You are on page 1of 64

Optimally Fuel Efficient Speed Adaptation

ASSAD AL ALAM

Masters’ Degree Project


Stockholm, Sweden March 2008

XR-EE-RT 2008:002
Optimally Fuel Efficient Speed Adaptation

ASSAD AL ALAM

Master’s Thesis at Automatic Control


Supervisor: Per Sahlholm
Examiner: Karl H. Johansson

TRITA xxx yyyy-nn


iii

Abstract

An optimal velocity trajectory for a heavy duty vehicle, obtained with the aid of
modern GPS and digital map devices, depends on several variables. Curvature
speed limitations, road grade, and posted road speed are common constraints
imposed by the road travelled. This thesis presents a method for modelling
and analysing a switching controller through the use of the former mentioned
constraints. A non-linear model for the heavy duty vehicle is derived, enabling
suitable control methods to be applied. Pontryagin’s Principal and LQR are
discussed to get a profound understanding of how the controller should be de-
signed. It is discovered that a switching controller based on optimal control and
engineering experience is most favourable for the problem at hand. The con-
troller is designed to address the main objectives set in this paper of minimising
fuel consumption, travelling time, and brake wear.
Gauss-Newtons’s algorithm for non-linear equations is used to estimate
curve radii. Other input parameters are presumed to be available. GPS data
error is discussed to perform a sensitivity analysis. An electronic horizon is pro-
duced on three road segments, entailed with data of the future road topology.
Finally the switching controller is applied to the road segments. Experimental
results show that the controller produces a velocity trajectory, which reduces
fuel consumption by 5-15% and brake wear by 15-35%, while the travelling
time is only increased by 1-2%.
iv

Acknowledgements
The work described in this master’s thesis has been conducted at the System Pre-
development Department, REP, at Scania CV AB in Södertälje, Sweden. It was
supervised by the Automatic Control Systems department, at the Royal Instititute
of Technology (KTH), in Stockholm. I would first and foremost like to thank my
supervisor Per Sahlholm at Scania, for all his help and guidance during this Master’s
Thesis. His input have been invaluable and inspiring. A deep gratitude is extended
to Jon Andersson for his significant input to this project. Along with Per Sahlholm
and Jon Andersson, other department members at REP namely Joseph Ah-King,
Rickard Lyberger, Håkan Gustavsson, Daniel Thuresson, Erik Persson and the se-
nior manager Nils-Gunnar Vågstedt, have been most helpful and supportive. I
would therefore like to extend my gratitude towards them as well. Finally I would
like to thank my supervisor Karl H. Johansson at KTH for his guidance, creative
input, time and support.
Contents

Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iv

Contents v

1 Introduction 1

2 Background 3
2.1 Look-Ahead . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
2.1.1 Premise . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
2.1.2 Limitations . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
2.1.3 Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2.2 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.3 Thesis Outline & Objective . . . . . . . . . . . . . . . . . . . . . . . 5

3 Vehicle Modelling 7
3.1 Powertrain . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
3.2 Longitudinal Forces . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
3.3 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12

4 Optimal Control 13
4.1 Pontryagin’s Principal . . . . . . . . . . . . . . . . . . . . . . . . . . 14
4.2 LQR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
4.3 Optimal Solution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20

5 Switching Controller 23
5.1 Road Speed . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
5.1.1 Coasting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
5.1.2 Fuel Cut-Off Point . . . . . . . . . . . . . . . . . . . . . . . . 25
5.1.3 Maximum Roll-Off . . . . . . . . . . . . . . . . . . . . . . . . 26
5.1.4 Most Favourable Braking . . . . . . . . . . . . . . . . . . . . 27
5.1.5 Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . 28
5.2 Curvature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
5.2.1 Curve-Radius Estimation . . . . . . . . . . . . . . . . . . . . 30
5.2.2 Maximum Curve Speed . . . . . . . . . . . . . . . . . . . . . 32
5.2.3 Method Strategy . . . . . . . . . . . . . . . . . . . . . . . . . 33

v
vi Contents

6 Analysis 39
6.1 Simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
6.1.1 Test-Road . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
6.1.2 Road 225 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
6.2 Sensitivity & Robustness Analysis . . . . . . . . . . . . . . . . . . . 46
6.2.1 GPS-Error . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
6.2.2 Drift . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
6.2.3 Grade Error . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48

7 Conclusion 51

8 Future Work & Extensions 53

Bibliography 55

A Appendix - LQR Mathematical Derivation 57


Chapter 1

Introduction

Fascinating, inspiring and at times petrifying could be said about the development
of the technology today. Day by day new breakthroughs are made, new inventions
are pursued, and new intriguing applications of old technology and methodology
are found.
A relatively new technology called the Global Positioning System (GPS) has
been the interest of many fields. The GPS is becoming increasingly relevant in
road-navigation systems. Traffic is becoming intense and more complex throughout
the world, making it increasingly difficult for the driver to focus on all the relevant
aspects of driving. A lot of research is therefore being conducted regarding the use,
reliability and accuracy of the information a GPS can provide. Future navigation
systems may not only guide the vehicle along the best route, but also direct it in a
cost-efficient manner.
By enriching the information to the driver and the vehicle with e.g. future
curvature information and traffic signs could most probably improve the driver
behaviour significantly. Hence implementing control systems as a next step by using
the GPS information and advanced digital maps as input could not only reduce
costs, but also increase safety and improve on environmental aspects. Emission
regulations are becoming more stringent making the preceding field of research a
top priority in the modern vehicle industry.

1
Chapter 2

Background

2.1 Look-Ahead
Look-Ahead could be described as a process of acquiring ’Preview Information’.
It is a concept for providing control strategies with information about future dis-
turbances and inputs. The information can be properties such as road topology,
curvature, and speed regulations, which are important input signals when designing
a speed controller. Look-Ahead provides information based on GPS and digital
maps data, which is then processed intelligently. The information is subsequently
conveyed to the driver or the vehicle, thus enhancing the perception of future dis-
turbances and enabling suitable future actions by extending the drivers information
input-horizon. It is mainly utilized in Advanced Driver Assistance Systems (ADAS).

2.1.1 Premise
It is assumed that the driver selects a drive mission, i.e. a given route. Information
regarding the topology, curvature and legal speed limitations is acquired through
the Look-Ahead system. Accounting for the fact that the system is a heavy truck,
it is possible to derive a mathematical model for the system and use the available
electronic information horizon to calculate a proper and possibly optimal velocity
trajectory.

2.1.2 Limitations
If the information provided by the GPS is accurate to a certain extent, proper esti-
mations such as curvature, grade, speed-limits, etc. can be made with a justifiable
reliability. However, physical limitations such as the time required to receive signals
from the satellite, signal distortion, delays in the actuator data processing subsys-
tems, amongst others make room for significant possibility of errors. The loss of
reliable information could lead to dangerous situations such as the heavy vehicle
not reducing the speed in time.

3
4 CHAPTER 2. BACKGROUND

Therefore this thesis is set to be a primary study, resulting in simulations and


not physical implementation on an actual vehicle. The simulations and results
are limited to smaller winding roads. Solely road topology constraints are focused
upon, i.e surrounding traffic and its behaviour is not accounted for in the objectives.
The effects of other vehicles in the traffic are discussed briefly, but an adaptive
methodology is nonetheless considered to be out of the scope of this thesis. Road
grade and posted road speeds are presumed to be given. Conducting research and
designing methods for measuring such data inputs would require more time than
what is allowed for this thesis.

2.1.3 Implementation
To implement a designed controller, the road topography information is obtained
by the combination of an on-board database with altitude information and a global
positioning unit (GPS). It is assumed that road information is available and the
current route is predicted or supplied by the driver as stated in 2.1.1. By creating
a model from existing parameters based upon the heavy vehicle’s characteristics,
a prediction of vehicle motion and energy consumption as a function of control
signals and known disturbances can be made. Road topology information, desired
constraints on comfortable lateral acceleration, and maximum deceleration limits,
are presumed to be available control inputs. The road grade is assumed available
through various mathematical tools. Finally the conventional cruise controller is
fed with set points from an optimisation algorithm.

Global position received


from a GPS unit

Laptop calculating optimal trajectories and


feeding the vehicle with set speeds

Road slope stored


in on−board database

Truck receiving set speeds and reporting


current velocity, gear and estimated mass

Figure 2.1. Information flow in the implementation procedure.[9]


2.2. RELATED WORK 5

A GPS unit connected to a laptop receives the global position as shown in Figure 2.1.
The position is then matched to the road database to obtain slope information about
the road ahead. The computer subsequently calculates a most favourable speed
trajectory for a desired electronic horizon ahead through an appropriate method.

2.2 Related Work


Using GPS positioning and digital maps have been proposed in various works. The
problem of how to drive a heavy truck over various road topographies such that
the fuel consumption is minimised has been addressed by Fröberg et. al., 2006 [4],
amongst others. An optimisation problem is formulated, which has the advantage
of enabling explicit analytical solutions. Another presented approach is developing
a predictive cruise controller for a heavy truck, using dynamic programming to
numerically solve the optimal control problem. Thereby, several strategies are found
that might be appropriate for various types of road segments. The results presented
in the article show that there is a potential to reduce fuel consumption by utilising
information about the topology ahead.
However, fuel consumption is not the only beneficial factor. Research has been
conducted to create advanced driver speed assistance in curves. One article pre-
sented by Aguilera et. al., 2007 [7]; discusses driver interaction modes, i.e. ways in
which a driver assistance may interact with the driver. The method uses a vehicle
infrastructure driver speed profile, which gathers information from the electronic
horizon, the driver behaviour, and the vehicle dynamics. A risk function is con-
structed to determine what action is suitable to convey to the driver. The risk
function is based on the maximum possible speed in a curve, i.e. a curve speed
warning system (CSWS) of sorts.
Another article presented by the Mitsubishi Motors Corporation in Japan [5],
investigates how cornering can create discomfort and, under extreme circumstances,
result in loss of vehicle control. It is believed that conventional implementations of
the active safety approach, e.g. antilock braking and traction-control systems can
be improved by complementing such control systems with strategies that perform
vehicle control prior to entering the turn. Therefore, an investigation is conducted in
the article involving empirical studies regarding location of brake activation before
various curves and mathematically based maximum speed limits based on the curve
radius and the preferred lateral acceleration.
The research presented above is however not based on deriving an optimal con-
trol strategy for speed reductions to posted speed limits and curvatures.

