Linear Quadratic Control

for Heavy Duty Vehicle Platooning

KUO-YUN LIANG

Master’s Degree Project

Stockholm, Sweden April 2011

XR-EE-RT 2011:017
 Linear Quadratic Control
for Heavy Duty Vehicle Platooning

KUO-YUN LIANG

Master’s Thesis at Automatic Control

Supervisor: Assad Alam
Examiner: Ather Gattami

XR-EE-RT 2011:017
 Abstract
Scania CV AB has estimated that the fuel consumption
constitutes 30 % of the operational costs of a heavy duty
vehicle. With the increasing oil price, reducing the fuel
consumption becomes highly profitable. A mitigation to
the high fuel consumption is to form a platoon. By intro-
ducing vehicle platooning, the fuel consumption is reduced
for the following vehicles. The fuel consumption is reduced
due to the slipstream effect, which is an atmospheric drag
reduction that occurs behind a traveling vehicle. Therefore,
by driving the vehicles at a close intermediate distance, the
overall resistive force is reduced.

First a study of robustness, of a finite long platoon based

upon the string stability criteria, is done. The main focus
of this thesis is to find a centralized control law for a finite
platoon, which considers the air drag reduction. A LQR
and a LQT controller are presented with respect to a given
information structure for a three-vehicle platoon. Simula-
tion results show an energy reduction of up to 11.4 % for
the second vehicle and up to 13.1 % for the third vehicle,
when driving 80 km/h in a platoon on a highway between
Södertälje and Norrköping with a time headway of 0.25 s.
Acknowledgements
With this thesis I complete the Master of Science degree in Electrical Engineering
at the Royal Institute of Technology (KTH) in Stockholm, Sweden. This work in
this master’s thesis has been conducted between November 2010 and April 2011 at
the Pre-development of Intelligent Transportation Systems Department (REPI) at
Scania CV AB in Södertälje, Sweden and was supervised at the Automatic Control
Department at KTH.

First, I want to express my deepest gratitude to my supervisor at Scania, Assad

Alam, for introducing this project, his valuable time, support, and guidance he has
given me throughout the whole project. Henrik Pettersson at Scania is acknowl-
edged for his outstanding help, support, and knowledge that were irreplaceable. I
also want to express my gratitude to my supervisor at KTH, Ather Gattami, for
the wise inputs and feedback. Many thanks go to everyone at the Pre-development
Department, REP, at Scania for welcoming me and making every day enjoyable.
Lastly, I want to express my gratitude for the support from family and friends.

Kuo-Yun Liang
Stockholm, 2011
Contents

I Introductory 1

1 Introduction 3

2 Background 5
2.1 Vehicle platooning . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.1.1 Cruise Control and Adaptive intelligent Cruise Control . . . . 5
2.1.2 Vehicle-to- vehicle/infrastructure communication . . . . . . . 6
2.2 String stability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.3 Related work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.4 Thesis objective & delimitations . . . . . . . . . . . . . . . . . . . . 7
2.5 Thesis outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8

II String stability 9

3 PID-controller 11
3.1 Velocity and relative distance . . . . . . . . . . . . . . . . . . . . . . 11
3.2 Velocity transfer function . . . . . . . . . . . . . . . . . . . . . . . . 12
3.3 Relative distance transfer function . . . . . . . . . . . . . . . . . . . 12
3.4 Simulation study I . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13

4 Scaled-varying PID-controller 17
4.1 Transfer functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
4.2 Simulation study II . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18

III Locally centralized three-vehicle platoon 21

5 Platoon model 23
5.1 Powertrain . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
5.2 Longitudinal forces . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
5.3 Linearized model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28

6 Optimal control 31
 6.1 Linear Quadratic Regulator . . . . . . . . . . . . . . . . . . . . . . . 31
6.2 Linear Quadratic Tracking control with integral action . . . . . . . . 33
6.2.1 Reference tracking system . . . . . . . . . . . . . . . . . . . . 33
6.2.2 Reference tracking platoon model . . . . . . . . . . . . . . . . 34
6.2.3 LQ optimal control . . . . . . . . . . . . . . . . . . . . . . . . 34
6.3 Cost function design . . . . . . . . . . . . . . . . . . . . . . . . . . . 35

7 Simulation 37
7.1 Simulation scenario . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
7.2 Different time headways . . . . . . . . . . . . . . . . . . . . . . . . . 37
7.3 Change in mass distribution . . . . . . . . . . . . . . . . . . . . . . . 43
7.3.1 40-30-40 tonnes platoon . . . . . . . . . . . . . . . . . . . . . 43
7.3.2 40-50-40 tonnes platoon . . . . . . . . . . . . . . . . . . . . . 44
7.4 Simulation on a measured road profile . . . . . . . . . . . . . . . . . 46
7.4.1 Södertälje to Norrköping . . . . . . . . . . . . . . . . . . . . . 46

8 Expanding the platoon 49

8.1 Sub-platoons within the platoon . . . . . . . . . . . . . . . . . . . . 49
8.2 Simulation verification . . . . . . . . . . . . . . . . . . . . . . . . . . 50

IV Discussion 51

9 Summary 53
9.1 String stability study with mass . . . . . . . . . . . . . . . . . . . . . 53
9.2 Locally centralized three-vehicle platoon . . . . . . . . . . . . . . . . 54

10 Conclusion 55

11 Future work 57

Bibliography 59

Appendices 61

A PID-controller setup 63

B Mathematical derivation of LQR 65

C Mathematical derivation of LQT with integral action 67

 Part I

Introductory

1
Chapter 1

Introduction

The commercially available technology in vehicles has grown rapidly in a few years,
offering enhanced driving experience and simplifying the daily driving. This devel-
opment may lead to full automation of the vehicle, enabling everyone to sit behind
the wheel without any driving skills. The possibilities can be enormous.

One of the services in progress today for vehicle automation is vehicle platoon-
ing. Platooning enables several vehicles to drive with a short intermediate distance
in between, acting together as one unit. Platooning serves mainly two purposes,
to increase the traffic flow and to reduce the emissions through the reduced fuel
consumption. The latter utilizes the atmospheric drag reduction behind a traveling
vehicle. It is generally referred to as the slipstream effect, which makes it possible
for the following vehicle to consume less fuel.

According to (Schittler, 2003), the average milage per year of an European truck is
150.000 km and has an average fuel consumption of 32.5 l/100 km. Scania CV AB
has estimated that the fuel consumption constitutes 30 % of the operational costs
of a heavy duty vehicle (HDV) (Scania CV AB, 2001). This means that reducing
the fuel consumption can be of vast profit. The fuel consumption can be reduced a
lot through HDV platooning with small intermediate distances. Empirical results
(Alam et al., 2010) show that the fuel consumption can be reduced up to 7.7 % with
a time gap of one second, when two identical HDVs travel as a platoon at 70 km/h.

However, a small distance between vehicles raises issues such as safety and com-
fort aspects for the driver. Higher risk for collisions is one issue that must be
prevented when introducing vehicle platooning. Normally, when an accident oc-
curs on the road, it takes a human driver seconds before reacting and braking,
but with vehicle-to-vehicle communication the time delay is reduced tremendously.
Therefore, a solution is to introduce vehicle-to-vehicle communication that allows
a vehicle to pass on information of its current and next coming behavior to other
vehicles in the vicinity.

3
Chapter 2

Background

2.1 Vehicle platooning

The nomenclature, used in this thesis, for vehicle platooning is defined as shown in
fig. 2.1. The platoon consists of i = {1, 2, ..., N } vehicles, where i = 1 is the lead
vehicle of the platoon and i = N is the tail vehicle. vi denotes the velocity of the
ith vehicle and di−1,i denotes the intermediate distance between the ith vehicle and
the vehicle ahead. Furthermore, the movement of the platoon is assumed to be only
longitudinal.

Figure 2.1: Vehicle platooning of N vehicles on a flat road.

2.1.1 Cruise Control and Adaptive intelligent Cruise Control

The Cruise Control (CC), also known as speed control, is a system that automati-
cally maintains the vehicle speed according to the speed reference set by the driver.
An extension to the CC is the Adaptive intelligent Cruise Control (AiCC), which
takes the vehicle ahead into account. The AiCC will lower its speed to the speed of
the vehicle ahead, if the reference speed of the AiCC is set higher than the current
speed of the vehicle in front. The distance between the vehicles is the vehicle speed
multiplied with a time gap set by the driver, ddesire = τdesire v. This is achieved
through radar measurements of the velocity and the intermediate distance of the
vehicle in front. The AiCC will act as a CC when there is no vehicle within the
range of the radar. The experiment in (Alam et al., 2010) utilized the AiCC to
accomplish the 7.7 % reduction of fuel consumption for two identical HDVs.

5
 CHAPTER 2. BACKGROUND

2.1.2 Vehicle-to- vehicle/infrastructure communication

Vehicle-to-vehicle (V2V) and vehicle-to-infrastructure communication (V2I), also
known together as V2X communication, allows wider spread of information than
the AiCC. V2X communication enables additional information being broadcasted
to multiple receivers. With V2V communication, information from several vehicles
ahead and the following vehicles can be obtained to make any driving adjustments
accordingly. It is also possible to spread its own information and behavior to nearby
vehicles. This two-way communication makes it possible for a robust vehicle pla-
tooning. V2I communication is a communication between vehicle and infrastruc-
ture. It allows the vehicle to gather information such as traffic light or any road
work ahead that needs to be considered. This allows smarter path planning, which
can be both more time and fuel efficient. In a vehicle platooning with close inter-
mediate distances, the possibilities of time delays, noises, and package losses have
to be taken into account such that no collision occurs.

2.2 String stability

One robustness analysis in vehicle platooning is string stability. It is possible that
a small change in velocity occurs and amplifies throughout the platoon, this could
lead to an unwanted platoon behavior, i.e. a slinky effect, and may result in vehicle
collisions. Thus it is of most importance to have a controller that prevents the
amplification and propagation. If a platoon has the ability to prevent any slinky
effects, it is called string stable. Human drivers in a traffic jam are an example
of string instability due to different reactions and time delays, that causes a slinky
effect.

The relation between vehicle i and i − 1 is expressed as:

vi (s) = Gvi (s)vi−1 (s) (2.1)
where Gvi (s)
is the velocity transfer function for the ith vehicle. The function
describes how the ith vehicle will behave according to the velocity of i − 1th vehicle.
In order to prevent amplification and propagation of velocity changes, the following
is defined (Yamamura and Seto, 2006):
kGvi (s)k∞ ≤ 1. (2.2)
It is difficult to satisfy eq. (2.2) for every vehicle in a platoon due to non-identical
vehicle dynamics and different controllers, thus another approach could be to define:
kGv2 (s)Gv3 (s)...GvN (s)k∞ ≤ 1 (2.3)
It is clear that eq. (2.3) is satisfied when eq. (2.2) is, for all i ∈ [2, N ]. Both equations
are string stability criteria, however eq. (2.2) is more conservative. Equation (2.2)
prevents amplification of velocity change from one vehicle to another. Meanwhile
eq. (2.3) assures the tail vehicle does not obtain a disastrous behavior due to a small
velocity change from the lead vehicle.

6
2.3. RELATED WORK

2.3 Related work

The term string stability has been known for the last decades with a large amount
of published articles. Articles describing the possibility to stabilize platoons with
different strategies and controllers. Variable time headway (Yanakiev and Kanel-
lakopoulos, 1995), obtaining information both from the vehicle ahead and behind
(Zhang et al., 1999), and utilizing non-identical decentralized control (Khatir and
Davison, 2004) are just three of the many strategies and controls for string stability
that have been published. One common factor of these articles is the assumption of
identical vehicles in the platoon. Therefore, the first part of this thesis is to analyze
the behavior of a platoon with a mass variation.

This thesis is closely related to the master thesis (Hammar and Ovtchinnikov, 2010),
where it has taken the air drag reduction into account and performed simulations
with different masses. Their work describes a possible strategy to utilize a decen-
tralized controller on a HDV platoon. Hence, this thesis focuses on finding the
possibilities with a centralized controller.

