Sie sind auf Seite 1von 8

PDdE-based analysis of vehicular platoons

with spatio-temporal decoupling


Ivo Herman

Dan Martinec

Zdenek Hurak

Michael Sebek

Faculty of Electrical Engineering, Czech Technical University in


Prague, Technicka 2, 166 27 Prague, Czech Republic (e-mail:
ivo.herman@fel.cvut.cz).
Abstract: The paper provides exact analytical solutions of wave-like PDdE-based (partial
dierential-dierence equation) models of several popular vehicular platooning problems with
bidirectional distributed control. Ecient evaluation of eigenvalues, eigenfunctions and wave
velocities is proposed for a nite vehicular platoon without an approximation by a continuum
model. Thanks to the decoupling of spatial and temporal dynamics, closed-loop stability
can be analyzed by looking at a characteristic polynomial using a root-locus-like graphical
technique. Although the paper only contains discussions for static state feedback controllers
(using measurements of inter-vehicular distances and relative velocities), extensiona to higher-
order (dynamic) controllers is straighforward and will be included in the extended version of
the paper.
Keywords: distributed control, eigenvalue, large-scale systems, partial dierential equations,
transportation, vehicles, waves
1. INTRODUCTION
Intelligent vehicle transportation systems (IVTS) are sup-
posed to be the future of personal transport. In these
systems, automatic cars travel in a long string (platoon)
maintaining tight gaps between the cars. An increased
highway capacity and safety are an obvious motivation.
The most common scenario is that the cars travel in
long platoons with one leader in the front which sets the
common speed.
IVTS have been an active area of research for several
decades, the rst results appearing in the literature as
early as in fties. One research direction considered fully
centralized approaches whereas the mainstream research
direction nowadays focuses on distributed (or decentral-
ized) control, where individual cars are controlled by their
onboard controllers, processing only the information that
is available locally. This information includes the physical
variables characterizing the state of the car itself and its
nearest neighbors (typically the predecessor and possibly
also the follower).
The advantage of information localization is in much easier
scaling of the platoon, higher robustness and lower commu-
nication demands, if any. However, several new phenomena
appear with the advent of these distributed/decentralized
control schemes. Well known is the so-called string insta-
bility, which imposes higher demands on the vehicles that
are far from the source of the regulation error, see e. g.
Seiler et al. (2004). Recent results by Bamieh et al. (2012)
show that a vehicular platoon controlled in a decentralized
way with a very high number of cars is not coherent, that

