transp. Syst., vol. 13, no. 2, pp. 644–654, Jun. 2012. for traffic flow forecasting,” IEEE Trans. Intell. Transp. Syst., vol. 8, no. 2,
[7] D. Park, L. R. Rilett, and G. Han, “Spectral basis neural networks for real- pp. 367–373, Jun. 2007.
time travel time forecasting,” J. Transp. Eng., vol. 125, no. 6, pp. 515–523, [33] S. Sun and X. Xu, “Variational inference for infinite mixtures of Gaussian
Nov. 1999. processes with applications to traffic flow prediction,” IEEE Trans. Intell.
[8] Q. Y. Q. Ye, S. C. Wong, and W. Y. Szeto, “Short-term traffic speed Transp. Syst., vol. 12, no. 2, pp. 466–475, Jun. 2011.
forecasting based on data recorded at irregular intervals,” in Proc. 13th [34] S. Sun, R. Huang, and Y. Gao, “Network-scale traffic modeling and
Int. IEEE Conf. ITSC, 2010, pp. 1541–1546. forecasting with graphical lasso and neural networks,” J. Transp. Eng.,
[9] I. Okutani, “The Kalman filtering approaches in some transportation and vol. 138, no. 11, pp. 1358–1367, Nov. 2012.
traffic problems,” Transp. Res. Rec., vol. 2, no. 1, pp. 397–416, 1987. [35] E. Cascetta, “Estimation of trip matrices from traffic counts and survey
[10] Y. Xie, Y. Zhang, and Z. Ye, “Short-term traffic volume forecasting using data: A generalized least squares estimator,” Transp. Res., vol. 18B,
Kalman filter with discrete wavelet decomposition,” Comput.-Aided Civil no. 4/5, pp. 289–299, 1984.
Infrastruct. Eng., vol. 22, no. 5, pp. 326–334, Jul. 2007. [36] E. Codina and J. Barceló, “Adjustment of O-D trip matrices from observed
[11] I. Okutani and Y. J. Stephanedes, “Dynamic prediction of traffic volume volumes: An algorithmic approach based on conjugate directions,” Eur. J.
through Kalman filtering theory,” Transp. Res. Part B Methodol., vol. 18, Oper. Res., vol. 155, pp. 535–557, 2004.
no. 1, pp. 1–11, 1984. [37] Y. Nie, H. M. Zhang, and W. W. Recker, “Inferring origin–destination trip
[12] H. J. H. Ji, A. X. A. Xu, X. S. X. Sui, and L. L. L. Li, “The applied matrices with a decoupled GLS path flow estimator,” Transp. Res. Part B
research of Kalman in the dynamic travel time prediction,” in Proc.18th Methodol., vol. 39, pp. 497–518, 2005.
Int. Conf. Geoinformat., 2010, pp. 1–5. [38] M. J. Maher, X. Zhang, and D. Van Vliet, “A bi-level programming
[13] Y. Wang and M. Papageorgiou, “Real-time freeway traffic state estima- approach for trip matrix estimation and traffic control problems with
tion based on extended Kalman filter: A general approach,” Transp. Res. stochastic user equilibrium link flows,” Transp. Res. Part B Methodol.,
Part B Methodol., vol. 39, no. 2, pp. 141–167, Feb. 2005. vol. 35, pp. 23–40, 2001.
[14] M. Van Der Voort, M. Dougherty, and S. Watson, “Combining Kohonen [39] X. Zhou, X. Qin, and H. S. Mahmassani, “Dynamic origin–destination
maps with ARIMA time series models to forecast traffic flow,” Transp. demand estimation using multi-day link traffic counts for planning appli-
Res. Part C Emerg. Technol., vol. 4, no. 5, pp. 307–318, Oct. 1996. cations,” Transp. Res. Forum, vol. 5, no. 2, pp. 30–38, 2003.
[15] B. M. Williams, M. Asce, L. A. Hoel, and F. Asce, “Modeling and fore- [40] X. Z. X. Zhou and H. S. Mahmassani, “Dynamic origin–destination de-
casting vehicular traffic flow as a seasonal ARIMA process: Theoretical mand estimation using automatic vehicle identification data,” IEEE Trans.
basis and empirical results,” J. Transp. Eng., vol. 129, no. 6, pp. 664–672, Intell. Transp. Syst., vol. 7, no. 1, pp. 105–114, Mar. 2006.
Nov. 2003. [41] D. de Ortúzar and L. G. Willumsen, Modelling Transport. Hoboken, NJ,
[16] D. Billings and J. S. Yang, “Application of the ARIMA models to urban USA: Wiley, 2011.
roadway travel time,” in Proc. IEEE Int. Conf. Syst., Man, Cybern., 2006, [42] D. P. Smith and B. G. Hutchinson, “Goodness of fit statistics for trip
pp. 2529–2534. distribution models,” Transp. Res. Part A, Gen., vol. 15, pp. 295–303,
[17] C. Chen, J. Hu, Q. Meng, and Y. Zhang, “Short-time traffic flow prediction 1981.
with ARIMA-GARCH model,” in Proc. IEEE Intell. Veh. Symp., 2011, [43] J. Desrosiers and M. E. Lübbecke, “A primer in column generation,”
pp. 607–612. Column Gener., vol. 3, pp. 1–32, 2005.
