Electric Vehicle Design

Electric Vehicle Design, Racing and Distance to Empty Algorithms
by
Lennon Patrick Rodgers

ARCHIES
B.S. Mechanical Engineering MASSACHUSETTS INs fliE
University of Illinois Urbana Champaign (2003) OFTCNOLOGY
M.S. Mechanical Engineering AJ 25

Massachusetts Institute of Technology (2006)
ARIES
Submitted to the Department of Mechanical Engineering
in Partial Fulfillment of the Requirements for the Degree of
Doctor of Philosophy in Mechanical Engineering
at the
MASSACHUSETTS INSTITUTE OF TECHNOLOGY
June 2013
0 2013 Massachusetts Institute of Technology. All rights reserved.
A uth o r............................................................................ . -............ .-:................................................

Department of Mechanical Engineering
) ,14 May 14 th, 2 0 13
C ertified by ............................................................. ..................... ..........................

Daniel D. Frey
Professor of Mec anical Engineering
Thesis Supervisor and Chair
Accepted by....................................................
David E. Hardt
Professor of Mechanical Engineering
Chairman, Committee on Graduate Students
1
2
Electric Vehicle Design, Racing and Distance to Empty Algorithms
by
Lennon Patrick Rodgers
Submitted to the Department of Mechanical Engineering

on May 14 th, 2013 in Partial Fulfillment of the
Requirements for the Degree of Doctor of Philosophy in
Mechanical Engineering
Abstract
This research began with the goal of designing and building an electric motorcycle to
compete in the Isle of Man TT Zero race. A set of parametric physics-based models was derived
to size the batteries and motors, predict vehicle speeds and predict the time required to finish the
race. In June 2011 the motorcycle design and simulations were tested in three races on the Isle of
Man. Post-race analysis showed that the predictions had less than 10% error.
The energy estimation methods that were developed for the motorcycle were
subsequently modified and applied to non-racing electric vehicles. Instead of predicting the
energy required to travel a known route, it is more useful for non-racing applications to consider
the reverse scenario, which is the distance the vehicle can travel before charging is required. This
is referred to as the Distance to Empty (DTE). Recent studies have shown that current DTE
algorithms are inadequate and cause "range anxiety" among users. This is because conventional
approaches only use past driving data to estimate DTE and thus are unable to accurately predict
changes in driving conditions. However, the algorithm developed in this thesis uses
measurements from the past along with knowledge of the future route. A multivariate linear
regression model is used to adjust a historical average of energy consumption based on estimated
changes in speed, traffic and temperature.
Finally, the new DTE algorithm was compared to conventional methods by simulating a
large number of full battery discharges under realistic driving conditions. A Markov-based
stochastic speed profile generator was used as input to the models. Example simulations show
that including future driving conditions in the DTE algorithm can significantly reduce error.
Thesis Supervisor: Daniel D. Frey

Title: Professor of Mechanical Engineering
3
4
Acknowledgments
This work would not have been possible without the support of my committee members
Dan Frey, Sanjay Sarma and Warren Seering. I am deeply appreciative of their guidance and
friendship. I would also like to thank the following members of the Electric Vehicle Team who
spent many late nights working on the motorcycle: Will Pritchett, Radu Gogoana, Mark
Jeunnette, Randall Briggs, Paul Karplus, Erick Fuentes, Manyu Belani, Dan Kelly, Richard Yoon
and Romi Kadri. Shane Colton was also a key team member, patient teacher and friend. Our
professional motorcycle rider Allan Brew risked his life for the sake of engineering while also
being a true gentleman. His wife Jan warmly hosted us on the Isle of Man with an unlimited
amount of English tea. Mark Gardiner and Bill Dube were important advisors to the team and I
am grateful for their help. Tom German shared his racing insights, which were captured in this
thesis. Yet-Ming Chiang's unwavering support kick-started the motorcycle project and inspired
us in many ways. The following organizations supported this research financially or with
equipment: Singapore University of Technology and Design (SUTD), MIT-SUTD International
Design Center (IDC), A123 Systems, BMW, MIT Energy Initiative (MITEI), MIT
Transportation, MIT Department of Mechanical Engineering, MIT School of Engineering, MIT
Edgerton Center, the Isle of Man government, Rahn's Motorcycling Engineering and Boston
Moto. The following colleagues were a major help in my Distance to Empty algorithm work:
Erik Wilhelm, Stephen Zoepf, Don MacKenzie, John Heywood and Sriram Krishnan. It was a
great honor to work with these talented friends.
And last but not least - a special thanks to my wife Jenn and other family members for their
enduring support.
5
6
Table of Contents
PART I: DESIGNING AN ELECTRIC MOTORCYCLE FOR THE ISLE OF MAN TT ZERO RACE.........11
1. INTRODUCTION TO PART I................................................13

1.1 MOTIVATION...........................................................................................................................................-.- - .... 13
1 .2 O V ERV IEW ...................................................................................................................................... .. ------------------........ 13
1.3 RELATED W ORK..............................................................................................................................--.. 14