2.4 Thesis objective & delimitations

There has been a large amount of research regarding string stability. The authors
mainly focus on finding a controller, which allows the vehicle platoon to be string
stable. Therefore, one approach in this thesis is to analyze how mass variation can
affect string stability of a platoon. This idea came from a spring-mass system where
a mass difference causes a damping effect.

Studies of platooning has gotten far with several ideas of how it can be done, es-
pecially from the string stability articles. Many suggests a decentralized controller,
whereas a centralized control has not been studied further. The AiCC today is a
good example of a decentralized controller used commercially. However, there is
no commercially available centralized controller. The reason is that there does not
exist a standard platform of how it can be obtained. The technology is available but
further studies of how communication in platooning should be done and platooning
itself need to be done before realizing it. This project has its goal to find a central-
ized control, which takes the air drag reduction into account and that can be used
in platooning. The goal is to analyze the potentials and how well it performs, both
in fuel consumption and the ability for maneuvers, compared to a decentralized
controller.

This project is divided into two parts, the first part focuses on how mass varia-
tions in a platoon affects string stability. Simple dynamics and controller are used
to grasp the essential behavior of platooning. This study will mainly focus on mass
variations with some simple simulations to verify the results. Meanwhile, the second

7
 CHAPTER 2. BACKGROUND

part introduces a centralized control that will be simulated using a more realistic
vehicle model. The movements of the platoon will be assumed longitudinal due
to simplicity and all the information being transferred between vehicles are instan-
taneous. The only information being broadcasted is the longitudinal position and
velocity. No additional functions such as preview information of the road, known
as look-ahead, are used. This is to ensure that the focus is on how the platoon be-
haves and the possibilities of reducing the fuel consumption by utilizing a centralized
control. Intruding vehicles and other outer disturbances are neglected.

2.5 Thesis outline

The thesis is structured as followed: chapters 3 and 4 are part II of the thesis,
which concerns the string stability study with mass variations. Chapter 3 gives the
functions of a simplified platoon and verifies them, meanwhile the following chapter
presents a solution to the problem.

Part III concerns the centralized controller. It starts with deriving and describ-
ing the vehicle and platoon model of three vehicles thoroughly in chapter 5. This
will serve as a premise for all the forthcoming calculations. Chapter 5 also de-
scribes a linearization of the non-linear platoon model, since chapter 6 presents an
optimal control solution based on a linear model. Chapter 7 shows the results of
the simulations based on a Simulink model made by Scania CV AB. A solution to
expand the platoon, from a platoon of three vehicles to a larger platoon consisting
of sub-platoons, is introduced in chapter 8.

Parts II and III can be read separately, due to the difference in modeling the sys-
tem. Lastly, chapters 9 to 11 summarize and conclude the whole project along with
future possibilities with centralized platooning.

8
 Part II

String stability

9
Chapter 3

PID-controller

A vehicle platoon can be simplified as a mass-spring-damper system (Yanakiev

and Kanellakopoulos, 1996), which this part of study will be using similarly. This
study mainly focuses on how a vehicle platoon behaves and how it affects string
stability with mass variations using identical control. With the mass-spring system
simplification, the vehicle platoon system follows as:

mi ẍi + bi ẋi = ui (3.1)

where the index i denotes the ith vehicle in the platoon, mi is the mass, bi is the
damping effect, xi is the longitudinal position, and ui is the input of the ith vehicle.

In this chapter, it is assumed that all the vehicles use the same PID-controller
but the masses of the vehicles vary. The following PID-controller is used:
Z
ui = Pi (xi−1 − xi − τd ẋi ) + Ii ( xi−1 − xi − τd ẋi dt) + Di (ẋi−1 − ẋi ) (3.2)

where τd denotes the time headway. The index i in the controller will be kept, even
if identical controller is used for all vehicles. This makes it easier to distinguish
which vehicle the controller belongs to.

3.1 Velocity and relative distance

The first two relations that will be derived are how a velocity change vi−1 affects the
relative distance di−1,i = xi−1 − xi and how that distance affects the velocity of the
following vehicle, vi . These two relations will serve as a premise for the forthcoming
calculations of the string stability part. The dynamic system is given by inserting
eq. (3.1) into eq. (3.2):
Z
mi ẍi + bi ẋi = Pi (xi−1 − xi − τd ẋi ) + Ii ( xi−1 − xi − τd ẋi dt) + Di (ẋi−1 − ẋi ). (3.3)

11
 CHAPTER 3. PID-CONTROLLER

By introducing the following notations xi−1 − xi = di−1,i and ẋi = vi , the Laplace
transform can be given as:

1 1
(mi s + bi )vi (s) = (Pi + Ii + Di s)di−1,i (s) − (τd Pi + τd Ii )vi (s).
s s
The first relation obtained from the calculations is how vehicle i reacts to a change
in relative distance di−1,i , which is given by the following transfer function:

vi (s) s2 Di + sPi + Ii
Hidv (s) = = 2 (3.4)
di−1,i (s) s mi + s(bi + τd Pi ) + τd Ii

The second relation is how velocity vi−1 affects the relative distance di−1,i . This can
be obtained with sdi−1,i (s) = vi−1 (s) − vi (s) giving the following transfer function:

di−1,i (s) s2 mi + s(bi + τd Pi ) + τd Ii

Hivd (s) = = 3 (3.5)
vi−1 (s) s mi + s2 (bi + τd Pi + Di ) + s(Pi + τd Ii ) + Ii

The steady state error of both transfer functions, eqs. (3.4) and (3.5), are dependent
on τd , which is reasonable due to which time headway is set by the driver.

3.2 Velocity transfer function

The more important part in a platoon is to analyze how a velocity change affects
other vehicles in the platoon. A small velocity change by the lead vehicle should
not amplify and propagate throughout the platoon and cause any collisions. By
multiplying both the earlier derived transfer functions, eqs. (3.4) and (3.5), the
velocity transfer function is obtained:

vi (s)
Gvi (s) = Hidv (s)Hivd (s) =
vi−1 (s)
(3.6)
s Di + sPi + Ii
2
= 3
s mi + s2 (bi + τd Pi + Di ) + s(Pi + τd Ii ) + Ii

With eq. (2.2), the string stability criteria is fulfilled when kGvi (s)k∞ ≤ 1. The
inequality fulfillment depends mostly on the values of Pi , Ii , and Di , and somewhat
on the mass mi . This study refers from being dependent on what PID-controller
is used, hence eq. (3.6) does not give any insight on how a mass variation affects
string stability since it also depends on the PID-controller.

3.3 Relative distance transfer function

Another aspect that should be analyzed in platooning is how a change in relative
distance propagates through the platoon. Relative distance between vehicles is a

12
3.4. SIMULATION STUDY I

straightforward answer whether the vehicles collide or not. The transfer function
for relative distance can be obtained with the calculation:
di−1,i (s)
Gdi (s) = Hi−1
dv
(s)Hivd (s) = (3.7)
di−2,i−1 (s)
s2 mi + s(bi + τd Pi ) + τd Ii s2 Di−1 + sPi−1 + Ii−1
= 2 .
s mi−1 + s(bi−1 + τd Pi−1 ) + τd Ii−1 s3 mi + s2 (bi + τd Pi + Di ) + s(Pi + τd Ii ) + Ii
Notice that the masses of two different vehicles are now included in the transfer
function. The first quotient could be simplified due to the identical PID-controller
and by doing the following assumptions:

Pi−1 Pi

 ≈



 m i−1 m i
 b

b
i−1 i
≈ (3.8)


 mi−1 mi
I Ii

i−1

≈


 .
mi−1 mi

Then eq. (3.7) can be approximated as:

di−1,i (s) mi s2 Di−1 + sPi−1 + Ii−1

Gdi (s) = ≈ . (3.9)
di−2,i−1 (s) mi−1 s3 mi + s2 (bi + τd Pi + Di ) + s(Pi + τd Ii ) + Ii

This approximation clarifies how the platoon should be arranged depending on the
vehicle masses. By arranging heavier vehicles first and with a decreasing order
in mass, the simplified quotient becomes less than one, mmi−1 i
< 1. This enables
the eq. (3.9) to satisfy the inequality kGi (s)k∞ ≤ 1. It should be noted that the
d

inequality could be satisfied with the approximation for eq. (3.9) but may not be
satisfied for eq. (3.7). Although the string stability criteria does only include the
velocity changes between the vehicles and does not include the amplification and
propagation between the relative distances, this area could be as important to study.

3.4 Simulation study I

To verify the result from section 3.3, a small simulation study is made in Matlab. A
PID-controller is used and the simulation scenario is a ramp from 0 m/s to 10 m/s
in 10 s. To be able to verify the result from eq. (3.9), three different platoon ar-
rangements were used; identical vehicles, increasing weight order, and in decreasing
order in mass. To make the three different platoon arrangements comparable, the
platoon length and the sum of the weight of all vehicles in the platoon are the same.
The reader is referred to appendix A for further information on the simulation setup.

The first simulation is done with a platoon of identical vehicles with the mass of 40
tonnes each. Figure 3.1 illustrates the result with two different platoon lenghts: 41

13
 CHAPTER 3. PID-CONTROLLER

vehicles in fig. 3.1a and 81 vehicles in fig. 3.1b. The platoon behavior depicted in
fig. 3.1 shows a velocity change amplifies throughout the platoon for each vehicle.
Furthermore, the relative distance between the vehicles gets influenced, with slinky
effect and amplifying distance error as a result. Such conduct is considered string
unstable and the H∞ -norm of the velocity transfer function is also above one, see
appendix A.

Mass of 40t for each vehicle, totally 41 vehicles Mass of 40t for each vehicle, totally 81 vehicles
20 25

20
15
Velocity (m/s)

Velocity (m/s)
15
10
10

5
5

0 0
0 50 100 150 200 250 300 350 400 450 500 0 50 100 150 200 250 300 350 400 450 500
Time (sec) Time (sec)

30 40

25
Relative distance (m)

Relative distance (m)

30
20

15 20

10
10
5

0 0
0 50 100 150 200 250 300 350 400 450 500 0 50 100 150 200 250 300 350 400 450 500
Time (sec) Time (sec)

(a) Platoon length of 41 identical vehicles (b) Platoon length of 81 identical vehicles

Figure 3.1: A ramp with identical vehicles of mass 40 tonnes each in a platoon. The
velocity and relative distance are plotted.

The second platoon configuration is with weight increasing order. The lead vehicle
has a mass of 20 tonnes and the tail vehicle has a weight of 60 tonnes. Two different
platoon lengths were simulated, which are presented in fig. 3.2. Figure 3.2a shows
the platoon behavior with 41 vehicles with 1000 kg weight gap and fig. 3.2b with 81
vehicles with a weight gap of 500 kg. The results show that the velocity behavior
remains growing, similar to the previous simulation, but the major difference is the
relative distance. Analysis from eq. (3.9) clearly explains that the H∞ -norm grows
with a lighter vehicle in front. This platoon configuration is not considered string
stable.

The third and last platoon simulation is a platoon with decreasing order in mass.
The mass will remain the same as the second simulation but in reverse order, from
60 tonnes to 20 tonnes. The platoon behavior is depicted in fig. 3.3. Notice that
the max velocity is the same for all three platoon configuration. However, the only
configuration is the weight decreasing order where the velocity amplification sub-
sides further back in the platoon. The relative distance for the decreasing weight
order is highly advantageous over the other platoon configurations. This confirms
eq. (3.9), that by arranging heavier vehicles first in a platoon with a descending
order the platoon can become string stable. This is with the assumptions that
identical controllers are used for every vehicle.

14
3.4. SIMULATION STUDY I

Mass of 20 to 60t with step: 1000kg, totally 41 vehicles Mass of 20 to 60t with step: 500kg, totally 81 vehicles
20 25

20
15
Velocity (m/s)

Velocity (m/s)
15
10
10

5
5

0 0
0 50 100 150 200 250 300 350 400 450 500 0 50 100 150 200 250 300 350 400 450 500
Time (sec) Time (sec)

35 50

30
40
Relative distance (m)

Relative distance (m)

25

20 30

15 20
10
10
5

0 0
0 50 100 150 200 250 300 350 400 450 500 0 50 100 150 200 250 300 350 400 450 500
Time (sec) Time (sec)

(a) 41 vehicles, weight gap of 1000 kg (b) 81 vehicles, weight gap of 500 kg

Figure 3.2: A ramp with weight increasing platoon order, 20 to 60 tonnes. The
velocity and relative distance are plotted.

