Chapter 9

Macroscopic traffic flow models

Summary of chapter. The previous chapters of these notes discussed macroscopic traffic flow
characteristics. It was show how the main variables density, flow and speed relate via the
continuity equation and the conservation of vehicle equation. It was also discussed how these
relations could be used to determine spatial-temporal dynamics of the traffic flow using shock
wave analysis.
In this chapter, macroscopic traffic flow models are discussed. The aim of these models is to
describe the dynamics of the traffic flow. Several macroscopic (or continuum) traffic flow models
have proven to be simple enough for the real-time simulation of large traffic networks, while
being sufficiently complex to realistically describe de principal aggregate traffic flow variables
and their dynamics. These models are applicable due to the following reasons:

1. Macroscopic models describe the most important properties of traffic flows, such as the
formation and dissipation of queues, shock waves, etc.

2. Macroscopic models are computationally less demanding. Moreover, the computational

demand does not increase with increasing traffic densities, i.e. does not depend on the
number of vehicles in the network.

3. They enable determination of average travel times, the mean fuel consumption and emis-
sions in relation to traffic flow operations.

4. These models can be written in closed-form and are generally deterministic and less sen-
sitive to small disturbances in the input.

In sum, macroscopic network traffic flow models are suitable for short-term forecasting
in the context of network-wide coordinated traffic management. They are applicable in the
development of dynamic traffic management and control systems, designed to optimise the
traffic system and can be used to estimate and predict average traffic flow operations.
The underlying assumption of a macroscopic traffic flow model is that traffic can be de-
scribed as a continuum, similar to a fluid or a gas. Macroscopic equations describing traffic flow
dynamics are hence very similar to models for continuum media.
At this point, let us emphasise that the macroscopic models described in this section are in
fact macroscopic in their representation of traffic. This does not necessarily imply that the traffic
processes or the behaviour of the flow is described in aggregrate terms. On the contrary, several
macroscopic models (i.e. the gas-kinetic models) consider the interaction between vehicles (or
rather, the average number of interactions between the vehicles and the consequent average
changes in the flow dynamics) explicitly. Although the representation of traffic is macroscopic,
the behavioural rules describe the dynamics of the flow is in fact microscopic.

149
150 CHAPTER 9. MACROSCOPIC TRAFFIC FLOW MODELS

List of symbols
k veh/m traffic density
q veh/s traffic volume
Q (k) veh/s equilibrium flow as function of density
v m/s traffic speed
V (k) m/s equilibrium speed as function of density
ω m/s shock wave speed
c m/s kinematic wave speed
qc veh/s capacity
r veh/s inflow (at on-ramp)
s veh/s outflow (at off-ramp)
Ñ (x, t) veh cumulative vehicle count (smoothed)
c (k) , λ∗ m/s kinematic wave speed / characteristic speed
C - characteristics
ki,j veh/m average density for [xi−1 , xi ] at instant tj
qi,j veh/s average flow during [tj , tj+1 ] at interface xi
Tj s time step
li m cell length
D, S veh/s local demand, local supply
τ s relaxation time
λ1,2 m/s characteristic speeds for higer-order model
κ veh/m phase-space density
f - speed probability density function
V0 m/s desired (free) speed
π - immediate overtaking probability
σ - transition probability
Θ m2 /s2 speed variance

9.1 General traffic flow modelling issues

Traffic operations on roadways can be improved by field research and field experiments of real-
life traffic flow. However, apart from the scientific problem of reproducing such experiments, the
problem of costs and safety play a role of dominant importance as well. Traffic flow and micro-
simulation models designed to characterise the behaviour of the complex traffic flow system
have become an essential tool in traffic flow analysis and experimentation. Depending on the
type of model, the application areas of these tools are very broad, e.g.:

• Evaluation of alternative treatments in (dynamic) traffic management.

• Design and testing of new transportation facilities (e.g. geometric designs).

• Operational flow models serving as a sub-module in other tools (e.g. traffic state estima-
tion, model-based traffic control and optimisation, and dynamic traffic assignment).

• Training of traffic managers.

Research on the subject of traffic flow modelling started some forty years ago, when Lighthill
and Whitham [34] presented a macroscopic modelling approach based on the analogy of vehicles
in traffic flow with the dynamics of particles in a fluid. Since then, mathematical description
of traffic flow has been a lively subject of research and debate for traffic engineers. This has
resulted in a broad scope of models describing different aspects of traffic flow operations, such
as microscopic and continuum (or macroscopic) models. The latter macroscopic (or rather
9.2. KINEMATIC WAVE MODEL AND APPLICATIONS 151

continuum models) consider the traffic flow as a continuum, i.e. like a fluid with specific
characteristics, and are the topic of this chapter.
The description of observed phenomena in traffic flow is not self-evident. General mathe-
matical models aimed at describing this behaviour using mathematical equations include the
following approaches:

1. Purely deductive approaches whereby known accurate physical laws are applied.

2. Purely inductive approaches where available input/output data from real systems are
used to fit generic mathematical structures (ARIMA models, polynomial approximations,
neu-ral networks).

3. Intermediate approaches, whereby first basic mathematical model-structures are developed

first, after which a specific structure is fitted using real data.

Papageorgiou [41] argues that it is unlikely that any microscopic or macroscopic traffic flow
theory will reach the descriptive accuracy attained in other domains of science (e.g. Newtonian
physics or thermodynamics). The author states that the only accurate physical law in traffic flow
theory is the conservation of vehicles equation; all other model structures reflect either counter-
intuitive idealisations or coarse approximations of empirical observations. Consequently, the
challenge of traffic flow researchers is to look for useful theories of traffic flow that have sufficient
descriptive power, where sufficiency depends on the application purpose of their theories.

9.2 Kinematic wave model and applications

As mentioned before, the most important equation in any macroscopic traffic flow model is the
conservation of vehicle equation
∂k ∂q
+ = r(x, t) − s(x, t) (9.1)
∂t ∂x
where k = k(x, t) denotes the traffic density, describing the mean number of vehicles per unit
roadway length at instant t and location x; q = q(x, t) denotes the traffic volume or flow rate,
and describes the mean number of vehicles passing the cross-section x per unit time at time t;
r and s respectively denote the inflow and outflow from and to on- and off-ramps.
The kinematic wave model (or LWR model, or first-order model) assumes that the traffic
volume can be determined from the fundamental relation between the density and the traffic
volume (or likewise, the speed), i.e.
q = Q(k) (9.2)
In the previous chapter, we have illustation how, using shock wave theory, we can construct
analytical solutions of the kinematic wave model for simple flow-density relations. However,
it turns out that shock wave analysis will not predict the correct behaviour of the flow for
any fundamental diagram, since it does not predict the occurence of so-called acceleration fans
(described in the ensuing of this chapter). These acceleration fans describe the behaviour of
traffic flowing out of a congested area, e.g. a queue of stopped vehicles.