2.3 Thesis Outline & Objective


A lot of research within the subject matter have already been done. A model pre-
dictive control strategy has previously been designed [10]. However, the constraint
of the speed limits inflicted from the legal road speed limitation and curvature were
6 CHAPTER 2. BACKGROUND

never accounted for. A curve speed assistant prototype with similar goals but a fo-
cus on hardware implementation has previously been built at Scania. Information
regarding topology were taken into account when designing a controller which gath-
ered information about curvature and road speed limitations through a software
tool known as Advanced Driver Assistance System Research Platform (ADASRP).
The controller was based upon the truck model and inversely calculated a fuel opti-
mal velocity trajectory. However, the physical limitations discussed in 2.1.2 caused
certain problems which the controller could not handle. In addition the road slope
data was used as an input to obtain a reasonable fuel cut-off point for slowing down,
but it was not sufficient in certain situations. The assumption that the heavy vehi-
cle always had the possibility to reduce its speed in time through coasting was not
always valid. In the situation of facing a long and steep downhill, the system would
reach its lower speed set point, rendering a non-existing solution. To obtain the
desired speed at the foot of the hill, the truck had to reduce its speed significantly.
This resulted in an unacceptably early control input and an unreasonable vehicle
road behaviour. Thus it was discovered that in some cases using the brakes can be
justified, despite the effect on fuel economy [12].
Hence, new relevant parameters such as driver comfort, time optimality, road
dynamics and driver behaviour must be highlighted. Also, the effects of the physical
limitations have to be accounted for. Most notably were the delays in the processing
and actuation parts, which consequently have to be addressed. Thus the objective
of the thesis will be to account for and evaluate these newly emphasised parameters.
Previous experience have also proven that it is very important to imply a failsafe
system of sorts in the case of estimation error. It has to be noted however that these
precautions will be applied according to the complexity of the case. Too complex
situations will not be dealt with in great depth due to the time constraint issued
upon this research. The main focus will lie on designing a controller which allows
for travelling an assigned route in a comfortable, fuel-, time-, and brake-efficient
manner.
The thesis is hence structured as follows: in chapter three a description of the
vehicle modelling is presented, which serves as a premise for all the forthcoming
calculations. Thereafter a study of optimal control approaches are undertaken.
However, the analytical control inputs derived from the optimal control strategies
are found to increase in complexity as additional constraints are implemented within
the model. Further the optimal control presented in this case is based on a linear
model and cannot account for the non-linear behaviour that arises primarily from
the road grade and gear shifting. The varying road grade and gear shifting have a
significant impact on the vehicle behaviour and cannot be neglected. Therefore a
new non-linear switching controller is designed based upon optimal control theory
and engineering experience, which is discussed in chapter five. Chapter six presents
results and sensitivity analysis based on the controller’s performance. Finally the
limitations and strengths of the controller are discussed in chapters seven and eight,
along with the future possibilities of the presented topic.
Chapter 3

Vehicle Modelling

The main parts of a heavy duty vehicle (HDV) consist of engine, clutch, transmission
shafts and wheels. The combination of all these parts creates the driveline or also
known as the powertrain. The powertrain is a fundamental part when evaluating
the dynamics of the vehicle. It can be modelled in various ways depending on the
purpose and use. The main interest of this study is to create a discrete model of
the powertrain, based upon the simple model depicted in Figure 3.1, which will be
used as the base for the controller design.

3.1 Powertrain
General powertrain modelling can be found in Vehicular Systems [3]. The main
parts of interest within this study of the powertrain are depicted in Figure 3.1.

Figure 3.1. A basic model of the powertrain. The engine utilized in this model is a
diesel engine.

Engine: The engine produces a torque through combustion of diesel mixed with
a surplus of air in a very high pressurised chamber. The highly explosive combustion

7
8 CHAPTER 3. VEHICLE MODELLING

drives the crank shafts, which in turn are connected to the clutch by a shaft, causing
a desired torque. The output torque from the engine is characterised by the torque
resulting from the combustion, the internal friction from the chamber walls, and the
external torque from the clutch. Thus if the inertia is obtained, Newton’s second
law gives:

Je ω̇e = Me − Mc (3.1)
where Me (ωe , δ) is the output engine torque obtained empirically through an
engine map, which depends on the angular speed ωe and the engine fuelling δ.

Clutch: The clutch involves two frictional discs, which are pressed together
and connects the flywheel of the engine with the transmission’s input shaft. Such
clutches are commonly found in vehicles equipped with manual transmission. The
connection between the transmission and the clutch is considered to be stiff, i.e.:

Mt = Mc (3.2)
ωt = ωc (3.3)
where Mt denotes the torque output and ωt denotes the output angular speed
from the transmission.

Transmission: The transmission is the connection between the clutch and the
propeller shaft. It consists of a set of cogwheels (gears) which are connected such
that the output torque is transformed depending on which gear is engaged. It is
modeled in this case as a conversion ratio it , which varies according to the specific
gearbox transmission characteristics. The transmission in this case is modelled as
an optimal procedure for gear changing during coasting, i.e. allowing the HDV
to only operate under the ideal range of operation for most of the gearboxes at
Scania CV AB. The ideal range is empirically determined to be between 1100 and
1400 RPM. Thus the 12 geared transmission box is mapped with respect to velocity
through (3.4).
60 v
RP M = (it if ) (3.4)
2π rw
The mapping is illustrated in Figure 3.2.
Another characteristic of the gear box is the efficiency ηt . The inertia of the trans-
mission is neglected and the gear shifts are assumed to be instantaneous, i.e. an
immediate change of conversion ratio and efficiency, hence

Mp = it ηt Mt (3.5)
ωp = it ωe (3.6)
where p denotes the subscript for propeller shaft.
3.1. POWERTRAIN 9

Figure 3.2. A mapping diagram between gears, RPM, and velocity. i = 1, . . . , 12


denotes the active gear. As the RPM decreases during coasting and reaches RP Mn =
1100, the gear is instantaneously shifted to a lower gear, increasing it to RP Mn+1 =
1400. The current velocity is unaffected since the gear shift is considered to be
instantaneous, i.e. vn = vn+1 .

Propeller Shaft: The propeller shaft connects the transmission to the final
drive. No friction is assumed and the connection is considered to be stiff.

Mp = Mf (3.7)
ωp = ωf (3.8)
Final Drive: Like the transmission, the final drive is characterised by a con-
version ratio if , and an efficiency ηf . The value for the ratio and the efficiency
depends on the final drive design. Neglecting inertia the following relation could be
made by the input and output.

Md = if ηf Mf (3.9)
ωd = if ωf (3.10)
Drive Shafts: The drive shafts connects the final drive to the wheels. In this
simplified model it is assumed that the wheel speed is the same for both wheels. It
should be kept in mind that the wheel speed differs when the vehicle enters a curve.
However, it is negligible compared to other simplifications within the model. The
connection between the wheels and the drive shafts is considered to be stiff and can
therefore be modelled as:

Mw = Md (3.11)
ωw = ωd (3.12)
Wheels: The connection between the road and the wheels is modeled by as-
suming no slip, i.e.:
10 CHAPTER 3. VEHICLE MODELLING

Jw ω̇w = Mw − Mb − rw Fw (3.13)
rw ωe
v = rw ω w = (3.14)
it if
The braking torque Mb is hard to measure, often difficult to model, and is most
often zero. Therefore, it is neglected in this model.

3.2 Longitudinal Forces


The external, i.e. longitudinal forces on the heavy vehicle is modelled according to
Figure 3.3.

Fairdrag Fgravity

Froll
Fbrake Fengine
α

Figure 3.3. The longitudinal forces inflicted upon a heavy vehicle in motion.

Thus applying the generalized Newton’s second law gives the state-equation:

mt v̇ = Fengine − Fbrake − Fairdrag − Froll − Fgravity (3.15)

where Fbrake is considered to be zero due to the explanation given in 3.1 and α
denotes the road grade. During coasting the engine exerts a brake force as opposed
to a driving force when the accelerator is applied. The total accelerated mass is:

Jw it 2 if 2 ηt ηf Je
mt =+ m + (3.16)
rw 2 rw 2
The aerodynamic force is given by:
1
cw Aa ρa v 2
Fairdrag = (3.17)
2
where cw denotes the airdrag coefficient, Aa denotes the maximum cross-sectional
area of the vehicle and ρa denotes the air density. The rolling resistance is given by:

Froll = cr mg cos(α) (3.18)


cr denotes the corresponding coefficient, g denotes the gravitational constant,
and m denotes the vehicle mass. Finally the gravitational force is given by:
3.2. LONGITUDINAL FORCES 11

Fgravity = mg sin(α) (3.19)

All of the constants are obtained empirically. Combining equations (3.1)-(3.17)


a final mathematical vehicle model could be derived as:

dv 1
v̇ = = (Fengine − Fairdrag (v) − Froll (α) − Fgravity (α)) =
dt mt
1
= it 2 if 2 ηt ηf Je
(Fengine − Fairdrag (v) − Froll (α) − Fgravity (α) =
Jw
2 + m + 2
rw rw (3.20)
rw 2 i i η η
t f t f
= Me (ωe , δ)−
Jw + mrw 2 + it 2 if 2 ηt ηf Je rw
1 
− cw Aa ρa v 2 − cr mg cos(α) − mg sin(α)
2

Note that (3.20) is a nonlinear time varying state-space equation. The data from
the GPS and the digital maps is spatially sampled rather than with respect to time
and must therefore be transformed using the chain rule:

dv dv ds dv dv 1
= =v ⇒ = (Fengine − Fairdrag (v) − Froll (α) − Fgravity (α))
dt ds dt ds ds vmt
(3.21)
where v > 0 is evident due to physical properties. Hence using a first order
Euler approximation results in a discrete spatially sampled model with sampling
distance ∆s, the difference equation is deduced as:

vk = vk−1 + ∆s∆vk−1 (3.22)

where

Mek−1 1 sin(αk−1 )
∆vk−1 = c1 − c2 vk−1 − c3 − c4 (3.23)
vk−1 vk−1 vk−1

which is the discretized version of equation (3.21). c1 , c2 , c3 , c4 is clarified in


(3.24). They denote the vehicle parameters, which are a function of the gear.

1 2
rw it if ηt ηf 2 rw cw Aa ρa
c1 = c2 =
Jw + mrw 2 + i2 t i2 f ηt ηf Je Jw + mrw 2 + i2 t i2 f ηt ηf Je
(3.24)
rw 2 mg rw 2 mg
c3 = cr c4 =
Jw + mrw 2 + i2 t i2 f ηt ηf Je Jw + mrw 2 + i2 t i2 f ηt ηf Je
12 CHAPTER 3. VEHICLE MODELLING

3.3 Summary
By modelling the driveline as presented in 3.1, an expression for the output force
from the engine can be derived. Inserting that expression into Newton’s second
law, after deriving expressions for the longitudinal forces acting upon the HDV, a
non-linear discretized equation is derived. Equation (3.22) presents a non-linear dis-
cretized equation, which allows for the estimation of a velocity trajectory throughout
a given stretch. Thus, changes in the varying road grade and gear shifts can be ac-
counted for, which partly characterises the non-linearity in the equation. Thereby a
continuous (3.21) and a discrete (3.22) model is derived, which will serve as a base
for the mathematical procedures in the forthcoming chapters. Equation (3.22) will
prove to be the most useful equation throughout this thesis.
Chapter 4

Optimal Control

Optimal control is one of the most useful systematic methods for control design.
Often control problems become very complex and hard to solve using pure intuition.
Optimal control presents a systematic method and logical reasoning to solve complex
control problems. A control problem could have many solutions. However, it is often
desirable to find the best solution according to certain criterion.
To obtain a better understanding of how an optimal controller is to be designed
a mathematical study is undertaken. Two possible optimisation strategies for op-
timal control are implemented to give a more profound perspective of the optimal
solution. Constraints are forced upon the solution space to obtain optimal results
with respect to fuel, time, driver comfort, and brake application. To solve this opti-
mal control problem, it needs to be stated in the standard form shown in equation
(4.1).