------------........
2. M OTORCYCLE SYSTEMS ENGINEERING ..................................................................................... 15
2.1 SUBSYSTEM MODELS...........................................................................................................................----------...... 16
2.1.1 Backward-lookingSimulations .................................................................................................................. 16
2.1.2 Forward-lookingSimulations........................................................................................................................ 21
2.2 FULL-THROTTLE SIMULATIONS .........................................................................................................---................... 24
2.3 CONSTANT VEHICLE SPEED SIMULATIONS ............................................................................................................ 26
2.3.1 Proofthat ConstantSpeed Minimizes Energy Consumption ..................................................... 26
2.4 CONSTANT POWER SIMULATIONS ..........................................................................................................----......-----. 29
2 .5 B RA KIN G LO SSES.........................................................................................................- - - -.... --....------------............. 3 0
2.6 ESTIMATING THE REQUIRED MOTOR POWER AND BATTERY CAPACITY ..................................................... 31
3. M OTORCYCLE DESIGN AND TESTING.............................................................................................. 33
3.1 ELECTRICAL DESIGN....................................................................................................-----.....---------..................... 34
3.2 M ECHANICAL DESGNIGN ..............................................................................................................--.......--------............. 36
3 .3 T ESTIN G.................................................................................................................... .. --------......--------.................... 38
4. M OTORCYCLE RACE RESULTS AND ANALYSIS.............................................................................. 43
4.1 ANALYZING THE MOTORCYCLE PERFORMANCE................................................................................................ 43
4.2 MODEL VALIDATION .........................................................................................--......--........-----------.......................... 45
4.3 FRAMEWORK FOR SPURRING INNOVATION THROUGH RACING ..................................................................... 47
4.3.1 Consider the HistoricalContext.................................................................................................................... 48
4.3.2 Utilize the Powerof Regulation....................................................................................................................48

4.3.3 Drive Technology ...............................................................................................-------------------------............... 48
4.3.4 Provide Valued Entertainment..................................................................................................................... 48
4.3.5 Inspire Consumer Demand.......................................................................................................................------- 49
PART II: ESTIMATING AN ELECTRIC VEHICLE'S DISTANCE TO EMPTY....................................... 51
7
5. IN TRO D U CTION TO PA RT II............................................................................................................. 5 3
5.1 M OTIVATION................................................................................................................................................................54
5.2 OVERVIEW .................................................................................................................................................................... 54
5.3 RELATED W ORK .......................................................................................................................................................... 55
6. STOCHASTIC VEHICLE SIMULATIONS FOR EVALUATING DTE ALGORITHMS.....................57

6.1 SIMULATION ARCHITECTURE ................................................................................................................................... 57
6.2 STOCHASTIC SPEED PROFILE GENERATOR ....................................................................................................... 5 8

6.2.1 Raw Driving Data............................................................................................................................................... 59
6.2.2 Categorizing and Grouping Data................................................................................................................. 59
6.2.3 M arkov M odel....................................................................................................................................................... 62
6.2.4 Simulating Highway, City and Aggressive Driving Conditions.................................................. 63
6.2.5 Creating a Stochastic Speed Profile............................................................................................................ 64
6.3 SUBSYSTEM M ODELS..................................................................................................................................................65
6.3.1 Vehicle M odel........................................................................................................................................................65

6.3.2 Transm ission M odel...........................................................................................................................................66
6.3.3 Motor M odel.......................................................................................................................................................... 66
6.3.4 Battery M odel ....................................................................................................................................................... 67
6.3.5 A uxiliary M odel....................................................................................................................................................67
6.4 GENERATING STOCHASTIC VEHICLE SIMULATIONS ......................................................................................... 68
7. FU N DA M ENTA L DTE CO N CEPTS......................................................................................................... 71

7.1 KEY DTE EQUATIONS .................................................................................................................................................. 71
7.2 M EASURING THE REMAINING BATTERY ENERGY ........................................................................................... 74
7.3 OTHER A PPLICATIONS W ITH A SIM ILAR FORMULATION................................................................................ 75
7.4 CONVENTIONAL DTE A LGORITHMS..........................................................................................................................75
7.5 USING ESTIMATES OF FUTURE CONDITIONS AND MODELS TO IMPROVE DTE PREDICTIONS ................... 79
8. A REGRESSIO N -BA SED DTE A LGO RITH M ....................................................................................... 83

8.1 M ULTIVARIATE LINEAR REGRESSION M ODEL .................................................................................................. 84
8.2 EXPLANATORY VARIABLES ......................................................................................................................................... 85
8.3 CREATING A TRAINING DATASET ............................................................................................................................ 86
8.4 A LGORITHM SIMULATION ......................................................................................................................................... 87
8.5 VALIDATING THE ASSUMPTIONS AND FIT OF THE REGRESSION MODEL ..................................................... 88
9. CO M PA RIN G D TE A LGORITH M S........................................................................................................ 91
8
9.1 M EASURING THE PERFORMANCE OF D TE A LGORITHM S .................................................................................. 91
9.2 RESULTS ....................................................................................................................................................................... 92
PART III: SUMMARY AND CONTRIBUTIONS ......................................................................................... 97
10. CO NCLU SIO NS .......................................................................................................................................... 99

