Received 4 August 2023; revised 30 October 2023; accepted 9 November 2023.

Date of publication 28 November 2023; date of current version 8 December 2023.

Digital Object Identifier 10.1109/OJITS.2023.3335303

Nonlocal Calculus-Based Macroscopic Traffic Model:

Development, Analysis, and Validation
PUSHKIN KACHROO 1 (Senior Member, IEEE), SHAURYA AGARWAL 2 (Senior Member, IEEE),
ANIMESH BISWAS3 , AND ARCHIE J. HUANG 2 (Member, IEEE)
1 Department of Electrical and Computer Engineering, University of Nevada Las Vegas, Las Vegas, NV 89154, USA

2 Civil, Environmental and Construction Engineering Department, University of Central Florida, Orlando, FL 32826, USA

3 Department of Mathematics, University of Nebraska at Lincoln, Lincoln, NE 68588, USA

CORRESPONDING AUTHOR: S. AGARWAL (e-mail: [email protected])

ABSTRACT Nonlocal calculus-based macroscopic traffic models overcome the limitations of classical
local models in accurately capturing traffic flow dynamics. These models incorporate “nonlocal” elements
by considering the speed as a weighted mean of downstream traffic density, aligning it more closely with
realistic driving behaviors. The primary contributions of this research are manifold. Firstly, we choose
a nonlocal LWR model and Greenshields fundamental diagram and prove that this traffic flow model
satisfies the well-posed conditions. Furthermore, we prove that the chosen model maintains bounded
states, laying the groundwork for developing numerically stable schemes. Subsequently, the efficacy of
the proposed nonlocal model is evaluated through extensive field validation using real traffic data from
the NGSIM dataset and developing a stable numerical scheme. These validation results highlight the
superiority of the nonlocal model in capturing traffic characteristics compared to its local counterpart and
establish its enhanced accuracy in reproducing complex traffic behavior. Therefore, this research expands
both the theoretical constructs within the field and substantiates its practical applicability.

INDEX TERMS Macroscopic traffic model, nonlocal calculus, nonlocal LWR model.

I. INTRODUCTION and the velocity still depends on completely local traffic

LASSIC macroscopic models such as the LWR model behavior at the same (t, x).
C ∂t ρ(t, x) + ∂x f (ρ(t, x)) = 0, where ρ : R+ × [a, b] →
[0, ρm ] is the time and space varying traffic density defined
The other method that can be used employs a nonlocal
velocity effect. An example is the model studied in [1],
on a road segment [a, b] and where the values of the function and [2]. The model is ∂t ρ(t, x) + ∂x [ρ(t, x)v(ρ(t, x) ∗d
output reside in [0, ρm ], ρm being the jam density. The wd (x))] = 0, where ∗dis the downstream convolution defined
x+d
density-dependent traffic flow is f (ρ) and is given as the as ρ(t, x) ∗d wd (x) = x ρ(t, y)wd (y − x)dy.  Here wd (x) is
product of density-dependent speed v(ρ) and density as a nondecreasing probability function, i.e., R wd (x)dx = 1,
f (ρ) = ρv(ρ). with support on [0, d]. This model clearly takes into account
The problem with this model is in the velocity term the lookahead nonlocal nature of driving.
v(ρ(t, x)) as the velocity at (t, x) is the function of traffic There is a great deal of recent development in the topic of
density ρ at (t, x). This means that there is no look-ahead extending traffic models and conservation laws to nonlocal
behavior of the driver in the model. There are two ways versions [3], [4], [5], [6], [7], [8], [9], [10], [11]. This
to deal with it. One way is to add a gradient term to the makes sense as the actual reality and the physics of the
velocity function as v(ρ, ∂x ρ) = v0 (ρ) + k∂x ρ. This converts phenomenon clearly indicate the driver’s view of integrating
the original hyperbolic LWR PDE into a diffusive (parabolic) traffic conditions in front, which corresponds to the spatial
one. However, the problem is that the model is still local, integration of local traffic density in front of the driver.
Research Gaps: The development of a nonlocal LWR
The review of this article was arranged by Associate Editor Yajie Zou. model and its usage for practical applications requires