The research was supported by the Grant Agency of the Czech


Republic within the projects GACR 13-06894S (Ivo Herman) and
GACR P103-12-1794 (D. Martinec and M. Sebek) .
is, it cannot be moved as a rigid body when subject to
random noise.
There are many control congurations used for decen-
tralized vehicular platoon control: predecessor following,
constant-time-headway policy, bidirectional control, leader
following or cooperative automatic cruise control. The
predecessor following algorithm, although the easiest to
implement, suers from the string instability, while the
others can be made string stable, see e.g. Seiler et al.
(2004); Tangerman et al. (2012). The last two control
congurations require inter-vehicle communication.
We focus on analysis of bidirectional control of platoon.
Our approach uses a solution of Partial Dierential-
Dierence Equation (PDdE), that is, the model is discrete
in the spatial variable (the index of a vehicle) and the time
is considered continuous. The mathematical background
for this was introduced in Ferrari-Trecate et al. (2006) and
later extended to double integrator dynamics in Galbusera
et al. (2007), although these papers dealt with consensus
tasks.
Several authors have used continuous approximation of ve-
hicular strings. Then they solve the problem in an innite-
dimensional setting assuming the number of vehicles is
large enough. One interesting result by Barooah et al.
(2009) is that when using mistuned bidirectional control,
that is, when the controllers processing the distance to
the predecessor and the follower are not identical, the
rate at which the least stable eigenvalue approaches zero
with an increasing number of cars is slower. Recent works
prove that asymmetric controllers assures uniform bound
on eigenvalues (see Hao and Barooah (2012a)) at the price
of string instability (in Tangerman et al. (2012)).
4th IFAC Workshop on Distributed Estimation
and Control in Networked Systems
September 25-26, 2013. Rhine-Moselle-Hall, Koblenz, Germany
978-3-902823-55-7/2013 IFAC 144 10.3182/20130925-2-DE-4044.00019
Another possible approach for platoon analysis is 2D
polynomial approach as in Sebek and Hurak (2011).
Many results presented in this paper have already ap-
peared in the literature. Spatial eigenvalues for symmetric
case can be found from Toeplitz matrices as in Hao and
Barooah (2012b). Their asymmetric counterparts were de-
scribed by da Fonseca and Veerman (2009). Despite this,
our approach oers a solution in the form of decoupled
ordinary dierential equations, which exhibit many system
characteristics in the convenient form of characteristic
polynomials. Moreover, a unied framework is presented,
where many previous results appear immediately. System
eigenvalues as well as spatial eigenvalues are essential parts
of the PDdE solution. String instability is detected in the
asymmetric case as a result of a very easy calculation of
eigenfunctions.
2. CONTROL ALGORITHMS
The task of making the platoon of vehicles travel at a
given speed and specied spacing could be accomplished
by decentralized (or distributed) control with a controller
onboard each vehicle. If the only other cars from which
a given car gets the relevant information (measurements)
are the nearest neighbors, that is, the so-called predecessor
and successor, the control algorithm is called bidirectional.
In the simplest case when the measurements needed from
the neighbors are the distances to them, that is, the
dierences in absolute positions, there is no need for
establishing extra communication links between the cars.
Besides from the nearest neighbors, each car can also
receive information from the car governing the whole
platoon, which is called the leader. Finally, the gains in the
bidirectional controllers can be symmetric and asymmetric
with respect to the measurements form the predecessor
and the follower. This brings us to the classication of a
few types of control strategies studied in this paper
Leader symmetric control. The task for every car of
the platoon is to travel with the preset spacing and at
the leaders speed. The controller gains for the spacing
errors ahead and behind are identical. The last car just
measures the distance to its predecessor and compares
it with desired distance d
ref
. Self-velocity feedback is
necessary.
Leader asymmetric control. This is very similar to the
leader symmetric control, only the spacing gains are
dierent for the predecessor and follower errors. The
motivation for the controller asymmetry is given in
Barooah et al. (2009) (where they call is mistuning).
They prove that the convergence of eigenvalues to zero
can be made slower.
Neighbours velocity feedback control. This diers from
the previous cases in using not only the vehicles own ve-
locity, but also the velocity of the immediate neighbours.
This can also be achieved by using a PD (proportional-
derivative) controller for the spacing error control. Both
symmetric and asymmetric cases are analyzed with a
PD controller in the paper.
2.1 Car model
We assume there are N + 1 cars in total in the string,
including also the externally controlled car. The vehicles
are numbered by a spatial index k = 0, 1, 2, . . . , N.
The individual cars are modelled as double integrators
x
k
(t) = u
k
(t). (1)
The position of the kth car is denoted by x(k, t) and its
velocity by v(k, t). Actual distances to the neighboring cars
along with the desired distance are input to the controller.
The control is given by a static state feedback
u(k, t) = av(k, t) +
pf
(x(k 1, t) x(k, t) d
ref
)

pb
(x(k, t) x(k + 1, t) d
ref
). (2)
We assume that the controller gains
pf
applied to the
distance errors to the predecessor vehicle and
pb
to
the follower vehicle can generally dier. d
ref
denotes the
reference distance among cars. The use of static feedback
is in accordance with Barooah and Hespanha (2005), so
the deviation in leaders velocity causes bounded errors in
spacing, regardless of the number of vehicles. The velocity
feedback av(k, t) is necessary for cancelling out the initial
conditions; on the other hand it causes nite steady state
error in spacing of the platoon.
In the last part of the paper a use of a PD controller is
discussed. The obvious benet of using a PD controller is
in the ability of keeping equal spacing and leaders velocity.
2.2 Partial dierential-dierence equations
Using the theory exposed in Ferrari-Trecate et al. (2006),
the scenarios described above can be modeled using par-
tial dierential-dierence equations (PDdE). The time t
is continuous, while the spatial variable k is discrete.
The advantage of PDdE is in a better insight into some
properties of the overall system and/or their easier cal-
culation. Unlike, for instance, the approach in Barooah
et al. (2009), one can derive exact PDdE solution of the
system equations without using continuum approximation.
The general PDdE derived for a controller (2) and double
integrator dynamics has a form
x
tt
(k, t) = ax
t
(k, t) +
c
x
kk
(k, t) (3)
+
1
(x
k
(k, t) d
ref
)
2
(x
k
(k, t) d
ref
).
We use the subscript notation for a partial derivative, that
is
x(k,t)
t
= x
t
(k, t). x
k
(k, t) = x(k 1, t) x(k, t) is a
forward dierence in positions, x
k
(k, t) = x(k, t) x(k +
1, t) is a backward dierence. Forward dierence in our
setting stands for a distance to the to cars predecessor,
i. e. to the car with lower index. The second-dierence term
x
kk
(k, t) is an instance of Laplacian dened in Ferrari-
Trecate et al. (2006)
x
kk
(k, t) = x(k 1, t) 2x(k, t) + x(k + 1, t). (4)
The gains are dened as
c
= min(
pf
,
pb
),
1
=
pf