10.1 SUM MARY OF THESIS............................................................................................................................................... 99
10.2 SUM MARY OF CONTRIBUTIONS........................................................................................................................... 101
10.3 O NGOING WORK.................................................................................................................................................... 102
10.3.1 IncreasingA verage Speeds at the Isle of M an TT Zero................................................................. 103
10.3.2 A dvancing D TE A lgorithms......................................................................................................................... 104
10.3.3 Com paring DTE A lgorithms....................................................................................................................... 104
R EFEREN CES ................................................................................................................................................. 105
9
10
Part I: Designing an Electric Motorcycle for
the Isle of Man TT Zero Race
11
12
1. Introduction to Part I
Part I describes the process of designing, building, testing and racing a high performance
electric motorcycle for the 2011 Isle of Man Tourist Trophy (TT) Zero race. Systems
engineering was used to size the batteries and motors, predict vehicle speeds and predict the time
required to finish the race. Forward- and backward-looking vehicle simulations were used to
predict the motorcycle performance under full-throttle conditions and estimate the total battery
energy required to traverse the entire course. The mechanical and electrical designs are described
along with the critical process of incremental testing. In June 2011 the motorcycle design and
simulations were tested in three races on the Isle of Man. Post-race analysis showed that the
predictions had less than 10% error.
1.1 Motivation
The Isle of Man Tourist Trophy (TT) race is the oldest existing motorcycle race and for
over 100 years has served as a proving ground for both riders and engineers to advance
motorcycle technology. Soichiro Honda, founder and then president of Honda Motor
Corporation, once declared that the innovation required to win the Isle of Man race would "rank
at the world's highest levels of engineering" [1]. The desire to create innovation through racing
continues today and in 2009 an electric class was added to the TT with the aim of advancing zero
emission vehicle technology. The Isle of Man TT Zero is an example of a new breed of "zero
emission" races, which aim to spur innovation that will reduce the environmental impact of
consumer vehicles. Racing has historically been a catalyst for innovation, particularly in the
early years of motorcycles and automobiles [2]. New concepts were tested on the track and the
desire to win drove companies to produce superior technology. Consumer demand for better
performance motivated companies to transfer the technology from the racetrack to the mass
market.
1.2 Overview
Systems engineering is used to ensure feasibility early in the design process, select the
proper motor and battery size, and predict the motorcycle performance in Chapter 2. Both the
13
electrical and mechanical subsystem designs and testing methods are described in Chapter 3.
Finally, the race data is used to evaluate the motorcycle performance and validate the integrated
models in Chapter 4. The final chapter also outlines a set of guidelines for designing zero
emission races with the aim of promoting innovation.
1.3 Related Work

The Isle of Man TT Zero race in 2009 served as the first ever large-scale electric
motorcycle race. Since the field is still in infancy, there is very little engineering literature on the
topic of designing electric motorcycles for racing. There is no literature on predicting energy
requirements for the Isle of Man race since the gasoline motorcycles have never been energy
limited. A university in Germany built an electric motorcycle for the TTXGP in 2012 and
published a paper on their design, but they did not describe the methods used to size the motor
and battery [3]. Two recent studies describe a motor controller [4] and motor [5] specifically
designed for electric motorcycle racing.
Though there are few publications related to electric motorcycle racing, there is an
extensive amount of literature for solar vehicle racing. Solar vehicle racing started in the mid
1980s and includes teams from international corporations and universities. Similar to electric
motorcycle racing on the Isle of Man, one of the central challenges in solar racing is to predict
the required battery energy and maximize vehicle performance. The problem of optimally
controlling the vehicle to minimize energy has been studied for solar cars [6][7] and more
generally for electric vehicles [8][9]. These methods could be applied to the TT Zero race,
though extensions would be required to include coming (braking). Also, it is questionable that
optimal control would yield large energy savings. More recent work describes designs for the
hybrid-electric and all-electric Formula Society of Automotive Engineers (FSAE) competitions
[10]. The most common approach is to simulate the vehicle using commercially available
software along with custom models as needed. For example, commercially available software
has been used to optimize a hybrid vehicle drivetrain [11] [12]. Finally, there has been a large
amount of work on general vehicle and subsystem (motors, etc.) modeling [13][14].
14
2. Motorcycle Systems Engineering
This section describes how system-level engineering was used in designing the electric
motorcycle. More specifically it explains the methods used to explore how changes in the design
parameters affect performance metrics (Table 2.1). Systems engineering was used to size the
batteries and motors, predict vehicle speeds and predict the time required to finish the race. The
process involved deriving a set of subsystem models (Section 2.1) and then simulating various
scenarios based on specific inputs. Two types of simulations were considered. The first estimated
the motor and battery power, and vehicle speed of the motorcycle when full-throttle is applied
(Section 2.2). The second simulation estimated the battery energy required to traverse the entire
course. The conventional approach to estimating a vehicle's energy consumption is to simulate
the vehicle using an estimated speed profile as input (e.g. EPA "driving cycle" [15]). But the
speed profiles are difficult to predict in this case because the TT course is complex with over 200
turns, varying elevation and is traversed at fluctuating high speeds. Thus a new approach was
derived, which assumed that the required battery energy was in between estimates of lower and
upper bounds. The lower bound, or best case, was determined by assuming the motorcycle
maintained a constant vehicle speed throughout the course, which was shown to be the most
energy efficient way to drive. The upper bound, or worst case, was determined by assuming the
motorcycle maintained a constant motor power throughout the course, which would yield large
speed fluctuations and is thus inefficient. The lower and upper bounds are derived in Sections 2.3
and 2.4, respectively. Section 2.5 adds braking to the speed, time and energy estimates. Finally,
Section 2.6 explains how the methods described were used in practice to estimate the required
motor power and battery energy.
15
Table 2.1: Design parameters and performance metrics.
Design Parameters Performance Metrics
* Mass * Acceleration
e Aerodynamics * Maximum and average speed
e Rolling resistance e Battery energy consumed
e Sprocket ratio e Power (vehicle, batteries, motors)
e Finishing time
eBattery specifications
eMotor specifications
eMotor controller
specifications and settings
e Drivetrain efficiencies
2.1 Subsystem Models