Mass of 60 to 20t with step: 1000kg, totally 41 vehicles Mass of 60 to 20t with step: 500kg, totally 81 vehicles
20 25

20
15
Velocity (m/s)

Velocity (m/s)

15
10
10

5
5

0 0
0 50 100 150 200 250 300 350 400 450 500 0 50 100 150 200 250 300 350 400 450 500
Time (sec) Time (sec)

30 35

25 30
Relative distance (m)

Relative distance (m)

25
20
20
15
15
10
10
5 5

0 0
0 50 100 150 200 250 300 350 400 450 500 0 50 100 150 200 250 300 350 400 450 500
Time (sec) Time (sec)

(a) 41 vehicles, weight gap of 1000 kg (b) 81 vehicles, weight gap of 500 kg

Figure 3.3: A ramp with weight decreasing platoon order, 60 to 20 tonnes. The
velocity and relative distance are plotted.

15
Chapter 4

Scaled-varying PID-controller

The problem, with identical controllers in a platoon, is the different dynamics each
vehicle has depending on the mass. Each vehicle with different weight has different
dynamics and therefore renders identical PID-controller useless for heavier vehicles.
A fixed PID-controller is useful for light weighted vehicles up to a point where the
controller is no longer considered string stable. A proposed solution is to introduce
a scalar to the PID-controller, which can take the mass of the vehicle into account.
The vehicle system keeps the same dynamics as in eq. (3.1), meanwhile the proposed
controller is now expressed as:
Z
ui = Ki [Pi (xi−1 − xi − τd ẋi ) + Ii ( xi−1 − xi − τd ẋi dt) + Di (ẋi−1 − ẋi )] (4.1)

The following sections will follow the same procedure as in chapter 3.

4.1 Transfer functions

The derivations of the four next coming transfer functions are done with exactly
the same steps as in sections 3.1 to 3.3, and are therefore omitted in this section.
The results of the transfer functions are:
vi (s) s2 Di + sPi + Ii
H̄idv (s) = = mi (4.2)
di−1,1 (s) s2 Ki + s( Kbii + τd Pi ) + τd Ii
s2 m Ki + s( Ki + τd Pi ) + τd Ii
bi
di−1,i (s) i

H̄ivd (s) = = mi (4.3)

vi−1 (s) s3 Ki + s2 ( Kbii + τd Pi + Di ) + s(Pi + τd Ii ) + Ii
vi (s) s2 Di + sPi + Ii
Ḡvi (s) = = mi (4.4)
vi−1 (s) s3 Ki + s2 ( Kii + τd Pi + Di ) + s(Pi + τd Ii ) + Ii
b

di−1,i (s) mi Ki−1 s2 Di−1 + sPi−1 + Ii−1

Ḡdi (s) = ≈
di−2,i−1 (s) Ki + s ( Ki + τd Pi + Di ) + s(Pi + τd Ii ) + Ii
mi−1 Ki s3 m i 2 bi

(4.5)

17
 CHAPTER 4. SCALED-VARYING PID-CONTROLLER

Equation (4.5) is an approximation using the following assumptions:


Ki−1 Pi−1 Ki Pi

 ≈



 mi−1 mi
bi−1 bi


≈ (4.6)


 mi−1 mi
 Ki−1 Ii−1 Ki Ii



 ≈
mi−1 mi

Similar to chapter 3, eqs. (4.2) to (4.4) do not add further information. The key
remains in the relative distance transfer function, eq. (4.5). It is now possible to
obtain kḠdi (s)k∞ ≤ 1 with any kind of mass distribution with the help of the scalars,
Ki−1 and Ki . This was not possible for Gdi (s) from eq. (3.9), since it was bounded
to be arranged in a weight decreasing order to achieve a stable behavior.

4.2 Simulation study II

To validate the results from section 4.1, same simulation from section 3.4 will be
done but with a scaled-varying PID-controller. The varying scalar is set to be:
Ki = 10000
mi
. For further information on the simulation setup, the reader is referred
to read appendix A.

Mass of 40t for each vehicle, totally 41 vehicles Mass of 40t for each vehicle, totally 81 vehicles
12 12

10 10
Velocity (m/s)

Velocity (m/s)

8 8

6 6

4 4

2 2

0 0
0 50 100 150 200 250 300 350 400 450 500 0 50 100 150 200 250 300 350 400 450 500
Time (sec) Time (sec)

12 12

10 10
Relative distance (m)

Relative distance (m)

8 8

6 6

4 4

2 2

0 0
0 50 100 150 200 250 300 350 400 450 500 0 50 100 150 200 250 300 350 400 450 500
Time (sec) Time (sec)

(a) Platoon length of 41 identical vehicles (b) Platoon length of 81 identical vehicles

Figure 4.1: A ramp with identical vehicles using scaled-varying PID. The velocity
and relative distance are plotted.

The results of the three different weight distributions are all illustrated in figs. 4.1
to 4.3. The three figures show similar platoon behavior and most importantly, the
relative distances between the vehicles do not get amplified or inherit any slinky

18
4.2. SIMULATION STUDY II

Mass of 20 to 60t with step: 1000kg, totally 41 vehicles Mass of 20 to 60t with step: 500kg, totally 81 vehicles
12 12

10 10
Velocity (m/s)

Velocity (m/s)
8 8

6 6

4 4

2 2

0 0
0 50 100 150 200 250 300 350 400 450 500 0 50 100 150 200 250 300 350 400 450 500
Time (sec) Time (sec)

12 12

10 10
Relative distance (m)

Relative distance (m)

8 8

6 6

4 4

2 2

0 0
0 50 100 150 200 250 300 350 400 450 500 0 50 100 150 200 250 300 350 400 450 500
Time (sec) Time (sec)

(a) 41 vehicles, weight gap of 1000 kg (b) 81 vehicles, weight gap of 500 kg

Figure 4.2: A ramp with increasing weight order using scaled-varying PID. The
velocity and relative distance are plotted.

Mass of 60 to 20t with step: 1000kg, totally 41 vehicles Mass of 60 to 20t with step: 500kg, totally 81 vehicles
12 12

10 10
Velocity (m/s)

Velocity (m/s)

8 8

6 6

4 4

2 2

0 0
0 50 100 150 200 250 300 350 400 450 500 0 50 100 150 200 250 300 350 400 450 500
Time (sec) Time (sec)

12 12

10 10
Relative distance (m)

Relative distance (m)

8 8

6 6

4 4

2 2

0 0
0 50 100 150 200 250 300 350 400 450 500 0 50 100 150 200 250 300 350 400 450 500
Time (sec) Time (sec)

(a) 41 vehicles, weight gap of 1000 kg (b) 81 vehicles, weight gap of 500 kg

Figure 4.3: A ramp with decreasing weight order using scaled-varying PID. The
velocity and relative distance are plotted.

effect. This confirms that eq. (4.5) can fulfill the inequality kḠdi (s)k ≤ 1 for any
mass distribution. Furthermore, the string stability criteria is also satisfied for all
mass variations used in the simulation, i.e. kḠvi (s)k ≤ 1.

19
 Part III

Locally centralized three-vehicle

platoon

21
Chapter 5

Platoon model

To enable an understanding of modeling a HDV platoon, the knowledge of the

systems and the dynamics of a HDV are required. The outcome of this chapter is to
derive a state-space model of the vehicle model, from the powertrain to the vehicle
dynamics of a HDV, with a platoon of three HDVs. This is necessary in order
to develop a controller for the platoon. General vehicle modeling can be found in
(Gillespie, 1992).

5.1 Powertrain
A basic model of a powertrain with the main parts of interests is depicted in fig. 5.1.
The powertrain model follows the same model defined in (Alam, 2008).

Figure 5.1: A basic model of a powertrain.

23
 CHAPTER 5. PLATOON MODEL

Engine
The engine utilized in the model is a diesel engine. The engine output torque,
represented as Te (ωe , δf ) where ωe is the angular velocity of the flywheel and δf is
the fueling level, is characterized by the driving torque, which is produced during
combustion, minus internal friction of the engine. The external load comes from the
clutch denoted with Tc . Newton’s second law of motion gives the following relation:

Je ω̇e = Te − Tc (5.1)

where Je is the mass moment of inertia of the engine.

When the engine works within the operating range, which is ideally between 1100
and 1400 RPM, the torque Te (ωe , δf ) can be linearized as:

Te (ωe , δf ) = k1 ωe + k2 δf + k3 (5.2)

where k1 , k2 , and k3 are constants of the engine properties. The conversion between
RPM and angular velocity is:
60
RP M = ωe .
2π

Clutch
The clutch consists of two frictional discs connecting the flywheel of the engine with
the input shaft of the transmission. This type of clutch is mostly found in vehicles
equipped with manual transmission. The clutch is considered to be stiff, which
yields in no torque and angular velocity change:

Tt = Tc
(5.3)
ωt = ωc

where Tt denotes the torque output from the clutch and input to the transmission,
and ωt denotes the angular velocity.

Transmission
The transmission consists of a set of gears that converts the output torque depending
on the engaged gear. Each gear has a conversion ratio it and an efficiency constant
ηt . The inertia is neglected and the gear shifts are assumed to be instantaneous,
which results in the following relations:

Tp = it ηt Tt
(5.4)
it ωp = ωe

where Tp denotes the output torque from the transmission and input to the propeller
shaft, and ωp denotes the angular velocity.

24
5.1. POWERTRAIN

Propeller shaft
The propeller shaft connects the transmission with the final drive. The shaft is
considered to be stiff with no friction, which gives:

Tf = Tp
(5.5)
ωf = ωp

where Tp denotes the output torque from the propeller shaft and input to the final
drive, and ωp denotes the angular velocity.

Final drive
The final drive is, like the transmission, characterized by a conversion ratio if and
an efficiency constant ηf . The inertia is likewise neglected, giving the following
relations:
Td = if ηf Tf
(5.6)
if ωd = ωf
where Td denotes the output torque from the final drive and input to the drive
shafts, and ωd denotes the angular velocity.

Drive shafts
The drive shafts connect the final drive with the wheels. It is presumed that the
wheel speed is the same for both wheels and the drive shafts are stiff. Therefore,
the powertrain equation is not influenced by the following relations:

Tw = Td
(5.7)
ωw = ωd

where Tw denotes the output torque to the wheels, and ωw denotes the angular
velocity.

Wheels
Lastly, by assuming no slip, the equation of motion for the wheel is described as:

Jw ω̇w = Tw − Tb − rw Fw (5.8)

rw ωe
v = rw ωw = (5.9)
it if
where Jw is the wheel inertia, rw is the wheel radius, v is the vehicle velocity, and
Fw is the resulting force that drives the vehicle forward. The braking torque Tb is
difficult to model and is therefore neglected in this model.

25
 CHAPTER 5. PLATOON MODEL

Complete powertrain
By combining eqs. (5.1) and (5.3) to (5.9) together, the complete powertrain equa-
tion can be expressed as:

it if ηt ηf Jw + i2t i2f ηt ηf Je
Fw = Te − 2
v̇ = Fengine − Finertia (5.10)
rw rw
where Fengine is the force produced from the engine and Finertia is the inner force
required for the engine to overcome in order to produce a driving force.

5.2 Longitudinal forces

When a HDV is in motion, there are several forces acting on the vehicle. Figure 5.2
illustrates the longitudinal forces on the HDV. Fbrake is neglected due to the expla-
nation in previous section, and α denotes the road angle.

Figure 5.2: The longitudinal forces acting on a HDV in motion.