Standard form:

Z tf
min L(x(t), u(t))dt + φ(x(tf ))
u:[0,tf ]→U 0
(4.1)
subject to:
ẋ(t) = f (x(t), u(t)), x(0) = x0
Remarks

• U ⊂ R set of admissible control


• Infinite dimensional optimisation problem:
Optimisation of functions u : [0, tf ] → U
• Constraints on x from the dynamics.
• Final time tf fixed

13
14 CHAPTER 4. OPTIMAL CONTROL

4.1 Pontryagin’s Principal


In optimal control, the problem at hand could be solved by using the Pontryagin
minimum principle (PMP). By investigating local properties and setting empirical
constraints, necessary conditions for optimality can be obtained and an optimal
trajectory can be calculated. One advantage amongst many of using PMP is that it
can be used even though dynamic programming fails due to lack of smoothness of
the optimal cost function. However, it should be noted that it only gives a necessary
condition for optimality.
The current problem could hence be stated in (4.2), using the same notation as
in [6]:
Z
J = min (u(t)T Ru(t))dt

ẋ = Ax + Bu, u = Feng
x(tf ) = vf , x(0) = v0 , ẋ(t) ≥ −kdecc (4.2)
x ∈ Xn , u ∈ Un
X = {x : vmin < v < vmax } , U = {u : Feng,min < Feng < Feng,max }

where the state vector x is chosen to be the vehicle velocity and the input-
control signal is the produced force on the vehicle, i.e. momentum, from the vehicle
engine. The produced force from the engine is allowed to be both positive and
negative, indicating whether work is being done on or by the engine. In other
words, the control signal could also be denoted as u = [Feng Feng,brake ]T , i.e. having
the characteristic of a driving and braking force.
The objective is to find the smallest possible input energy, which will make the
vehicle decelerate to the desired final velocity within a desired distance. This ap-
proach requires a fixed start point and a fixed final point. The calculated non-linear
state equation (3.20) needs to be linearized to apply the PMP. It is evident that
the equation (3.20) is rather linear within small segments. The optimal vehicle
behaviour for a speed decrease from 90km/h to 70km/h is of vast interest for the
problem at hand, since it is very common event on the Swedish highways. There-
fore, investigating a linear equation within the segment 70-90 km/h should give
satisfactory results. Linearizing the equation yields:

dv(t) 1
= (Fengine − Fairdrag (v) − Froll (α) − Fgravity (α)) ≈
ds(t) vmt
1 c1 M0 c3 c1
≈ (− 2 − c2 + 2 )∆x + ∆u = (4.3)
mt v0 v0 mt v0
1 c1 M0 c3 c1
= (− 2 − c2 + 2 )x̃ + ũ = Ax̃ + B ũ
mt v0 v0 mt v0
4.1. PONTRYAGIN’S PRINCIPAL 15

where
x̃ = x − x0 = x − v0 , ũ = u − u0 = u − M0
v0 = 80km/h, α0 = 0
c2 v0 2 + c3 cr + c3 sin(α0 )
M0 =
c1

Thus the optimization problem can be stated accordingly:


Z
J ∗ = min (ũ(s) + u0 )2 dt

subject to: (4.4)


˙
x̃(s) = Ax̃(s) + B ũ(s)
x(0) = x0 , x(sf ) = xf

The input is minimized in (4.4) and not the deviation from the linearization point.
There is no restriction on setting the starting point to s0 = 0, since the system allows
for arbitrary time translation of its solutions. Hence, if u(s) ∈ U transfers the initial
state x(0) = xi to the final state x(sf ) = xf , then ũ = u(s − si ) transfers x̃(si ) = xi
to x̃(si + sf ) = xf and the trajectories are related accordingly as x̃(s) = x(s − si ).
(4.4) is now in the standard form. The optimal input can thus be calculated
analytically by first stating the hamiltonian:

H(x, u, λ̃) = λ0 (ũ + u0 )2 + λT (Ax̃ + B ũ) (4.5)

The adjoint equation is hence:

dH
λ̇(s) = − λ(s) = −AT λ(s) ⇒ λ(s) = e−As λ(0) (4.6)
dx
Choosing λ0 = 1, due to the fact that the system is controllable, the calculation
could be furthered accordingly:

arg min H(x, u, λ̃) = arg min λ0 (ũ + u0 )2 + λT (Ax̃ + B(ũ + u0 ) − Bu0 )
ũ+u0 ũ+u0
1 1
⇒ ũ + u0 = − B T λ(s) = − B T e−As λ(0) (4.7)
2 2
1 1
⇒ x̃˙ = Ax̃ + B ũ = Ax̃ − B( B T e−As λ(0) + u0 )
2 2

Hence, the known general solution for such a differential solution is:
16 CHAPTER 4. OPTIMAL CONTROL

sf 1
Z
x(sf ) = eAsf x(0) − eA(sf −τ ) B( B T λ(τ ) + u0 )dτ = . . . =
0 2
1 sf A(sf −τ ) sf
Z Z
T T
Asf
= e x(0) − e BB T eA (sf −τ ) dτ eA sf λ(0) − eA(sf −τ ) dτ Bu0 =
2 0 0
1 T
= eAsf x(0) − W (sf , 0)eA sf λ(0) − G(sf , 0)Bu0
2
It is evident that λ(0) can be solved for, thus the optimal input is:
T (s
ũopt,P M P = −B T eA f −s)
W −1 (sf , 0)(eAsf x(0) − G(sf , 0)Bu0 − x(sf )) − u0 (4.8)

Implementing ũopt,P M P and simulating the state equation gives a trajectory


shown in Figure 4.1. The distance for which no input is required, i.e. the vehicle
simply decelerates to the final desired velocity due to the existing frictional forces
with no fuelling, is determined to be sd = 1094m. The results for simulating the
system for a deceleration stretch longer than sd are shown in Figure 4.1.
xtrajectory xtrajectory
90 100
Velocity [km/h]

Velocity [km/h]

85
90
80
80
75

70 70
0 200 400 600 800 1000 1200 1400 1600 0 500 1000 1500 2000 2500 3000 3500
Distance [m] Distance [m]
uopt uopt
140 500

120 400
Torque [Nm]

Torque [Nm]

100 300

80 200

60 100
0 200 400 600 800 1000 1200 1400 1600 0 500 1000 1500 2000 2500 3000 3500
Distance [m] Distance [m]

Figure 4.1. Plot of the velocity trajectory and the actual optimal input from the
engine for a longer fixed deceleration stretch than sd . The figure to the left is simu-
lated for a fixed final deceleration strecth of sf = 1500 and the figure on the right for
sf = 3000

It can be observed that a deceleration distance longer than sd , mandates a positive


declining control input, i.e. the engine has to work to maintain the speed initially
since the deceleration stretch is too long. It becomes evident by studying the right
figure in Figure 4.1, that if a much longer deceleration stretch is mandated, the
optimal solution is to keep the top speed as long as possible and successively de-
celerating the velocity as late as possible. Intuitively it is understood that the fuel
consumption is minimised by slowing down to the final desired speed without using
the brakes. Also, by maintaining the maximum allowed speed as long as possible
the travelling time is minimised.
The results for simulating the system for a deceleration stretch shorter than sd
is illustrated in Figure 4.2.
4.2. LQR 17

xtrajectory
90

Velocity [km/h] 80

70

60
0 100 200 300 400 500 600
Distance [m]
uopt
−350
Torque [Nm]

−400

−450

−500
0 100 200 300 400 500 600
Distance [m]

Figure 4.2. Plot of the velocity trajectory and the actual optimal input from the
engine for a shorter fixed deceleration stretch than sd , i.e. for sf = 500

The trajectories depicted in Figure 4.2 illustrates the case when the deceleration
distance is shorter than sd . As expected, the optimal control input is a negative
declining force input, i.e. a relatively large "braking" force is initially required to
slow down to the final speed during the short stretch.

4.2 LQR

The optimal controller calculated through the PMP disclosed a deeper insight and
revealed key characteristics of the optimal trajectory. However, the cost function
in (4.4) only constrains the input by minimising it. Consequently all the objec-
tives presented in 2.3 are not fulfilled. Mandating additional constraints in (4.4)
increases the complexity of the optimal solution in the PMP, rendering an analytical
expression difficult.
Thus the discrete linear quadratic regulator (LQR) follows as an alternative
approach. Hence, the non-linear discrete model (3.22) is linearized in (4.9) through
a first order Taylor approximation. It is conducted around a desired equilibrium
point. As explained in 3.2, each gear exerts a braking force of different order of
magnitude when coasting. This is denoted as a mode in the discrete linear model
(4.9).
The transition matrix is given by:
18 CHAPTER 4. OPTIMAL CONTROL

df
Am = 1 + |x=xm ∆s = 1 + ∆vm ∆s
dx
c1
Bm = ∆s
vm

c M 2
where ∆vm = − 1,m v0 2
m
− c2 + vc032 (cr + sin(α0 )), and Mm = c2 v0 +c3 (cc1r +sin(α0 )) .
x̃ = x − xm denotes the linearized state variable, and ũ = M − Mm denotes the
relative engine torque. Subsequently the linearized discrete state space model is:

x̃k+1 = Am x̃k + B ũk (4.9)


Having a linear discrete difference equation, the costfunction can be stated and
modified to address additional constraints. By changing the equilibrium to xm =
70km/h, a constraint could be put on the final state. A constraint could also be put
on maximum allowed deceleration, which tends to the driver comfort requirement.
By adding these new constraints along with the constraint on the input force from
the engine, the issues of minimizing fuel consumption, driver comfort, and brake
efficiency are addressed. The mathematical representation of the cost function is
given by (4.10).
sf −1
(x̃˙ T Qx̃˙ + ũT Rũ) + x̃(sf )T Qf x̃(sf )
X
min (4.10)
s=0

subject to: x(0) = x0


x̃k+1 = Am x̃k + B ũk

The problem could be solved in several ways. A Dynamic Programming (DP)


solution is applied, since it gives an efficient, recursive method to solve LQR prob-
lems [11]. Accordingly a value function is stated in (4.11).

sf −1
(x̃˙ T (τ )Qx̃(τ
˙ ) + ũT (τ )Rũ(τ )) + x̃(sf )T Qf x̃(sf )
X
Vs (z) = min (4.11)
u(s),...,u(sf −1) τ =s

subject to x̃(s) = z, and x̃(τ + 1) = Am x̃(τ ) + B ũ(τ )

Vs (z) gives the min-cost-to-go starting from state z at point s, hence V0 (x0 ) is
the minimum LQR cost. Thus the dynamic programming principle could be stated
as:

Vs (z) = min(z T Qz + wT Rw + ż T Q1 ż + Vt+1 (Az + Bw)) (4.12)


w
where w = ũ + u0 . Solving the Hamilton-Jacobi equation, a Riccati recursion is
formed (4.13)
4.2. LQR 19

Ps−1 = Q + AT P̃s A − AT P̃s B(R + B T P̃s B)−1 B T P̃s A) (4.13)


By solving the recursive Riccati equation, the optimal input solution is given in
(4.14)

uopt,LQR (s) = −(R + B T Ps+1 B)− 1B T Ps+1 Ax(s) (4.14)