[18] B. M. Williams, P. K. Durvasula, and D. E. Brown, “Urban freeway traffic [44] M. R. Garey and D. S. Johnson, “Computers and intractability: A guide
flow prediction application of seasonal autoregressive integrated,” Transp. to the theory of NP-completeness (series of books in the mathemati-
Res. Rec., vol. 1644, pp. 132–141, 1998. cal sciences),” in Computers and Intractability. New York, NY, USA:
[19] E. J. Schmitt and H. Jula, “On the limitations of linear models in pre- Macmillan, 1979, p. 340.
dicting travel times,” in Proc. IEEE Intell. Transp. Syst. Conf., Sep. 2007, [45] B. Y. R. Shebata, “Selection of the order of an autoregressive model by
pp. 830–835. Akaike’s information criterion,” Biometrika, pp. 117–126, 1976.
[20] F. Guo, J. W. Polak, and R. Krishnan, “Comparison of modeling ap- [46] J. M. Mendel, Lessons in Estimation Theory for Signal Processing,
proaches for short term traffic prediction under normal and abnormal Communications, Control. Englewood Cliffs, NJ, USA: Prentice-Hall,
conditions,” in Proc. 13th Int. IEEE Conf. Intell. Transp. Syst., 2010, 1995.
pp. 1209–1214. [47] H. Akaike, “Fitting autoregressive models for prediction,” Inn. Inst.
[21] B. L. Smith, B. M. Williams, and R. Keith Oswald, “Comparison of para- Statist. Math., vol. 21, no. 1, pp. 243–247, 1969.
metric and nonparametric models for traffic flow forecasting,” Transp. [48] California Partners for Advanced Transportation TecHnology. [Online].
Res. Part C Emerging Technol., vol. 10, no. 4, pp. 303–321, Aug. 2002. Available: www.path.berkeley.edu
[22] Y. Kamarianakis and P. Prastacos, “Space–time modeling of traffic flow,” [49] PTV Group. [Online]. Available: http://vision-traffic.ptvgroup.com/
Comput. Geosci., vol. 31, no. 2, pp. 119–133, Mar. 2005. en-us/products/ptv-visum/
Authorized licensed use limited to: Sathyabama Institute of Science and Technology. Downloaded on January 30,2025 at 08:03:33 UTC from IEEE Xplore. Restrictions apply.
 662 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 16, NO. 2, APRIL 2015

Afshin Abadi received the M.S. degree in electrical Petros A. Ioannou (S’80–M’83–SM’89–F’94) re-
engineering from University of Southern California, ceived the B.Sc. degree (with first-class honors) from
Los Angeles, CA, USA, where he is currently work- University College London, London, U.K., in 1978
ing toward the Ph.D. degree. and the M.S. and Ph.D. degrees from University
His research focuses on container terminal op- of Illinois at Urbana—Champaign, Champaign, IL,
eration, transportation planning, and car navigation USA, in 1980 and 1982, respectively.
systems. In 1982 he joined the Department of Electri-
cal Engineering—Systems, University of Southern
California, Los Angeles, CA, USA, where he is
currently a Professor in the same department and the
Director of the Center of Advanced Transportation
Technologies. He also holds courtesy appointments with the Department of
Aerospace and Mechanical Engineering and the Department of Industrial and
Systems Engineering. He is the Associate Director for Research with the
METRANS Transportation Center, University of Southern California. He was
a Visiting Professor with the University of Newcastle, Newcastle, Australia,
and the Australian National University, Canberra, Australia, in the Fall of 1988
and the Technical University of Crete, Chania, Greece, in the Summer of 1992
and Fall of 2001. He served as the Dean of the School of Pure and Applied
Science with University of Cyprus, Nicosia, Cyprus, in 1995. In 2009 he was
with the Department of Electrical Engineering and Information Technologies,
Cyprus University of Technology, Limassol, Cyprus, while on sabbatical leave
from University of Southern California. He is the author or coauthor of eight
books and over 300 research papers in the areas of controls, vehicle dynamics,
neural networks, nonlinear dynamical systems, and intelligent transportation
systems.
Dr. Ioannou received the Outstanding Transactions Paper Award from the
IEEE Control System Society in 1984 and the 1985 Presidential Young In-
Tooraj Rajabioun received the B.Sc. degree in elec- vestigator Award for his research in adaptive control. In 2009 he received the
trical engineering from University of Tehran, Tehran, IEEE ITSS Outstanding ITS Application Award and the 2009 IET Heaviside
Iran, in 2006. He is currently working toward the Medal for Achievement in Control from the Institution of Engineering and
Ph.D. degree with University of Southern California, Technology (former IEE). In 2012 he received the IEEE ITSS Outstanding ITS
Los Angeles, CA, USA. Research Award. He has been an Associate Editor for IEEE T RANSACTIONS
His research area includes controls, intelligent
ON AUTOMATIC C ONTROL , International Journal of Control, Automatica, and
transportation systems, estimation, and statistics. IEEE T RANSACTIONS ON I NTELLIGENT T RANSPORTATION S YSTEMS. He
is a member of the Board of Governors of the IEEE Intelligent Transportation
Society. He is a Fellow of the International Federation of Automatic Control
and the Institution of Engineering and Technology.

Authorized licensed use limited to: Sathyabama Institute of Science and Technology. Downloaded on January 30,2025 at 08:03:33 UTC from IEEE Xplore. Restrictions apply.