9.2.1 Analytical solutions using method of characteristics

This section presents the so-called method of characteristics that can be applied to construct
analytical solutions to the kinematic wave model. We can rewrite the conservation of vehicle
equation (assuming that r = s = 0) as follows

∂k ∂q
+ Q0 (k) =0 (9.3)
∂t ∂x
152 CHAPTER 9. MACROSCOPIC TRAFFIC FLOW MODELS

x=L
z>0

Initial conditions k(x,0)

Parameterised curve
(x0(z),t0(z))

z=0 z<0
x=0

Boundary conditions k(0,t) t

Figure 9.1: Parameterised curve (x0 (z) , t0 (z)) describing the boundary of the region of interest.

where Q0 (k) = dQ dk . The kinematic wave model is a first-order partial differential equation. It
is non-linear since Q0 (k) depends on k. The quantity c(k) = Q0 (k) will be referred to as the
characteristic speed, since it describes the speed of the characteristic curves that play a pivotal
role in the solution of hyperbolic equations, such as the kinematic model. For one, it turns
out that the density k is constant along any characteristic curve of the kinematic wave model,
and that the characteristic curves are therefore straight lines. It is also important to note that
in the kinematic model the speed of the traffic u (x, t) is always larger than the characteristic
speed c (k (x, t)), i.e. traffic information never travels faster than the traffic that carries it. This
implies that traffic flow is described as an anisotropic fluid: what happens behind a vehicle
generally does not affect the behaviour of that vehicle.

Method of characteristics overview

Let us now take a closer look at solving equation (9.3). Let us first consider the boundary of
the region of interest. Generally, the boundary consist of

1. the initial conditions (at t = 0) and

2. boundary conditions (at the start x = 0 and end x = L of the roadway section).

We assume that we can construct a parameterised curve (x0 (z), t0 (z)) that describes these
boundary of the region of interest. Fig. 9.1 shows an example of such a curve. Note that along
this curve, the density is known!
For a specific value z, consider another parameterised curve Cz , starting from this boundary
at location (x0 (z) , t0 (z))

Cz = (x(s; z), t(s; z)) with x(0; z) = x0 (z) and t(0; z) = t0 (z) (9.4)

For the example shown in Fig. 9.1, for z > 0, the curve Cz emanates from t0 (z) = 0 (the initial
road conditions), while for z < 0, the curve emanates from x0 (z) = 0 (the upstream boundary
of the region). We will define the curve by the following ordinary differential equations in t
dt
= 1 with t(0; z) = t0 (z) (9.5)
ds
dx
= c (k) = Q0 (k) with x(0; z) = x0 (z) (9.6)
ds
for s > 0. Fig. 9.2 depicts two curves Cz , respectively emanating from the initial road conditions
(in this case, for some z > 0) and from the boundary (for some z < 0). The curves Cz are
9.2. KINEMATIC WAVE MODEL AND APPLICATIONS 153

x=L Curve Cz emanating from

~
k(s,z) = k(x,0) (x0(z),t0(z)) for z > 0

Initial conditions k(x,0)

Curve Cz emanating from
(x0(z),t0(z)) for z < 0

~
k(s,z) = k(0,t)
x=0

Boundary conditions k(0,t) t

Figure 9.2: Curves Cz emanating from boundary describes by (x0 (z) , t0 (z)).

called characteristic curves or z-characteristics.

By defining the curve Cz according to the equations (9.5) and (9.6), it turns out that the
density k̃(s; z) := k(x(s; z), t(s, z)) along the curve Cz satisfies

dk̃ ∂k dt ∂k dx
= + (9.7)
ds ∂t ds ∂x ds
∂k ∂k
= ·1+ · Q0 (k) = 0 (9.8)
∂t ∂x

Since ddsk̃ = 0, the density k along Cz is constant. More specifically, the density along Cz is equal
to the density at the point where the curve Cz originates, i.e. at s = 0. As a result, we have

k̃(s; z) = k (x(s; z), t(s; z)) = k (x0 (z), t0 (z)) (9.9)

Note that this implies that the curve Cz is a straight line, travelling at speed c̃(z) = c (k (x(s; z), t(s; z))) .
To construct the solution of the kinetmatic wave model, initial and boundary conditions
need to be established. By considering the curves emanating from the initial conditions as well
as the boundaries, we can construct the solution k(x, t) for t > 0 and 0 < x < L.
If we, for the sake of argument, neglect the influence of the boundaries of the considered
roadway section, we have seen that the formal solution of Eq. (9.3) is in fact

k (x + t · ck(x, 0), t) = k (x, 0) for all t > 0 (9.10)

The speed of the characteristics (also referred to as waves of constant density) is equal to
the derivative of the equilibrium flow Q(k) = kV (k) and is hence positive as long as k <
kc (unconstrained flow or free flow) and is negative as long as k > kc (constrained flow or
congestion). Furthermore, the speed of the characteristics is bounded by the mean vehicle
speed, i.e.
dQ d (kV (k)) dV
c(k) = = = V (k) + k < V (k) (9.11)
dk dk dk
since dV
dk ≤ 0 (see Fig. 9.3). Only in the region when the average vehicle speed V (k) is constant
(the so-called stable region) and thus dV
dk = 0, we have c(k) = V (k).

Effect of non-linearity of Q (k)

The objective of this section is to illustrate the effects of non-linearity of the flow-density curve,
and particular the differences that will result when applying shock wave analysis and the LWR
154 CHAPTER 9. MACROSCOPIC TRAFFIC FLOW MODELS

slope = car-speed
slope = wave-speed

k kc k

Figure 9.3: Interpretation of the wave speed

k(x,0) = given

k(x,t) = k(x2 ,0)

x2
k(x,t) = k(x1 ,0)
x1

0 t

Figure 9.4: Lines of constant density

theory using non-linear flow-density curves. Waves of different density k will travel with different
wave speeds c (k) = Q0 (k), and thus waves may either fan out or converge and intersect.
If k(x, 0) is a decreasing function of x (or equivalently, if the speed is an increasing function
of x, so that cars are accelerating), then from Fig. 9.3 we notice that Q0 - assumed to be a
decreasing function of k - will be an increasing function of x. This corresponds to the situaiton
illustrated in Fig. 9.4 in which the higher speed waves are initially ahead of the slower speed
waves and so the region of acceleration tends to expand linearly with time as waves spread
further and further apart. Fig. 9.5 shows a decreasing density k(x, 0) at time zero. A short
time later the high density region has moved forward less than the low density region: a sharp
initial change in density thus tends to disperse. At this point, we note that in this situation, we
can determine the density k (x, t) for all t > 0 is k (x, 0) is known. To this end, we only need to
determine the slope of the characteristic emanating from k (x, 0) which intersects point (x, t) ,
as s shown in Fig. 9.4.
If k(x, 0) is an increasing function of x, i.e. cars are decelerating, then the higher speed
waves at low density are initially behind the slower speed waves for the higher density. A
gradually increasing k(x, 0) tends to become steeper as shown in Fig. 9.6, and will eventually
have a vertical tangent. This occurs as soon as any two waves of neighbouring values of k
actually intersect. At this time, the solution (9.10) to the kinematic wave model breaks down
9.2. KINEMATIC WAVE MODEL AND APPLICATIONS 155

k(x,0) k(x,t)

0 x

Figure 9.5: Dispersion of an acceleration wave

k(x,0) k(x,t1) k(x,t2)