For a more detailed explanation of how the optimal solution was derived from
(4.12)-(4.14), the reader is kindly referred to Appendix A. Implementing ũopt,LQR
and simulating the state equation gives a trajectory shown in the figures below. The
main focus lies on what new information the simulation with additional constraints
will prevail. Therefore, similar cases presented in 4.1 are studied for comparison.
Figure 4.3 shows the simulation for when given deceleration stretch is longer than
stretch for which the vehicle simply decelerates to the final desired velocity due to
the existing frictional forces.

xlqr
90
Velocity [km/h]

85

80

75

70
0 500 1000 1500
Distance [m]
ulqr
140
Torque [Nm]

120

100

80
0 500 1000 1500
Distance [m]

Figure 4.3. Plot of the velocity trajectory and the actual optimal input from the
engine for a longer fixed deceleration stretch than sd , i.e. for sf = 1500

It can be seen in Figure 4.3 that the optimal input is positive and slightly increased,
i.e. roughly kept constant in the beginning. The velocity decrease trajectory shifts
from a non-linear to a linear deceleration as the final velocity is reached. Figure 4.4
shows the optimal trajectories for a fixed deceleration stretch shorter than sd .
It is noticed in Figure 4.4 that the input is constantly negative. Due to the fact
that the deceleration stretch is too short, a braking force must be applied. The neg-
ative force is increased in a manner that makes the optimal velocity trajectory seem
rather linear. A constraint is put on the deceleration, i.e. driver comfort, in the
20 CHAPTER 4. OPTIMAL CONTROL

xlqr
90

Velocity [km/h]
85

80

75

70
0 100 200 300 400 500
Distance [m]
ulqr
−250
Torque [Nm]

−300

−350

−400
0 100 200 300 400 500
Distance [m]

Figure 4.4. Plot of the velocity trajectory and the actual optimal input from the
engine for a shorter fixed deceleration stretch than sd , i.e. for sf = 500

cost-function (4.10). Consequently the optimal input reveals an increasing deceler-


ating force, i.e. it is favourable to slowly increase the braking force as deceleration
is commenced.

4.3 Optimal Solution


Both the continuous linearized model in 4.1 and the discrete linearized model in 4.2
shows an optimal behaviour of the control input. The optimal input is based on a
linearized model which utilizes no dynamic information regarding the topology of
the road. Variations in the road grade have significant effect on the control input. It
is however concluded that implementing a road grade dependence in the controller
is out of the scope of this thesis. Also additional necessary conditions are very hard
to implement in the cost-function to give analytical solutions.
An essential factor concerning optimality in the problem at hand is a well de-
fined time-constraint. Given a long deceleration stretch, the optimal control input
according to the LQR will allow the speed to slowly decrease on the given stretch for
fuel-efficiency, while the PMP would suggest maintaining the current speed. Main-
taining the maximum legal speed a little longer and then decelerating "comfortably"
would result in a higher average speed, hence decreasing the total travelling time,
which is a significant factor in some cases. Such behaviour is also more accepted by
experienced drivers.
Another known crucial factor is driver comfort. The vehicle cannot be allowed
to decelerate too fast. Studies have shown that a maximum deceleration of 0.5
4.3. OPTIMAL SOLUTION 21

m/s2 is considered to be comfortable. This issue puts another constraint on the


controller, which was hard to implement in the PMP but possible in the LQR. Ac-
knowledging the fact that there are many constraints on the optimal input signal,
it can be understood that a favourable solution does not lie in one single optimal
control strategy, but in the combination of several strategies. Therefore, a switch-
ing control algorithm based upon the constraints should give a more advantageous
control behaviour.
Chapter 5

Switching Controller

As discovered in the previous chapter, an optimal controller addressing all the objec-
tives discussed in 2.3 is very difficult to implement analytically. The road topology
must be taken into consideration when designing an optimal controller. Therefore,
grade, road speed, and curvature must be utilised as input parameters to produce
an optimal velocity trajectory through a controller as depicted in Figure 5.1.

Figure 5.1. Block diagram for the controller design

The non-linear vehicle model (3.20) presents a method to calculate the velocity
with the grade as an input parameter. Hence, an optimal speed selection algorithm
can be designed with the aid of an electronic horizon.

5.1 Road Speed


5.1.1 Coasting
The point of deceleration explicitly due to frictional forces can be calculated by
modifying (3.22). The modified formula is stated in (5.1).

23
24 CHAPTER 5. SWITCHING CONTROLLER

vk = vk−1 + ∆s(−∆vk−1 ) (5.1)


Thus information regarding the topology, i.e. the grade up to the point of
interest, varying vehicle parameters, and the desired speed at a given location can be
used to calculate a fuel cut-off point. Wasting energy through unnecessary braking
is thereby avoided and the fuel consumption is minimised. The method utilises the
input of a future decrease in posted road speed and calculates the appropriate speed
at every time-step inversely from that point until the current maximum speed is
reached.

95
vinv
90
Velocity [km/h]

vRoad
85
80
75
70
65
200 400 600 800 1000 1200 1400 1600 1800 2000
Distance [m]

0
Altitude [m]

−5

−10

200 400 600 800 1000 1200 1400 1600 1800 2000
Distance [m]

Figure 5.2. Speed adaption for a road speed decrease from 90 km/h to 70 km/h

Figure 5.2 shows a simple case for the applied method, where a legal road speed
decrease from 90 km/h to 70 km/h occurs. It clearly shows that cutting off the fuel
at approximately 900 meters, causes the HDV to decelerate by coasting down to
the desired velocity. Fuel is saved and braking is reduced during coasting, creating
a cost efficient and safe solution.
However, when faced with a steep downhill the magnitude of the gravitational
force will increase with the increased magnitude of the grade. If it is increased to a
certain extent, it will prohibit the vehicle from deceleration. The consequence might
be several different scenarios. The speed has to be decreased to an unacceptable
level before the downhill in order to obtain the legal speed. In such a scenario, the
speed decreases far below the final desired velocity resulting in an uncomfortable
driver experience. This is also unacceptable road behaviour according to the traffic
behind the heavy vehicle. Another scenario is when the grade of the hill produces a
gravitational force upon the HDV, which is equal to the frictional forces. The vehicle
would therefore keep a constant speed and never decelerate. A third scenario is when
no solution exists at all. It could occur if the vehicle is travelling on a road with a
5.1. ROAD SPEED 25

speed limit of 90 km/h and there is a decrease in the limit at the bottom of a hill to
70 km/h. In this case the steep grade of the road causes the vehicle to accelerate past
the legal speed limit which is an unacceptable solution. Hence,merely using (5.1) as
a control method might have either no solution or unacceptable solutions such as
an undesirably long deceleration stretch. Owing to the fact that the inverse method
cannot be implemented all the time, an alternative method must be switched to.
The alternative methods have to address and optimise three major areas of concern,
which is depicted in Figure 5.3.

Figure 5.3. Three decision regions when facing a steep downhill

When the inverse method is not obtainable, the new method must determine:

(I) When to initiate the fuel cut-off point: Maximum fuel reduction is obtained
by implementing the fuel cut-off point as soon as possible.

(II) How long is coasting possible: It is most favourable to coast as long as possible
without breaking the legal speed limit.

(III) If braking is deemed to be necessary, when is it most advantageous to initiate


braking: Braking instigates wasting energy and should thus be minimised.

5.1.2 Fuel Cut-Off Point


Maximum fuel reduction is obtained by implementing the fuel cut-off point as soon
as possible. The fuel consumption will be minimized by allowing the vehicle to
coast as long as possible. However, as presented in 5.1.1, it might imply that the
vehicle decelerates during a long stretch. If the road grade produces a gravitational
force which is slightly less than the opposing frictional forces, the inverse method
might suggest a cut-off point as far as 3 km from the speed change. Reducing
26 CHAPTER 5. SWITCHING CONTROLLER

the speed over a 3 km long stretch will certainly agitate the held up traffic behind
the truck, induce driver discomfort, and increase the total travelling time. The
controller behaviour will therefore most probably not be perceived as warranted by
the driver. Hence the driver will disengage the controller, making the controller
obsolete. On the other hand a short fuel cut-off point will not minimise the fuel
consumption.
Driver experience shows that it is very difficult for a driver to predict the length
of the deceleration stretch if the road is a road grade present. However, the experi-
enced driver estimates the deceleration stretch of a vehicle on a level ground rather
easily and intuitively. Hence, calculating the distance it takes for a heavy vehicle to
decelerate on a level ground could be used as a reference fuel cut-off point. Thus,
the vehicle will not be allowed to apply any decelerating control input before the
calculated cut-off point. Modifying equation (3.15) by removing the gravitational
force contribution and setting the roll force to a constant, the distance of interest
could easily be calculated and is referred to as SMax . Therefore the vehicle will not
be allowed to commence decelerating, i.e. cutting off the fuel until it reaches SMax .
This preserves the use of the inverse velocity estimation as an optimal solution while
driving in an incline, since SMax will clearly be longer.

5.1.3 Maximum Roll-Off


It is most favourable to coast as long as it does not break the legal Speed limit. Setting
the fuel cut-off point to SMax , will enable the vehicle to start coasting from that
point and onwards. Consequently a forward velocity estimation method is applied
through equation (3.22), making use of the topology information and allowing the
possibility of a fuel cut-off point. However, due to the topology, a few previously
mentioned problems might arise, which are depicted in Figure 5.4.

Figure 5.4. Three common problems that arises within the forward method.

Figure 5.4 shows three identified scenarios. 5.4 a) illustrates the case when a
5.1. ROAD SPEED 27

forward simulation results in a speed decrease followed by a larger speed increase


due to the topology. In such a scenario the driver might experience discomfort
due to the sudden decrease followed by an increase in speed. It is counterintuitive
to suddenly have a drop in speed only to regain it moments after. Hence, SMax
must be redefined from that point which will avoid such fluctuations in the desired
velocity. 5.4 b) shows the scenario when the topology demands that the speed be
decreased below the final desired velocity. This is a similar situation to which the
inverse controller failed. A proper action would be to shorten SMax to that point
and restarting the forward simulation from the new point until the situation does
not arise. Hence, the number of iterations would be minimised while maintaining a
cost-effective method. 5.4 c) simply shows the consequence of implementing SMax .
Naturally the velocity at the final point where the mandated speed reduction occurs
will be overshot due to the short deceleration stretch. Therefore, the controller will
have to switch to a favourable deceleration method by an intelligent application of
the several braking systems.