By applying Netwon’s second law of motion:

mv̇ = Fw − Fbrake − Fairdrag − Froll − Fgravity

with eq. (5.10) the equation can be rewritten as:

mt v̇ = Fengine − Fbrake − Fairdrag − Froll − Fgravity (5.11)

where m denotes the vehicle mass and mt is the accelerated mass given by:

Jw + i2t i2f ηt ηf Je
mt = m + 2
. (5.12)
rw

26
5.2. LONGITUDINAL FORCES

Aerodynamic force
The aerodynamic force is given by:
1
Fairdrag (v, d) = cD Aa ρa v 2 (5.13)
2
where Aa is the maximal cross sectional area of the vehicle, and ρa is the air density.
cD (d) = cd (1 − f100
i (d)
) is the air drag coefficient with the air drag reduction produced
by the preceding vehicle. d is the intermediate distance between the vehicles and
fi (d) is a non-linear function depicted in fig. 5.3 with i denoted as the ith vehicle
in the platoon.
Mapping of Cd reduction
80
Lead HDV
One HDV ahead
70 Two HDVs ahead

60

50
Cd reduction [%]

40

30

20

10

0
0 10 20 30 40 50 60 70
Relative distance in convoy [m]

Figure 5.3: Empirical result of air drag reduction of HDV platooning (Alam et al.,
2010).

Due to the non-linearity of fi (d), the function is approximated with a first order
least square approximation giving the following:
f1 (d)ls = −0.9379d + 12.8966 0 ≤ d ≤ 15
f2 (d)ls = −0.4502d + 43.0046 0 ≤ d ≤ 80
(5.14)
f3 (d)ls = −0.4735d + 51.5027 0 ≤ d ≤ 80
otherwise fi (d)ls = 0
where fi (d)ls is the least square approximation of the air drag reduction. It is
assumed that the air drag reduction does not increase any further beyond the third
vehicle.

27
 CHAPTER 5. PLATOON MODEL

Roll resistance
The roll resistance force is modeled as:
Froll (α) = cr mg cos(α) (5.15)
where cr is the roll coefficient, and g is the gravitational constant.

Gravity
The gravitational force is given by:
Fgravity (α) = mg sin(α). (5.16)

Combined equations
By combining eqs. (5.10) to (5.16) together, a mathematical expression for the
vehicle model can be expressed as:
2
rw it if ηt ηf 1

v̇i = Te − cd Aa ρa v 2
Jw + mrw
2 + i2 i2 η η J
t f t f e rw 2
(5.17)
1 fi (d)

+ cd Aa ρa v 2 − cr mg cos(α) − mg sin(α) .
2 100
By introducing the following five constants:
1 2
rw it if ηt ηf 2 rw Aa ρa cd
κ1 = κ2 =
Jw + mrw
2 + i2 i2 η η J
t f t f e Jw + mrw2 + i2 i2 η η J
t f t f e
2 mg
cr rw 2 mg
rw
κ3 = κ4 = (5.18)
Jw + mrw
2 + i2 i2 η η J
t f t f e Jw + mrw
2 + i2 i2 η η J
t f t f e

fi (d)
 
φi (d) = 1 −
100
the vehicle model from eq. (5.17) can be simplified to:
v̇i = κ1 Te − κ2 φi (d)v 2 − κ3 cos(α) − κ4 sin(α). (5.19)

5.3 Linearized model

A linear quadratic solution will be utilized, therefore the non-linear model from
eq. (5.17) needs to be linearized. The linearization point are chosen around a set
reference velocity vref , a constant slope αref , a set distance dref and an engine
torque that maintains the velocity. The equilibrium point are denoted by:
v0 = vref
α0 = αref
d0 = dref = τref v0 (5.20)
κ2 φi (d0 )ls v02 + κ3 cos(α0 ) + κ4 sin(α0 )
T0 = .
κ1

28
5.3. LINEARIZED MODEL

By applying a first order Taylor approximation around the equilibrium point in

eq. (5.20), the following linearized model is obtained:

v̇i ≈ g 0 (v0 , α0 , d0 , T0 ) + gv0 (v0 , α0 , d0 , T0 )∆vi + gd0 (v0 , α0 , d0 , T0 )∆di + gT0 (v0 , α0 , d0 , T0 )∆Ti
≈ −2κ2 φi (d0 )ls v0 ∆vi − κ2 φ0i (d0 )ls v02 ∆di + κ1 ∆Ti (5.21)
= ai ∆vi + bi ∆di + ci ∆Ti

where:
 0.009379 i=1

d fi (d)ls
  
φ0i (d)ls = 1− = 0.004502 i = 2
dd 100
0.004735 i = 3.



The change in relative distance between two vehicles is described by:

∆d˙i−1,i = ∆vi−1 − ∆vi . (5.22)

The platoon model can now be written in a state-space equation through eqs. (5.21)
and (5.22):

∆v̇1 a1 b1 0 0 0 ∆v1 c1 0 0
      

∆d˙1,2 1 0 −1 0 0 ∆d1,2 0 0 0  ∆T
 
     
       1
∆v̇2 = 0 b2 a2 0 0 ∆v2 + 0 c2 0   ∆T2 
       
 
∆d˙2,3 0 0 1 0 −1 ∆d2,3 0 0 0  ∆T3
      
     
∆v̇3 0 0 0 b3 a3 ∆v3 0 0 c3
(5.23)
⇒ ẋ = Ax + Bu.

29
Chapter 6

Optimal control

The optimal control theory is a mathematical optimization method to solve control

theory problems. The problem to solve with optimal control is to find a control law
for a given system with respect to given cost functions that is penalized if the states
or inputs of the system get too large. If the dynamical system is described by linear
equations and the costs is a quadratic function of the states and inputs, then the
optimal control problem can be solved as a linear quadratic (LQ) problem.

The centralized controller for the platoon will be solved with two different LQ
solutions. Firstly, a linear quadratic regulator (LQR) solution is used. This solu-
tion does not directly determine a reference signal that the vehicles can track, i.e. a
velocity variation that can be followed. Therefore, to enable velocity tracking, the
lead vehicle is equipped with the CC. The CC is set such that velocity changes can
be done and the lead vehicle will follow the reference. This disables the property to
control the lead vehicle with the LQR controller, hence only the two latter vehicles
are controlled. The second centralized controller utilized is a linear quadratic track-
ing (LQT) with integral action. This solution allows a reference signal to be tracked.
This enables a control law for the whole vehicle platoon, consisting of three vehicles.

The reason for two different LQ solutions is to achieve a better understanding of

vehicle platooning. Is it better to let the lead vehicle drive with the CC and let the
two following vehicles catch up or let the lead vehicle consider the vehicles behind
and remain as a platoon? With that question in mind, the two control laws were
obtained.

6.1 Linear Quadratic Regulator

The standard cost function setup for a LQR problem is shown in eq. (6.1):
Z tf  
J= xT (τ )Qx(τ ) + uT (τ )Ru(τ ) dτ + xT (tf )Qf x(tf ) (6.1)
0

31
 CHAPTER 6. OPTIMAL CONTROL

together with dynamic system constraints and initial conditions, the objective to
minimize the cost function can be written as:
Z tf  
∗
J = min xT (τ )Qx(τ ) + uT (τ )Ru(τ ) dτ + xT (tf )Qf x(tf )
u 0 ( (6.2)
ẋ(t) = Ax(t) + Bu(t)
Subject to:
x(0) = x0 .

This is called a finite horizon LQR and requires information ahead of time, such
as the look-ahead. In order to solve the LQR problem, three requirements need to
be satisfied. The pair (A, B) needs to be controllable, the cost matrices Q and Qf
need to be positive semi definite, and R needs to be positive definite.

The platoon model described in chapter 5 is linearized around an equilibrium point.

Therefore, this LQR setup penalizes when the states or inputs deviate from the
equilibrium point. By minimizing the input, i.e. the torque, the fuel will also be
minimized by the relation shown in eq. (5.2). For a detailed explanation of the LQR
derivation, the reader is referred to read appendix B.

A solution for a finite horizon LQR is to apply dynamic programming (DP). First
set P (tf ) = Qf and work from time tf and backward to the initial time, zero. This
is done by solving the Riccati differential equation, numerically:

−Ṗ (t) = AT P (t) + P (t)AT − P (t)BR−1 B T P (t) + Q. (6.3)

At time t, the optimal control uopt is given by:

uopt (t) = K(t)x(t) (6.4)

where
K(t) = −R−1 B T P (t).
Repeat the steps and work the way back for all t ∈ [0, tf ] for an optimal control
over all t. This DP method extends to time-varying A, B, Q and R matrices.

A more simple case is when the matrices A, B, Q and R are constant together
with an infinite horizon problem. The cost function and conditions is then given
by: Z ∞
J ∗ = min xT (τ )Qx(τ ) + uT (τ )Ru(τ ) dτ
u 0 ( (6.5)
ẋ(t) = Ax(t) + Bu(t)
Subject to:
x(0) = x0 .
This system requires no preview information and is considered to be in steady state.
Therefore, the Riccati differential equation, eq. (6.3), can now be considered as an
Algebraic Riccati Equation (ARE):

AT Pss + Pss AT − Pss BR−1 B T Pss + Q = 0. (6.6)

32
6.2. LINEAR QUADRATIC TRACKING CONTROL WITH INTEGRAL ACTION

Equation (6.6) needs to be solved numerically and with the numerical solution, the
optimal control uopt is given by:

uopt (t) = −R−1 B T Pss x(t). (6.7)

6.2 Linear Quadratic Tracking control with integral action

The idea of this solution is to let the lead vehicle follow a reference signal, in this
case a velocity reference. This allows the lead vehicle to track a set reference velocity
smoothly with a LQ solution. Since it is a centralized control, the lead vehicle will
also take the following vehicles into account, which the LQR solution lacks. The
following method follows the same derivation as in (Groves et al., 2005).

6.2.1 Reference tracking system

The concept is to introduce an error: ė = r − x, where r is one or several reference
signals. By including an integrating action, the feedback control eliminates the
steady state error of the system. Hence, the integral error is given by e = ė dt,
R

which is introduced as a new state to the dynamic system:

" # " #" # " # " #

ė 0 −I e 0 I
= + u+ r (6.8)
ẋ 0 A x B 0

⇒ x̃˙ = Ãx̃ + B̃u + Gr.

The error system for eq. (6.8) can then be defined as:

" #
e
ẽ = = M r + H x̃ (6.9)
r−x

where:
" # " #
0 I 0
M= H= .
I 0 −I

This will serve as a premise for the LQT cost function in section 6.2.3.

33
 CHAPTER 6. OPTIMAL CONTROL

6.2.2 Reference tracking platoon model

By adapting the platoon model, eq. (5.23), with the reference tracking system de-
scribed in section 6.2.1, the new dynamical system can be described as:

∆ė 0 −1 0 0 0 0 ∆e
    
 ∆v̇1   0 a1 b1 0 0 0  ∆v1 
∆d˙1,2
    
   0 1 0 −1 0 0  ∆d1,2 
= +
    
∆v̇2 0 0 b2 a2 0 0 ∆v2
 
    
∆d˙2,3 0 0 0 1 0 −1 ∆d2,3
    
    
∆v̇3 0 0 0 0 b3 a3 ∆v3
(6.10)
0 0 0 1
   

 c1 0 0 
 ∆T
 
 0 

1
 0 0 0   0 
+   ∆T2  +  r
    
0 c2 0 0
 ∆T3
   
0 0 0 0
  
   
0 0 c3 0

⇒ x̃˙ = Ãx̃ + B̃u + Gr

where ∆e = (r − ∆v1 ).
R

The new error system is defined as:

∆e 0 1 0 0 0 0 0 ∆e
      

 r − ∆v1  
  1 


 0 −1 0 0 0 0 
 ∆v1 

 ∆d1,2   0   0 0 1 0 0 0  ∆d1,2 
ẽ =  = r +  (6.11)
      
∆v2 0 0 0 0 1 0 0 ∆v2
 
      
∆d2,3 0 0 0 0 0 1 0 ∆d2,3
      
      
∆v3 0 0 0 0 0 0 1 ∆v3

⇒ ẽ = M r + H x̃.

6.2.3 LQ optimal control

The cost function for a LQ problem with finite horizon now follows as:

Z tf  
J= ẽT (τ )Qẽ(τ ) + uT (τ )Ru(τ ) dτ + ẽT (tf )Qf ẽ(tf )
0 (
˙ (6.12)
ẽ(t) = H Ãx̃(t) + H B̃u(t) + HGr(t)
Subject to:
ẽ(0) = ẽ0

34
6.3. COST FUNCTION DESIGN

with the assumption that ṙ = 0 and by inserting eq. (6.9), the expanded cost
function is:
Z tf  
J= x̃T H T QH x̃ + 2rT M T QH x̃ + rT M T QM r + uT Ru dτ +
0
+x̃ (tf )H T Qf H x̃(tf ) + 2rT (tf )M T Qf H x̃(tf ) + rT (tf )M T Qf M r(tf )
T
(6.13)
˙
(
x̃(t) = Ãx̃(t) + B̃u(t) + Gr(t)
Subject to:
x̃(0) = x̃0 .