c 2023 The Authors. This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License.
For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
900 VOLUME 4, 2023
addressing some issues. Firstly, the well-posedness is an
essential property of a partial differential equation model. It
entails a theorem of the existence, uniqueness of its solution,
and the continuity of the solution to its initial condition.
The chosen model must be well-posed for it to be useful.
Secondly, a prevailing limitation in numerous numerical
schemes for nonlocal models is the inability to ensure that
the original bounds imposed on the traffic flow are preserved FIGURE 1. Nonlocal Traffic Density Field.
throughout its evolution [11]. This constraint is critical for
practical applications, particularly the upper bound on traffic
jam density, which originates from the physical principles on x ∈ [0, ], where ρ(t, x) ∈ [0, ρm ], ρm being the jam
governing traffic modeling. Lastly, field validations are density, i.e., the traffic density when the traffic speed is zero.
required to compare the efficacy of local and nonlocal models The variable ρd (t, x) is the smoothed look-ahead traf-
in capturing the complex traffic phenomenon. fic density obtained by a convolution of ρ(t, x) with a
Contributions: This paper addresses the research gaps look-ahead kernel function η(x). The kernel function η(x)
mentioned earlier. Specifically, it contributes to the field of generalizes the probability density function wd (x), defined
traffic flow modeling as follows. earlier.
 x+d
1) We demonstrate the well-posedness of the adopted ρd (t, x) = ρ(t, x)  η(x) = ρ(t, y)η(y − x)dy (1)
nonlocal LWR model and the Greenshields fundamen- x
tal diagram by proving (Theorems 2 and 3) that they where symbol  has been used for a look-ahead convolution.
satisfy the well-posedness conditions provided in [1], The kernel η(x) ∈ Cn ([0, d];R+ ) is a non-increasing function
and [2]. and has the following property.
2) We resolve the boundedness issue through Lemma 1.  d
By proving the Lemma, we demonstrate that the
η(x)dx = 1 (2)
nonlocal LWR model adheres to the boundedness 0
without violation, enabling the construction of stable The conservation law is
numerical schemes.
3) We implement an adapted Lax-Friedrichs algorithm ∂t ρ(t, x) + ∂x [ρ(t, x)v(ρd )] = 0 (3)
to generate numerically stable, approximate solutions.
As mentioned previously, we can use one of the models for
The schemes were designed using two look-ahead
v(ρd ) as a function of ρd . Without loss of generality, we
kernels.
will use Greenshields relationship in this paper:
4) Finally, the practical effectiveness and robustness of  
our proposed nonlocal model are substantiated through ρd
v(ρd ) = vf 1 − (4)
a thorough field validation using real-world traffic data ρm
from the NGSIM dataset. With additional analysis of
where vf is the free low speed, and ρm is the jam density.
the convolution window size and model performance
comparison, we showcase the developed models’
superior performance and utility in reconstructing the B. A GENERAL INITIAL BOUNDARY VALUE PROBLEM
field data. In this section, we define the sense in which the solution of
the conservation law with its initial condition and boundary
conditions has to be considered for it to be well-posed [22].
II. BACKGROUND
However, the model used in [22] is generic, and we
In this section, we will illustrate a nonlocal version of the
will adapt it specifically for developing a nonlocal initial
traffic model originally developed by Lighthill, Whitham,
boundary value problem for LWR paired with Greenshields
and Richards called the LWR model [12], [13], [14]. The
fundamental diagram.
speed relationship with density can be linear (Greenshield’s
In order for the model that we present to be useful for
model, [15]), logarithmic model (Greenberg model, [16]),
applications, the model must be well-posed; otherwise, the
exponential (Underwood model, [17]), bell-shaped (Drake
model becomes useless for any meaningful analysis. Well-
model, [18]), etc. The models can also exhibit multiple
posedness is an important property of a partial differential
phases [5], [19] and use statistical Physics [20]. There also
equation (PDE). A well-posed PDE entails a theorem of
exist higher-order traffic flow dynamics and traffic network
the existence and uniqueness of its solution as well as the
models [21].
continuity of the solution to its initial condition. Specifically,
the model must have at least one solution for the given
A. NONLOCAL LWR MODEL initial condition, which refers to the existence of the
The nonlocal model that we will use is illustrated in Figure 1. solution. Moreover, there should only be one solution for
Traffic density ρ(t, x) is a time-varying scalar field defined the given initial condition. The solution should be stable

VOLUME 4, 2023 901

KACHROO et al.: NONLOCAL CALCULUS-BASED MACROSCOPIC TRAFFIC MODEL

 ∞ 
to initial conditions, which means that small changes to + sgn(ρb − κ) f (t, x, κ, r(t, b))
initial conditions should produce only small changes to the 0 
evolution of the system trajectories. To this end, Definition 1 − f (t, x, ρ(t, b−), r(t, b)) φ(t, b)dt ≥ 0 (8)
is a weak entropy (Kruzhkov) solution. Definition 1 and
The initial and boundary conditions are understood in
Theorem 1 illustrate the requirements for the model to be
a weak sense consistent with entropy and satisfy trace
well-posed.
property [23], [25], [26].
1) MODEL DESCRIPTION
(sgn(ρ(t, a+) − κ) − sgn(ρa (t) − κ))
The general initial boundary value problem is
(f (t, a, ρ(t, a+), r(t, a)) − f (t, a, κ, r(t, a))) ≤ 0,
∂t ρ(t, x) + ∂x f (t, x, ρ, ρ(t, x) ∗ η(x)) = 0, (t, x) ∈ R+ × (a, b)
∀κ ∈ R, x = a
ρ(0, x) = ρ0 , x ∈ (a, b)
(sgn(ρ(t, b−) − κ) − sgn(ρb (t) − κ))
ρ(t, a) = ρa (t), t ∈ R+
(f (t, b, ρ(t, b−), r(t, b)) − f (t, b, κ, r(t, b))) ≥ 0,
ρ(t, b) = ρb (t), t ∈ R+ (5)
∀κ ∈ R, x = b (9)
f ∈ C2 (R+ × [a, b] × R × R; R) satisfies
f (t, x, 0, r) = 0, ∀t, x, r, 2) WELL-POSEDNESS AND PROPERTIES