min(
pf
,
pb
),
2
=
pb
min(
pf
,
pb
) and only one of
the gains
1
,
2
is therefore nonzero.
The PDdE (3) is very similar to a classical damped wave
equation known in continuous case. If both gains
pf
,
pb
were identical (i.e. symmetric control), the physical anal-
ogy would be a mass-spring system where two masses are
NecSys 2013
September 25-26, 2013. Koblenz, Germany
145
connected by a massless spring. If the gains were dierent
(asymmetric control), we could think of a controller as
a spring which has dierent stiness in each direction.
Using the physical analogy, we easily see the velocity
feedback is necessary, since it acts as a damping factor in
the wave equation, thus the eect of the initial conditions
can be cancelled out.
2.3 Solutions of the overall system
Our goal in this paper is to nd the complete solution
x(k, t) for all the mentioned platooning control schemes.
The general PDdE (3) remains the same, only the gains

c
,
1
,
2
and boundary conditions dier among control
schemes. We use a classical assumption that the function
x(k, t) can be separated into a product of two functions
x(k, t) = (k)(t). (5)
Separation of variables allows us to nd the spatial eigen-
values, eigenfunctions as well as the complete solution.
By the term spatial eigenvalues
n
we understand the
eigenvalues of the spatial second dierence operator (and
also the eigenvalues of the Laplacian matrix associated
with the information graph of the formation). These must
be distinguished from the system eigenvalues
n
given
directly by the matrix A in the equation
x(t) = Ax(t) +Bu(t). (6)
The matrix A describes a complete platoon, taking into
account all the couplings among the cars (through relative
measurements). Here x is a stacked vector of all the states
of the individual cars, i.e. x = (x
0
, v
0
, x
1
, v
1
, . . . , x
N
, v
N
)
T
.
3. LEADER SYMMETRIC CONTROL
The behavior of the rst car (leader) is regarded as
given and together with the given (and known) reference
distance at the last car they set the boundary conditions.
Considering the the equal gains
pf
=
pb
=
p
and the
initial conditions f
0
(t), f
1
(t) we obtain the following PDdE
x
tt
(k, t) =
p
x
kk
(k, t) ax
t
(k, t) (7)
x(0, t) = x
L
(t) (8)
x
k
(N, t) = d(t) (9)
x(k, 0) = f
0
(k) (10)
x(k, 0) = f
1
(k). (11)
The boundary function x
L
(t) describes a movement of the
platoon leader. Spatial dierence x
k
(N, t) = d(t) at the
last car has the meaning of desired distance to the previous
car and is usually constant, i. e. x
k
(N, t) = d
ref
. As we do
not have homogeneous boundary conditions, we separate
the sought function x(k, t) into two parts
x(k, t) = y(k, t) + y(k, t). (12)
Function y(k, t) is chosen such that it only satises bound-
ary conditions, so it can be a linear function in k between
x
L
(t) and the dierence at the last car
y(k, t) = x
L
(t) + d(t)k. (13)
Plugging (12) to (7)(11), an inhomogeneous partial die-
rential-difference equation with homogeneous boundary
conditions is obtained
y
tt
(k, t) =
p
y
kk
(k, t) ay
t
(k, t) + q(k, t), (14)
where q(k, t) = y
tt
(k, t) a y
t
(k, t). This is similar to
a plucked string equation with external force. Supposing
a separable function of the form (5), equation (7) can be
written as

(t)(k) =
p
(t)

(k) a

(t)(k) (15)
since the time derivatives act only on the function of
time and the spatial derivatives only on the function of
space. If both functions and are not identically zero,
the separated formula for solution of the homogeneous
PDdE reads

(t) + a

(t)

p
(t)
=

(k)
(k)
=
n
, (16)
where
n
is a separation constant and also a spatial
eigenvalue of the solution. To achieve equality for any
combination of k and t, both fractions must be constants.
Then we get two equations, one for each fraction

(k)
n
(k) = 0 (17)