The requirement remains the same as for the LQR case. The pair (Ã, B̃) must be
controllable, and the cost matrices Q and Qf are positive semi definite while R is
positive definite. This gives the following differential equations (readers are referred
to appendix C for the derivation):

Ṗ = −P Ã − ÃT P − H T QH + P B̃R−1 B̃ T P (6.14a)

   
ġ = P B̃R−1 B̃ T − ÃT g − H T QM + P G . (6.14b)

Equation (6.14a) is the same standard Riccati equation as eq. (6.3), while eq. (6.14b)
is an auxiliary vector equation that defines the feedforward gain.

By assuming constant Ã, B̃, Q and R with infinite horizon, the controller can be
assumed to be in steady state, i.e. Ṗ = 0 and ġ = 0. The following optimal control
uopt is obtained as:
uopt (t) = Kx x̃(t) + Kr r(t) (6.15)
where:
Kx = −R−1 B̃ T Pss
 −1  
Kr = −R−1 B̃ T Pss B̃R−1 B̃ T − ÃT H T QM + Pss G .

6.3 Cost function design

Normally, the difficult part with utilizing a LQ solution is the choice of cost matrices,
in this case Q and R. The matrices determine how the dynamical system performs
and should therefore be chosen such that the desired performance is achieved. The
costs and cost matrices will vary for the two LQ solutions, due the additional states
i.e. the reference signal and the accumulated error. However, a general idea for the
vehicle platoon, from eq. (5.23), is described.

It is important that the vehicles in a platoon do not collide within the platoon,
hence the vehicle should not deviate from the set time gap. Therefore, the differ-
ence between the relative distance and set distance gap should be penalized:

di−1,i − τ vi = (∆di−1,i + dref ) − τ (∆vi + vref ) = ∆di−1,i − τ ∆vi .

35
 CHAPTER 6. OPTIMAL CONTROL

The vehicle performance can be smoother by also adjusting the velocity according
to the velocity of the vehicle ahead, hence:
vi−1 − vi = (∆vi−1 + vref ) − (∆vi + vref ) = ∆vi−1 − ∆vi .
Lastly, the torque inputs should be penalized from its equilibrium point to maintain
a low fuel consumption:
∆Ti .

The LQR cost function can now be described as:

Z ∞
J= ϕd1 (∆d1,2 − τ ∆v2 )2 + ϕd2 (∆d2,3 − τ ∆v3 )2 + ϕv1 (∆v1 − ∆v2 )2 +
0  (6.16)
+ϕv2 (∆v2 − ∆v3 )2 + ϕT1 ∆T22 + ϕT2 ∆T32 dτ.

Notice that ∆T1 is not penalized in the LQR solution, which was mentioned in the
introduction of chapter 6. The lead vehicle utilizes the CC to maintain its own
speed without regarding the following vehicles. The cost function for the LQT is
slightly different, with additional penalties for torque input for the lead vehicle, the
accumulating error, the reference deviation r − ∆v1 , and relative distance between
vehicle 1 and 3.

Equation (6.16) can now be written in matrix form where the cost matrices are
as follows:
ϕv1 0 −ϕv1 0 0
 

 0 d
ϕ1 −τ ϕd1 0 0 

QLQR =  −ϕ1 −τ ϕ1 ϕ1 + τ ϕ1 + ϕ2
v d v 2 d v 0 −ϕv2
 

0 0 0 d −τ ϕd2
 
ϕ2
(6.17)
 
0 0 −ϕv2 −τ ϕd2 ϕv2 + τ 2 ϕd2
" #
ϕT1 0
RLQR = .
0 ϕT2
Similar method gives the cost matrices for the LQT solution:
ϕe 0 0 0 0 0
 

 0 ϕv1 0 0 0 0 

 0 0 ϕd1 + ϕd3 −τ (ϕd1 + ϕd3 ) ϕd3 −τ ϕd3 
QLQT = 
 
0 0 −τ (ϕ1 + ϕ3 ) τ (ϕ1 + ϕ3 ) + ϕ2
d d 2 d d v −τ ϕ3d −ϕv2 + τ 2 ϕd3

 
0 0 ϕd3 −τ ϕd3 ϕd2 + ϕd3 −τ (ϕd2 + ϕd3 )
 
 
0 0 −τ ϕ3d −ϕ2 + τ ϕ3
v 2 d −τ (ϕ2 + ϕ3 ) ϕ2 + τ 2 (ϕd2 + ϕd3 )
d d v

0 0
 
ϕT1
RLQT = 0 ϕT2 0 .
 
0 0 ϕT3
(6.18)

36
Chapter 7

Simulation

The simulations were done in Simulink with a vehicle model made by Scania CV
AB. A few modifications were made to enable platooning and the implementation of
the proposed controllers. To analyze the effect a platoon has on fuel consumption,
a few fuel saving functionalities such as downhill speed control and look-ahead were
removed. This is to obtain more accurate simulation results, and also due to the
reason that the proposed controllers cannot take those features into account. The
obtained Simulink model did not have a fuel consumption output, however the
consumed energy in the engine was obtainable. The consumed energy is almost
proportional to the fuel consumption, and is therefore a valid result to analyze.

7.1 Simulation scenario

A scenario with velocity changes were simulated on a 4.5 km long flat road. The
three vehicles are identical and the weight is set to 40 tonnes each. The scenario
starts with the platoon driving at 80 km/h, thereafter brake to 70 km/h at t = 50 s.
After holding the speed for 50 s, the platoon accelerates to 85 km/h and lastly, at
t = 150 s, a short deceleration back to 80 km/h occurs before reaching the goal. The
reason for the small velocity changes is that the linearized model with a LQ solution
cannot guarantee optimal performance for large deviations from the equilibrium
point. The set scenario is to examine if the LQ controllers can manage sudden
velocity changes such that the vehicles in the platoon do not collide. Different time
headways are used to analyze the saved energy potentials compared to the safety
aspects. The time headways used are τ : 0.25, 0.5, 0.75, and 1 s. The reference
result that both LQ controllers will be compared to, is a single vehicle with CC on
the same road profile.

7.2 Different time headways

Different time headways are used in order to compare the energy saved with the
relative distance. A closer distance is more sensitive to time delays and package

37
 CHAPTER 7. SIMULATION

losses, due to the short time headway for any possible maneuvers. The result from
(Alam et al., 2011) indicates a minimum relative distance of 1.2 m between two
identical HDVs can be obtained without endangering a collision. This is assumed
with no delays in the system and that the HDV can achieve maximum brake ca-
pacity instantaneously. A brief calculation of a time delay of 500 ms shows that the
intermediate distance only has to be increased to 2 m to avoid collision. Although
this project does not consider any disturbances, the risks must be considered for
real life appliciation
Normalized input energy consumption [%]

First vehicle, 40t

101

100

99

98

97

96

95 Fully centralized LQT

Centralized LQR with CC
94
0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
tau [s]
Normalized input energy consumption [%]

Second vehicle, 40t

90
89
88
87
86
85
84
Fully centralized LQT
83 Centralized LQR with CC

0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

tau [s]
Normalized input energy consumption [%]

Third vehicle, 40t

88

86

84

82
Fully centralized LQT
Centralized LQR with CC
80
0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
tau [s]

Figure 7.1: Energy saved for each vehicle separately compared to a single vehicle
with the CC driving on a flat road with vehicle changes.

Figures 7.1 and 7.2 show the results from the simulations utilizing two different LQ
controls, normalized against a single vehicle driving with the CC on the same road
profile. Figure 7.1 shows the consumed energy for each vehicle in the platoon. It
can be seen that the first vehicle with the LQR only reduces a small amount of the
energy when the following vehicle is relatively close. In this case, only when the
time headway is set to τ ≤ 0.75 s, thus allowing the following vehicle be close enough
to enable a small air drag reduction for the lead vehicle. The saving potentials are
vast for the second and third vehicle for all simulated time headways. Figure 7.2
shows the results as a whole platoon compared to a single vehicle using the CC,
by comparing consumed energy, time difference, and average velocity. The time

38
7.2. DIFFERENT TIME HEADWAYS

Normalized input energy consumption [%]

Average of the platoon

93
92
91
90
89
88
87
Fully centralized LQT
86 Centralized LQR with CC

0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

tau [s]

Time difference

100.5
Travelling time [%]

100

Fully centralized LQT

Centralized LQR with CC
99.5
0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
tau [s]

Average velocity
80.4

80.2
Avg velocity [km/h]

80

79.8

79.6 Fully centralized LQT

Centralized LQR with CC
Single vehicle with CC
79.4
0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
tau [s]

Figure 7.2: The average consumed energy, time, and velocity differences compared
to a single vehicle with the CC driving on a flat road with velocity changes.

Table 7.1: Summary of figs. 7.1 and 7.2, comparing different controllers with a single
vehicle with the CC on a flat road with velocity changes.
LQR control LQT control
Tau, τ 0.25 s 0.5 s 0.75 s 1s 0.25 s 0.5 s 0.75 s 1s
Avg velocity
79.92 79.92 79.91 79.91 79.86 79.89 79.90 79.89
[km/h]
Time saved [%] 0.00 0.00 0.00 0.00 -0.06 -0.02 -0.02 -0.02
Vehicle 1
1.78 0.74 0.01 0.00 5.31 4.31 3.50 3.34
energy saved [%]
Vehicle 2
14.33 13.77 12.42 10.54 16.82 16.32 14.88 13.74
energy saved [%]
Vehicle 3
16.52 15.31 11.87 11.89 19.46 18.53 17.36 17.57
energy saved [%]
Avg energy
10.87 9.94 8.10 7.48 13.86 13.04 11.91 11.55
saved [%]

39
 CHAPTER 7. SIMULATION

and average velocity do not differ noticeably, however the obtained energy saving
potentials are large.

Table 7.1 summarizes both the results from figs. 7.1 and 7.2. The analysis shows
that the LQT controller is more energy efficient for all vehicles with barely any time
or velocity differences. The main reason lies in the cruise control, it is not optimized
to be an energy efficient function. The CC only maintains the set vehicle velocity
and has a slight overshooting behavior, which cannot be considered energy efficient.
The LQR control does not control the first vehicle, which uses the CC to maintain a
velocity. The following vehicles are imitating the behavior of the lead vehicle, hence
the whole platoon suffers the same loss from the CC behavior.

Velocity, avg speed is 79.9203km/h

86

84

82

80
Velocity [km/h]

78

76

74

72 Vehicle 1, 40t
Vehicle 2, 40t
70 Vehicle 3, 40t

0 20 40 60 80 100 120 140 160 180 200

Time [s]

Relative distance, with tau = 0.25s

5.8
Relative distance [m]

5.6

5.4

5.2

4.8 Between vehicle 1 and 2

Between vehicle 2 and 3

0 20 40 60 80 100 120 140 160 180 200

Time [s]

Figure 7.3: Three-vehicle platoon utilizing LQR control with CC on a flat road,
with time headway τ = 0.25 s. The velocity and relative distance are plotted.

The performance and behavior of the platoon are also important to analyze. The
velocity and relative distance profile utilizing the LQR control on a flat road are
depicted in fig. 7.3. Figure 7.4 illustrates the similar, however with the LQT con-
trol. Both controllers display a relatively good following behavior where the platoon
hardly deviates from the set time headway, in this case τ = 0.25 s. As earlier men-
tioned, the CC is not optimized to be an energy efficient function and this can be
seen in fig. 7.3, where it produces an overshoot during brake and acceleration. This
can be compared to fig. 7.4 where the tracking control is utilized, it has smooth

40
7.2. DIFFERENT TIME HEADWAYS

Velocity, avg speed is 79.8634km/h

85

80
Velocity [km/h]

75

Vehicle 1, 40t
Vehicle 2, 40t
70 Vehicle 3, 40t

0 20 40 60 80 100 120 140 160 180 200

Time [s]

Relative distance, with tau = 0.25s

5.8
Relative distance [m]

5.6

5.4

5.2

4.8 Between vehicle 1 and 2

Between vehicle 2 and 3

0 20 40 60 80 100 120 140 160 180 200

Time [s]

Figure 7.4: Three-vehicle platoon utilizing LQT control on a flat road, with time
headway τ = 0.25 s. The velocity and relative distance are plotted.

transitions between the velocity changes without any overshoots. The drawback
with the smooth transition is the slightly lower deceleration when braking.