sup |∂ρ f (t, x, ρ, r)| < K1 , The well-posedness result and solution properties of the
t,x,ρ,r model given in Section II-B1. (titled, Model) are presented
sup |∂x f (t, x, ρ, r)| < K2 |ρ|, in the theorem below [22].
t,x,r Theorem 1: Given the hypothesis and conditions of
sup |∂r f (t, x, ρ, r)| < K2 |ρ|, Equation (5), Equation (6), and Equation (7), ρ0 ∈ (L∞ ∩
t,x,r
BV)((a, b); R+ ), ρa , ρb ∈ (L∞ ∩BV)(R+ ; R+ ), then ∀t > 0,
sup |∂xx f (t, x, ρ, r)| < K2 |ρ|,
t,x,r ρ ∈ BV((0, t) × (a, b); R+ ) is a weak entropy solution in
sup |∂xr f (t, x, ρ, r)| < K2 |ρ|, the sense of Kruzhkov satisfying the conditions prescribed
t,x,r in Equation (8), and having the properties stated here.
sup |∂rr f (t, x, ρ, r)| < K2 |ρ| (6) The properties provide bounds on the ρ(t, ·) L∞ (a,b) ,
t,x,r
ρ(t, ·) L1 (a,b) , and ρ(t, ·) TV(a,b) norms.
We have K1 > 0, K2 > 0 which are constants, and η ∈
(C1 ∩ W1,∞ )(R; R) with ρ(t, ·) L∞ (a,b) ≤ eLt ρ0 L∞ (a,b)
 
η(x)dx = 1 (7) ρ(t, ·) L1 (a,b) ≤ ρ0 L1 (a,b) + α ρa
L1 [0,t] + ρb L1 [0,t]
R κ 2  κ1 t
TV(ρ(t, ·); (a, b)) ≤ eκ1 t TV(ρ0 ; (a, b)) + e −1
Conservation laws generally have discontinuities due to κ1
shock waves and contact discontinuities, as well as rar- + TV(ρa ; [0, t]) + TV(ρb ; [0, t]) (10)
efaction waves, which require the notion of weak and
Continuity of the solution in L1 (a, b) with respect to time
vanishing viscosity or entropy solutions. The initial and
is shown by
boundary values also need to be understood in the weak
sense [22], [23], [24] as shown below. For the sake of ρ(t2 , ·) − ρ(t1 , ·) L∞ (a,b) ≤ C(t2 )(t2 − t1 ) (11)
completeness, we borrow the following definition for this
notion of the solution from [22]. The uniqueness of the solution is obtained by the continuity
Definition 1: Given the initial condition ρ0 ∈ property of the solution with respect to the initial and
L∞ ((a, b); R), and the two boundary conditions ρa , ρb ∈ boundary data as shown in Equation (12) where ρ and σ are
L∞ (R+ ; R), ρ ∈ BV(R+ × (a, b); R) is a weak entropy two solutions with initial data ρ0 and σ0 , and the boundary
solution in the sense of Kruzhkov, if ∀φ ∈ C10 (R2 ; R+ ), and data ρa , ρb , σa , and σb .
∀κ ∈ R, we have ρ(t, ·) − σ (t, ·) L1 (a,b)
 ∞  b  

|ρ − κ|∂t φ + sgn(ρ − κ) f (t, x, ρ, r(t, x)) ≤ eS t ρ0 − σ0 L1 (a,b) + κ4 ρa − σa L1 [0,t]

0 a
 + ρb − σb L1 [0,t] (12)
− f (t, x, κ, r(t, x)) ∂x φ

d The terms α, L, S, κ1 , κ2 , κ3 , and κ4 are all defined
− sgn(ρ − κ) f (t, x, κ, r(t, x))φ dxdt in [22].
dx
 b
+ |ρ0 − κ|φ(0, x)dx III. ANALYSIS OF NONLOCAL INITIAL BOUNDARY
 ∞
a VALUE LWR

+ sgn(ρa − κ) f (t, x, ρ(t, a+), r(t, a)) In this section, we will apply the existing well-posedness
0  results for the model shown in Section II-B2. to our chosen
− f (t, x, κ, r(t, a)) φ(t, a)dt initial boundary value problem shown in Section II-A. In
902 VOLUME 4, 2023
other words, we prove that the chosen model - nonlocal Theorem 2: The solution to the nonlocal initial boundary
LWR paired with Greenshields fundamental diagram - is value LWR presented in Equation (13) satisfies
well-posed.
sup |∂ρ f (t, x, ρ, r)| < K1 (19)
t,x,ρ,r
∂t ρ(t, x) + ∂x f (t, x, ρ, ρ(t, x)  η(x)) = 0,
(t, x) ∈ R+ × (a, b) Proof: We obtain the formula for f (t, x, ρ, r) from
Equation (14).
ρ(0, x) = ρ0 , x ∈ (a, b),  
r
ρ(t, a) = ρa (t), ρ(t, b) = ρb (t), t ∈ R+ (13) f (t, x, ρ, r) = ρv(r) = vf ρ 1 − (20)
ρm
f ∈ C2 (R+ × [a, b] × R × R; R) is given by We have already shown in Lemma 1 that r ∈ [0, ρm ],
f (t, x, ρ, ρ(t, x)  η(x)) and hence
   