0 x

Figure 9.6: Focussing of deceleration waves

and , if it where continued beyond this time, would assign to some (x, t) point more than one
value of k, becomes more than one wave passes through this point. The existence of a vertical
tangent, however, means that ∂k ∂x becomes infinite and the differential equation from which the
solution (9.10) was derived is no longer valid.
A phenomena such as this is familiar in fluid dynamics. The failure of the differential
equation is corrected by allowing discontinuities in k(x, t) which are called shocks. Shock wave
theory was described in chapter 8 of these notes and is needed to complement the method
of characteristics to find analytical solutions to kinematic wave model. In chapter 8 we have
proven that the speed of the shock equals

q2 − q1 Q (k2 ) − Q (k1 )
ω= = (9.12)
k2 − k1 k2 − k1

where k1 and q1 respectively denote the density and flow upstream of the shock S, and k2 and
q2 denote the density and flow downstream of the shock. Recall that in the limit of very weak
shocks, the shock line converges to the tangent of the Q curve (kinematic wave).
To evaluate k(x, t), Eq. (9.12) is used to determine the path of the shock. Fig. 9.7 illustrates
how in a typical problem one can construct the solution k(x, t) graphically. If one is given k(x, 0)
then one can draw the waves of constant k, or use the solution (9.10), to determine k(x, t) at
least until such time t0 when two waves first intersect at a point (x0 , t0 ). At this moment, a
shock starts to form. Let us now determine the time t0 and location x0 at which this occurs.
Noe that except for very exceptional inital conditions k(x, 0) there is generally no perfect
focussing of waves (waves converge on a single point). Therefore, the shock starts as an ‘infin-
itesimal shock’, meaning that the intersecting waves have nearly the same speed. The shock
156 CHAPTER 9. MACROSCOPIC TRAFFIC FLOW MODELS

shock path

x0
x+a
x

0 t0 t

Figure 9.7: Construction of the shock path

will start with a speed approximately equal to those of the two intersecting waves. Futhermore,
it is clear that the speed of the shock will have a speed which is between the speeds of the
intersecting waves1 .
Let t0 and x0 be the first time two waves intersect and the location they intersection respec-
tively. From Fig. 9.7 we observe that two waves starting at the points x and x + a will intersect
at this time t0 and position x0 given that the following condition is met:

x0 = x + c(x)t0 = x + a + c(x + a)t0 (9.13)

where c(x) = Q0 (k(x, 0)). Since t0 is the first time any two waves intersect, it follows that

1 c(x) − c(x + a)
= max (9.14)
t0 x,a a

If c(x) has a continuous derivative, it follows from the mean value theorem that (c(x) − c(x + a)) /a
is equal to the derivate of c at some point between x and x + a. Therefore
1 dc
= max (9.15)
t0 x dx

The values of x for which the maximum is realised also give the values of x in Eqn. (9.13) that
determine the position of x0 where the shock originates.
We now know where the shock starts as well as its initial speed, thus also the position of
the shock a short period later, at time t0 + ∆t. If no new shocks form between t0 and t0 + ∆t,
the values of k on either side of the shock at time t0 + ∆t are still determined by the waves
starting at t = 0. The density on either side of the shock is determined by the wave which
intersects the schock at time t0 + ∆t and approaches the shock from the same side. Another
wave will approach from the other side and determine the density of that side. We now know
the densities on either side of the shock at time t0 + ∆t and we can reevaluate the shock speed
at t0 + ∆t. We can then estimate the shock position at a still later time (e.g. t0 + 2∆t), etc. In
essesnce, we are performing a graphical integrtion of Eqn. (9.12).
1
For weak shocks, the shock speed is approximately the average of the wave speeds at either side of the shock.
9.2. KINEMATIC WAVE MODEL AND APPLICATIONS 157

τ τ+τ’
0
t

Figure 9.8: Cars stopped at a traffic signal

If some other shock should form elsewhere we proceed similarly to find their paths. If two
shocks should intersect, we simple combine them into a single shock as shown in Fig. 9.7. The
shock speed of the single shock is determined by the densities in the regions adjacent to the
shock not including the region annihilated in the collision.

Example application of the kinematic wave model

As a practical example of how the kinematic wave model can be applied, let us consider what
happens when a steady flow of traffic is suddenly stopped, for instance at an intersection, and
then released again, as would occur at a traffic signal. The trajectories of cars which the theory
should predict are shown in Fig. 9.8. The stopping point is at x = 0, and the trajectory
of the first car that is stopped is designated by x1 (t). The kinematic wave model does not
describe in detail the trajectory of car 1. If we are given the density ki , the intensity qi or the
speed vi of the initial approach traffic strean, all other quantities are also determined by the
Q(k) relation. The speed at which car 1 approaches the stop line can be determined, since
the traffic conditions are known. At the stopping line, the speed of the lead car is zero. The
kinematic wave model does not describe the details of deceleration and acceleration of car 1,
and additional information is required in order to correctly describe its behaviour. However, on
a scale of time and distance in which the kinematic wave model aught to be applied, we may
expect that deceleration and acceleration of car 1 is of negligible length of time.
The deceleration of the first car, being nearly instantaneous, creates a shock wave immedi-
atly. It is a shock from the initial state (qi , ki ) to the state q = 0 and k = kj . The slope of the
shock line in Fig. 9.9 between the states determines the shock speed. In Fig. 9.8, the shock
starts at x = 0 and t = 0 and travels upstream with constant speed (shown by the dashed
line). The shock line represents in effect the rear of the queue caused by the traffic signal. The
continuum approximation implies that the decelerations occur practically instantaneously.
When the lead car accelerates again, it sends out a fan of acceleration waves for all car
speeds between 0 and u0 . The slowest wave travels backwards with the wave speed c associatd
with car speed u = 0. Clearly, this wave speed is given by the slope of Q(k) at k = kj , as shown
in Fig. 9.9. When this wave starting at x = 0, t = τ as shown by the dashed line in Fig. 9.8
158 CHAPTER 9. MACROSCOPIC TRAFFIC FLOW MODELS

q slope = shock speed

slope = wave speed

ki kj k

Figure 9.9: Evaluation of flows from a traffic signal

intersects the shock line at point A, this signals the car at the tail at the queue to move. The
shock, however, does not apprear: the waves that intersect the shock only assign net values for
the density or car speed on the front side of the shock. Since

q1 − q2 k1 u1 − k2 u2
ω(t) = = (9.16)
k1 − k2 k1 − k2
where ω again denotes the shock speed, and where state 1 and 2 respectively denote the state
upstream and the state downstream of the shock. Since the flow q1 upstream of the shock
increases, the shock gains forward speed. It eventually moves with a positive speed and crosses
the intersection at x = 0.
The time at which this occurs can be determined easily. We konw that the wave with the
wave speed zero is the wave corresponding to qc since the tangent to the Q(k) curve is horizontal
at q = qc . The number of cars that cross the intersection before the shock arrives is therefore
qc multiplied by the time τ 0 for the shock to arrive. This must, however, also be equal to the
total number of cars that have arrived by this time, i.e.

qi (τ + τ 0 ) = qc τ 0 (9.17)

in which τ is the length of time the traffic is stopped.

The shock, as it moves forward, becomes weaker and weaker. The car speeds at the rear of
the shock remains at vi but the car speeds at the front keep increasing. Eventually, the shock
will overtake all waves for car speeds less thant vi and the shock will degenerate into a wave
or a shock of zero jump. The waves of car speeds larger than vi however, move forward faster
than te shock and the shock never reaches them.