The notch at t = 108 s in figs. 7.3 and 7.4 represents the vehicle gear shifting. Al-
though the gear shifting does not seem to cause issues for the two displayed figures,
it can actually have an impact on the platooning behavior. Figure 7.5 illustrates the
same road profile with the LQR controller, however the time headway is increased
to τ = 1 s. This clearly shows that gear shifting is delayed for the following vehi-
cles, which in turn causes a wider intermediate distance. The following vehicles are
therefore accelerating longer in order to compensate for the distance loss. However,
a platoon with the LQT control responds differently, which is depicted in fig. 7.6.
Instead of having the following vehicles drive in a higher velocity, the lead vehicle
accelerates slower in order to compensate for the increased distance gap. This can
be noticed in the saved energy for the third vehicle in table 7.1, where the saving
potential is dropped by a lot with a higher time headway. This ultimately shows the
main difference between considering and not considering the following vehicles in
a platoon. With information from the following vehicles and vehicles ahead allows
the platoon to act more as one unit.

41
 CHAPTER 7. SIMULATION

Velocity, avg speed is 79.914km/h

88

86

84

82
Velocity [km/h]

80

78

76

74

72 Vehicle 1, 40t
Vehicle 2, 40t
70 Vehicle 3, 40t

0 20 40 60 80 100 120 140 160 180 200

Time [s]

Relative distance, with tau = 1s

26

25
Relative distance [m]

24

23

22

21

20 Between vehicle 1 and 2

Between vehicle 2 and 3

0 20 40 60 80 100 120 140 160 180 200

Time [s]

Figure 7.5: Three-vehicle platoon utilizing LQR control with CC on a flat road,
with time headway τ = 1 s. The velocity and relative distance are plotted.

Velocity, avg speed is 79.8945km/h

85

80
Velocity [km/h]

75

Vehicle 1, 40t
Vehicle 2, 40t
70 Vehicle 3, 40t

0 20 40 60 80 100 120 140 160 180 200

Time [s]

Relative distance, with tau = 1s

24

23.5

23
Relative distance [m]

22.5

22

21.5

21

20.5

20
Between vehicle 1 and 2
19.5 Between vehicle 2 and 3

0 20 40 60 80 100 120 140 160 180 200

Time [s]

Figure 7.6: Three-vehicle platoon utilizing LQT control on a flat road, with time
headway τ = 1 s. The velocity and relative distance are plotted.

42
7.3. CHANGE IN MASS DISTRIBUTION

7.3 Change in mass distribution

Different vehicle weights have different dynamics, therefore, it is necessary to analyze
the platoon behavior with mass variations. Although this is more difficult to analyze
numerically, it is instead simulated with two different weights on the second vehicle
in the platoon. Simulations of a lighter middle HDV with the weight of 30 tonnes
and a heavier HDV with 50 tonnes are done.

7.3.1 40-30-40 tonnes platoon

A platoon, consisting of a 40 tonnes lead vehicle followed by a 30 tonnes HDV and
lastly a 40 tonnes HDV as the tail vehicle, is simulated with both the LQ controllers.
The time headway is set to τ = 0.25 s. This is done to analyze how mass differences
in the platoon affect the platoon behavior.

Velocity, avg speed is 79.9203km/h

86

84

82

80
Velocity [km/h]

78

76

74

72 Vehicle 1, 40t
Vehicle 2, 30t
70 Vehicle 3, 40t

0 20 40 60 80 100 120 140 160 180 200

Time [s]

Relative distance, with tau = 0.25s

5.8
Relative distance [m]

5.6

5.4

5.2

4.8 Between vehicle 1 and 2

Between vehicle 2 and 3

0 20 40 60 80 100 120 140 160 180 200

Time [s]

Figure 7.7: A 40-30-40 tonnes platoon with LQR control traveling on a flat road.
The velocity and relative distance are plotted.

Figures 7.7 and 7.8 illustrate the velocity and relative distance behaviors for both
LQ controllers. The figures do not show any irregular behavior, which was expected.
A lighter vehicle with the same truck properties as a 40 tonnes HDV should not
have problems maintaining the velocity and the intermediate distance.

43
 CHAPTER 7. SIMULATION

Velocity, avg speed is 79.8684km/h

85

80
Velocity [km/h]

75

Vehicle 1, 40t
Vehicle 2, 30t
70 Vehicle 3, 40t

0 20 40 60 80 100 120 140 160 180 200

Time [s]

Relative distance, with tau = 0.25s

5.8
Relative distance [m]

5.6

5.4

5.2

4.8 Between vehicle 1 and 2

Between vehicle 2 and 3

0 20 40 60 80 100 120 140 160 180 200

Time [s]

Figure 7.8: A 40-30-40 tonnes platoon with LQT control traveling on a flat road.
The velocity and relative distance are plotted.

7.3.2 40-50-40 tonnes platoon

The mass of the second vehicle of the platoon has now been increased to 50 tonnes,
and the time headway still remains at τ = 0.25 s. A heavier HDV in the middle of
the platoon should establish velocity issues, unless the engine is powerful enough to
maintain the same velocity as the preceding vehicle.

The simulation results are depicted in fig. 7.9 for the LQR control and fig. 7.10
for the LQT control. The heavier HDV cannot keep up with the lighter HDV in
front during acceleration, and therefore the distance gap widens. To compensate
the distance loss, the heavier HDV accelerates longer to obtain a higher velocity.
In this case, a velocity over the allowed limit for HDVs in Sweden, i.e. 90 km/h.
The LQ control cannot take the speed limit into account, and is therefore possible
for the vehicle to accelerate past the speed limit. A solution to solve this issue is
to have a parallel controller calculating what speed the HDV obtains. If the HDV
is past the speed limit, the parallel controller would then cut off the torque input,
but this is not taken into account in this project. However, the platoon with the
LQT controller acts differently. The first vehicle accelerates slower, as in section 7.2,
which can be noted in the additional notch in the velocity in fig. 7.10.

44
7.3. CHANGE IN MASS DISTRIBUTION

Velocity, avg speed is 79.9196km/h

90

88

86

84
Velocity [km/h]

82

80

78

76

74
Vehicle 1, 40t
72
Vehicle 2, 50t
70 Vehicle 3, 40t

0 20 40 60 80 100 120 140 160 180 200

Time [s]

Relative distance, with tau = 0.25s

14

13

12
Relative distance [m]

11

10

6
Between vehicle 1 and 2
5 Between vehicle 2 and 3

0 20 40 60 80 100 120 140 160 180 200

Time [s]

Figure 7.9: A 40-50-40 tonnes platoon with LQR control traveling on a flat road.
The velocity and relative distance are plotted.

Velocity, avg speed is 79.7433km/h

85

80
Velocity [km/h]

75

Vehicle 1, 40t
Vehicle 2, 50t
70 Vehicle 3, 40t

0 20 40 60 80 100 120 140 160 180 200

Time [s]

Relative distance, with tau = 0.25s

5.8
Relative distance [m]

5.6

5.4

5.2

4.8
Between vehicle 1 and 2
Between vehicle 2 and 3
4.6
0 20 40 60 80 100 120 140 160 180 200
Time [s]

Figure 7.10: A 40-50-40 tonnes platoon with LQT control traveling on a flat road.
The velocity and relative distance are plotted.

45
 CHAPTER 7. SIMULATION

7.4 Simulation on a measured road profile

There are rarely any long flat roads in Sweden. Therefore, simulations on a real
road profile is done to see how the topography affects the platoon behavior. The
highway from Södertälje to Norrköping is approximately 120 km long, which will be
used in the simulation. Identical HDVs are used, with a mass of 40 tonnes each.

7.4.1 Södertälje to Norrköping

Figures 7.11 and 7.12 show the results of consumed energy and average velocity
against the set time headway for both LQ controllers. The benchmark is a single
HDV driving the same road with the CC. Table 7.2 summarizes the results from
both figures.
Normalized input energy consumption [%]

First vehicle, 40t

101

100

99

98

Fully centralized LQT

97
Centralized LQR with CC

0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

tau [s]
Normalized input energy consumption [%]

Second vehicle, 40t

93

92

91

90

89
Fully centralized LQT
88 Centralized LQR with CC

0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

tau [s]
Normalized input energy consumption [%]

Third vehicle, 40t

91

90

89

88

87 Fully centralized LQT

Centralized LQR with CC
86
0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
tau [s]

Figure 7.11: Energy saved for each vehicle separately compared to a single vehicle
with CC driving on the highway from Södertälje to Norrköping.

The energy reduction for each vehicle is large, but not as large as with the pre-
vious simulations on a flat road. The gap of saving potentials between the LQR
and LQT control has decreased. The LQT control is still more energy efficient,
with approximately 1 percentage per vehicle compared to 3 percentage on the flat
road. The explanation is due to the frequent topography change, which causes over-
shooting behaviors for both controls. Since the LQR control is already overshooting
during velocity change, the topography variations may not cause a large impact as

46
7.4. SIMULATION ON A MEASURED ROAD PROFILE

on the LQT control. Otherwise, both platoons present good vehicle following behav-
ior when the platoon consists of identical vehicles, indicated in figs. 7.13 and 7.14.
Normalized input energy consumption [%]

Average of the platoon

95

94

93

92

91 Fully centralized LQT

Centralized LQR with CC
90
0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
tau [s]

Time difference

100.4
Travelling time [%]

100.2

100

99.8

99.6 Fully centralized LQT

Centralized LQR with CC
99.4
0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
tau [s]

Average velocity
80.4

80.2
Avg velocity [km/h]

80

79.8

79.6 Fully centralized LQT

Centralized LQR with CC
Single vehicle with CC
79.4
0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
tau [s]

Figure 7.12: The average consumed energy, time, and velocity differences compared
to a single vehicle with CC driving on the highway from Södertälje to Norrköping.

Table 7.2: Summary of figs. 7.11 and 7.12, comparing different controllers with a
single vehicle with CC on the highway from Södertälje to Norrköping
LQR control LQT control
Tau, τ 0.25 s 0.5 s 0.75 s 1s 0.25 s 0.5 s 0.75 s 1s
Avg velocity
79.86 79.86 79.86 79.86 79.94 79.95 79.95 79.95
[km/h]
Time saved [%] 0.00 0.00 0.00 0.00 0.10 0.11 0.11 0.12
Vehicle 1
1.40 0.58 -0.00 -0.00 2.56 1.92 1.27 1.24
energy saved [%]
Vehicle 2
10.58 9.72 8.48 7.49 11.37 10.77 9.47 8.43
energy saved [%]
Vehicle 3
11.99 10.80 9.90 9.42 13.11 12.07 11.08 10.51
energy saved [%]
Avg energy
7.99 7.03 6.13 5.64 9.01 8.25 7.27 6.73
saved [%]

47
 CHAPTER 7. SIMULATION

Velocity, avg speed is 79.8702km/h

82
Velocity [km/h]

80

78

Vehicle 1, 40t
76
Vehicle 2, 40t
Vehicle 3, 40t

1000 1100 1200 1300 1400 1500 1600

Time [s]

Relative distance, with tau = 0.25s

6
Relative distance [m]

5.8

5.6

5.4

5.2

5 Between vehicle 1 and 2

4.8 Between vehicle 2 and 3

1000 1100 1200 1300 1400 1500 1600

Time [s]

Altitude

50
Altitude [m]

40

30

20

1000 1100 1200 1300 1400 1500 1600

Time [s]

Figure 7.13: A part of the highway between Södertälje and Norrköping. The platoon
utilizes the LQR control with CC. The velocity, relative distance and altitude is
plotted.