ρd r
= ρv(ρd ) = vf ρ 1 − fρ (t, x, ρ, r) = v(r) = vf 1 − ≤ vf (21)
ρm ρm
 x+d
ρ(t, x)η(y − x)dy
= vf ρ 1 − x (14) Theorem 3: The solution to the nonlocal initial boundary
ρm
value LWR presented in Equation (13) satisfies
Now, we check the conditions on the flow f to be satisfied
sup |∂x f (t, x, ρ, r)| < K2 |ρ|,
for our model. t,x,r
sup |∂r f (t, x, ρ, r)| < K2 |ρ| (22)
t,x,r
A. FLOW CONDITIONS
The conditions to be satisfied and the corresponding analysis Proof: We again use the formula for f (t, x, ρ, r) from
is presented next. Equation (14). We easily see that fx (t, x, ρ, r) = 0 and hence,
the first condition is satisfied. For the second condition, we
1) ZERO CONDITION perform the following analysis:
 
The condition to be satisfied is r vf
|∂r f (t, x, ρ, r)| = ∂r vf ρ 1 − ≤ |ρ| (23)
f (t, x, 0, r) = 0, ∀t, x, r (15) ρm ρm

which is clearly satisfied by substituting ρ = 0 into the

formula for f provided in Equation (14). 3) SECOND ORDER CONDITIONS
The conditions to be satisfied are
2) FIRST ORDER CONDITIONS
The conditions to be satisfied are sup |∂xx f (t, x, ρ, r)| < K2 |ρ|,
t,x,r
sup |∂ρ f (t, x, ρ, r)| < K1 , sup |∂xr f (t, x, ρ, r)| < K2 |ρ|,
t,x,ρ,r t,x,r
sup |∂x f (t, x, ρ, r)| < K2 |ρ|, sup |∂rr f (t, x, ρ, r)| < K2 |ρ| (24)
t,x,r t,x,r
sup |∂r f (t, x, ρ, r)| < K2 |ρ| (16) We see by using the formula for f (t, x, ρ, r) from
t,x,r
Equation (14) that all the three-second order partial deriva-
In order to prove supt,x,ρ,r |∂ρ f (t, x, ρ, r)| < K1 , we first tives are zero and, hence, trivially satisfy the second order
prove the following lemma. conditions for the model.
Lemma 1: Given ∀x ∈ (a, b), ρ0 (x) ∈ [0, ρm ] =⇒ ∀t >
0, x ∈ (a, b), ρ(t, x) ∈ [0, ρm ]. B. NUMERICAL SCHEME AND KERNEL DESIGN
Proof: On any characteristic parameterized as (t, x(t)), In [22], an adapted Lax-Friedrichs algorithm has been used to
we have produce approximate solutions to the general model in order
to prove the well-posedness by the application of Helly’s
∂t ρ(t, x) + ∂x f (t, x(t), ρ(t, x(t)), ρd (t, x(t))) = 0 (17)
compactness theorem. We divide our space interval (a, b)
On the initial point of this characteristic, we obtain into n steps with the space increment x taken as x =
(b − a)/n and time increment t. Then based on (13), the
|ρd (0, x(0))| = |[ρ(0, x)  η(x)](x(0))| time update difference equation is given by,
 x+d 
= ρ(0, x)η(y − x)dy (x(0))   1    
x ρ n + 1, j = ρ n, j + 1 + ρ n, j − 1
 x+d  2
t    
≤ ρm η(y − x)dy (x(0)) = ρm (18) − f n, j + 1, ρd n, j + 1
x 2 x    
− f n, j − 1, ρ n, j − 1 , (25)

VOLUME 4, 2023 903

KACHROO et al.: NONLOCAL CALCULUS-BASED MACROSCOPIC TRAFFIC MODEL

FIGURE 3. Spatio-temporal binning strategy.

IV. FIELD VALIDATION

A. DATA DESCRIPTION
For this paper, we use traffic flow data of the US-101 freeway
from the NGSIM dataset. The NGSIM program captured
high-resolution vehicle trajectory video data at four locations
between 2005 and 2006 [28]. In this program, freeway and
FIGURE 2. Convolution Kernels. urban arterial data were collected to assist the FHWA Traffic
Analysis Tools Program. The two freeway segments are
I-80 and US-101, whereas the two arterial segments are
where ρd (t, x) is the nonlocal density at the point (t, x) as Lankershim Boulevard and Peachtree Street. A network of
defined via look-ahead convolution. If η(x) is the kernel the synchronized digital camera was used to capture video
defined on the interval [0, d] where d < (b − a), then to data, which was translated into microscopic vehicle trajectory
compute (1) numerically we use, data via a customized software application called NGVIDEO.
 n
d −1  The temporal resolution of the NGSIM trajectory data is 100
   