The subsystem models derived here are used in later subsections to explore the
motorcycle performance based on design parameters (Table 2.1). When the speed profile is
specified as input, the simulation is referred to as backward-looking and when the throttle profile
is specified the simulation is referred to as a forward-looking. The following subsections
describe each of these approaches.
2.1.1 Backward-looking Simulations

A backward-looking simulation is shown in Figure 2.1. The speed is specified as a
function of time, i(t), and is the input to the vehicle model. Environmental Protection Agency
(EPA) "Driving Cycles" are commonly used to specify the speed [15], though this approach is
not applicable to racing since the speeds and general driving conditions are different from those
represented by the standard EPA specifications. A vehicle model determines the wheel force, F,
required to follow the prescribed speed profile based on a set of design parameters. Similarly, the
transmission and motor models determine how much electric power is needed by the motor, Pb.
The battery model translates the total electrical need to an estimated amount of battery energy
(Eb), which includes internal losses. These models are derived in the following subsections.
16
Vehicle Wheel Force Motor Torque Power from
Speed Vehicle Speed Motor Speed Battery
xi Vehicle Fw,; xi Transmission Tm,h m,i Motor Pb,i

I- Model Model 10Model
Start
Battery Model
Losses
Battery Losses
Loop Eb,i At
Battery Energy
Consumed during the
ith increment
Figure 2.1: A backward-looking vehicle simulation.
Vehicle Model
The electric motorcycle can be modeled as a mass on an incline with externally applied
forces (Figure 2.2). It is propelled forward by an electric motor acting through the rear wheel
with a force, F,, but slowed down by the aerodynamic drag, Faero, the rolling resistance from the
two wheels, Froiing, and the horizontal component of weight, Fgrainy. Newton's second law states:
F = mz = F, - Faero - Froing - Fgravity 2.1
Which can be written as [30]:
mx = F, - 1/2 pCdA i + w) 2 - mgCrrCosfl - mgSinfl 2.2
Where m is the total mass of the vehicle and rider, w is the wind speed, Crr is the rolling
resistance coefficient, p is the density of air, CdA is the aerodynamic coefficient, g is gravity, and
fl is the angle of the road. Solving for F,,:
F, = mx + 1/2 pCdA(i + w) 2 + mgCrrCos3 + mgSin3 2.3
The unknown parameters CdA and Crr can be found experimentally through wind tunnel and
road testing. The acceleration, Y, is determined by differentiating the known (specified) speed
profile.
17
4000- tO\VVN
Figure 2.2: The motorcycle can be modeled as a mass on an incline with externally applied forces.
Transmission Model
The force on the rear wheel, F., originates from a motor acting through a chain and two
sprockets. Summing the torques around the wheel yields (Figure 2.3):
T = Iww = Tw - FwGr 2.4
Where -r,, is the torque on the rear wheel, I,, is the mass moment of inertia and #w is the
rotational acceleration of the wheel(s). Assuming the wheel is the only rotating mass and is a
uniformly distributed cylinder of radius r,:
I = 2 2.5
Where m, is the total mass of all the wheels. The following relationship can be determined
through geometry:
6=- 2.6
Taking the derivative of Equation 2.6 yields:
Ow = - 2.7
Combining Equations 2.4, 2.7 and 2.8:
w 2 + Fw 2.8
18
Conservation of energy can be used to relate the wheel torque to the motor torque (Figure 2.3):
TWOw =1?dTmam 2.9
Where r7d is the transmission efficiency, which can be estimated or determined experimentally.
Solving for the motor torque:
$w1
Tm = Tw 6 m ?d.1 2.10
Defining the "sprocket ratio", z, as:
2.11
Combining Equations 2.8, 2.10 and 2.11:

=(y+ Fw w
Tm 2.12
2 )zr7d
And the motor speed can be determined by combining Equations 2.6 and 2.11:
.zx
6m =-- 2.13
Figure 2.3: A free body diagram of the rear wheel (left) and the entire drivetrain (right). The
torque on the motor results in a force on the wheel.
Motor Model
Conservation of power yields a relationship for the power going to the motor from the battery, Pb
(Figure 2.4):
TmOm
Pb = 77mc 17m 2.14
19
Where 1 mc is the efficiency of the motor controller, which is typically assumed to be constant
and ~90 to 95%. The efficiency of the motor, ?7m, can vary greatly based on the motor torque and
speed. A motor efficiency map relates the motor efficiency to torque and speed and is typically
obtained experimentally using a dynamometer [17]. Combining Equations 2.12, 2.13 and 2.14:
"M
+F .c
Pb= 2.15
77mclmd
ilmc Mechanical
Pb Power Out
Battery ]Motor Controll~er Motor

m ,0m
Figure 2.4: Energy flows from the Battery -> Motor Controller -> Motor -> Motor Torque.
Battery Model
A battery cell can be modeled as a resistor in series with a voltage source. The cells are
connected in parallel and series to form a battery module or pack (Figure 2.5). The battery losses
can be modeled as Ohmic:
Eb,boss = I Rb 2.16
Where Ib is the current through the battery pack and Rb is the total internal resistance, which can
be determined using:
k
Rb = -Rc 2.17
Where Re is the internal resistance of a single cell and k and I are the number of cells in series
and parallel, respectively (Figure 2.5). The power from the battery can be written as:
Vbnlb - Ib R = Pb 2.18
Where Vbfl is the voltage on the battery pack when no current is being drawn. The only
remaining unknown is Ib, which can be determined by solving Equation 2.18 using the quadratic
equation:
V E2 4Ro~
Vb n Pb 2.19
b - 2
Rb 2
Rb
20
The total power (including losses) from the battery can be written as:
V2 V
Vb
4R
2z 2 Rb Pb
V
l= 2.20
Pbn = V n
T a f gl 2 Rb 2 Rb
The total energy (including losses) from the battery can now be determined using:
Eb = f Pbn dt 2.21
Single Cell Module or Pack
T-
T- 1T-
1= number of cells in parallel

Figure 2.5: A battery cell was modeled as a resistor in series with a voltage source.
2.1.2 Forward-looking Simulations

The motor torque, Tm, for a DC permanent magnet motor is proportional to the current
passing through the motor:
Tm = KiIm 2.22
where K, is the torque constant, and I, is the motor current. Rearranging Equation 2.12 and
combining the result with Equation 2.22:
Z7dK My ..
F, = r KIm 2.23
rwTM 2 x
Now the acceleration of the vehicle can be determined by combining Equations 2.2 and 2.23:
z = Im ('IdKi,m - 'I2 pCdA(i + w) 2

- mgCrrCosfl - mgSin )
2.24
M+ W rw
The only unknown in Equation 2.24 is the motor current, In, which is based on the throttle input.
Once Im is known, the speed and position for the next increment can be determined using
numerical integration:
xi+1 = *i + zi(ti+1 - ti) 2.25
Xi+ 1 = Xi + !i+ 1 (ti+1 - ti) 2.26
21
DeterminingI, based on the throttle input is the focus for the remainderof this subsection.
A DC permanent magnet motor can be modeled as a voltage-generating source (Vemf) in
series with a resistor (Rm) (Figure 2.6). By summing the voltages around the motor model circuit,
the following relationship can be derived:
Im = (Vm - Vemf)/Rm 2.27
Veinf is proportional to the rotational speed of the motor, 6 m, which can be related to the linear
speed of the rear wheel, i, since they are connected. Rewriting Equation 2.27:
Vm - K8m Vm zKo
Im R R R X 2.28
Rm Rm Rmrw
M Mcotor
Mechanical
Power Out
VM Vemf
Tm A6 m
Figure 2.6: It was assumed that a Direct Current (DC) Permanent Magnet (PM) can be modeled as
a resistor in series with a back-EMF generating device.
As shown in Equation 2.28, the motor controller continually increases Vm as * increases
in order to maintain a constant I.. This is done until (1) the motor current reaches a maximum
value, IS"*, which is a value specified by the engineer based on performance and/or motor
thermal considerations, or (2) the motor controller reaches "all-on" conditions (100% duty
cycle). When (2) occurs, the batteries are essentially directly connected to the motor and thus In
= Ib, Vm V ' and I, decreases with increasing speed as shown in Figure 2.7 as a solid line. In
other words, the motor current is equal to the minimum value of these two scenarios:
Im = minimum{s -Immi, Ib 2.29
Scenarios (1) and (2) are referred to as current-limited" and voltage-limited, respectively.
' When the motor controller is at 100% duty cycle ("all on") there is still a small voltage drop across the transistors.
22
Vb zKO
Rm Rmnr,
Increasing Vr,.
sI limit Vb
Sm\
X (Speed)
Figure 2.7: This curve shows the relationship between vehicle speed (i) and motor current (Im). The
motor controller continually increases the motor voltage (Vm) as i increases in order to maintain a
constant Im. Eventually the back-EMF becomes large enough that the motor is voltage limited and
thus the current decreases with vehicle speed.
The exact values of Vb, I, V,, Im can be determined by simultaneously solving four
equations obtained from the models shown in Figure 2.8. First, conservation of power through
the motor controller yields:
lmcVblbb = VmIm 2.30
Second, summing the voltage drops through the motor model yields:
Vm - ImRm = K 0em 2.31
Combing this result with Equation 2.13:
zi
Vm - ImRm = K6 - 2.32
Third, summing the voltage drops within the battery model yields:
Vbn - IbRb = Vb 2.33
Thus the four equations are Equations 2.30, 2.32 and 2.33 and:
I = s -Itmi 2.34
Where s is the throttle signal and varies from 0 to 1. Equations 2.30, 2.32 and 2.33 can be solved
simultaneously again, but instead of using 2.34 for the forth equation, the following is assumed
(100% duty cycle of motor controller):
Im = Ib 2.35
The solutions for Vb, Ib, Vm associated with the minimum motor current are used (Equation 2.29).
The total energy consumed from the battery can now be determined:
Eb = f Vbnlb dt 2.36
23
Battery
+ |
V - Motor Controller em!
*- Im
Figure 2.8: Power electronics model consisting of the battery, motor controller and motor.
2.2 Full-throttle Simulations