5.1.4 Most Favourable Braking


Braking instigates wasting energy and should thus be minimised. It is not within the
objectives of this paper to assess how the various vehicle brake systems should be
applied, but rather what the effects upon the driver comfort will be. Applying the
brakes will lead to loss of the produced energy from the fuel. Therefore, they should
be applied as late as possible. However, there is a trade-off between the time when
brakes are applied and the magnitude of the decelerating force required to bring the
vehicle to the desired velocity. If the brakes are applied at a relatively late stage,
the driver will experience great discomfort due to the large deceleration. Experts in
the field state that a maximum deceleration of 0.5m/s2 is acceptable, without the
driver feeling any discomfort. Braking procedure with high performance regarding
driver discomfort and fuel-cost can easily be calculated. The resulting speed change
from a certain deceleration can be determined from (5.2).

dv dv ds dv dv adec
adec = = =v ⇒ =
dt ds dt ds ds v (5.2)
adec
⇒ vk = vk-1 + ∆s(−∆vk ) = vk-1 + ∆s(− )
vk-1

It has to be noted that situations may arise when a larger deceleration is induced
by the topology. In such cases the controller must allow the vehicle to decelerate
according to what the road grade mandates. If a larger deceleration through coast-
ing e.g. due to a steep incline before the point of the speed reduction arises, the
controller should switch to coasting. Such a behaviour is deemed to be intuitive by
the driver.
28 CHAPTER 5. SWITCHING CONTROLLER

5.1.5 Implementation
The controller switches between several methods which are optimal with respect to
various objectives mentioned in 2.3. Within every method several unique problems
that may arise are solved. By not allowing the vehicle to decelerate until it reaches
Smax , the traffic behind the vehicle will not get upset, since the vehicle deceleration
is kept relatively short. However, by not reducing the speed in time, the vehicle
will overshoot the desired end velocity. Therefore the brakes have to be applied in
order to maintain the legal speed limit. Consequently the vehicle will have a fuel
cut-off point, but will have to apply the brakes eventually. Such vehicle behaviour
is acceptable according to the surrounding traffic dynamics and it also increases
the average speed, i.e. reduces the travelling time. The vehicle coasts as long as
possible, minimising the fuel consumption. The overall procedure is demonstrated
in Figure 5.5

Figure 5.5. An overview the Switching Controller methodology

If an electronic horizon is available, the controller will be able to produce an optimal


trajectory with respect to time, fuel consumption, driver comfort, and braking.
Figure 5.6 shows the optimal velocity trajectory for when the vehicle faces a road
speed decrease at the bottom of the hill.
The hill is too steep for the inverse method to be applied. The velocity vinv
(dotted green line) shows that an inverse method solution exists but is unacceptable.
5.2. CURVATURE 29

100

95

Velocity [km/h] 90

85

80

75 Actual Velocity
vbrake
70 vmax
vforward
65
vinv
60
1100 1200 1300 1400 1500 1600
Distance [m]

0
Altitude

−2

−4

−6

−8

1100 1200 1300 1400 1500 1600


Distance [m]

Figure 5.6. Optimal velocity trajectory for a steep downhill. The red (dotted)
line depicts the posted road speed. The green (dotted) line depicts the velocity
trajectory from inverse method. A blue (dash-dotted) line depicts the favourable
braking strategy. The purple crosses depicts the forward method trajectory. A (solid)
black line illustrates the resulting total velocity strategy produced by the switching
controller.

The controller switches to the forward method after calculating the optimal point
of control action, i.e. Smax . The optimal velocity is therefore indicated with vforward
(purple crosses) in Figure 5.6 Evidently the shortened deceleration stretch compels
the vehicle to overshoot the final desired velocity. Thus, the optimal braking velocity
trajectory (blue dashed line) is switched to in the end. The estimated optimal
velocity is finally given by the black line.

5.2 Curvature

The grade and road speed reductions are not the only significant inputs to the
switching controller shown in Figure 5.5. The curvature dictates a maximum speed.
There are several ways to determine a maximum allowed speed through road curves.
In this study, it is determined by setting a speed that is mandated by a comfortable
centripetal force. Curve speed warning systems (CSWS) is a vast and current field
of research [8]. The centripetal force along with the banking of the road is taken
into consideration within many of those studies. There are scientific methods based
on models of various complexity that determine what exact velocity is required for
the HDV to tilt.
30 CHAPTER 5. SWITCHING CONTROLLER

5.2.1 Curve-Radius Estimation


The issue that arises in determining a maximum velocity through a curve is whether
the radius of the curve can be obtained. There is a database table within ADASRP,
which has a measured radius attribute for certain roads within Europe. However,
the applicability of the data is questionable. The radius can vary significantly
depending on the path chosen through the curve, i.e. middle or the side. It is
known that a driver usually cuts through the curve to maintain a higher velocity
throughout the curve. Therefore the geographically mapped road curvature is not
the always actual travelled path. A method of estimating the road curvature can
thereby be established.
The radius is estimated from a GPS-trace, which contains the actual positioning
of the vehicle. Mind that there is an inaccuracy with GPS-positioning, but the
error is handled through averaging that seems to be adequate. By estimating a best
curve fit on five data points, a circle can be estimated. Gauss-Newton’s method is
modified and applied in order to find the desired curve-radius in all GPS points.
The modification is carried out in (5.3).

(x − xc )2 + (y − yc )2 = R2
⇒x2 + y 2 = R2 + 2xc x + 2yc y − (x2c + yc2 )
(5.3)
 
"
x2 + y 2
# "
1 x y
# R2 − (xc 2 + yc 2 )
⇒ .. = .. .. .. 2xc
 
 
. . . . 2yc

The least square method is used to serve as a guess for the value of R. Due to the
non-linear nature of this problem, the error of the estimate can further be reduced
by the use of iteration. Gauss-Newton’s algorithm [2] is as follows:

• Write the non-linear equation on the standardform f (z) = 0.

• Determine the analytic expression for the elements in the Jacobian J(z)

• Use reasonable values as starting points (in this case the values obtained in
the least square method above).

• (*) Calculate f and J with the measured zi -values, i.e. the five points obtained
in the GPS-map data.

• Solve the linear equation Jδz = −f

• Update: z = z + δz

• Repeat (*) until desired accuracy is obtained.

Hence, the circle radius is presented as a minimization problem. The error of the
curve fit is reduced iteratively by find using a least square estimation as an initial
5.2. CURVATURE 31

solution to the minimization. Applying the method on data measured from a known
test road at Scania, the result depicted in Figure 5.7 is obtained. The radii of various
lengths are depicted as (red dashed-dotted) lines with the magnitude corresponding
to the length of the radius.

1600

1400

1200

1000

800

600

400

200

−200

−400
−400 −200 0 200 400 600 800 1000 1200 1400 1600

Figure 5.7. Radii plotted on a known 10.3 km long test-road.

A straight road could be perceived as having an infinite radius. Therefore, the


vehicle faces a rather long radius entering a curve.

500

450

400
Y−coordinate

350

300

250

200

150
900 950 1000 1050 1100 1150 1200 1250 1300
X−coordinate

Figure 5.8. Radii plotted for a S-curve of the test-road.

The radii decreases as the vehicle proceeds through the curve and increases as it
exists as shown in Figure 5.8. Slightly bending curves does not have any impact on
32 CHAPTER 5. SWITCHING CONTROLLER

the maximum speed limit constraints. Therefore, only radii of less than 700 meters
are taken into consideration. Hence, the true curvature experienced by the vehicle
is obtained through measuring the actual path travelled by the vehicle. Data is
also collected from a test-run around an island of known radius. By comparing the
calculated radius with the measured radius, the error is thereby estimated to be
±0.5m, which is an acceptable tolerance.

5.2.2 Maximum Curve Speed


For the purpose of this study a maximum speed is derived simply by determining
a maximum centripetal force exerted on the driver. A threshold of 0.15g − 0.2g is
determined in this case [12]. Thus a maximum velocity when travelling through a
curve can be calculated according to equation 5.4.

v2 C 0.2g
r
0.2 = − E ⇒ vmax = (5.4)
g C
where g is the gravitational constant, C = R1 is the curvature, and R is the
radius of the road. The road crossfall E is neglected due to lack of data. Hence an
additional maximum speed limit is set on the road. The constraints set from the
curvature is illustrated for a single curve in Figure 5.9.

80
Velocity [km/h]

60

40

20
2400 2500 2600 2700 2800 2900 3000 3100 3200
Distance [m]

400
Y−coordinate

200

−200
700 800 900 1000 1100 1200 1300 1400 1500 1600
X−coordinate

Figure 5.9. Radii plotted for a segment of the known 10.3 km long test-road.

Figure 5.9 shows the maximum comfortable velocity calculated by (5.4), for
the curve segment depicted in the lower part of the figure. Unlike the mandated
posted road speed, which instantaneously changes the maximum comfortable speed
for entering a curve is smooth. Clearly, the speed only has to be reduced to ap-
proximately 60km/h when entering the first curve. However, entering the narrower
curve subsequently, a much lower speed is necessary not to feel discomfort. Hence,
5.2. CURVATURE 33

applying (5.4) on the test-road depicted in Figure 5.7, the combined constraint of
the mandated legal speed and the maximum curve velocity is depicted in Figure
5.10.

80
vcurve
vlegal

70

60

50
Velocity [km/h]

40

30

20

10
0 2000 4000 6000 8000 10000 12000
Distance [m]

Figure 5.10. The maximum speed allowed on a given road inflicted by legal con-
traints and curvature. A (solid) blue line depicts the speed mandated by the curva-
ture. The (dotted) red line illustrates the posted road speed.

Figure 5.10 clearly shows that the original speed limit shown by the dashed red
line, is highly altered by the constraints mandated by the road curvature. The new
smooth velocity constraint implemented on the controller induces new problems
upon the strategy of the switching controller’s forward method.

5.2.3 Method Strategy


The aim is to reach the maximum allowed velocity set by the shortest radius, i.e.
the min-point, of every encountered curve on the given electronic horizon in an
efficient and strategically sound manner. Naturally the switching controller will
apply the same procedure as suggested in 5.1.5. First and foremost the inverse
method is chosen if the deceleration stretch is shorter than Smax and the maximum
speed constraint set by the legal road speed in addition to the curvature constraint
is not exceeded. However, if the afore mentioned criteria is not fulfilled, the forward
method is switched to.
Approaching a curve, the objective is to be able to decrease the speed of the
vehicle in time in an energy efficient manner so that it does not overshoot the
maximum velocity set by the shortest radius.
Figure 5.11 shows an attempt to reach the desired final velocity through the inverse
method. However, the maximum allowed velocity is set by the curvature is trans-
gressed, creating an unacceptable velocity trajectory. The bound set by the legal
road velocity and the curvature must never be exceeded. Therefore the minimum of
34 CHAPTER 5. SWITCHING CONTROLLER

60

50

Velocity [km/h]
40

30

20 vmax
vinv

2600 2650 2700 2750 2800 2850


Distance [m]

95
90
85
80
Altitude

75
70
65
60
55
50
2600 2650 2700 2750 2800 2850
Distance [m]

Figure 5.11. Plot of the maximum speed allowed along with the inverse method
velocity trajectory. The (dotted) red line illustrates the maximum allowed speed
mandated by the curvature and posted road speed. The (dashed-dotted) green line
depicts the velocity trajectory produced by the inverse method.

the velocity trajectory and the velocity constraint should always be chosen. On the
other hand resorting to that option might create an undesirable fast decelerating
trajectory as depicted in Figure 5.12

60

55

50
Velocity [km/h]

45

40

35
vbrake
30 vmax

25 vinv

2550 2600 2650 2700 2750 2800 2850


Distance [m]

95
90
85
80
Altitude

75
70
65
60
55
50
2550 2600 2650 2700 2750 2800 2850
Distance [m]

Figure 5.12. Plot of the maximum speed allowed and the inverse method velocity
trajectory, along with the comfortable deceleration (braking) trajectory. The (dotted)
red line depicts the curvature and posted road speed. A (dashed-dotted) green line
illustrates the velocity trajectory calculated by the inverse method. The (dashed-
dotted) blue line illustrates a favourable braking trajectory.