ρd n, j = ρ n, j + i η[i] x, (26) ms. That translates to a sampling frequency of 10 Hz. These
i=0 high-resolution vehicle trajectory data offer the exact location
and detailed lane positions of each vehicle. In addition, it
where nd x = d. Finally we use (14) to compute the flux.
provides a vehicle’s relative locations with respect to other
For our simulations, we use the following two types of the
vehicles in the study area.
kernel.
Freeway Segments: Vehicle trajectories on a 1640ft long
1) The constant kernel, η(x) = Kχ[0,d] (x) such that K = segment of I − 80 in Emeryville, California, were captured
1/d. Here χ[0,d] (x) is the characteristics function. from the seven synchronized digital video cameras on April
2) Linearly decreasing kernel, η(x) = K(d − x)χ[0,d] such 13, 2005. The 45 minutes long video data has three equal
that K = 2/d2 . time segments from three different periods: 1) 4:00 p.m. to
d
In both cases, K is such that 0 η(x)dx = 1. We notice 4:15 p.m., 2) 5:00 p.m. to 5:15 p.m., and 3) 5:15 p.m. to
that to compute ρd at the space variable j; we need 5:30 p.m. Vehicle trajectories on a 2100ft long segment of
to know the local density data ρ in the interval [j, j + US101 (Hollywood Freeway) in Los Angeles, California,
nd − 1]. Due to this reason, the right boundary conditions were captured from the eight cameras positioned on a 36-
need to be ‘thick.’ We call this boundary condition to story building near to the freeway. The 45-minute video data
be nonlocal. Another important aspect of this numerical has three equal time segments: 1) 7:50 a.m. to 8:05 a.m.,
algorithm is stability. We observe that if we have t 2) 8:05 a.m. to 8:20 a.m., and 3) 8:20 a.m. to 8:35 a.m.
sufficiently small compared to x, then we almost always get June 15, 2005. There is an on-ramp and an off-ramp
a numerically stable solution due to the Courant-Friedrichs- at approximately 590ft (Ventura Boulevard) and 1280ft
Lewy condition [27]. While working with the real data, we (Cahuenga Boulevard) locations, respectively, on the studied
implement a smaller time discretization than given in the segment of US101. There is an on-ramp at approximately
dataset using a simple interpolation scheme. That scheme 400ft location on the studied segment of the I − 80 highway.
helps us to get numerically stable solutions for the nonlocal US101 has five main lanes and one auxiliary lane between
equation. The illustration of the convolution kernels is shown the ramps in the studied areas, while I−80 has six mainlines,
in Figure 2. including a high occupancy lane.

904 VOLUME 4, 2023

FIGURE 4. The NGSIM Dataset, US-101 Freeway, 7:50pm - 8:35pm June 15, 2005.

the density and velocity data points and then applied linear
regression. The y-intercept of the linear regression is the jam
density, and the x-intercept is the free flow velocity. The
estimated free-flow speed vf is 80 feet per second, and the
estimated jam density ρm is 0.12 vehicle per ft.

B. RESULTS AND DISCUSSION

We evaluate the effectiveness of the nonlocal LWR
model paired with Greenshields fundamental diagram by
FIGURE 5. Parameter Estimation of ρm and vf . performing the reconstruction of field data similar to
our previous works using physics-informed deep learning
(see [31], [33], [34]).
Binning Strategy: NGSIM provides high-resolution vehicle
Now, we present the reconstructed density field by five
trajectory data, which is in a Lagrangian setting. To construct
models - one local LWR model and four nonlocal LWR
velocity and density, the trajectories need to be converted into
models using the modified Lax-Friedrichs scheme described
Eulerian bins. Some popular binning methods are provided
earlier. The nonlocal LWR models vary in configuration
in [29], [30]. Next, we briefly describe how to convert the
based on the kernel type (constant or linearly decreasing)
vehicular traces to extract their spatio-temporal pattern. For
and the width of the kernel window (60 ft and 200 ft).
example, let’s define the 2D spatio-temporal domain or grids
The numerical solution setup of nonlocal LWR models
as [0, l]×[0, t] (see Figure 3). Any bin in the defined domain
using the modified Lax-Friedrichs scheme is shown in
is characterized by
   Figure 6. We use the initial condition (t = 0), the lower
bi,j = [i x, (i + 1) x] × j t, (j + 1) t i∈(1···n ),j∈(1···n ) boundary condition (x = 0), and a thick upper boundary
x t
(27) condition (for the computation of ρd ) as the sole training
input. For nonlocal models, the convolution kernel acts in
where, the number of spatial bins nx = l
x and number of the rolling look-ahead window shown in the figure.
temporal bins nt = t t We use the relative squared error (RSE) defined in (29)
Ṽi,j = Meantrace∈bi,j (V(trace)) (28) to measure the inaccuracy in the reconstructed density field.
In (29), ρ(x( j) , t( j) ) is the ground truth value of vehicle
We assume that the traffic variables, such as velocity, density at location x( j) and time t( j) , and ρ̂(x( j) , t( j) ) is the
density, or flow, are constant in each bin and that trajectories estimated value of ρ(x( j) , t( j) ) from the model output. N1
leave a trace on the spatio-temporal bins [29]. Considering is the total number of spatial locations, and N2 is the total
x = 20ft/bin and t = 5sec/bin, we have constructed number of timestamps in the dataset:
velocity and density fields from NGSIM data of US-101 N1 ·N2  ( j) ( j)  2
freeway, which is shown in Figure 4(a) and Figure 4(b). j=1 ρ̂ x , t − ρ x( j) , t( j)
RSE = N1 ·N2  ( j) ( j) 2 × 100% (29)
Figure 4(c) shows the corresponding flow field.
j=1 ρ x , t
For estimating traffic parameters such as free-flow velocity
and jam density, we utilized the density and velocity data set The model with the smallest RSE is the nonlocal LWR
discussed earlier (see Figure 5). More details on the data can model with a linearly decreasing kernel and a 60-ft look-
be found in [31], [32]. We first created a mapping between ahead window. The evaluation of reconstruction results is