(t) + a

(t)
n
(t) = 0. (18)
The spatial equation (17) must comply to the zero bound-
ary conditions and using (4) one gets
(k 1) 2(k) + (k + 1) =
n
(k). (19)
The only solution of this dierence equation which satises
the boundary conditions (0) = 0,
k
(N) = 0 is

n
(k) = sin
(2n + 1)k
2N + 1
(20)

n
=2
_
cos
_
(2n + 1)
2N + 1
_
1
_
=4 sin
2
(2n + 1)
4N + 2
. (21)
Thus, there are N +1 dierent space modes (the index of
the spatial mode being n = 0, 1, . . . N, so this number is
nite) and the value of the separation constant depends
non-linearly on the mode. The function
n
(k) is called an
nth eigenfunction and
n
is an nth spatial eigenvalue. We
denote
n
=
(2n+1)
2N+1
the frequency of the spatial sinusoid

n
(k). Note that all spatial eigenvalues are real and non-
positive, for n = N can be equal to 0.
Now we know the values of spatial eigenvalues
n
, so
we can calculate also the solution of temporal part of
the equation (t). To solve it, the so-called eigenfunction
expansion is used. As the eigenfunctions for dierent
modes are mutually orthogonal, we can expand (14) into
eigenfunctions
N

n=1

n
(t)
n
(k) =
p
N

n=1

n
(t)

n
(k) a
N

n=1

n
(t)
n
(k)
+
N

n=1
q
n
(t)
n
(k), (22)
All the sums start from one, since sine of zero is zero. Using
the fact that

n
(k) =
n

n
(k) we can eliminate
n
(k).
The function q
n
(t) is given by
q
n
(t) =

N
i=1
q(i, t) sin
(2n+1)i
2N+1

N
i=1
sin
2 (2n+1)i
2N+1
. (23)
3.1 Decoupled ODE
The terms in sums in (22) must be equal for each n, so we
get N + 1 ordinary inhomogeneous dierential equations

n
(t) + a

n
(t)
p

n
(t) = q
n
(t). (24)
NecSys 2013
September 25-26, 2013. Koblenz, Germany
146
1 0.8 0.6 0.4 0.2 0
10
5
0
5
10
Real
I
m
a
g
N=35


Pole
1 0.8 0.6 0.4 0.2 0
10
5
0
5
10
Real
I
m
a
g
N=3500


Pole
Fig. 1. Locations of eigenvalues of platoons with two dierent sizes. One easily recognizes that the case with more
vehicles is just more densely sampled version of the case with lower number of cars. Eigenvalues lie on the same
root-locus-like curve.
First consider homogeneous case, where q = 0. Since
n
is negative, the equation (24) has all coecients positive
and the system is stable for all modes. One
n
= 0, which
corresponds to unobservability of global motion by the
members of the platoon (see Fax and Murray (2004)). By
separability assumption we obtained a decoupled system,
for which stability and eigenvalues can be found by exam-
ining only second order polynomials, instead of taking into
account whole and possibly very large system matrix A.
It is worth of noting that the characteristic polynomial of
a single position-controlled car, which is not part of the
platoon, has a form
s
2
+ as +
p
, (25)
whereas the interconnected characteristic polynomial is
s
2
+ as
n

p
. (26)
Therefore only the last coecient is multiplied by
n
in an
interconnected system (26) and the interconnection does
not inuence the overall stability, although the individual
modes approach zero as
n

p
goes to zero.
From (24), the eigenvalues of the overall system (7) are
given as

1,n;2,n
=
1
2
_
a
_
a
2
+ 4
p

n
_
. (27)
We remark that these eigenvalues were obtained just
as roots of second order polynomial in (26). To get all
eigenvalues, it is only needed to calculate N +1 times the
roots. This way of calculation is only linear in N from
computational complexity point of view.
The equation (26) has yet another explanation, which even
easily generalizes to any static or dynamic controller. As
the spatial eigenvalue acts as a gain of the controller part of
the equation (
n

p
), we can think of each of the decoupled
ODEs as a closed loop system with feedback gain equal to

n
. This is schematically depicted in Fig. 2.
As
n
varies with n, the gain of the closed loop varies as
well and for each spatial eigenvalue there are two poles
placed in the s plane. Thus the poles lie on curves de-
scribed by a root locus curve for independently controlled
car. The location of poles is shown for two dierent N in
Fig. 1. The fact that
n
is in the interval 4, 0 causes
the eigenvalues to lie in a certain region regardless of what
the number of cars is. As the number of cars increases, the
eigenvalues are sampled more densely on a curve region set
1
s +as+1
2
C
Platoon
Controller