9.2.2 Application of kinematic wave model for bicycle flows at signalised

intersection
Consider a road on which a vehicle flow is moving in a certain direction. For the sake of
argument, we assume that the flow is constant and equal to q1 . We assume that for the flow,
the relation between density and flow can be approximated satisfactory using Greenshield’s
function µ ¶
k
q (k) = u0 k 1 − (9.18)
kj
where u0 is the free speed of the vehicles and kj is the jam-density. Assume that at the start of
our analysis, no queue is present: traffic conditions are stationary and homogeneous (region 1).
The density equals k1 while the flow is equal to q1 = q (k1 ). At a certain time (say, t = −tr ), the
traffic signal that is present at x = 0 turns to red, causing the drivers to stop at the stopping
9.2. KINEMATIC WAVE MODEL AND APPLICATIONS 159

x (m)
F'

Free flow
region 3

Undisturbed
region 1
Acceleration D'
region 4

t (m) A B D
-t r 0 tg
Queue C E
region 2
M Queue E'
region 2’
Undisturbed
region 1

Figure 9.10: Emerging regions for controlled intersection example

line. Downstream of the stopping line, no vehicles will be present (region 3). As a result,
a queue forms at the stopping line, which extend upstream (region 2, x < 0). At t = 0, the
green phase starts and the first cyclists drive away from the stopping line. We assume that the
drivers at the head of the queue react and accelerate instantaneously (without delay) to the free
speed u0 . We have seen that the transition from the jam conditions for x < 0 to the free flow
conditions downstream x > 0 causes a so-called acceleration fan (region 4). The acceleration
fan described how the vehicles inside the congestion drive away from the queue.
At time t = tg , the second red phase starts and cyclists are again held back at the stopping
line. This causes the formation of a new queue (region 20 ). Fig. 9.10 shows qualitatively the
predicted traffic flow conditions resulting when using the kinematic wave model. It also shows
the different emerging regions as well as some points of interest. Furthermore, Fig. 9.11 shows
a number of vehicle trajectories.
Let us now use what we have learnt so far to determine a mathematical solution to this
problem. Let us first consider the shock wave between region 1 and region 2, described by the line
xAC (t). The speed of the shock wave cannot be determined from the method of characteristics.
Rather, we need to apply shock wave theory to determine the wave speed ω AC . Since the traffic
conditions upstream of the shock wave and downstream of the shock wave are constant, the
shock speed is also constant. We find
q2 − q1 q1
ω AC = =− (9.19)
k2 − k1 kj − k1
which, by (tA , xA ) = (−tr , 0)
q1
xAC (t) = ω AC · (t + tr ) = − (t + tr ) (9.20)
kj − k1
The next step is to determine the line xAC (t) describing the boundary between the jam
region 2 and the acceleration region 4. The latter is described by the characteristics that are
emanated from the point (0, 0) (see Fig. 9.10). These characteristics are straight lines described
by the kinematic wave speed
µ ¶
0 k
c (k) = q (k) = u0 1 − 2 (9.21)
kj
160 CHAPTER 9. MACROSCOPIC TRAFFIC FLOW MODELS

x (m)
F'

D'

t (m) A B D

C E

M E'

Figure 9.11: Example vehicle trajectories for controlled intersection example

The characteristic ‘c’ is then described by xc (t) = c · t, where c ∈ [−u0 , u0 ]. It is very important
to note that we can use this expression to determine the density k4 (t, x) (and the speed and
the flow) for all (t, x) in region 4:
kj ³ x´
k4 (t, x) = u0 − (9.22)
2u0 t
1³ x´
u4 (t, x) = u0 + (9.23)
2 t
∙ ¸
1 kj 2 ³ x ´2
q4 (t, x) = u − (9.24)
4 u0 0 t
for all (t, x) in region 4.
The characteristic that moves upstream (in the direction x < 0) at the highest speed moves
upstream with speed −u0 . This is precisely the characteristic that separates region 2 and region
4. Thus, the line xBC (t) moves with speed −u0 and is thus given by the following expression
(since xBC (0) = 0)
xBC (t) = −u0 t (9.25)
We can then determine the coordinates of point C easily. Some straightforward computations
show that
k1 k1
tC = tr and xC = −u0 tC = −u0 (9.26)
kj − k1 kj − k1
The next step is to determine a relation for the shock wave xCE (t) separating undisturbed
region 1 and the acceleration region 4. On the contrary to the shock wave xAC (t), the speed
of the shock xCE (t) is not constant since the downstream conditions (in region 4) are non-
stationary (since they satisfy Eqns.(9.22)-(9.23)). We can still however apply shock wave theory
on the non-stationary conditions. This provides us with the following expression for the shock
wave speed ωCE (t) describing the slope of line xCE (t)
∙ ³ ´2 ¸
2 xCE (t)
kj u0 − t − 4q1 u0
q4 − q1 1
ω CE (t) = = h ³ ´i (9.27)
k4 − k1 2 k u − xCE (t) − 2k u
j 0 t 1 0
9.2. KINEMATIC WAVE MODEL AND APPLICATIONS 161

Since Z t
xCE (t) = xCE (tC ) + ω CE (s) ds (9.28)
tC
we can determine xCE (t) by solving the following ordinary differential equation
d
xCE = ω CE (t) with xCE (tC ) = xC (9.29)
dt
where ω CE (t) depends on xCE (t) according to Eq. (9.27). Solving this differential equation is
not straightforward and requires application of the method of variation of constants. We can
determine that: √
xCE (t) = c (k1 ) t − (u0 + c (k1 )) tC t (9.30)
is the solution that satisfies Eq. (9.29).
Proof. Consider the ordinary differential equation
h ¡ ¢ i
k u2 − x 2 − 4q u
dx 1 j 0 t 1 0
= £ ¡ x ¢¤ (9.31)
dt 2 kj u0 − t − 2k1 u0
Define x = ty. Then
dx dy
=t +y (9.32)
dt dt
and thus Eq. (9.31) becomes
¡ ¢
dy 1 kj u20 − y 2 − 4q1 u0
t +y = (9.33)
dt 2 kj (u0 − y) − 2k1 u0
Then, we can easily derive that
¡ ¢
dy kj u20 − y2 − 4q1 u0 − 2kj (u0 − y) y + 4k1 u0 y
t = (9.34)
dt 2kj (u0 − y) − 4k1 u0
³ ´
y 2 − 2u0 1 − 2 kkj1 y + u20 − 4 kq1j u0
= kj (9.35)
2kj (u0 − y) − 4k1 u0
 
³ ´ k
k1 u0 1− k1
2 − 2u k1 2 j
1 y 0 1 − 2 kj y + u0 − 4 kj u0
= ³ ´ (9.36)
2 k
u0 1 − 2 k1j − y
³ ´ ³ ´2
2 − 2u k1 2 1 − 2 k1
1 y 0 1 − 2 kj y + u0 kj
= ³ ´ (9.37)
2 u0 1 − 2 k1 − y
kj
³ ´ ³ ´2
Note that c (k1 ) = u0 1 − 2 kk1j and c2 (k1 ) = u20 1 − 2 kk1j and thus

dy 1 y 2 − c (k1 ) y + c2 (k1 ) 1 (y − c (k1 ))2 1 1

t =− =− = − y + c (k1 ) (9.38)
dt 2 y − c (k1 ) 2 y − c (k1 ) 2 2
We can again use Eq. (9.32) to see that the equation above becomes
dy 1 1 dx 1x 1
t + y − y = c (k1 ) → = + c (k1 ) (9.39)
dt 2 2 dt 2t 2
This is a linear differential equation, which can be solved by variation of constants. To this end,
we first determine the solution to the reduced differential equation
dx 1x
= (9.40)
dt 2t
162 CHAPTER 9. MACROSCOPIC TRAFFIC FLOW MODELS

which equals √
x=C t (9.41)
The variation of constant approach involves trying the following solution in the non-reduced
differential equation (9.39) √
x = C (t) t (9.42)
which yields
dx 1 1 dC √ 1 C (t) 1
= C (t) √ + t= √ + c (k1 ) (9.43)
dt 2 t dt 2 t 2
dC √ 1
t = c (k1 ) (9.44)
dt 2
√
and thus C (t) = c (k1 ) t + C0 , which in turn yields
³ √´
x (t) = c (k1 ) t + C0 t (9.45)