VOLUME 4, 2023 905

KACHROO et al.: NONLOCAL CALCULUS-BASED MACROSCOPIC TRAFFIC MODEL

FIGURE 6. Reconstruction by nonlocal LWR with Look-ahead.

TABLE 1. Performance analysis: Inaccuracy in reconstruction.

TABLE 2. Sensitivity analysis: Convolution window size on model performance.

tabulated in Table 1, and the best performance is marked in
bold.
For comparison, the original NGSIM data is plotted in
Figure 7(a), and the reconstructed density fields are shown in
the following figures: Figure 7(b) is the reconstruction by the
local LWR model. Figure 7(c) and Figure 7(d) are generated
by the nonlocal (constant kernel) LWR models with a look-
ahead window of 60-ft and 100-ft, respectively. Similarly,
Figure 7(e) and Figure 7(f) are individually produced by
nonlocal (linearly decreasing kernel) LWR models with a
look-ahead window of 60-ft and 100-ft.
Additionally, we perform the sensitivity analysis on the
size of the convolution window and investigate its impact
on the performance of the nonlocal models. The results are
summarized in Table 2 and visualized in Figure 8.
From the sensitivity analysis result, we observe that V. CONCLUSION
the best-performing convolution window size for the linear In conclusion, this work propels the field of macroscopic
kernel is 60 feet. The model with the linearly decreasing traffic flow modeling forward by addressing the boundedness
convolution kernel, which gradually discounts the weight issue inherent in previous models. A nonlocal LWR model
of traffic information further away from the driver within paired with Greenshields fundamental diagram was chosen
the 60-ft convolution window, achieved the best RSE results for analysis, and the well-posedness of the model and bound-
among all models. If the convolution kernel is substituted edness of the states were established within the previously
with the constant version, then equal weights are assigned established framework in the literature. The resolution of
to the traffic state values in front of the driver during the this limitation in the chosen model paves the way for
convolution process. The most suitable window size is 160 more realistic and practical traffic modeling. The imple-
feet in this case. mentation of the Lax-Friedrichs algorithm, augmented with

906 VOLUME 4, 2023

FIGURE 7. The NGSIM Dataset and Reconstructions.

in conjugation with nonlocal traffic flow models. The Well-

posedness of these models, boundedness of the states, and
field validation will be promising future research directions.

REFERENCES
[1] S. Blandin and P. Goatin, “Well-posedness of a conservation law
with non-local flux arising in traffic flow modeling,” Numerische
Mathematik, vol. 132, no. 2, pp. 217–241, 2016.
[2] P. Goatin and S. Scialanga, “Well-posedness and finite volume
approximations of the LWR traffic flow model with non-local
velocity,” Netw. Heterogeneous Media, vol. 11, no. 1, pp. 107–121,
2016.
[3] Q. Du, J. R. Kamm, R. B. Lehoucq, and M. L. Parks, “A new approach
for a nonlocal, nonlinear conservation law,” SIAM J. Appl. Math.,
vol. 72, no. 1, pp. 464–487, 2012.
[4] E. Abreu, R. De la Cruz, J. Juajibioy, and W. Lambert,
FIGURE 8. Convolution Window Size on Nonlocal Model Performance. “Lagrangian-Eulerian approach for nonlocal conservation laws,”
J. Dyn. Differ. Equ., pp. 1–47, Jul. 2022. [Online]. Available:
https://doi.org/10.1007/s10884-022-10193-8
two look-ahead kernels, provides a scheme for numerically [5] M. Treiber, A. Hennecke, and D. Helbing, “Derivation, properties,
and simulation of a gas-kinetic-based, nonlocal traffic model,” Phys.
stable and accurate traffic field reconstruction. Through Rev. E, Stat. Phys. Plasmas Fluids Relat. Interdiscip. Top., vol. 59,
validation using the NGSIM dataset, we demonstrate that the no. 1, p. 239, 1999.
proposed nonlocal model outperforms its local counterparts [6] F. A. Chiarello, “An overview of non-local traffic flow models,” in
Mathematical Descriptions of Traffic Flow: Micro, Macro and Kinetic
in capturing real-world traffic flow patterns. Furthermore, the Models. Cham, Switzerland: Springer, 2021, pp. 79–91.
design of two look-ahead kernels and the sensitivity analysis [7] A. Bressan and W. Shen, “On traffic flow with nonlocal flux: A
to identify the optimal look-ahead window size provide relaxation representation,” Arch. Ration. Mech. Anal., vol. 237, no. 3,
pp. 1213–1236, 2020.
valuable insights. Thus, this research makes important [8] A. J. Huang, A. Biswas, and S. Agarwal, “Incorporating nonlocal
theoretical and practical contributions to the field of traffic traffic flow model in physics-informed neural networks,” 2023,
flow modeling. In future work, other fundamental diagrams arXiv:2308.11818.
[9] R. M. Colombo, M. Garavello, and M. Lécureux-Mercier, “A class of
such as the inverse-lambda shaped fundamental diagram [35], nonlocal models for pedestrian traffic,” Math. Models Methods Appl.
and triangular fundamental diagram [36] can be considered Sci., vol. 22, no. 4, 2012, Art. no. 1150023.