1
,
2
, ...
N+1

Fig. 2. Closed loop diagram for a decoupled system. The
car system consists of gains
n
and an integrator with
velocity feedback. The spatial eigenvalue
n
acts as a
gain change known in normal root-locus analysis.
by a root locus curve. Thus, the controller can be designed
by taking into account only the root locus curve to achieve
the desired properties.
Although the approach is dierent, the role of spatial
eigenvalues corresponds to the the role of eigenvalues of
graph Laplacian as described in Fax and Murray (2004).
Again, here the controller must stabilize the formation for
any gain set by
n
. Instead of a Nyquist criterion, the
stability in our approach is analyzed in a complex plane
by root-locus of a closed loop with dierent gains.
3.2 Total solution
Having the homogeneous solution, the particular solution
is obtained by variation of constants formula

n
(t) = e

1,n
t
_
t
0
e

2,n

q
n
()
W(
1,n
,
2,n
)
d + a
n
e

1,n
t
+e

2,n
t
_
t
0
e

1,n

q
n
()
W(
1,n
,
2,n
)
d + b
n
e

2,n
t
, (28)
where W(
1,n
,
2,n
) is the Wronskian of the homogeneous
solution for the mode n of (24). The constants a
n
, b
n
are
determined by Fourier series expansion of the modied
initial conditions

f
0
(k) = f
0
(k) y(k, 0),

f
1
(k) = f
1
(k)

y(k, 0) into the eigenfunctions


n
(k).
At this stage both separated functions were found, that is
y
n
(k, t) =
n
(k)
n
(t). Finally the solution to the original
PDdE (7) is a sum of all solutions of individual modes
NecSys 2013
September 25-26, 2013. Koblenz, Germany
147
0 5 10 15 20 25 30
0
0.5
1
1.5
Modes
S
p
e
e
d
s
Wave speeds for different modes, N=30


Fig. 3. Wave speeds for individual modes. Slower modes
are not traveling.
x(k, t) = y(k, t) + y(k, t) (29)
= x
L
(t) + d(t)k +
N

n=1
sin
_
(2n + 1)k
2N + 1
_

n
(t)
Correctness of this solution can be checked by plugging it
into the original PDdE (7).
3.3 Wave speed
For determining the wave speed, we start with basic
response to the initial conditions with zero boundary
conditions and omitting Fourier coecients a
n
, b
n
. Then
the solution for one mode n is
x
n
(k, t) = sin
(2n + 1)k
2N + 1
e
1
2
_
a
_
a
2
16
p
sin
2
(2n+1)
4N+2
_
t
(30)
This results in two cases
(1) the mode is overdamped, i. e. the square root is real.
(2) the mode causes oscillating response, i. e. the square
root is complex.
In the rst case the exponential term in (30) is real.
Thus, the time response for individual car is exponentially
decaying. Therefore, the spatial sinusoid decreases its
amplitude without any shifting in space. This mode does
not cause travelling wave in space. In the second case the
term in the square root of (30) is negative, so we get
a complex number in the time part of the solution
x
n
(k, t) = e

1
2
at
sin
(2n + 1)k
2N + 1
_
sin
_
1
2

n
t
_
+cos
_
1
2

n
t
__
, (31)
where
n
=
_
a
2
+ 16
p
sin
2 (2n+1)
4N+2
is a positive con-
stant. Now the goniometric terms can be put together by
product-to-sum formula to obtain
x
n
(k, t) =
1
2
e

1
2
at
sin
_
(2n + 1)
2N + 1
_
k
2N + 1
2(2n + 1)

n
t
__
.
(32)
Only the sin cos part of (31) was written, the others nally
have the same wave speed. The wave speed c
n
is then
obviously (in cars per second)
c
n
=
2N + 1
2(2n + 1)
_
a
2
+ 16
p
sin
2
(2n + 1)
4N + 2
(33)
Therefore, in (32) we obtained two waves, one travelling to
the left (with the + sign), the other travelling to the right.
So the mode splits into two waves with half amplitudes.
When N is large and a = 0, sine can be substituted by its
argument and we get c =

p
, which is in correspondence
with a speed in a continuous medium. Generally written
for the controller of any order (including PI, PID etc.)
c
n
=
Im{
n
}

n
, (34)
where
n
is interconnected system eigenvalue correspond-
ing to the mode n and
n
is the frequency in the spatial
sinusoid, in this case
n
=
(2n+1)
2N+1
. Wave speeds are
plotted as a function of mode in Fig. 3.
To conclude, we have the following conditions to decide if
the mode causes a travelling wave
(1) If a
2
> 16
p
sin
2 (2n+1)
4N+2
, then the mode is just real
and does not cause a travelling wave. If a
2
> 16
p
,
then none of the modes cause a travelling wave.
(2) If a
2
< 16
p
sin
2 (2n+1)
4N+2
, the response is oscillating
and the wave travels.
Obviously the higher the mode, the higher the chance to
cause a travelling wave. So there will usually be a travelling
wave only on the higher modes. The result that dierent
modes in discrete systems spread with dierent velocities
is known in physics, see for example Morin (2009). One
interesting illustration of wave behavior is in Fig. 4.
4. LEADER ASYMMETRIC CONTROL
Barooah et al. (2009) proved some advantages of using
dierent gains for spacing error to the predecessor (front)
vehicle
pf
and to the follower (rear, back) vehicle
pb
, as
discussed above. Here we assume all front and rear gains
are the same for all the cars. The system is described by
the following PDdE
x
tt
(k, t) =
c
x
kk
(k, t) +
1
(x
k
(k, t) d
ref
)