Figure 5.12 illustrates the maximum comfortable deceleration trajectory (blue


5.2. CURVATURE 35

line, dash-dotted line) as well as the inverse method trajectory (green, dash-dotted
line) and the maximum allowed velocity (red, dotted line). The maximum allowed
velocity is lower than the inverse method trajectory and should therefore be cho-
sen at the 2770 meter marker in Figure 5.12. However, that would imply that the
requested deceleration between the 2770 m and the 2790 m markers for the up-
per limit would clearly be greater than the comfortable deceleration. Thus if the
maximum allowed velocity is lower than the inverse method trajectory and it decel-
erates uncomfortably fast, a suitable strategy is applied. The final desired velocity
is reached by applying the inverse method to a point one discretisation step ahead
and then keeping a constant velocity during the last part. Thereby the vehicle will
still coast as long as possible and not exceed any of the constraints. It implies a
safe, comfortable and economic vehicle behavior. Note that if the grade of the road
induces a faster deceleration than the calculated comfortable deceleration speed, it
is still considered to be intuitive, and therefore allowed.
In many cases a mandated speed reduction set by the curvature from e.g. 70
km/h to 20 km/h will result in an unacceptably long deceleration stretch Smax
presented in 5.1.2. An upper limit of 700 meters is therefore set on the maximum
allowed deceleration stretch.
Hence, many large speed reductions will result in switching to the forward
method. In that case the inverse method will no longer apply and can therefore not
be shifted to resolve the problem of too fast deceleration due to the curvature con-
straints. Thus, an appropriate braking strategy is formed. By calculating the slope
of the mandated curvature speed, a new point of interest can be determined. The
point where the slope of the upper limit exceeds the comfortable braking trajectory
will thereby serve as a new point of reference to where a new comfortable braking
trajectory will be calculated. The final desired speed will therefore be reached in
advance and subsequently held until the original point of interest is reached. If the
inverse method could not be applied in the case depicted in Figure 5.12 due to the
topology, the intersection point at the 2770 meter marker would serve as a new
reference point to which the braking velocity trajectory would be shifted.
As seen earlier in Figure 5.4 b), there might arise a situation where an action
would be to decrease the velocity below the final velocity. The action adapted in
that situation was to reduce Smax to that point and continue the velocity trajectory
from that point onward. However, when the maximum velocity profile changes due
to the dynamics entering a curve, the action of reducing Smax to that point is no
longer an effective choice in certain situations as illustrated in Figure 5.13.
The top plot in Figure 5.13 clearly shows that if the forward simulation starts
just before 7000 m, the estimated velocity trajectory of the forward method (purple
crosses) will drop below the final speed at approximately 7050 m. Hence, the forward
method will shift Smax to that point and starts over, which is not the best solution.
A proper action in this case would be to shorten Smax an arbitrary fixed distance
and then starting the velocity trajectory from that point onward. Doing so will
allow the forward method to eventually follow the more efficient trajectory (dotted-
green line). That trajectory is considered to be a better solution, since the fuel
36 CHAPTER 5. SWITCHING CONTROLLER

Figure 5.13. Plot of the trajectory resulting from the forward method. The purple
(crosses) depicts the velocity trajectory calculated by the forward method in each
sampling point. A red (dotted) line illustrates the mandated maximum speed. The
(dotted) green line illustrates a more fuel efficient alternative. The blue (dashed-
dotted) line depicts the comfortable braking trajectory. The (solid) black line illus-
trates the final produced velocity trajectory.

cut-off point arises earlier and the increase of speed is smaller from the last minima,
i.e. both the fuel consumption and the brake wear is lowered.
It should also be noted that the forward method trajectory might drop below the
final velocity on the flat part of the comfortable braking trajectory due to the road
grade, as can be seen after the 7400 m marker in the top plot of Figure 5.13. If such
situations arises, Smax should obviously not be shortened and the higher velocity
of the brake-trajectory should be the natural choice. Such a situation might also
arise if the upper limit is chosen as a part of the optimal trajectory and not deemed
to create an unacceptable deceleration. An incline in the road might then create
a larger deceleration, which would in turn create a trajectory from that forward
method that drops beneath the final desired velocity. Smax should not be reduced
in such a situation. Instead, the braking trajectory should be maintained.
Figure 5.14 shows the importance of separating the three cases presented in 5.1.1. In
this case the forward method is applied, since the inverse method is not applicable.
The switching controller is always calculating suitable strategies to undertake in each
discritization step. Before the velocity trajectory drops beneath the tolerance set on
the final desired velocity at the 9110 meter marker, the road grade creates a higher
deceleration than the comfortable braking trajectory, making the coasting trajectory
an optimal choice for the braking strategy. Evidently the controller switches to the
braking method and consequently makes an unacceptable instantaneous increase in
velocity of approximately 30 km/h. Such behaviour is obviously not tolerated since
5.2. CURVATURE 37

Figure 5.14. Plot of the trajectory resulting from a controller misunderstanding.


The purple (crosses) depicts the velocity trajectory calculated by the forward method
in each sampling point. A red (dotted) line illustrates the mandated maximum speed.
The blue (dashed-dotted) line depicts the comfortable braking trajectory. The (solid)
black line illustrates the final produced velocity trajectory.

it creates a physically impossible velocity request. A misunderstanding of which


method, i.e. inverse, forward, or braking, has priority, can lead to an unacceptable
velocity output. Therefore, the current active method strategies are - must be -
isolated and prioritised from other methods.
Hence, the controller first have to make a choice of whether to choose the inverse
method velocity profile or the forward method depending on the predicted decel-
eration distance for every legal speed reduction and every maximum allowed curve
velocity set by the minimum radius within that curve. It then has to decide how far
the vehicle can coast before activating proper braking action. If the road grade dic-
tates that the vehicle must lower its speed below the final desired velocity along the
calculated velocity trajectory, the deceleration stretch must be reduced and a new
optimal trajectory must be calculated from a new reference point. The switching
controller constantly compares suitable strategies and chooses the most appropriate
one. If the topology compels the controller to switch to the braking method, several
new strategies discussed above are investigated. However, it is imperative that the
method strategies are separated as discussed above.
Chapter 6

Analysis

The switching controller outputs a velocity trajectory. The trajectory serves as an


input to the cruise controller onboard the heavy duty vehicle, which thereby can
serve as a control to the accelerator command. STARS is a simulation tool designed
at Scania CV AB. It too, can utilize a velocity trajectory as an input to the vehicle
accelerator command. Other input parameters to the highly developed simulation
environment are the road grade, altitude, distance sampling points, and whether
it is discretized with respect to time or distance. STARS is based upon a highly
advanced vehicle modelling for fuel consumption, simulations and measurements.
For efficient simulations the model uses production code for speed and gear control,
exchangeable data sets to allow simulation of the whole production range of engine
types, and advanced models for the longitudinal forces exerted upon the vehicle
[14].

6.1 Simulation
6.1.1 Test-Road
Based upon the maximum velocity constraints illustrated in Figure 5.10 derived
from the test road depicted in Figure 5.7, a favourable velocity trajectory can be
produced from the switching controller. An electronic horizon is created for the 10.3
km long stretch, containing all the required input data required for the controller
to produce the trajectory illustrated in Figure 6.1.
The controller calculates a velocity trajectory based upon fuel efficiency, time,
driver comfort, and minimal braking for each maximum curve speed invoked by the
minimum curve radius, illustrated in the top part of Figure 6.1 (black line). It then
calculates a trajectory for the posted speed decreases, shown in the lower part of
Figure 6.1. An increase of legal road speed is best handled by the internal control
system of the vehicle and therefore not handled by the switching controller, since
the speed request to the cruise controller can be instantaneous.
The optimal trajectory was used as the speed request input in STARS. The
simulation revealed that the total fuel consumption is 2.37 liters, the total energy

39
40 CHAPTER 6. ANALYSIS

80
vcurve
70 vmin
vref
60 vopt

Velocity [km/h]
50

40

30

20

10
0 1000 2000 3000 4000 5000 6000 7000 8000 9000 10000
Distance [m]

80

70

60

50

40

30

20

10
0 1000 2000 3000 4000 5000 6000 7000 8000 9000 10000

Figure 6.1. Optimal velocity trajectory produced by the switching controller for
the test-road.

produced by the brakes 5056 kJ, and the total travelling time is 710.98 seconds. An
experienced driver was also asked to drive the same stretch, with a relatively light
Scania Truck, which was the only truck available. The truck weighed approximately
11 tons, which was accounted for when calculating the controller velocity trajectory.
The driver was unaware of the purpose of the test and not influenced in any way.
By minimising the bias of the experiment, the velocity trajectory is recorded and
depicted in Figure 6.2.

100
vdriver
90 vlegal

80

70
Velocity [km/h]

60

50

40

30

20

10

0
0 2000 4000 6000 8000 10000
Distance [m]

Figure 6.2. Velocity trajectory produced by an experienced driver on the test-road.

The velocity trajectory from the driver was simulated in STARS. Naturally
6.1. SIMULATION 41

the input was not allowed to exceed the maximum posted road velocity as the
driver chose to do at times. The simulation resulted in a fuel consumption of 2.49
liters, 6053.8 kJ produced brake energy, and a travelled time of 700.52 s. This
implies that the velocity trajectory produced by the switching controller reduced
the fuel consumption by 4.8%, and the brake usage by 16%, while only increasing
the travelling time by 1.5%.
The preceding results were obtained by comparison with an experienced driver
who knew the oncoming road topology and posted speed limits. Therefore a slight
bias in the driving behaviour must be acknowledged. An inexperienced driver would
try to reduce the speed comfortably according to the sighted oncoming speed limi-
tations. Thus an inexperienced driver is simulated in STARS by creating an input
velocity trajectory, which only has a constraint of comfortably reducing the speed
according to the maximum allowed deceleration.

80
vcurve
vmin
70
vref
vDriver
60
Velocity [km/h]

50

40

30

20

10
0 2000 4000 6000 8000 10000 12000
Distance [m]

Figure 6.3. Velocity trajectory produced by an inexperienced driver profile on the


test-road.

Figure 6.3 shows the velocity trajectory used in STARS. A speed adaptation man-
dated by the road constraints is not commenced until deemed absolutely necessary.
Such driver behaviour is very common for inexperienced drivers who do not decel-
erate until the posted speed limit or curve is sighted, due to the lack of Look-Ahead
information. The total fuel consumption in this case was 2.59 litres. The brake wear
had increased to 7650.5 kJ and the total travelling time was 698.98 s. According
to previous results, the switching controller consequently reduced fuel consumption
by 8.5%, brake wear by 33.9%, while merely increasing the travelling time by 1.7%,
which is a substantial improvement.
The vehicle used in this experiment was relatively light. Usually commercial
trucks weigh between 11-60 tons. The lower limit of 11 tons represents a truck with
no cargo and the upper limit of 60 tons being a truck with full cargo. A heavier
42 CHAPTER 6. ANALYSIS

truck will be able to coast during longer stretches, since a downhill road section will
result in a larger accelerating gravitational force. Thus, the fuel consumption and
brake usage could probably be decreased further for heavier vehicles.

vupper limit
80 vController
vDriver

70

60
Velocity [km/h]

50

40

30

20

2000 3000 4000 5000 6000 7000


Distance [m]

Figure 6.4. A segment of the vehicle’s velocity trajectory for both the driver and
the switching controller.