VOLUME 4, 2023 907

KACHROO et al.: NONLOCAL CALCULUS-BASED MACROSCOPIC TRAFFIC MODEL

[10] B. Piccoli and F. Rossi, “Transport equation with nonlocal velocity [36] C. Daganzo, Fundamentals of Transportation and Traffic Operations,
in Wasserstein spaces: Convergence of numerical schemes,” Acta vol. 30. Oxford, U.K.: Pergamon, 1997.
Applicandae Mathematicae, vol. 124, no. 1, pp. 73–105, 2013.
[11] P. Amorim, R. M. Colombo, and A. Teixeira, “On the numerical PUSHKIN KACHROO (Senior Member, IEEE)
integration of scalar nonlocal conservation laws,” ESAIM Math. Model. received the first Ph.D. degree in mechanical
Numer. Anal., vol. 49, no. 1, pp. 19–37, 2015. engineering from the University of California at
[12] M. J. Lighthill and G. B. Whitham, “On kinematic waves. I: Flow Berkeley performing research in vehicle control,
movement in long rivers. II: A theory of traffic on long crowded the second Ph.D. degree in mathematics from
roods,” Proc. Roy. Soc. London Series A, vol. A229, no. 1178, Virginia Tech in the area of hyperbolic system
pp. 281–345, 1955. of partial differential equations with applications
[13] G. B. Whitham, Linear and Nonlinear Waves. Hoboken, NJ, USA: to traffic control and evacuation, and the third
Wiley, 2011. Ph.D. degree in physics from the University of
[14] P. I. Richards, “Shockwaves on the highway,” Operat. Res., vol. 4, Nevada Las Vegas, on the topic of Quantum Logic
pp. 42–51, Feb. 1956. and Computation, where he is the Lincy chaired
[15] B. D. Greenshields, “A study in highway capacity,” in Proc. Highway Professor with the Department of Electrical and Computer Engineering.
Res. Board, 1935 pp. 448–477. He has been a Professor with Virginia Tech and a Visiting Professor with
[16] H. Greenberg, “An analysis of traffic flow,” Operat. Res., vol. 7, no. 1, University of California at Berkeley. He has also been a Visiting Professor
pp. 78–85, 1959. with IIT Delhi, IIT Bombay, and IIT Jammu. His M.S. degrees are from Rice
[17] R. T. Underwood, Speed, Volume, and Density Relationships: Quality University, Virginia Tech, and UNLV, and his B.Tech. from IIT Bombay
and Theory of Traffic Flow. New Haven, CT, USA: Yale Bureau in Civil Engineering. He is finishing another M.S. degree from IIT Delhi
Highway Traffic, 1961, pp. 141–188. in Traffic Safety. He has published more than 200 journal and conference
[18] J. Drake, J. Schofer, and A. May, “A statistical analysis of speed- papers, including eleven books and four collected volumes.
density hypotheses,” in Proc. 3rd Int. Symp. Theory Traffic Flow, 1967,
pp. 53–87. SHAURYA AGARWAL (Senior Member, IEEE)
[19] A. Schadschneider, “Statistical physics of traffic flow,” Physica A Stat. received the B.Tech. degree in electronics and
Mechan. Appl., vol. 285, nos. 1–2, pp. 101–120, 2000. communication engineering from the Indian
[20] B. S. Kerner, “The physics of traffic,” Phys. World, vol. 12, no. 8, Institute of Technology Guwahati, Guwahati,
p. 25, 1999. and the Ph.D. degree in electrical engineering
[21] M. Garavello and B. Piccoli, Traffic Flow on Networks, vol. 1. from the University of Nevada, Las Vegas, in
Springfield, MO, USA: Amer. Inst. Math. Sci., 2006. 2015. He is currently an Assistant Professor
[22] C. De Filippis and P. Goatin, “The initial–boundary value problem for with the Civil, Environmental, and Construction
general non-local scalar conservation laws in one space dimension,” Engineering Department, University of Central
Nonlinear Anal., vol. 161, pp. 131–156, Sep. 2017. Florida. He is the founding Director of the Urban
[23] F. Otto, “Initial-boundary value problem for a scalar conservation law,” Intelligence and Smart City (URBANITY) Lab.
Comptes rendus de l’Académie des sciences. Série 1, Mathématique, He was previously an Assistant Professor with the Electrical and Computer
vol. 322, no. 8, pp. 729–734, 1996. Engineering Department, California State University, Los Angeles, from
[24] C. Bardos, A.-Y. LeRoux, and J.-C. Nédélec, “First order quasilinear 2016 to 2018. He was a Postdoctoral Research Fellow with New York