2
(x
k
(k, t) d
ref
) ax
t
(k, t) (35)
x(0, t) = x
L
(t) (36)
x
k
(N, t) = d(t) (37)
x(k, 0) = f
0
(k) (38)
x(k, 0) = f
1
(k), (39)
where the constants
c
,
1
,
2
and forward and backward
dierences were dened in the section 2.2. Although physi-
cal analogy is not as clear as in the previous case, majority
of the properties are still valid here. To let the spatial
behavior depend only on the level of asymmetry (as was
done in da Fonseca and Veerman (2009)), we introduce
=

pb

pf
. (40)
Then the gains become part of the time function of PDdE,
therefore part of the controller.
Asymmetric control diers from the symmetric one in
presence of the rst spatial dierence. It is not possi-
ble to analytically express the spatial eigenfunctions and
eigenvalues, however, they can be calculated numerically.
Eigenfunctions are not harmonic any more

n
(k) =
_
_
1

_
k
sin
n
k, (41)
NecSys 2013
September 25-26, 2013. Koblenz, Germany
148
0 50 100 150 200 250
5
0
5
10
15
Car number
P
o
s
i
t
i
o
n
t=0


Calculated
Simulated
0 50 100 150 200 250
0.4
0.2
0
0.2
0.4
0.6
0.8
1
Car number
P
o
s
i
t
i
o
n
t=45


Calculated
Simulated
Fig. 4. Comparison of the state-space evolution and the mode-shifting of the initial values with the damped model. There
are two travelling and damped waves consisting of higher modes, and the sine arc in the middle corresponding to
the rst mode, which is the only non-travelling. The system exhibits a dierent reection at the ends, with opposite
phase at the leaders side, in-phase at the end. N = 250,
p
= 10, a = 0.1. As the dierent modes spread with
dierent speeds, the shape of the wave changes and as the result the small oscillations behind the wave are apparent.
where
n
is the nth solution (measured by the distance
from zero on the interval 0, ) of the nonlinear equation
sin(N
n
)
_
1

sin((N + 1)
n
) = 0. (42)
This equation follows from the boundary condition at the
end of the platoon. Clearly, if
pf
=
pb
, then = 1 and
eigenfunctions are identical to (20).
Notice that the eigenfunctions (41) are spatially increasing
when < 1, so
pf
>
pb
. This results in spatial instabil-
ity, which in turn causes string instability (see Fig. 5). If
on the other hand the square root is lower than one, the
cars further in the platoon will not react to movements
caused by the leader fast enough, which results in increase
of regulation error norms. Our results correspond to those
in Tangerman et al. (2012), where harmonic instability of
asymmetric controller with static state feedback is proved.
We conjecture that string instability is a network (mea-
surement) topology property, it is not inuenced by the
design of the controller. Indeed, the gain of P controller
(either
p
or
pf
) can be the same in both cases, while
string instability arises only in the asymmetric case.
The spatial eigenvalues are given by

n
= 2

cos
n
1 . (43)
These eigenvalues are uniformly bounded from zero if the
gains are dierent. Clearly, the eigenvalue gets closest to
zero when cos
n
= 1 and then
n
= (1

)
2
, which is
always smaller than zero.
Having the spatial solution, one can easily calculate time
eigenvalues
n
, which are given as coecients in exponen-
tials of solution of the ODE

n
(t) + a

n
(t)
pf

n
(t) = q
n
(t), (44)
which is formally the same equation as in Leader symmet-
ric control in (24).
As
n
acts as a gain in the closed loop of a root-locus
analysis and it is bounded from zero, also the region for
time eigenvalues is uniformly bounded from zero. Thus,
platoon controllability is guaranteed for any number of
cars. This is a principally dierent proof of results in Hao
and Barooah (2012a).
The rest of the solution follows the approach presented
above. The coecients a
n
, b
n
, q
n
(t) are calculated from
generalized Fourier series expansion of

f
0
(k),

f
1
(k) and
0 100 200 300 400 500
350
300
250
200
150
100
50
0
Time [s]
P
o
s
i
t
i
o
n
s

pf
=101.2,
pb
=100


Car 1
Car 51
Car 101
Car 151
Car 201
Car 251
Car 300
Fig. 5. The leader decelerated immediately, the others also
have to stop. One clearly sees string instability of the
system, as the last car has the highest overshoot. The
level of asymmetry is very small, about 1 %
q(k, t) into the eigenfunctions (41). Let us dene functions
h
n
(i), s
n
as
h
n
(i) =
1
__
1