It is common when studying dynamic systems to excite the system with a "step input" to
simulate how it will respond to a maximum input signal. In the case of motorcycle design, a
synonymous concept is referred to as a full-throttle simulation. It is assumed that the rider starts
at zero speed and applies full-throttle. The motorcycle's speed is then simulated all the way until
acceleration is zero, which is the point of maximum speed. It is assumed that the motorcycle is
on a straight road and does not experience braking. The forward-looking simulation described in
Section 2.1.2 can be used to simulate the motorcycle based on a throttle profile (a step input was
used here, though a saw-tooth, ramp, etc. could also be used):
1. Specify the throttle profile (e.g. si = 1 for full-throttle), calculate Im,i by
simultaneously solving Equations 2.29, 2.30, 2.32 and 2.33. Then calculate the motor
torque using Equation 2.22.
2. Calculate the wheel force (Fj,) using Equation 2.23, obtain a road angle (fli) from
the GPS data, and use Equation 2.24 to calculate the vehicle acceleration (xi).
3. Calculate the battery energy using Equation 2.36.
4. Use Equations 2.25 and 2.26 to step forward in time (i+]). Repeat the algorithm for
the next increment.
Full-throttle simulations show how maximum acceleration and speed largely depend on
the drag forces (e.g. sum of the gravity, rolling, and aero forces) and the motor output power.
The motor power that is in excess of the drag power causes the vehicle to accelerate. As the
motorcycle increases in speed, the drag forces increase as shown in Figure 2.9a. The motor has a
constant torque until the back-EMF from the motor becomes limiting, which is when the
24
available power from the motor starts to decrease. Once the motor and drag power are equal, the
vehicle no longer accelerates and this represents the maximum speed. Thus the intersection of
the motor and drag power curves is the maximum speed of the motorcycle and where the
vehicle's acceleration is zero.
The behavior of the motor power curve is heavily influenced by both the sprocket ratio
and the maximum current from the motor controller (Iimi,). A larger sprocket ratio and/or 'limit
will yield higher motor power at a given speed, and thus greater acceleration (Figure 2.9b and c).
However, increasing the sprocket ratio decreases the maximum speed. As shown in Figure 2.9d,
more current is needed during vehicle acceleration than at steady state/cruise conditions.
Simulations like these were used to estimate the appropriate current limit and sprocket ratio.
35 35
30- 30 Constant
30 3imit
25 25
Increasing
20 20 Sprocket
~15 a 5 5Ratio
15I -
10 1 10
Max
Seed 5
0 0
0 5 10 15 20 25 30 35 40 45 0 5 10 15 20 25 30 35 40 45
Speed (m/s) Speed (m/s)
(a) (b)
45 450
40
35 w 400
E
- 30 *
E 350
25 C
20 *20 ~300
15 Increasing Sprocket Ratio U
10 W 250- Increasing Iit
with constant gear
5 ratio
0 _______________ _ 200L---- - --
0 10 20 30 40 50 60 70 80 0 10 20 30 40 50 60
Time (sec) Time (sec)
(c) (d)
Figure 2.9: Simulations like the ones shown here were used to investigate design tradeoffs.
25
The main limitation of the algorithm described above is the difficulty in knowing the
throttle input, si, as a function of distance or time. It is typically infeasible to know the throttle
profile for the entire course, since the rider adjusts the throttle in real-time based on personal
judgment. Thus this algorithm was not used to estimate the battery energy required to traverse
the entire course. However, it is useful in exploring maximum power, acceleration and speed
during short full-throttle simulations.
2.3 Constant Vehicle Speed Simulations

The backward-looking simulation described in Section 2.1.1 can be used to simulate the
motorcycle when it is assumed that a constant speed is maintained throughout the course:
3 i) from the GPS
1. Obtain a road angle (# data and calculate the wheel force (F",i) using
Equation 2.3 with 2 = 0 and i= constant.
2. Calculate the power from the battery to the motor controller using Equation 2.15 and
the power and energy consumed from the battery using Equations 2.20 and 2.21.
The following subsection proves that it is most energy efficient for a vehicle to maintain a
constant speed.
2.3.1 Proof that Constant Speed Minimizes Energy Consumption