equations with boundary conditions,” Commun. Partial Differ. Equ., University in 2016. As of March 2023, he has published a book, over
vol. 4, no. 9, pp. 1017–1034, 1979. 30 peer-reviewed journal publications, and multiple conference papers. His
[25] F. Dubois and P. Le Floch, “Boundary conditions for nonlinear research focuses on interdisciplinary areas of cyber–physical systems, smart
hyperbolic systems of conservation laws,” J. Differ. Equ., vol. 71, and connected transportation, and connected and autonomous vehicles.
no. 1, pp. 93–122, 1988. Passionate about cross-disciplinary research, he integrates control theory,
[26] K. H. Karlsen, K.-A. Lie, and N. H. Risebro, “A front track- information science, data-driven techniques, and mathematical modeling in
ing method for conservation laws with boundary conditions,” his work. His work has been funded by several private and government
in Hyperbolic Problems: Theory, Numerics, Applications. Basel, agencies. He serves as an Associate Editor for IEEE TRANSACTIONS ON
Switzerland: Birkhäuser, 1999, pp. 493–502. INTELLIGENT TRANSPORTATION SYSTEMS.
[27] R. Courant, K. Friedrichs, and H. Lewy, “Über die partiellen ANIMESH BISWAS received the bachelor’s degree
differenzengleichungen der mathematischen Physik,” Mathematische in electronics and telecommunication engineer-
Annalen, vol. 100, no. 1, pp. 32–74, 1928. ing from Jadavpur University, India, in 2007,
[28] (U.S. Dept. Transp., Washington, DC, USA). “FHWA, 2008a. the first master’s degree from Indian Institute
NGSIM–next generation simulation.” Accessed: May 24, 2021. of Technology Madras, Chennai, in 2009, the
[Online]. Available: https://datahub.transportation.gov/stories/s/i5zb- second master’s degree in electrical engineering
xe34 from Iowa State University in 2015, and the
[29] A. M. Avila, Applications of Koopman Operator Theory to Highway Ph.D. degree in mathematics from Iowa State
Traffic Dynamics. Santa Barbara, CA, USA: Univ. California, 2017. University in 2020. He is currently a Postdoctoral
[30] D. Ni, Traffic Flow Theory: Characteristics, Experimental Methods, Faculty Fellow with the Mathematics Department,
and Numerical Techniques. Oxford, U.K.: Butterworth-Heinemann, University of Nebraska–Lincoln. His research
2015. areas include the analysis of partial differential equations, specifically in
[31] A. J. Huang and S. Agarwal, “Physics informed deep learning for the field of nonlocal modeling. He is a member of American Mathematical
traffic state estimation,” in Proc. IEEE 23rd Int. Conf. Intell. Transp. Society.
Syst. (ITSC), 2020, pp. 1–6.
[32] A. M. Avila and I. Mezić, “Data-driven analysis and forecasting of ARCHIE J. HUANG (Member, IEEE) received the
highway traffic dynamics,” Nat. Commun., vol. 11, no. 1, pp. 1–16, B.Eng. degree in electrical engineering from
2020. Tsinghua University, the M.S. degree in applied
[33] A. J. Huang and S. Agarwal, “Physics-informed deep learning for urban science and informatics from New York
traffic state estimation: Illustrations with LWR and CTM models,” University, and the Ph.D. degree in civil engi-
IEEE Open J. Intell. Transp. Syst., vol. 3, pp. 503–518, 2022. neering from the University of Central Florida.
[34] A. J. Huang and S. Agarwal, “On the limitations of physics-informed He was a Senior Researcher with the Electrical
deep learning: Illustrations using first-order hyperbolic conservation and Computer Engineering Department, California
law-based traffic flow models,” IEEE Open J. Intell. Transp. Syst., State University, Los Angeles. His research
vol. 4, pp. 279–293, 2023. and work interests include the fleet rebalancing
[35] M. Koshi, “Some findings and an overview on vehicular flow problem of shared mobility services, simulation
characteristics,” in Proc. 8th Int. Symp. Transp. Traffic Theory, 1983, with connected and autonomous vehicles, traffic modeling with variable
pp. 403–451. speed limits, and physics-informed deep learning for traffic state estimation.

908 VOLUME 4, 2023