_
i
sin
n
i; s
n
=
1
_

N
i=1
sin
2

n
i
_ (45)
Then the Fourier series coecients are obtained as
(a
n
+ b
n
) = s
n
N

i=1

f
0
(i)h
n
(i)
(
1,n
a
n
+
2,n
b
n
) = s
n
N

i=1

f
1
(i)h
n
(i) (46)
q
n
(t) = s
n
N

i=1
q(i, t)h
n
(i).
The total solution is the same as in symmetric control up
to the eigenfunctions
x(k, t) = y(k, t) + y(k, t) = x
L
(t) + d(t)k
+
N

n=1
_
_
1

_
k
sin (
n
k)
n
(t), (47)
where
n
(t) is a solution of (44).
The wave velocity of the mode is obtained from the formula
(34) as
c
n
=
_
a
2
+ 4(2

pf

pb
cos
n

pf

pb
)

n
(48)
NecSys 2013
September 25-26, 2013. Koblenz, Germany
149
3 2.5 2 1.5 1 0.5 0
1
0.5
0
0.5
1
Re
I
m
System eigenvalues


(a) Symmetric controller
3 2.5 2 1.5 1 0.5 0
1
0.5
0
0.5
1
Re
I
m
System eigenvalues


(b) Asymmetric controller
Fig. 6. Locations of eigenvalues of symmetric and asymmetric platoons with 200 vehicles and PD controller. In both
cases all eigenvalues lie on the same root-locus like curve. In asymmetric case, the uniform boundedness from zero
is apparent. Level of asymmetry = 0.6. PD gains in both cases
p
= 5,
d
= 5.
5. PD CONTROLLER
The most common controller found in the literature for a
car in a platoon is the classical PD controller. Although the
PD controller is generally not realizable, if the vehicle re-
ceives the velocity measurements from its neighbors, a PD
controller acts just as a static state feedback. The deriva-
tive of spacing error stands for relative velocity dierence
between cars. Since PD controller keeps two integrators
in a closed loop, such platoon is capable to travel with
zero spacing error. On the other hand, due to the lack of
absolute velocity feedback the coherence is deteriorated
(see Bamieh et al. (2012)). The velocity feedback adds
also a time derivative of second spatial dierence, therefore
the resulting PDdE becomes of third order with a mixed
derivative term. The continuous counterpart of a platoon
with a PD controller is a wave equation with a Kelvin-
Voigt damping. The PDdE with symmetric controllers has
a form
x
tt
(k, t) =
p
x
kk
(k, t) +
d
x
kkt
(k, t) (49)
and its asymmetric counterpart
x
tt
(k, t) =
pf
x
k
(k, t)
pb
x
k
(k, t) (50)
+
df
x
kt
(k, t)
db
x
k,t
(kt). (51)
Boundary conditions are the same as given in P controller
cases. No self-velocity feedback a is necessary to keep the
system stable.
Both equations can be solved by separation of variables ap-
proach as was done above. The eigenfunctions are identical
to those derived in (20) and (41), respectively. Also spatial
eigenvalues are the same as in (21) and (43). Therefore
eigenfunctions and spatial eigenvalues are properties of
network topology, not of the controller structure. This
result is similar to that in Fax and Murray (2004), where
spatial eigenvalue corresponds to eigenvalue of Laplacian
matrix. Once again, asymmetric controller exhibits string
instability, as the values of eigenfunction are spatially
increasing.
The time equation in this case has the same form for both
spatial scenarios

n
(t)
df

n
(t)
pf

n
(t) = q
n
(t). (52)
The gains
df
=
d
and
pf
=
p
for symmetric controller.
This has a structure of root-locus with gains equal to
spatial eigenvalue. The system eigenvalues are given as
roots of characteristic polynomial
s
2