The objective of an electric race is to minimize the time it takes to cross the finish line for
a fixed amount of battery energy and a specified course. Once the race begins, the only variable
the driver can control is the speed of the vehicle, and this section will show that it is most energy
efficient to maintain a constant speed. In practice it is impossible to maintain a constant speed
since braking is required to safely traverse the corners of the course. But if the braking losses are
thought of as a fixed loss that only depend on the average speed, then this derivation is valid.
Battery and motor losses are not included in this proof, and it is assumed that the vehicle is
traveling at speeds above the terminal velocity (Figure 2.10).
The time it takes to traverse the entire course can be written as:
tf = /7 2.37
26
Where xfis the distance of the course, and 7 is the average speed. The minimum time, tfrnin, can
be written as:
tf,min = min f/ 2.38
Since x1 is fixed, t1 is minimized by maximizing the average speed. Thus we want to minimize the
energy use, E, but maximize the average speed, 7. The energy required by the vehicle to travel
between positions xO and xf can be written as:
Xf
E = fXf - dx 2.39
Combining Equations 2.3 and 2.39 yields:
E = f(mz + mgSin8 + mg pCosO + C(i + w) 2 )dx 2.40
And expanding the integration:
E = fmidx
XO xoxO
+ J
f 0XO
mgSinOdx + J
0
mgpCosOdx + ICC
Xf
xo
+ w) 2 dx 2.41
The first term can be rewritten as:

xdk Xd Xf
midx = m - dx = m idt = M idi = 1/2 m -f20 2.42
XO XO XO X0
Which is the kinetic energy of the vehicle. The second term can be rewritten as:
I
xo
Xf
mgSinOdx =
.YO
f mgdy
Y
=mg(y
= - yo) 2.43
Which is the potential energy of the vehicle. The third term of Equation 2.41 can be rewritten as:
f~f
f mgIpCosedx = mgyi Xf
Cosdx 2.44
Substituting Equations 2.42, 2.43 and 2.44 into Equation 2.41 yields:
Xf
E =* _;2)+ M~Xf
E = 1/2 m(*j - 0) + mg(yf - Yo) + mgy fXfCosodx + fXfC(' + w) 2 dx 2.45
xo XO
Assuming that ±k = 20 and yy = yo,, the only remaining term to be minimized based on the driver
control is:
X
Emin = min If C(i + w)2 dx} 2.46
27
Defining a new variable (Figure 2.11):
*x - X 2.47
Where ** is the speed above the average speed. Thus:
i = X +** 2.48
Now combining Equations 2.46 and 2.48 and expanding the terms:
Emin = mi Cj *2 + 72 + w2 + 21** + 2 w + 2**w)dx} 2.49

0
Since w cannot be controlled and 7 is a constant, they can be ignored in the minimization. The
remaining terms are:
min C J(V* + 2**I + 2i*w)dx} 2.50
Which is minimum when:

*= 0 2.51
Combining Equations 2.48 and 2.51:
2.52
This shows that the vehicle should maintain a constant speed in order to minimize the energy it
takes to complete the course.
140
120
E
100
80
0
.E
60
a,
40
20
0
0 5 10 15
Hill Decline (Degrees)
Figure 2.10: The terminal velocity for different hill decline angles. For example, on a hill of 10
degrees, the vehicle could accelerate (without power) all the way until reaching the terminal
velocity of 100 km/h. At the terminal velocity, the drag forces will equally balance the gravity hill
force and so the vehicle will no longer accelerate. If the vehicle is powered to speeds above the
terminal velocity, the vehicle will decelerate down to the terminal velocity when the power is
removed. The motorcycle is always above the terminal velocity so it can never accelerate down a
hill without motor torque.
28
U
0
x (Distance) Xf.t
Figure 2.11: The variable i* is defined as the difference between the instantaneous (x) and average
(k) speed.
2.4 Constant Power Simulations

The following equations were used for the assumption that the motor maintains a
constant power output throughout the course. Combining Equations 2.6, 2.8 and 2.9 yields:
PmUld m ..
F, = .M7 xw 2.53
2
And defining the motor output power, Pm as:
Pm = T m m 2.54
b = 2.55
flmcflm
The following steps were used:
1. Calculate the wheel force (Fw,1 ) for a specified constant motor power, Pm, using
Equation 2.53. Assume an initial vehicle speed.