We can determine the integration constant C0 by using the boundary condition x (tC ) = xC
√ √
C0 = − (u0 + c (k1 )) tC → x (t) = c (k1 ) t − (u0 + c (k1 )) tC t (9.46)

Let us assume that the next red phase starts at t = tg before the shock xCE has reached the
stopping line. Let us first describe the shock wave between region 4 and region 30 . Again, this
shock will not move at a constant speed, since the traffic conditions upstream of the shock are
non-stationary. We have
q4 − q30 q4
ωDD0 (t) = = = u4 (9.47)
k4 − k30 k4
that is, the speed of the shock equals the speed of the last vehicle emanating from the queue
before the start of the second red phase. Again we need to solve a differential equation to
determine the shock wave xDD0 (t)
d 1³ x 0´
xDD0 = u0 + DD with xDD0 (tg ) = 0 (9.48)
dt 2 t
which results in ¡ p ¢
xDD0 (t) = u0 t − t · tg (9.49)
Proof. Eq. (9.48) is an ordinary linear differential equation of the general form a (t) dx dt =
b (t) x+c (t). To solve differential equations of this sort, we first consider the reduced differential
equation
dx x
= (9.50)
dt 2t
which is easily solved by separating the variables x and t
1 1
dx = dt (9.51)
x 2t
which is solved by
1
ln x =ln x + C (9.52)
2
for some integration constant C; x thus satisfies
√
x=C t (9.53)

To solve the non-reduced differential equation, the variation of constant method is used. That
is, we try the following solution √
x = C (t) t (9.54)
9.2. KINEMATIC WAVE MODEL AND APPLICATIONS 163

Substituting this solution in Eq. (9.48) yields

√
C (t) = u0 t + C0 (9.55)

The integration constant C0 can be determined by considering the initial condition x (tg ) = 0
which yields the solution ¡ p ¢
x (t) = u0 t − t · tg (9.56)

In a similar way, we can determine the shock wave xDE (t) separating region 4 and region
20 . The speed of the shock wave equals
q20 − q4 q4
ω DE (t) = =− (9.57)
k20 − k4 kj − k4
1 h ³ xDE ´i
= − u0 − (9.58)
2 t
and thus
d 1h ³ x ´i
DE
xDE = − u0 − with xDE (tg ) = 0 (9.59)
dt 2 t
which has the following solution
¡ p ¢
xDE (t) = −u0 t − tg t (9.60)

Finally, the shock xEE 0 (t) separating region 1 and region 20 moves at constant speed ω EE 0
which equals
ω EE 0 = ωAC (9.61)

9.2.3 Implications for fundamental diagram

The previous section showed how the method of characteristics can be applied to determine
solutions to the kinematic wave model. We have seen how shocks are formed in regions where
the speed decreases with space. We have also seen how traffic leaving a queue gives rise to a
d
so-called acceleration fan, consisting of characteristics with different speeds c (k) = dk Q (k).
Implicitly, during the analysis, we have used special assumptions regarding the shape of the
fundamental diagram. More precisely, we assumed that the fundamental diagram Q (k) is a
concave function, i.e.
d2 Q
≤0 (9.62)
dk2
Let us illustrate why this property is important to represent traffic behaviour correctly. To
this end, consider a flow-density curve that is non-concave in the congested branch. Consider
the case illustrated by Fig. 9.13. In this case, the initial conditons k(x, 0) describe the case
of smoothly decreasing densities. Note that when the Q(k) function is concave, this region
would smooth out further. Fig. 9.13 shows that, since the congested branch of Q (k) is convex
instead of concave, the slopes of the characteristics emanating from the initial conditions at
t = 0 become more negative as x increases. In the end, the waves even intersect, causing a
shock wave. In other words, although dense traffic travels slower, the waves the dense traffic
carries travel faster than the waves in light traffic, given rise to the shock. This shock is thus
formed by traffic that drives out of the high into the lower density region.

Exercise 55 Describe yourself what will happen in the convex region of the fundamental dia-
gram in case the initial k (x, 0) describes conditions in which the density k(x, 0) increases with
increasing x. Also discuss why this is not realistic.

Remark 56 Appendix A discusses an alternative solution approach for the kinematic wave
model that is based on application of Green’s theorem.
164 CHAPTER 9. MACROSCOPIC TRAFFIC FLOW MODELS

k2

k(x,0) k(x,t)

k1

Low density region

x
Q(k)

Q(k1)

Q(k2) High density region

k1 k2 k t

Figure 9.12: Effect of convex congested branch of Q(k).

9.3 Numerical solutions to the kinematic wave model

The advantage of numerical results presented thus far is that they visually depict the effects
of downstream disturbances on upstream flow. Thus they provide a good insight into the
formation and dissipation of queues and congestion in time and space. The major disadvantage
of analytical approaches is that they are generally only applicable to simple situations, e.g. cases
where no on-ramps or off-ramps are present, where initial conditions and arrival patterns are
simple, cases in which there is no interaction between traffic lights, etc. However, in practise,
such ideal situations are seldom encountered and other approaches are needed, namely using
numerical approximations.

9.3.1 Approach to numerical approximation

Numerical solution approaches come in all sizes and shapes. In most cases, however, the ap-
proaches are based on discretisation in time and space. For one, this means that the roadway
is divided in small cells i. In most cases, these cells have a fixed length li (see Fig. 9.13). Cell
i has interfaces at xi−1 and xi , i.e. li = xi − xi−1 . Furthermore, let tj denote a discrete time
instant.
Considering time instant tj , the number of vehicles in cell i at tj can be determined from
using the cumulative flow function Ñ(x, t), or by integration of the density along the x. The
latter yields that the number of vehicles in cell i at instant tj equals
Z xi
k (x, tj ) dx = li · ki,j (9.63)
xi−1

where ki,j denotes the average vehicle density (space-averaged density) in cell i at time instant
tj , with Tj = tj+1 − tj . At the same time, the number of vehicles passing the interface at xi
during a period [tj , tj+1 ) can be determined directly from the cumulative flow function Ñ(x, t),
9.3. NUMERICAL SOLUTIONS TO THE KINEMATIC WAVE MODEL 165

Cell i

qi , j
ki-1 ki ki+1

li-1 xi

Figure 9.13: Discretisation of roadway into cells i of length li

or by integration of the flow rate during this period, i.e.