df

n
s
pf

n
(53)
which are

n1,2
=
1
2
_

df

n

_

df
2

2
n
+ 4
pf

n
_
. (54)
To achieve stability, is suces to have both controller
coecients positive, i. e.
pf
> 0,
df
> 0. The reason for
this is non-positivity of spatial eigenvalues. The location
of system eigenvalues (which appear in the matrix A in
(6)) is shown in Fig. 6.
6. CONCLUSION
Full analytical solutions of spatio-temporal models of a few
vehicular platooning control problems were derived. In
particular, bidirectional distributed control was investi-
gated. The problem was approached through the similari-
ties of the bidirectional distributed control with continuous
wave equations. Hence the tools for partial dierential or
actually partial dierential-dierence equations were used,
namely the principle of separation of spatial and temporal
dynamics.
This allowed ecient computation of the eigenvalues of
overall system matrix in a closed form as a solution
of N + 1 independent second-order polynomials. This
turns out particularly useful when the system is large-
scale (comprising a large number of vehicles) because the
computational eort of eigenvalue determination grows
only linearly with N.
The immediate use of the temporal part of the solution
is for a convenient analysis of (closed-loop) stability of
the whole vehicular platoon. It was shown that the char-
acteristic polynomial of the platoon is built as a minor
modication of a characteristic polynomial for the dy-
namics of a single car, with the modication consisting
in multiplying some coecients by value of the spatial
eigenvalue. This suggests a that root-locus-like graphical
technique can be used to plot the curves on which the
platoon eigenvalues are located.
The particular case of a leader asymmetric control seems
to be either string unstable or very slowly responding
to leaders movement, which follows explicitly from the
NecSys 2013
September 25-26, 2013. Koblenz, Germany
150
computed spatial solution. We think that this is true for
any controller, provided there are two integrators in the
open loop. On the other hand, uniform bound on eigen-
values distance from zero was proved, which guarantees
controllability of the whole platoon for any number of
vehicles.
It was proved that if the individual mode causes travelling
wave, the wave speed is nonlinearly dependent on the num-
ber of mode. We believe that understanding of some wave
aspects in the platoons can help in design of controller
with desired wave characteristics.
We emphasize that many results which are related to the
PDdE nature of the platooning problem have already been
shown in the literature. Mainly in the works by Barooah
and Hespanha (2005); Barooah et al. (2009); Hao and
Barooah (2012a,b); Tangerman et al. (2012). Nonetheless,
our approach achieves all these results in a unied way.
Among the conceptually simple extensions, which could
not t into this short paper, but which have already been
explored by the authors, are considerations of dynamic
(or higher-order) rather then proportional or PD only
controllers, symmetric and asymmetric. This makes the
analysis a bit involved since the resulting PDdE becomes
of order three or higher. The convenience of the root-locus-
like technique for shaping the eigenvalues can then be
even more appreciated.
REFERENCES
Bamieh, B., Jovanovic, M., Mitra, P., and Patterson, S.
(2012). Coherence in Large-Scale Networks: Dimension-
Dependent Limitations of Local Feedback. Automatic
Control, IEEE Transactions on, 57(9), 22352249.
Barooah, P., Mehta, P., and Hespanha, J. (2009).
Mistuning-Based Control Design to Improve Closed-
Loop Stability Margin of Vehicular Platoons. Automatic
Control, IEEE Transactions on, 54(9), 21002113.
Barooah, P. and Hespanha, J. (2005). Error amplication
and disturbance propagation in vehicle strings with
decentralized linear control. In Decision and Control,
2005 44th IEEE Conference on, 49644969.
da Fonseca, C. and Veerman, J. (2009). On the spectra
of certain directed paths. Applied Mathematics Letters,
22(9), 13511355.
Fax, J. and Murray, R. (2004). Information ow and
cooperative control of vehicle formations. Automatic
Control, IEEE Transactions on.
Ferrari-Trecate, G., Bua, A., and Gati, M. (2006). Anal-
ysis of Coordination in Multi-Agent Systems Through
Partial Dierence Equations. Automatic Control, IEEE
Transactions on, 51(6), 10581063.
Galbusera, L., Marciandi, M.P.E., Bolzern, P., and Ferrari-
Trecate, G. (2007). Control schemes based on the
wave equation for consensus in multi-agent systems with
double-integrator dynamics. Decision and Control, 2007
46th IEEE Conference on, 14981503.
Hao, H. and Barooah, P. (2012a). On achieving size-
independent stability margin of vehicular lattice forma-
tions with distributed control. Automatic Control, IEEE
Transactions on, 57(10), 26882694.
Hao, H. and Barooah, P. (2012b). Approximation error
in PDE-based modelling of vehicular platoons. Interna-
tional Journal of Control, 85(8), 11211129.
Morin, D. (2009). Waves. Unpublished. URL
http://www.people.fas.harvard.edu/ djmorin/
book.html.
Sebek, M. and Hurak, Z. (2011). 2-D Polynomial Approach
to Control of Leader Following Vehicular Platoons. In
Preprints of the 18th IFAC World Congress, 60176022.
Seiler, P., Pant, A., and Hedrick, K. (2004). Disturbance
propagation in vehicle strings. Automatic Control, IEEE
Transactions on, (10), 18351841.
Tangerman, F., Veerman, J., and Stosic, B. (2012). Asym-
metric decentralized ocks. Automatic Control, IEEE
Transactions on, 57(11), 28442853.
NecSys 2013
September 25-26, 2013. Koblenz, Germany
151

Das könnte Ihnen auch gefallen