It can clearly be seen in Figure 6.4 that the velocity trajectory produced by the
controller is very similar to what an experienced driver would use. The simulated
optimal velocity trajectory given by the designed controller suggests almost identical
maximum speed in the curves. Therefore, a maximum lateral acceleration of 0.2g
proved to be a very reasonable constant. However, not having access to the same
extent of Look-Ahead information, the driver cannot predict a proper fuel cut-off
point. Therefore the switching controller allows for a reduction of fuel consumption
and brake wear.
Figure 6.4 also reveals the implemented hysteresis band of sort to filter negligible
disturbances. It has also been applied due to the need of a failsafe, precautionary
device, so that the speed never reaches dangerous speed-limits. If the topology
differs from the estimated electronic horizon, the velocity is allowed to drift 4 km/h
away from the set reference, which can be observed between the 4500 meter and
the 5000 meter markers. Such a band must also be invoked due to the fact that the
controller would otherwise send erroneous control signals. A slight incline in the
topology might result in a minor decrease of the current set reference speed. If a
slight decline is followed, the speed would increase and might exceed the reference
speed to a minor extent. The lack of a hysteresis band would invoke the controller
to compel the vehicle to accelerate in the first case and brake in the latter, which
implies a heavy strain on the controller, unnecessary fuel consumption, and an
increased wear on the brakes in the long run. Due to the uneven nature of the
roads, slight fluctuations in the speed are considered to be negligible disturbances.
Past experience discussed in 2.3, showed that there were delays in the control
6.1. SIMULATION 43

signal due to physical limitations. Consequently, a Look-Ahead actuating action


is implemented. When facing a speed reduction, the reference signal acts an ad-
justable time ahead. In this study, the reference signal was set to act 2 seconds
ahead of time to account for the delays. No Look-Ahead action is applied to the
reference signals demanding an increase of speed, since such an action would force
the vehicle to transgress the maximum allowed speed.

6.1.2 Road 225


To further investigate the efficiency of the switching controller, experiments were
conducted on a less winding and actual Swedish commercial road. The road is
frequented by HDVs and runs between Södertälje and Nynäshamn. It is depicted
in Figure 6.5.

18000

16000

14000

12000

10000

8000

6000

4000

2000

0
0 1000 2000 3000 4000 5000 6000 7000 8000

Figure 6.5. The figure shows the two road segments experimented on Road 225.
The road segment with the starting point marked by a ♦ and a stop point marked by
a  is referred to as Road 1. The road segment with a starting point marked by a ×
and a stop point marked by a is referred to as Road 2. Road 2 is a slightly shorter
but nevertheless same road segment as Road 1. However, the travelled direction is
opposite for the two runs.

Longitude, latitude, driver speed, road grade, and altitude amongst other data
were recorded when these experiments were conducted on the roads above. The
data collected is naturally distorted with noise of varying nature. It is therefore
filtered through a low pass filter and a cubic interpolation process. A curve radii es-
timation described in 5.2.1 is conducted on the recorded electronic horizon. Hence,
all the input parameters have been formed to calculate a velocity trajectory from
the switching controller. The results when simulating the velocity trajectories along
44 CHAPTER 6. ANALYSIS

with the measured topology in STARS are listed in table 6.1.

Table 6.1. Simulation results for the switching controller velocity trajectory applied
on the two testroads.

Fuel Consumption [l/10 km] Brake Wear [kJ] Total travelling time [s]
Road 1 1.93 4084.4 1454.36
Road 2 1.74 4987.1 1116.39

Travelling a total distance of 26.08 km on Road 1 and a total distance of 20.42 km


on Road 2, the velocity kept by the driver was also recorded and used as an input
to STARS. The driver was simply told to maintain the legal speed limits and drive
as intuitively as possible. The results are presented in table 6.2.

Table 6.2. Simulation results for the recorded experienced driver velocity trajectory
derived from the two testroads.

Fuel Consumption [l/10 km] Brake Wear [kJ] Total travelling time [s]
Road 1 2.02 4568.6 1500.92
Road 2 1.71 4837.5 1144.5

The results from Road 1 show that the controller decreased the fuel consumption by
4.4%, the brake wear by 10.6%, and the travelling time by 3.1%. An approximate
1 dl/10 km reduction in the fuel consumption is a vast amount. The brake wear
is also reduced significantly. Thus, a mere increase of the travelling time by 46.56
seconds during a 26.08 km stretch implies that the fuel reduction is more than just
a reduction caused by lowering the average speed. The simulation for Road 2 on
the other hand resulted approximately in a 1.7% increase in fuel consumption, a
3% increase of brake wear, but a 2.5% decrease in the total travelling time. Road 2
compared to Road 1 has steeper inclines and no posted speed changes, which shows
the limitations of the controller. Going in the opposite direction and being familiar
with the road topology, the driver maintained a higher speed throughout the curves.
The driver also let the vehicle drop in speed when travelling inclines and picking up
the speed again in downhills, even though the posted speed was set to a constant
70 km/h. Such driving behaviour saves fuel consumption. Upon interviewing other
drivers at Scania CV AB, they confessed that such driving behaviour would be more
economical, but is never applied since they believe that maintaining a constant
speed regardless of the topology is a more comfortable behaviour. Therefore, the
inexperienced driver profile is simulated, which embodies such driver behaviour.
Table 6.3 clearly shows that the switching controller produces a more economical
driver behaviour compared to an inexperienced driver or a careless experienced
driver. Such a driver would maintain the posted speed as long as possible and only
6.1. SIMULATION 45

Table 6.3. Simulation results for the inexperienced driver profile velocity trajectory
applied on the two testroads.

Fuel Consumption [l/10 km] Brake Wear [kJ] Total travelling time [s]
Road 1 2.6 6659.9 1412.2
Road 2 1.88 7574.8 1101.8

reduce it comfortably when a speed constraint is detected. The switching controller


have reduced the fuel consumption by 25.8%, the brake wear by 38.7%, and the
travelling time is increased by 3%, for Road 1. Also according to table 6.3, the
switching controller have reduced the consumption by 7.45%, the brake wear by
34.2%, and increased the total travelling time by 1.3%. Unlike the experienced
driver, the inexperienced driver would not like to drop below the posted speed limit
even if it is warranted by the road topology, causing an increased fuel consumption.
The controller will likewise keep a constant speed and not calculate an appropriate
velocity trajectory unless a constraining curve or a posted speed decrease arises.
Therefore, the vehicle will brake on straight stretches upon facing a downhill or
accelerate during an incline to maintain the reference speed, which will cause higher
fuel consumption and an increase of brake wear.
During the run on Road 2, traffic in front of the driver slowed down before ex-
iting the road, which compelled the driver to slow down when there was no other
reason to slow down. The switching controller cannot account for such external
factors. It should be noted that Road 2 had far more disturbances than Road
1. It also entailed driver behaviour which is not commonly adapted by commer-
cial drivers. Road 2 demonstrated certain aspects which the switching controller
was not designed to handle. The fuel consumption and brake wear were increased
slightly as shown in the results from the experienced driver. However, the travelling
time was also reduced, resulting in no substantial loss or gain. The results from
the experienced driver in table 6.2 and the inexperienced driver in table 6.3, clearly
show how sensitive the fuel consumption and brake wear is to different driver be-
haviour. Thus the results concur with the hypothesis of this thesis, i.e. adjusting
the speed with respect to the road topology should result in a significant fuel re-
duction. It becomes evident that an efficient velocity trajectory, which is produced
in a favourable manner as suggested by the switching controller, can have a great
effect on the fuel economy and brake wear of the vehicle. Even though the driver
used fuel consumption reduction techniques which the controller cannot mimic, an
effective controller strategy of proper fuel cut-off point combined with coasting and
braking still proved to be efficient, which is evident from the results of Road 1.
46 CHAPTER 6. ANALYSIS

6.2 Sensitivity & Robustness Analysis


The switching controller has proven to be very effective. Results presented in 6.1.2
clearly demonstrate fuel savings and brake wear reduction. However, the controller
is based upon ideal input data. The results from the comparison between the
switching controller’s velocity trajectory and the recorded experienced driver’s ve-
locity trajectory are summarised in Table 6.4 for the convenience of this section.

Table 6.4. Simulated results from the comparison between the switching controller’s
and the recorded experienced driver’s velocity trajectories derived from the test-road.

Switching Fuel Consumption [%] Brake Wear [%] Total increase of [%]
Controller Reduction Reduction travelling time
Test-Road 4.8 16 1.5

In reality the input data is always contaminated with some form of noise. Hence, a
natural succession is to perform a sensitivity and robustness analysis.

6.2.1 GPS-Error
The current means of measuring road grade and exact GPS-position require im-
provement. The position acquired in GPS-navigation needs a clear signal from at
least three satellites to correctly estimate a three dimensional position fix. Basically,
positioning with the GPS system involves receiving a signal from the satellites with
information regarding the current time at the satellites [1]. The speed of the signal
is known; hence the distance from each satellite can be estimated. However, there
is a slight offset in all satellite clocks due to physical properties, and therefore a
fourth satellite is required to calibrate the clocks. Thus, three satellites are used
to calculate a horizontal position plus altitude while the fourth satellite is used to
synchronise the receiver’s clock with the satellites’ atomic clocks. Due to delays in
processing the transmitted signal of 2-3 seconds, the vehicle could have travelled
10-40 meters depending on the velocity of the vehicle. The density and the compo-
sure of the atmosphere also create a varying resistance for the transmitted satellite
signal. Therefore, the velocity of the signal might differ from the expected value.
A delay might cause the vehicle to believe that a downhill is approaching, when
it in fact already arrived. The vehicle would pick up speed unexpectedly, forcing
the controller to initiate the brakes. Similarly, if the electronic horizon predicts an
upcoming road incline and the vehicle already commenced the uphill due to de-
lay, the vehicle would accelerate to maintain the reference speed. Brake wear and
fuel consumption would increase due to the delay. Hence, a delay is induced in
the sampling for the test-road presented in 6.1.1, simulating the effect of an actual
delay.
A delay of one sample in Table 6.5 corresponds to approximately 10 m, de-
pending on the vehicle speed. It is evident that a delay will cause the reduction
6.2. SENSITIVITY & ROBUSTNESS ANALYSIS 47

Table 6.5. Simulation results for the applied delay disturbance.

Delay Fuel Consumption [%] Brake Wear [%] Total increase of [%]
Reduction Reduction travelling time
1 Sample 3.6 17 0.53
2 Sample 3.23 11.2 0.44
3 Sample 4.5 18 0.5

in fuel consumption to decrease and a higher average velocity will be maintained,


reducing the travelling time. However, the controller is fairly insensitive to delay
disturbances. This is probably due to the hysteresis band, which allows the actual
velocity to differ 4 km/h from the reference velocity. Therefore, no action from the
engine or the brake system is induced, i.e. the delay is compensated. It should be
noted that the results in the experiment for 3 samples, i.e. an approximate delay of
30 meters, actually increased the brake wear reduction. This can only be explained
by the produced controller velocity trajectory serendipitously mimicking the eco-
nomical driver behaviour discussed in 6.1.2. By delaying the road ahead, the vehicle
have in this case maintained a lower speed in many cases during the road inclines
and later demanded acceleration during the downhills, allowing for a reduction in
brake wear and strain on the engine. It is a unique case and highly circumstantial,
and the general effect is naturally a degeneration of fuel and brake reduction as the
delay increases.