Z tj+1
q(xi , t)dt = Tj · qi,j (9.64)
tj

where qi,j denotes the average flow (time-averaged flow ) across the interface at xi during the
period [tj , tj+1 ). By conservation of vehicles (or by integration of the conservation of vehicle
equation (9.1) over the region [xi−1 , xi ) × [tj , tj+1 )), we have

li (ki,j+1 − ki,j ) + Tj (qi,j − qi−1,j ) = 0 (9.65)

(assuming that the on-ramp inflows and off-ramp outflows are zero, i.e. r = s = 0) or

ki,j+1 − ki,j qi,j − qi−1,j

+ =0 (9.66)
Tj li

Note that the discrete representation (9.66) is still exact, i.e. no numerical approximation has
been applied so far. The remaining problem is to find appropriate expressions for the time-
averaged flow qi,j . This will require two types of approximation, namely space averaging and
time evolution.

9.3.2 Fluxes at cell-interfaces

In a numerical approximation scheme, in general the spaced-average densities ki,j are computed
for all cells i and all time instants tj . Hence, when establishing approximations for the time-
averaged flow qi,j , we need to express the flow qi,j as a function of the densities in the upstream
cell i, the downstream cell i + 1 at time instants tj and tj+1 , i.e.

qi,j = f (ki,j , k i+1,j , ki,j+1 , ki+1,j+1 ) (9.67)

Clearly, when the flows qi,j have been determined, eqn. (9.66) can be solved, either rightaway
(when qi,j = f (ki,j , ki+1,j )) or using an iterative approach (generally, when qi,j = f (ki,j , k i+1,j , ki,j+1 , ki+1,j+
Schemes where the function f only depends on the ‘current time’ tj are called explicit schemes;
schemes where the function f depends also on the future time tj+1 are called implicit schemes.
In general, explicit schemes are computationally more efficient, but less stable. For now, only
explicit schemes will be considered.

9.3.3 Simple explicit schemes

The most simple scheme imaginable describes the flow qi,j out of cell i as a function of the
conditions in that cell only, i.e.
qi,j = f (ki,j ) = Q (ki,j ) (9.68)
166 CHAPTER 9. MACROSCOPIC TRAFFIC FLOW MODELS

It is obvious that using this expression will not provide a realistic representation of upstream
moving shock waves. Image a situation where the downstream cell i + 1 is congested (e.g. due
to a blockade upstream of the cell). The analytical solution of the kinematic model states that
the flow out of cell i will be equal to zero, and that the region of jam-density will move further
upstream. However, using Eq. (9.68) shows that traffic will keep flowing out of cell i. As a
result, the density in the recieving cell i + 1 will become unrealistically high.
An easy remedy to this problem, would be to assume that the flow out of cell i + 1 is only
a function of the conditions in the receiving cell i + 1, i.e.

qi,j = f (ki+1,j ) = Q (ki+1,j ) (9.69)

Although this would clearly remedy the problem of traffic flowing into congested cells, the
problem now is that the transmitting cell i will also transmit vehicles when its empty. Some
researchers have proposed combining both approaches using weighted averages, i.e.

qi,j = αQ (ki,j ) + (1 − α) Q (ki+1,j ) (9.70)

The problem is then to determine the correct values for α.

9.3.4 Flux-splitting schemes

A class of well-known schemes are the so-called flux-splitting schemes. Recall that the conser-
vation of vehicle equation can be written as follows
∂k ∂k
+ c (k) =0 (9.71)
∂t ∂x
where c (k) = Q0 (k). The wave speeds c (k) in effect describe how information is transmitted
in the flow: when c (k) > 0, information is transmitted downstream, i.e. from cell i to cell
i + 1. When c (k) < 0, the waves propagate in the upstream direction. The flux-splitting
approach is aimed at splitting the flux (flow) into contributions into the downstream direction
and contributions into the upstream directions. These are respectively described by
1
c+ (k) = max {0, c (k)} = (c (k) + |c (k)|) (9.72)
2
1
c− (k) = min {0, c (k)} = (c (k) − |c (k)|) (9.73)
2
Note that ½ ½
c(k) k < kc 0 k < kc
c+ (k) = and c− (k) = (9.74)
0 k ≥ kc c (k) k ≥ kc
R
Note that Q (k) = f (k) = c (κ) dκ. Now, let
Z
±
f (k) = c± (κ) dκ (9.75)

which in turn yields Z ½

k
+ + Q(k) k < kc
f (k) = c (κ) dκ = (9.76)
0 Q (kc ) k ≥ kc
and Z ½
k
− − 0 k < kc
f (k) = c (κ) dκ = (9.77)
0 Q (k) − Q (kc ) k ≥ kc
Note that f + (k) is always positive, while f − (k) is always negative. Furthermore, note that

f (k) = f + (k) + f − (k) = Q (k) (9.78)

9.3. NUMERICAL SOLUTIONS TO THE KINEMATIC WAVE MODEL 167

250

200

density k (veh/k m )
150

100

50

0
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
location x (km)

Figure 9.14: Numerical solution using flux-splitting approach and exact solution.

∗ at the cross-section x is approx-

The flux-splitting scheme infers that the numerical flux fi,j i
imated by considering the downstream moving contribution f + (ki,j ) of cell i and the upstream
moving contribution f ∗ (ki+1,j ) of cell i + 1 as follows

fij∗ = f + (ki,j ) + f − (ki+1,j ) (9.79)

It turns out that this relatively simple scheme provides excellent results, which can be
applied easily in any practical application of the kinematic wave model. From a theoretical
viewpoint, the method has some undesirable properties. It is well known that the approach
tends to smooth shocks. An example of this is shown in Fig. 9.14. The extent to which this
occurs depends on the time-step and the cell lengths.

9.3.5 Godunov schemes

The flux-splitting approach has some practical and theoretical problems, one of which is the fact
that in theory, the flux may become negative. The Godunov scheme, which closely resembles
the flux-splitting scheme, does not have these drawbacks. In fact, it turns out that this scheme
is both simple, and can be interpreted from the viewpoint of traffic flow theory.

Riemann problem2
The Godunov solution scheme is based on the solution of a specific problem, which is know as
the so-called Riemann problem. The Riemann problem is defined by an intial value problem
with the following starting conditions
½
uL x < 0
u (x, 0) = (9.80)
uR x > 0

Note that the LWR model assumes that the flows are in equilibrium, implying that the speeds
determine the densties and the flows.
We can now define four different cases and determine the exact solution for these case. In
the fist case 1, depicted by Fig. (9.15), both the downstream and upstream flow conditions
2
In this section, we assume that the flow-density curve is concave.
168 CHAPTER 9. MACROSCOPIC TRAFFIC FLOW MODELS

kR kR

Location x

Location x
Interface Interface

kL
kL

Time t Time t

Figure 9.15: Case 1: downstream flows (kR , uR ) and upstream flows (kL , uL ) undercritical.

kR
Location x

Interface

kL

Time t

Figure 9.16: Case 2: downstream flows (kR , uR ) undercritical and upstream flows (kL , uL )
overcritical.