Velocity, avg speed is 79.9396km/h

82
Velocity [km/h]

80

78

Vehicle 1, 40t
76
Vehicle 2, 40t
Vehicle 3, 40t

1000 1100 1200 1300 1400 1500 1600

Time [s]

Relative distance, with tau = 0.25s

6
Relative distance [m]

5.8

5.6

5.4

5.2

5
Between vehicle 1 and 2
4.8 Between vehicle 2 and 3

1000 1100 1200 1300 1400 1500 1600

Time [s]

Altitude

50
Altitude [m]

40

30

20

1000 1100 1200 1300 1400 1500 1600

Time [s]

Figure 7.14: A part of the highway between Södertälje and Norrköping. The platoon
utlizes the LQT control. The velocity, relative distance and altitude is plotted.

48
Chapter 8

Expanding the platoon

8.1 Sub-platoons within the platoon

It is desirable to have a longer platoon consisting of more than three vehicles. There-
fore, a method is introduced to expand the centralized three-vehicle platoon to a
longer platoon consisting of several sub-platoons. The solution is to utilize the both
proposed controllers with a switcher between the LQR and the LQT control. To
avoid a rash following behavior after a gear shift from the LQR control, mentioned
at the end of section 7.2, a transition phase is introduced when switching between
the controllers. This will also limit the velocity of the following sub-platoon to the
set velocity reference.

The idea is illustrated in fig. 8.1. The sub-platoons utilize the LQT control when
it is far from the sub-platoon ahead, thus traveling independently. When another
sub-platoon is within range, it changes to a transition phase where the following sub-
platoon slowly decreases the velocity and adapting the velocity to the sub-platoon
ahead. This is done by ramping down the velocity reference signal of the LQT con-
trol. Lastly, when the sub-platoon is close enough, it switches to the LQR control
and follows the tail vehicle of the sub-platoon ahead. This allows arbitrary long pla-
toons, where the platoon is decentralized but consists of centralized sub-platoons.

Figure 8.1: The general idea of a transition phase between the switchings of two
LQ controls.

49
 CHAPTER 8. EXPANDING THE PLATOON

8.2 Simulation verification

To verify the switching control, a simple simulation on a flat road is done. The
lead sub-platoon starts at 70 km/h and is 300 m ahead of the following sub-platoon
that starts at 85 km/h. When both sub-platoons have merged, the lead sub-platoon
accelerates to 80 km/h. Both sub-platoons consist of identical vehicles of 40 tonnes
each and the time headway used for both sub-platoons is τ = 0.25 s. This is to
analyze if the both sub-platoons can act together as one unit even when two locally
centralized control are utilized.

2 subplatoons merging into 1 platoon

86

84

82

80
Velocity [km/h]

78

76

74 Vehicle 1
Vehicle 2
Vehicle 3
72 Vehicle 4
Vehicle 5
70 Vehicle 6

0 20 40 60 80 100 120 140 160 180 200

Time [s]

Relative distance, with tau = 0.25s

6.5
Between vehicle 1 & 2
Between vehicle 2 & 3
Between vehicle 3 & 4
Between vehicle 4 & 5
6 Between vehicle 5 & 6
Relative distance [m]

5.5

4.5
0 20 40 60 80 100 120 140 160 180 200
Time [s]

Figure 8.2: Velocity and relative distance of both sub-platoons when merging.

Figure 8.2 shows that both sub-platoons merged successfully and are acting as
one long platoon even with a velocity change. The reason for a bigger distance
gap at around t = 150 s is the gear shifting, which caused some acceleration loss.
Therefore, the following sub-platoon had to accelerate slightly longer to make up
for the distance loss. The velocity will not exceed the reference velocity set in the
LQT control. Altogether, the overall performance in merging and acting as one unit
is satisfying.

50
 Part IV

Discussion

51
Chapter 9

Summary

9.1 String stability study with mass

The term string stability has been known for decades. With the relation that
vi (s) = Gvi (s)vi−1 (s), the string stability can be defined in two different ways:

kGvi (s)k∞ ≤ 1

kGv2 (s)Gv3 (s)...GvN (s)k∞ ≤ 1.

The equations do not directly give information about the intermediate distances
between the vehicles, although collisions must be prevented in platooning. Hence,
a suggestion of also investigating the relative distance was proposed. The focus is
to analyze how mass variations affects the string stability.

To understand the overall behavior of platooning, a simple mass-spring system

was introduced: mi ẍi + bi ẋ = ui . Two transfer functions were derived, with the
assumption that every vehicle in the platoon utilize the same control, a standard
PID-control with respect to the difference in longitudinal position. The first func-
tion derived was the velocity transfer function Gvi (s), which did not give any further
insight into string stability with mass variations. The second function derived was
the relative distance transfer function Gdi (s) and by assuming three relations, an ap-
proximation was obtained. The approximated Gdi (s) holds the information on how
the platoon should be arranged depending on the weight of the vehicles, namely
in weight decreasing order with the heaviest vehicle first in the platoon. Simple
simulations also suggest the same weight arrangement.

The issue, with utilizing the same controller with different vehicle weight, is the
different dynamics each vehicle has depending on the mass. A fixed PID-control is
string stable for a set of lighter vehicles up to a certain limit, where it then is no
longer considered string stable. The control is rendered useless for that mass and
above. Therefore, a mass depending PID-control is suggested. Algebraic derivations
for both Ḡvi (s) and Ḡdi (s) and simulations are indicating string stable behaviors.

53
 CHAPTER 9. SUMMARY

9.2 Locally centralized three-vehicle platoon

There already exists a decentralized control operating on vehicles, but there is no
existing centralized control. Therefore, a study of a centralized control has been
done, to analyze its potential possibilities in platooning.

A platoon model consisting of three HDVs is derived, from the powertrain to the ex-
ternal longitudinal forces. The aerodynamic drag reduction in platooning has been
taken into concern in the forces acting on a HDV. The platoon model is non-linear,
hence the model is linearized around an equilibrium point to enable a state-space
equation.

Two LQ controls were proposed for centralized platooning. A LQR controller, where
only the two last vehicles are controlled while following the lead vehicle, which is
operated by the CC. A LQ tracking (LQT) controller with integral action was also
proposed, which is truly a fully centralized control. The main difference between
the two control laws is, that the lead vehicle does not take the following vehicles
into account with LQR control while the LQT control does.

Simulations imply that both LQ controllers operate properly on a flat road and
on the highway from Södertälje to Norrköping, with identical HDVs of 40 tonnes
each. Different time headways were used to analyze the saved energy with safety
aspects, i.e. relative distance. The results show that the reduced energy contra the
time headway is almost linear. To verify how the information of the vehicle behind
affects platooning, different masses were set on the second vehicle. The simulation
proved the LQT controller to be better, hence staying within the platoon is more
beneficial.

Lastly, it is desirable to expand the platoon to more than three vehicles. The
idea is to use a switching control between the LQR and the LQT controller. Utilize
the LQT control when the platoon drives independently from other platoons. When
a platoon is within range, let the reference velocity slowly drop to the same velocity
as the vehicle ahead. Finally, switch to the LQR control when the desired distance
is achieved. This way, several sub-platoons can merge into one larger platoon and
still retain the locally centralized control.

54
Chapter 10

Conclusion

This project was given two tasks concerning vehicle platooning. The first objective
was to study the impact mass has on the robustness of a platoon, i.e. string stability.
A vehicle platoon can be described as a mass-spring system and by introducing a
mass difference into the system, a damping effect is instated. All types of automatic
control can be described approximately as a PID-control. This turned out to be a
simple analysis for string stability. Intuitively, it feels logical to have a heavier ve-
hicle ahead. A heavier vehicle accelerates and decelerates slower and therefore, the
following vehicle can easily follow the heavy vehicle and achieve a shorter braking
distance. This is valid with the assumption that all the vehicle properties are iden-
tical besides the mass. Theoretical analysis and simulations confirm the same. A
proposed solution to avoid a fixed mass arrangement is to introduce a mass depen-
dent controller. This gives the possibility to obtain a string stable platoon regardless
of the vehicle mass.

Given a realistic non-linear model of a HDV, the second objective was to design
a centralized controller that addresses consumed energy and vehicle safety when
driving in a three-vehicle platoon. A LQR control was proposed but due to the
lack of a proper outer reference signal, a second control was introduced. A LQ
tracking control with integral action allows an outer reference signal that enables
the vehicles to follow. The difficult part was to tune the cost matrices such that a
desirable performance is achieved. The focus was on the safety, to hold the correct
relative distance between the vehicles. By holding a close intermediate distance,
the energy consumption for the second vehicle can be reduced with 8.4 − 11.4 %
depending on the set time gap and 10.5 − 13.1 % for the third vehicle when driving
80 km/h on a highway. This was simulated in Simulink utilizing a real road profile
between Södertälje and Norrköping in Sweden. There are benefits with the LQR
control despite the drawback of not having information of the following vehicles for
the lead vehicle. The LQR control enables the possibility to extend longer platoons
while keeping the locally centralized control properties. This is possible by introduc-
ing a switching control, which utilizes the LQT and LQR controls whenever feasible.

55
 CHAPTER 10. CONCLUSION

It is possible to obtain a smoother and comfort ride, but the reasoning was im-
mediately dropped. A driver is driving most comfortably with a distance of several
meters away on a highway. The introduction of a centralized control with a short
time headway, reduces the distance to nearly a tenth. This could create comfort
issues for the driver if a short time headway was utilized. The progress to a short
time headway should proceed slowly for the driver to adapt. Hence, the focus was
mainly to avoid vehicle collisions.

A platoon, with a LQ control, considers how each vehicle in the platoon drives
and can therefore drive as one unit. A problem occurs when merging several sub-
platoons. Each sub-platoons only consider their own platoon to drive as one unit,
hence there is no information on the following sub-platoon, exactly like the AiCC
today. This is a trade-off between quick calculations and the amount of vehicles in a
sub-platoon. The proposed solution with a switching control allows fast calculations
and enables the controller to remain locally centralized within the sub-platoon.

56
Chapter 11

Future work

There is room for further improvements with the proposed centralized controllers.
Although the shortest intermediate distance proposed is 5 m at 80 km/h and the fact
that a 500 ms time delay and the lowered air drag only reduce the relative distance
with 2 m during hard brake, it is appropriate to analyze how the LQ controller han-
dles time delays, noises, and package losses. Also, a switching control could cause a
sudden jump or unwanted behavior to the platoon and should therefore be analyzed
further.

A slight problem that occurred in this project was when the platoon was accel-
erating and gear shifting, a solution could be to broadcast the information that a
gear shift will occur. Other information that benefit the platoon behavior should
also be broadcasted. Since this project is limited to information of only velocity
and distance being broadcasted, it could be an area of research to investigate what
necessary information is needed for platooning.

It is possible to further decrease the consumed energy by introducing other fuel

saving functions such as downhill speed control, look-ahead, and so forth. This
could then be compared with a decentralized controller, such as the AiCC, and
investigate the potential use for the future.

57
Bibliography

Assad Al Alam. Optimally fuel efficient speed adaptation. Master’s thesis, Royal
Institute of Technology, KTH, March 2008.

Assad Al Alam, Ather Gattami, and Karl H. Johansson. An experimental study

on the fuel reduction potential of heavy duty vehicle platooning. In 2010 13th
International IEEE Conference on Intelligent Transportation Systems (ITSC),
September 2010. doi: 10.1109/ITSC.2010.5625054. ISSN: 2153-0009.

Assad Al Alam, Ather Gattami, Karl H. Johansson, and Claire J. Tomlin. Estab-
lishing safety for heavy duty vehicle platooning: A game theoretical approach. In
18th IFAC World Congress, August 2011.

Thomas D. Gillespie. Fundamentals of vehicle dynamics. SAE International, Febru-

ary 1992. ISBN: 1-56091-199-9.

Kevin P. Groves, David O. Sigthorsson, Andrea Serrani, Stephen Yurkovich,

Michael A. Bolender, and David B. Doman. Reference command tracking for a
linearized model of an air-breathing hypersonic vehicle. In Proceedings of AIAA
Guidance, Navigation, and Control Conference and Exhibit, AIAA-2005-6144,
August 2005.

Gustav Hammar and Vadim Ovtchinnikov. Structural intelligent platooning by a

systematic LQR algorithm. Master’s thesis, Royal Institute of Technology, KTH,
October 2010.