6.2.2 Drift
6.2.1 states that a minimum of 4 satellites are required to give a proper altitude
triangulation. As the vehicle travel along a given route, the satellites also move in
orbit. Satellite signal qualities will change over time and the set of active satellites
may gain or lose members. When new members are acquired, a new triangulation
is made. Other circumstances, such as a clear line of sight, might be blocked due
to the surrounding environment. Similarly, line of sight might be cleared for other
satellites in orbit. The new satellites are often positioned to give a more accurate
triangulation and a sudden change of altitude position is observed.
It can give rise to a "Drift" in observed altitude as the GPS filtered output adapts
to a change in the set of source satellites. This is illustrated in Figure 6.6. Figure
6.6 shows the result of a modelled change of satellite constellation at the 2100 m,
4200 m, and 6200 m marker. It is a source of error since the road grade differs from
the actual value along the drift. Therefore the drift phenomenon is simulated by a
percentage increase in digression from the true altitude.
Table 6.6 clearly shows that the controller is very sensitive to the drift distur-
bance, unlike from the delay disturbance. A drift of 38.5m during the 1 km long
stretch actually results in an increase of brake wear and fuel consumption. Lowering
the drift to 7.7 meters resulted in a fuel reduction degeneration of approximately
48 CHAPTER 6. ANALYSIS

Actual Altitude
Drift Altitude

90

85

80
Altitude [m]

75

70

65

1000 2000 3000 4000 5000 6000


Distance [m]

Figure 6.6. The figure shows a model of real altitude and the measured altitude
during drift.

Table 6.6. Simulation results for the drift disturbance.

Drift during a Fuel Consumption [%] Brake Wear [%] Total increase of [%]
1km stretch [m] Reduction Reduction travelling time
0.8 2.47 9.45 1.55
3.85 2.74 13.2 1.52
7.7 2.82 6.55 1.71
38.5 -1.56 -5.66 0.46

40%. Such a magnitude in drift is not uncommon and could possibly present a
problem. The drift in altitude could strongly affect the road grade depending on
what procedure is used for estimation. Therefore, the method of grade estimation
is of great importance for the designed switching controller.

6.2.3 Grade Error


Studies have conducted within the field of grade estimation on travelled roads. Both
online and offline methods are researched and tested at the moment. A road grade
estimation algorithm for fusion of GPS and vehicle real-time sensor data, with mea-
surements from previous runs over the same road segment have been applied in [13].
That approach has shown success compared to known road grade measurements.
However, there is still enough error and uncertainty which might serve as a cause
for alarm in the switching controller.
The nature of the disturbance is unknown. However, each measurement is as-
sumed to be uncorrelated in this study. Thus the noise is modelled as random vari-
6.2. SENSITIVITY & ROBUSTNESS ANALYSIS 49

ables with mean µ and standard deviation σ. The random variables are presumed
to be independent and the distribution function is also presumed to be identical.
The Central Limit Theorem therefore states that the sample mean is asymptotically
Gaussian. To give an understanding of to what extent the noise affects the con-
troller, it is modelled as Gaussian white noise. Various magnitudes of the variance,
σ 2 , are investigated to deduce the sensitivity for the switching controller.

Table 6.7. Simulation results for the Gaussian White Noise disturbance applied to
the road grade for various standard deviations (σ).

Gaussian Fuel Consumption [%] Brake Wear [%] Total increase of [%]
White Noise Reduction Reduction travelling time
σ = 10−3 3.17 12.12 1.46
σ = 10−2.5 3.17 15.1 1.41
σ = 10−2 2.6 9.95 1.26

The results in table 6.7 are derived by adding a matrix containing pseudo-
random values drawn from a normal distribution with mean zero and standard
deviation σ. The switching controller seems unexpectedly, to be robust to Gaussian
road grade related errors. It can be deduced that adding a Gaussian white noise with
a standard deviation of 10−3 , the fuel consumption reduction is decreased to 3.7%.
A reduction of 3.7% corresponds to a degeneration of the original saving by 23%.
Increasing the standard deviation to 10−2.5 , does not affect the fuel consumption
significantly. However the brake wear reduction continues to degenerate. Finally
applying a Gaussian white noise with standard deviation of 10−2 , which is the same
order of magnitude as the actual grade, an approximately 50% degeneration in both
fuel and brake reduction is observed in the simulated results. Hence the controller
is fairly insensitive to road grade errors of the applied nature, which could possibly
explained by the HDV acting as a low pass filter due to its inertia, i.e. physical
reluctance to instantaneous changes. However, it is nonetheless important to derive
an accurate road grade estimation method, to obtain efficient controller behaviour.
Chapter 7

Conclusion

The objective of this master’s thesis was to design a controller that addresses fuel
consumption, travelling time, driver comfort, and brake wear. A switching controller
with switching rules based on analytical optimisation and engineering experience
controller was designed, since the constraints were hard to implement in a pure an-
alytical controller algorithm. Although the current means of measuring road grade
and exact GPS-position still require improvement, supplying an electronic horizon
has proven to be very effective. The switching controller is very versatile and intu-
itive. Several comfort parameters that were derived empirically, e.g. deceleration
comfort level, lateral acceleration amongst others, can easily be adjusted, allowing
for the driver to personalise the controller settings. Fuel consumption and brake
wear have been reduced significantly relative to the time loss due to a lower average
speed, which was established in the simulations that were made.
The simulations were conducted with a very light truck due to availability. A
heavier truck would allow for increased coasting and thereby further reductions in
the fuel consumption and brake wear. Thus the reduction percentages obtained in
the experiments conducted on country roads in this study could most probably be
increased significantly by applying the controller to a heavier vehicle.
Good fuel economy is one of the top priorities when purchasing a new vehicle.
Recent findings in the Scania Product Identity report show that the fuel cost con-
stitutes approximately one third of total operational cost in European long haulage.
Hence, saving 4-10% fuel through the switching controller will have a tremendous
effect on the fuel economy. Repair costs will also decrease as a result of minimising
brake usage through the velocity trajectory produced by the controller. The vast
strain and tear that is exerted on the HDV will also be reduced by diminishing
the brake application. Subsequently the overall lifespan of the heavy duty vehicle
will increase. The switching controller especially invokes two means for good fuel
economy. Firstly it induces an intuitive, comfortable, and very efficient economic
driving style. Secondly the utilisation of the kinetic energy produced or consumed
by the topology along with the vehicle properties is maximised with respect to the
constraints set in this master’s thesis. Safety is also increased by adapting the speed

51
52 CHAPTER 7. CONCLUSION

in time to the upcoming curves and posted speed limits, which effectively turns one
aspect of the switching controller to a curve speed warning system.
Owing to the fact that this study was successful in obtaining its objectives, it
still has to be kept in mind that the switching controller is sensitive to drift in road
grade. The input signal has to entail road grade information with relatively small
drift magnitude to obtain satisfactory controller behaviour. However, if the input
signal is properly delivered, the switching controller designed in this master’s thesis
can serve as strong platform to future control.
Chapter 8

Future Work & Extensions

It was duly noted in the driver simulations experiments conducted in this study, that
fuel consumption can further be reduced by maintaining a relatively lower speed
during road inclines and then letting the vehicle gain speed later during declines.
Therefore, an interesting field of research for the continuation of this study would
be to further develop the switching controller to account for such situations. Adding
such a feature to the versatile controller should not be difficult in practicality.
The switching controller can produce an automatic speed control through the
use of Look Ahead information. However, it is not currently designed to account for
any other external factors than the topology, curvature, and posted speed limits.
Hence, integrating the controller with the current adaptive intelligent cruise con-
troller (AiCC) at Scania CV AB, could result in a fully automated system in the
long run. However, currently the controller could serve as an advanced ECO-drive
if combined with a proper interface.

53
Bibliography

[1] Ahmed El-Rabbany. Introduction to GPS - The Global Positioning System,


Second Edition. Artech House, Inc, 2006.

[2] Gerd Eriksson. Numeriska Algoritmer med MATLAB. NADA,KTH, 2002.

[3] Lars Eriksson and Lars Nielsen. Vehiclur Systems. ISY Linköping Institute of
Technology, 2006.

[4] Anders Fröberg et al. Explicit fuel optimal speed profiles for heavy trucks on
a set of topographic road profiles. 2006.

[5] Kenneth Yoshioka et. al. Improvements in cornering safety through deceleration
control and road preview.

[6] Ulf Jönsson et. al. Optimal control, 2006.

[7] V. Aguilera et. al. An advanced driver speed assistance in curves: risk function,
cooperation modes, system architecture and experimental validation. 2005.

[8] Wilhelm Vogt et al. Navigation-based driver assistance systems.

[9] E. Hellström. Look-ahead Control of Heavy Trucks utilizing Road Topography.


PhD thesis, Licentiate thesis LiU-TEK-LIC-2007:28, Linköpings universitet,
2007.

[10] Erik Hellström. Look-ahead control of heavy trucks utilizing road topography.
Master’s thesis, Linköpings universitet, Department of Electrical Engineering,
Linköpings universitet, SE-581 83 Linköping, Sweden, 2007.

[11] Sanjay Lall. Engr 207 modern control, 2006-2007.

[12] Per Sahlholm, Henrik Jansson, Magnus Östman, and Karl Henrik Johansson.
Automated speed selection for heavy duty vehicles. 2007.

[13] Per Sahlholm, Ermin Kozica, and Karl Henrik Johansson. A sensor and data
fusion algorithm for road grade estimation. 2005.

55
56 BIBLIOGRAPHY

[14] Tony Sandberg. Heavy Truck Modeling for Fuel Consumption, Simulations,
and Measurements. PhD thesis, Licentiate thesis LiU-TEK-LIC-2001:61,
Linköpings universitet, 2001.
Appendix A

Appendix - LQR Mathematical


Derivation

Vs (z) = min(z T Qz + wT Rw + ż T Q1 ż + Vt+1 (Az + Bw)) (A.1)


w
We know that the final cost in the cost function is:

Vsf = z T Psf z (A.2)


where Psf = Qf

Thus, Dynamic Programming gives:

Vsf −1 (z) = z T Qz+min(wT Rw+(Az+Bw)T Q1 (Az+Bw)+(Az+Bw)T Psf (Az+Bw))


w
(A.3)
Differentiate equation above with respect to w = ũ + u0 in order to minimise:

2wT R + 2(Az + Bw)T Q1 B + 2(Az + Bw)T Psf B =


=2wT R + 2(Az + Bw)T (Q1 + Psf )B = (A.4)
T T
=2w R + 2(Az + Bw) (P̃sf )B = 0

⇒ wopt = −(R + B T P̃sf B)−1 B T P̃sf Az (A.5)

⇒ Vsf −1 (z) = z T Qz + wT opt Rwopt + (Az + Bwopt )T P̃sf (Az + Bwopt ) =


= z T (Q + AT P̃sf A − AT P̃sf B(R + B T P̃sf B)−1 B T P̃sf A)z = (A.6)
T
= z P̃sf −1 z

57

You might also like