are undercritical. When kR > kL (and thus qR > qL ), an upstream moving shock wave will
occur; when kR < kL , an acceleration fan will occur that describes the interface between the
two regions. In either case, the flow f ∗ at the interface at x = 0 equals the downstream flow,
i.e f ∗ = qL .
Case 2 depicts the situation where the downsstream conditions kR are undercritical, while
the upstream conditions are overcritical. In this situation, shown in Fig. 9.16, an acceleration
fan is transmitted from x = 0 at t = 0. Since along the interface at x = 0, the acceleration fan
transmits a characteristic with slope 0, we can use the fact that the slope of a characteristic
satisfies c (k) = dQ/dk to see that the (constant) density along this characteristic equals the
critical density kc , and hence the flow f ∗ at the cross-section at x = 0 equals the capacity qc .
Case 3 represents the situation where the downstream flow conditions are overcritical, while
the upstream flows are undercritical. In this situation, shown in Fig. 9.17, a shock wave will
emerge at x = 0 at time t = 0. The flow f ∗ at the cross-section at x = 0 depends on the
direction of the shock: if the shock moves downstream, i.e. when qR < qL , the flow f ∗ at the
interface equals the upstream flow qL . When the shock moves upstream, which happens when
qR > qL , the flow f ∗ at the interface equals the downstream flow qR . Clearly, the following
relation is valid in this situation: f ∗ = min {qL , qR }.
The final case 4 (Fig. 9.18) is the case where the downstream conditions kR and the upstream
conditions kL are overcritical. Similar to case 1, either an acceleration fan or a shock wave
9.3. NUMERICAL SOLUTIONS TO THE KINEMATIC WAVE MODEL 169

kR
kR

Location x

Location x
Interface Interface

kL
kL

Time t Time t

Figure 9.17: Case 2: downstream flows (kR , uR ) overcritical and upstream flows (kL , uL ) under-
critical.

kR kR
Location x

Location x

Interface Interface

kL kL

Time t Time t

Figure 9.18: Case 4: downstream flows (kR , uR ) and upstream flows (kL , uL ) overcritical.

emerge. In either case, the flow f ∗ across the cross-section at the interface at x = 0 equals the
flow qR in the downstream region.
Table 9.1 summarises the results for cases 1 to 4.
Let us now defined the traffic demand DL of the upstream region and the traffic supply SR
of the downstream region as follows
½
qL kL < kc
DL = (9.81)
qc kL > kc

and ½
qc kR < kc
SR = (9.82)
qR kR > kc
The traffic demand describes the traffic flow aiming to move out of the upstream region L (as-
suming that the flow conditions in the downstream region R allows this flow). For undercritical
conditions, the demand DL equals the flow qL ; for overcritical conditions the demand is limited
to the capacity qc .

Flow f ∗ at interface kR < k c kR > kc

kL < kc qL min {qL , qR }
kL > kc qc qR

Table 9.1: Flow at interface between upstream and downstream region

170 CHAPTER 9. MACROSCOPIC TRAFFIC FLOW MODELS

The traffic supply describes the maximum traffic flow the downstream region R can admit.
When the conditions in the downstream region are undercritical, this flow is limited by the
capacity of the road in this region; when the conditions in the downstream region are overcritical,
the flow is restricted by the flow qR in the downstream region. In a way, this flow describes the
space becoming available per unit time.
Using these definitions of demand and supply, we can use the results depicted in table 9.1
to show that the realised flow f ∗ at the cross-section equals

f ∗ = min {DL , SR } (9.83)

That is, the flow f ∗ at the interface equals the minimum between demand and supply.

Godunov’s scheme
The Godunov scheme is based on the solution to the Riemann problem. To this end, it is
assumed that the conditions in the (flow) transmitting cell i and the recieving cell i + 1 are
homogeneous with the cell, i.e.

k(x, tj ) = kij for all xi−1 ≤ x ≤ xi (9.84)

It needs no further clarification that to determine the flow at the cell interface xi between cell
i and cell i + 1, one in fact has to solve the Riemann problem. Thus, for cell i we define the
demand Di by ½
qi = Q (ki ) ki < kc
Di = (9.85)
qc = Q (kc ) ki > kc
while for cell i + 1 we define the supply
½
qc = Q (kc ) ki+1 < kc
Si+1 = (9.86)
qi+1 = Q (ki+1 ) ki+1 > kc

It the time period is short enough (i.e. assuming that the conditions in cells i − 1 and i + 2
∗ during the period [t , t
will not reach the interface at xi , the flow fi,j j j+1 ) equals

∗
fi,j (ki,j , ki+1,j ) = min {Di , Si+1 } (9.87)

Note that, given the assumption that the conditions in both cells i and i + 1 are homogeneous,
∗ between the cells is an exact solution of the LWR model; the approximation lies in
the flow fi,j
the assumption that the homogeinity assumption holds for each time step tj (which is obviously
not true).
It turns out that the Godunov scheme has excellent properties, such as accurate represen-
tation of shocks, ability to incorporate flow-density relations that are a function of x, etc.

9.4 Higher-order models

The kinematic wave model is a simple yet sufficient theory of traffic if one only cares about the
size and the end of a queue, i.e. the time-space trajectory of a shock. However, traffic flow
phenomena are very complex and some rather important phenomena elude the kinematic wave
model, as we have seen in chapter 5. The kinematic wave model described in the preceding
chapter has been criticised by many authors, mainly because of the following reasons:

• Speeds in the kinematic wave model are described by stationary speed-concentration rela-
tions implying that the speed reacts instantaneously to the traffic concentration without
any delay. Dynamic fluctuations around the equilibrium speed are observed in real-life
traffic flow (hysteresis; recall the acceleration and deceleration curves from chapter 5) but
are not described by the kinematic wave model.
9.4. HIGHER-ORDER MODELS 171

• The kinematic wave theory shows (speed-) shock formation by steeping speed jumps to
infinite sharp discontinuities in the concentration. That is, it produces discontinuous
solutions irrespective of the smoothness of the initial conditions. These are in contradiction
with smooth shocks observed in real-life traffic flow.
• The kinematic wave theory is not able to describe regular start-stop waves with amplitude
dependent oscillation times which have been observed in real-life traffic (recall transition
of synchronised flow into wide moving jams). Besides being a source of stress and irritation
for drivers, stop-and-go traffic also produces higher amounts of pollutants, while leading
to a higher fuel consumption due to the frequency of acceleration and deceleration.
• In real-life traffic flow, hysteresis phenomena have been observed, showing that the average
headways of vehicles approaching a jam are smaller that those of vehicles flowing out of
a jam (in case of relaxation dominant phase; see chapter 5). These phenomena are not
described by the kinematic wave theory.
• Kinematic wave models do not address the issue of traffic instability. It has been often
observed that in a certain traffic concentration range, small perturbations can induce
violent transitions.We again refer to chapter 5 where the issue of traffic instability was
discussed.

To resolve these problems, different approaches have been considered in the past.