Maziar E. Khatir and Edward J. Davison. Decentralized control of a large platoon

of vehicles using non-identical controllers. In Proceedings of the 2004 American
Control Conference, 2004, June 2004. doi: 10.1109/ACC.2004.182526. ISSN:
0743-1619.

Scania CV AB. Annual report, 2001.

Michael Schittler. State-of-the-art and emerging truck engine technologies for opti-
mized performance, emissions and life cycle costs. In 9th Diesel Engine Emissions
Reduction Conference, Rhode Island, USA. U.S. Department of Energy, August
2003.

59
 BIBLIOGRAPHY

Yoshinori Yamamura and Yoji Seto. A study of string-stable ACC using vehicle-to-
vehicle communication. In 2006 SAE World Congress, 2006-01-0348, April 2006.
doi: 10.4271/2006-01-0348.

Diana Yanakiev and Ioannis Kanellakopoulos. Variable time headway for

string stability of automated heavy-duty vehicles. In Proceedings of the
34th IEEE Conference on Decision and Control, 1995, December 1995. doi:
10.1109/CDC.1995.479245.

Diana Yanakiev and Ioannis Kanellakopoulos. A simplified framework for string

stability analysis in AHS. In In Proceedings of the 13th IFAC World Congress,
1996.

Youping Zhang, Elias B. Kosmatopoulos, Petros A. Ioannou, and C.C. Chien. Us-
ing front and back information for tight vehicle following maneuvers. In IEEE
Transactions on Vehicular Technology, January 1999. doi: 10.1109/25.740110.
ISSN: 0018-9545.

60
Appendices

61
Appendix A

PID-controller setup

The following PID-controller setup was used in chapters 3 and 4.

Pi = 500 N/m, Ii = 3 N/ms, Di = 20000 Ns/m

with b = 172 Ns/m and time headway τd = 1 s. With these values, the velocity
transfer function is, which was obtained in eq. (3.6):

20000s2 + 500s + 3
Gvi (s) = (A.1)
mi s3 + 20672s2 + 503s + 3
For the scaled-varying PID-controller, the same controller was used but together
with a mass depending scalar of Ki = 10000mi
. This gives the following velocity
transfer function, from eq. (4.4):

20000s2 + 500s + 3
Ḡvi (s) = (A.2)
10000s3 + ( 1720000
mi + 20500)s2 + 503s + 3

The H∞ -norms of eqs. (A.1) and (A.2) with a vehicle mass of 40 tonnes are:

kGvi (s)k∞ = 1.0029

(A.3)
kḠvi (s)k∞ = 1

Gvi (s) is considered string unstable while Ḡvi (s) is string stable according to eq. (2.2).

The simulation was done in Matlab and the scenario is from standing still to ramp
from 0 to 10m/s in 10s and maintaining the speed on a flat road. Three different
mass distribution were used but the total mass of the platoon remained the same
to be able to make a comparison, which follows:

1. Identical vehicles, with mass of 40 tonnes each.

2. Increasing mass distribution, from 20 to 60 tonnes.

63
 APPENDIX A. PID-CONTROLLER SETUP

3. Decreasing mass distribution, from 60 to 20 tonnes.

The simulation was done twice with two different platoon lengths:

• Platoon length of 41 vehicles, i.e. a 1000kg weight gap between vehicles for
the case with non-identical vehicles.

• Platoon length of 81 vehicles, i.e. a 500kg weight gap between vehicles for the
case with non-identical vehicles.

The results of the simulations can be read in sections 3.4 and 4.2.

64
Appendix B

Mathematical derivation of LQR

The standard LQR cost function:

Z tf  
∗
J = min xT (τ )Qx(τ ) + uT (τ )Ru(τ ) dτ + xT (tf )Qf x(tf )
u 0 ( (B.1)
ẋ(t) = Ax(t) + Bu(t)
Subject to:
x(0) = x0

where tf is the time-horizon, Q ≥ 0 is state-cost matrix, Qf ≥ 0 is the final state-

cost matrix, and R > 0 is the input cost matrix. It is required that the matrix pair
(A, B) is controllable.

One solution for the LQR problem is with dynamic programming. For t ∈ [0, tf ], a
value function Vt is defined by:
Z tf  
Vt (z) = min xT (τ )Qx(τ ) + uT (τ )Ru(τ ) dτ + xT (tf )Qf x(tf )
u t ( (B.2)
ẋ = Ax + Bu
Subject to:
x(t) = z

where Vt (z) gives the minimum LQR cost-to-go, starting from state z at time t.
Cost-to-go with no time left is just the final cost, hence Vtf = z T Qf z and there-
fore V0 (x0 ) is the minimum total cost. The value function is quadratic, hence the
Lyapunov function is:
Vt (z) = z T Pt z (B.3)
where Pt ≥ 0 and is symmetric. Pt can be obtained from a differential equation
running backwards in time for t ∈ [0, tf ].

By assuming that u(t) = w, constant over a small time interval [t, t + h], the
incurred cost over the interval is:
Z t+h    
Jt,t+h (z) = x(τ )T Qx(τ ) + wT Rw dτ ≈ h z T Qz + wT Rw (B.4)
t

65
 APPENDIX B. MATHEMATICAL DERIVATION OF LQR

and the new landed state is: x(t + h) ≈ z + h(Az + Bw).

Min-cost-to-go from where it lands is approximately:

Vt+h (z) = (z + h(Az + Bw))T Pt+h (z + h(Az + Bw))

 
≈ (z + h(Az + Bw))T Pt + hṖt (z + h(Az + Bw)) (B.5)
 
≈ z T Pt z + h (Az + Bw)T Pt z + z T Pt (Az + Bw) + z T Ṗt z

where h2 and higher terms were dropped. The cost-incurred together with min-
cost-to-go is approximately:

Vt (z) ≈ Vt+h (z) + Jt,t+h (z) (B.6)

 
= z T Pt z + h z T Qz + wT Rw + (Az + Bw)T Pt z + z T Pt (Az + Bw) + z T Ṗt z .

By minimizing over w, the optimal input is obtained:

2hwT R + 2hz T Pt B = 0
(B.7)
⇒ wopt = −R−1 B T Pt z.

By substituting wopt from eq. (B.7) into eq. (B.6):

 
z T Pt z ≈ z T Pt z + h z T Qz + wT Rw + (Az + Bw)T Pt z + z T Pt (Az + Bw) + z T Ṗt z
 
⇒ 0 = h z T Qz + z T AT Pt z + z T Pt AT z − z T Pt BR−1 B T Pt z + z T Ṗt z
 
= hz T Q + AT Pt + Pt AT − Pt BR−1 B T Pt + Ṗt z.

This yields to the following Riccati differential equation:

−Ṗt = AT Pt + Pt AT − Pt BR−1 B T Pt + Q. (B.8)

The LQR problem can be solved using the final condition Ptf = Qf .

66
Appendix C

Mathematical derivation of LQT with

integral action

With the concept of utilizing an error as: ė1 = r−x1 , where r is a reference signal, an
integrating action is introduced into the states as: e1 = ė1 dt. The new state-space
R

equation follows as:

" # " #" # " # " #
ė1 0 −I e1 0 I
= + u+ r (C.1)
ẋ1 0 A1 x1 B1 0

⇒ ẋ = Ax + Bu + Gr.
The new system including the error as a state can be defined as:
" #
e1
e= = M r + Hx. (C.2)
r − x1

The LQT cost function can now be expressed as:

1 1
Z tf  
J= eT (τ )Qe(τ ) + uT (τ )Ru(τ ) dτ + eT (tf )Qf e(tf )
20 2
( (C.3)
ė(t) = HAx(t) + HBu(t) + HGr(t)
Subject to:
e(0) = e0

where ṙ = 0 is assumed. By inserting eq. (C.2) in eq. (C.3):

1 tf  T T
Z 
J= x H QHx + 2rT M T QHx + rT M T QM r + uT Ru dτ
2 0
1 T 
+ x (tf )H T Qf Hx(tf ) + 2rT (tf )M T Qf Hx(tf ) + rT (tf )M T Qf M r(tf )
2
( (C.4)
ẋ(t) = Ax(t) + Bu(t) + Gr(t)
Subject to:
x(0) = x0

67
 APPENDIX C. MATHEMATICAL DERIVATION OF LQT WITH INTEGRAL
ACTION

where tf is the time-horizon, Q ≥ 0 is state-cost matrix, Qf ≥ 0 is the final state-

cost matrix, and R > 0 is the input cost matrix. It is required that the matrix pair
(A, B) is controllable and H 6= 0.

One solution for the LQT problem is by utilizing Lagrange multiplier. By introduc-
ing the Lagrange multiplier function λ and relaxing the problem by forming:
Z tf
L=J+ λT (τ ) (Ax(τ ) + Bu(τ ) + Gr(τ ) − ẋ(τ )) dτ (C.5)
0

where J is from eq. (C.4). The optimal control that minimizes eq. (C.4) will be
obtained by solving the following three equations:

∇x(t) L = 0 (C.6a)
∇u(t) L = 0 (C.6b)
∇λ(t) L = 0. (C.6c)

Note that r is assumed to be constant. Also the state r can be seen as noise, where
it is an unknown value and therefore L is not optimized and hence has no partial
derivative with repsect to r.

To find a solution eq. (C.6a), the following is derived:

d
Z tf Z tf Z tf
λ (τ )x(τ ) dτ =
T
λ̇ (τ )x(τ ) dτ +
T
λT (τ )ẋ(τ ) dτ
dt 0 0 0

= λT (tf )x(tf ) − λT (0)x(0)

Z tf Z tf
⇒ λT (τ )ẋ(τ ) dτ = λT (tf )x(tf ) − λT (0)x(0) − λ̇T (τ )x(τ ) dτ. (C.7)
0 0

By applying eqs. (C.6) and (C.7) on eq. (C.5), the following is obtained:

∇x(t) L = 0 = H T QHx + H T QM r + AT λ + λ̇ (C.8a)

−1
∇u(t) L = 0 = Ru + B λ T
⇒ u = −R T
B λ (C.8b)
∇λ(t) L = 0 = Ax + Bu + Gr − ẋ. (C.8c)

Equation (C.8c) is the dynamical system, which must always be satisfied. From
∇x(tf ) L, the following final condition is obtained:

∇x(tf ) L = H T Qf Hx(tf ) + H T Qf M r(tf ) − λ(tf ) = 0

(C.9)
⇒ λ(tf ) = H T Qf Hx(tf ) + H T Qf M r(tf ).

68
By adding the Lagrange multiplier as a state, from eq. (C.8a), to the dynamic
linear system and replacing u described in eq. (C.8b), the new system with the final
condition from eq. (C.9) can be described as:
" # " #" # " #
ẋ A −BR−1 B T x G
= + r
λ̇ −H T QH −AT λ −H T QM
(C.10)
x(0) = x0
(
Subject to:
λ(tf ) = H T Qf Hx(tf ) + H T Qf M r(tf ).

The new conditions are known as a two-point boundary value problem. Similar
with the case of LQR problem, the solution can be calculated backwards, from the
final condition to initial condition. The Lagrange multiplier solution can be given
asserted as follow:
λ = Pt x + gt r. (C.11)
By inserting this into the second state in eq. (C.10):

λ̇ = Ṗt x + Pt ẋ + g˙t r = −H T QHx − AT λ − H T QM r

whereas ẋ is known and gives the following equation:

Ṗt x + Pt Ax − Pt BR−1 B T Pt x − Pt BR−1 B T gt r + Pt Gr + ġt r =

= −H T QHx − AT Pt x − AT gt r − H T QM r.

Given that x 6= 0 and r 6= 0, this yields to two differential equations:

Ṗt = −Pt AT − AT Pt − H T QH + Pt BR−1 B T Pt (C.12a)

−1
ġt = (Pt BR B − A )gt − (H QM + Pt G)
T T T
(C.12b)

that must be satisfied. Equation (C.12a) is the standard Riccati differential equation
and eq. (C.12b) is an auxiliary vector equation defining the feedforward gain. The
LQT problem can be solved using the final conditions:

Ptf = H T Qf H
gtf = H T Qf M.

69