9.4.1 Derivation of the model of Payne

The model of [42] consist of the conservation of vehicle equation, and a dynamic equation
describing the dynamics of the speed u (x, t). The model can be derived in several ways, but
the most intuitive one is by considering the following, simple car-following model
µ ¶
1
vα (t + τ ) = Ṽ (sα (t)) = V (9.88)
sα (t)
where sα (t) = xα−1 (t) − xα (t) denotes the distance headway with between vehicles α − 1 and α.
Eq. (9.88) describes how vehicle α adapts a speed Ṽ that is a function of the distance between
him and his predecessor. However, the reaction of α will not be instantaneous, but will be
delayed by the reaction time τ . Let us define

xα (t) = x (9.89)

and
vα (t) = u(xα (t), t) = u (x, t) (9.90)
Let us first consider the left-hand-side of Eq. (9.88). We have

vα (t + τ ) = u (xα (t + τ ) , t + τ ) (9.91)

Since, by the theorem of Taylor, we have

dxα ¡ ¢
xα (t + τ ) = xα (t) + τ + O τ2 (9.92)
dt
Moreover
µ ¶
dxα ¡ 2¢
u (xα (t + τ ) , t + τ ) = u xα (t) + τ + O τ ,t + τ (9.93)
dt
dxα ∂u (x, t) ∂u (x, t) ¡ ¢
= u (x, t) + τ +τ + O τ2 (9.94)
dt ∂x ∂t
∂u ∂u ¡ ¢
= u (x, t) + τ u (x, t) +τ + O τ2 (9.95)
∂x ∂t
172 CHAPTER 9. MACROSCOPIC TRAFFIC FLOW MODELS

Neglecting the higher-order terms thus yields

∂u ∂u
vα (t + τ ) ≈ u (x, t) + τ u (x, t) +τ (9.96)
∂x ∂t
Let us now consider the right-hand-side of Eq. (9.88). First, note that the inverse of the
distance headway equals the density (evaluated at the location between vehicle α and vehicle
α − 1), i.e. µ ¶
1 1 1
= = k xα (t) + sα (t), t (9.97)
sα (t) xα−1 (t) − xα (t) 2
Again, the use of Taylor’s theorem yields
µ ¶
1
k xα (t) + (xα−1 (t) − xα (t)) , t (9.98)
2
1 ∂k ¡ ¢
= k (xα (t), t) + sα (t) + O sα (t)2 (9.99)
2 ∂x
µ ¶
1 ∂k 1
= k (x, t) + +O (9.100)
2k ∂x k2

Then
µ ¶ µ µ ¶¶
1 1
V = V k xα (t) + sα (t), t (9.101)
sα (t) 2
µ µ ¶¶
1 ∂k 1
= V k (x, t) + +O (9.102)
2k ∂x k2
µ µ ¶¶
1 ∂k 1 dV
= V (k (x, t)) + +O 2
(9.103)
2k ∂x k dk

Neglecting the higher-order terms yields

1 dV ∂k
Ṽ (sα (t)) ≈ V (k (x, t)) + (9.104)
2k dk ∂x
Combining Eqn. (9.96) and (9.104) yields the Payne model, describing the dynamics of the
speed u = u (x, t)
∂u ∂u 1 dV ∂k
u + τu +τ = V (k) + (9.105)
∂x ∂t 2k dk ∂x
or, in the more familiar form

∂u ∂u V (k) − u 1 dV 1 ∂k
+u = + (9.106)
∂t ∂x τ 2τ dk k ∂x
V (k) − u D (k) 1 ∂k
= − (9.107)
τ τ k ∂x
V (k) − u 1 ∂k
= − c2 (k) (9.108)
τ k ∂x
¯ ¯
1 ¯ dV ¯
where D (k) := − 12 dV 2
dk = 2 dk > 0, and where c (k) = D (k) /τ . For reasons which will
become clear later on, c (k) is generally referred to as the wave speed. Furthermore, if many
applications of the Payne model, the wave speed c (k) is chosen constant, i.e. c (k) = c0 .
Let us briefly describe the different terms that are present in the Payne model, and how
these terms can be interpreted from a traffic flow point-of-view:

• In hydrodynamics, the term u ∂u

∂x is referred to as the transport or convection term. It
describes changes in the traffic speed u due to the inflow and outflow of vehicles with
different speeds.
9.4. HIGHER-ORDER MODELS 173

• The term − D(k) ∂k

kτ ∂x is the anticipation term, since it can describe the reaction of drivers
to downstream traffic conditions.

• The relaxation term V (k)−u

τ describes the exponential adaptation of the speed to the
concentration-dependent equilibrium speed V (k).

Payne [42] has shown that his model predicts instability of the traffic flow in a certain density
range. This means that small disturbances in the flow can lead to the formation of a local traffic
jam. This will be illustrated in the remainder. Note that a number of high-order models fall
under the same model family.

9.4.2 Mathematical properties of the Payne-type models

This section discusses some of the important properties of the Payne-type models; the derivation
of these properties from the model equations (B.1) and (B.2) is explained in the subsequent
sections.
We have seen that the LWR model has a single family of characteristic curves, which are
dQ
described by the lines dxdt = λ∗ , where λ∗ = λ∗ (k) = dk . The Payne model has two families
of characteristics along which the properties of the flow are transported. These characteristic
curves are defined by the following ordinary differential equations

dx
C+ : = λ1 = u + c (k) (9.109)
dt

dx
C− : = λ2 = u − c (k) (9.110)
dt
It was shown that for the LWR model, the density was constant (conserved) along the
characteristic. This is not the case in the Payne model; in the Payne model, the dynamics of
the so-called Riemann or characteristic variables z1,2 are described. Assuming that c (k) = c0 ,
these are defined by
z1 = u + c0 ln k and z2 = u − c0 ln k (9.111)
The characterstics C+ and C− are sometimes refered to as the second-order waves, whereas the
waves C∗ described by dx
dt = λ∗ are called the first-order waves.
It can be shown that solutions to the Payne-model are unstable, if they satisfy the following
condition.
¯ ¯
¯ dV ¯
k¯¯ ¯ = −k dV ≥ c (k) (9.112)
dk ¯ dk
For small densities k, condition (9.112) will generally not hold, implying that the Payne model
would predict stable traffic flow. In other words, small perturbations δk (x, t) and δu (x, t) will
dissipate over time. This holds equally for regions in which dV /dk = 0. When the concentration
and the rate of decrease dV /dk is large enough, the second-order models will predict that traffic
flow becomes unstable. Under unstable conditions, small perturbations will grow over time and
will eventually turn into a traffic jam.
For the Payne mode, we have s ¯ ¯
1 ¯¯ dV ¯¯
c (k) = (9.113)
2τ ¯ dk ¯
implying that traffic flow conditions are unstable when
¯ ¯
¯ dV ¯
¯
k¯ ¯≥ 1 (9.114)
dk ¯ 2kτ
174 CHAPTER 9. MACROSCOPIC TRAFFIC FLOW MODELS

Consider, for instance, Greenshields’ fundamental diagram, we get

dV V0
=− (9.115)
dk kj

yielding that the Payne model would predict unstable traffic conditions when
s
1 kj
k≥ (9.116)
2τ V0

Note that the stability of the traffic conditions increase when the relaxation time τ of the flow
decreases. That is, the more timely drivers react to the prevailing traffic conditions, the more
stable the resulting traffic flow conditions are. For the Payne model, traffic flow conditions are
always stable in case τ → 0. For details on the mathematical analysis, we refer to appendix B
of these notes.

9.4.3 Numerical simulation and high-order models

Since the mathematical properties of higher-order models are still not well understood, reliable
numerical analysis is also still a problem. In appendix B, we propose a numerical scheme for
general high-order models. Another numerical approximation is implemented in the METANET
model, which is discussed in detail in Chapter 15.

9.5 Alternative modelling approaches and generalisations

Appendix C discusses a specific type of continuum model, namely gas-kinetic models. Appendix
D discusses generalisations of macroscopic models.