2. Obtain a road angle (fli) from the GPS data, and use Equation 2.2 to calculate the
vehicle acceleration (1j).
3. Calculate the battery energy by integrating Equation 2.55 or 2.20 depending on if
battery losses are included.
4. Use Equations 2.25 and 2.26 to step forward in time (i+1). Repeat the algorithm for
the next increment.
29
2.5 Braking Losses
The calculations so far have not included the energy lost from braking, which will
increase the estimates for the battery energy required and decrease the average speed. Thus a
braking efficiency, 4, is defined as:
avg. speed with braking
4' =
avg. speed without braking 2.56
For example, it can be assumed that the braking reduces the average speed by 15%, thus 4
=0.85. The average speed when braking is included can now be written as:
7
B=
=41x 2.57
where 7 is the average speed without braking and the superscript "B" denotes values after
braking estimates are included. Also by definition:
7
B _ Xf
X-B 2.58
Where x1 is the distance of the course and tB is the time to finish the course when braking is
included. Also:
7-Xf
t -
tf 2.59
Where tj is the time required to finish the course when braking is not included. Combining
Equations 2.57, 2.58 and 2.59 yields:
4't 2.60
Thus this higher speed would be used in the constant speed simulations.
An alternative approach is to estimate the amount of energy lost because of changes in
kinetic energy while braking through the turns. For example, let us assume that the motorcycle
enters n turns at 150 km/h and exits after braking at 100 km/h. The change in kinetic energy from
braking is:
1
AKE = n m(v 2 - v 22)
2 2.61
Substituting in values for this course and motorcycle (assuming n = 50 turns):
1
AKE = 50 300(1502 _ 1002) ~ 2.6 kWh 2.62
2
Thus -2.6 kWh of energy could be attributed to braking losses.
30
2.6 Estimating the Required Motor Power and Battery
Capacity
The following algorithm uses the methods described in Sections 2.2 through 2.5 to
estimate the required motor power and battery capacity:
1. Choose a desired average finishing time, tfj and then use Equation 2.60 to determine the
corresponding average speed without braking, x.
2. Use the value of 7 with the algorithm described in Section 2.3 to estimate the battery
energy, Eb. This serves as a lower-bound (best-case) estimate for the actual battery
energy.
3. Put the equations described in Section 2.4 into a spreadsheet, and iterate (or numerically
solve for) the value of P,, until the average speed equals 7 (the same value used above).
This serves as an upper-bound (worst-case) estimate for the actual battery energy.
4. Perform the full-throttle simulations in Section 2.2 to ensure that the maximum
acceleration, speeds and power are adequate.
For the final estimates it was assumed that an average speed of 137 km/h (27 minutes)
would be competitive. Based on the simulations described above, this would required 25 to 30
kW of average power from the motors and -10 to 11 kWh of battery energy (Figure 2.12). The
braking constant, y, was assumed to be 0.8 for the constant power driving cycle and 0.82 for the
constant speed driving cycle2 .
Table 2.2: Parameter specifications used for the vehicle simulations.

Description Variable Value Unit
Total Mass m 311 kg
Drag Area CdA 0.41 m
Rolling Resistance Cr 0.025
Motor Speed Constant Ko 42 RPM/V
Motor Torque Constant K 0.207 Nm/A
2 It was assumed that the constant power simulation would require more braking since there are more speed
fluctuations.
31
12.045
-- Simulated Energy (const. speed) 450
-Simulated Energy (const. power) j 400
10.0 ----- Altitude
-4 350
~8.03-0 8.0 b300
- 250
6.0
200 -o
m4.0 150 'P
100
m 2.0
50
0.0 -0
0 10 20 30 40 50 60
Distance (km)
Figure 2.12: Simulation of the battery energy consumed during the race assuming a 137 km/h
average speed.
32
3. Motorcycle Design and Testing
A 2010 BMW SI OOORR motorcycle was used as the base chassis. The CAD of the final
design is shown in Figure 3.1, and the component specifications are listed in Table 3.1. Two
motors were combined with a rigid shaft to provide a maximum of 32 kWh of continuous power.
The energy storage consisted of 106 kg of lithium-ion batteries totaling 11.9 kWh. This chapter
describes the electrical (Section 3.1) and mechanical designs (Section 3.2) including the
justification for critical design decisions.
Figure 3.1: CAD of the final motorcycle design.

Table 3.1: Final component specifications.
Masses
Rolling Battery Motors Other
Chassis Modules
77kg 106kg 25kg 23kg
Mass Distribution w/out Rider
Front Wheel -Rear Wheel Total Mass
122kg 109kg 231kg
Mass Distribution w/ 80kg Rider
157kg | 154kg | 311kg
Battery Specifications
Paral./series Voltage Capacity
6/30 99V 11.9 kWh
Motor Specifications (2 motors combined)
Cont. Power Cont. Current Model Number
32 kW 400A D135RAG
Motor Controller Specifications
Voltage Cont. Current Model Number
18-136V 540A KDH12121E
Drivetrain Specifications
Sprocket Ratio Chain Size O-Rings
53/16T 428 No
33
3.1 Electrical Design
The systems engineering described in Section 2.6 estimated that 25 to 30 kW of average
power would be required from the motors and ~10 to 11 kWh of battery energy would be
consumed over the entire course. Due to cost, availability, volume, and mass constraints, it was
decided to use two DC, permanent magnet, air-cooled motors from the Lynch Motor Company.
The motor selection set the maximum voltage on the battery pack to ~100V. The battery pack
was designed with help from A123 Systems; a custom version of their prismatic lithium-ion
Nanophosphate@ modules was used (Figure 3.2). The corresponding volumes of the batteries
and motors were modeled in CAD to ensure adequate space on the motorcycle chassis (Figure
3.3). Note that the lean angle and ground clearance surfaces were added to the CAD. The battery
and motor masses were also calculated and included within an overall mass budget. A single
motor controller from Kelly Controls was used to regulate the energy to both motors.
The motorcycle was instrumented with sensors to measure acceleration, speed, location,
current, voltage, and temperature (Table 3.2). A battery management system (BMS) was used to
measure the voltage of each cell and perform cell balancing. A safety monitoring system from
A123 Systems checked for vitals such as ground faults, the condition of the fuse, and over
current. All of the sensing was integrated via a CAN-bus and an open-source microcontroller
made by Arduino. The data was transmitted to a laptop via xBee wireless transmitters/receivers.
The data was displayed off-board in real-time via a Graphical User Interface written in an open-
source environment by Processing.org, and logged on-board. The rider display consisted of a
series of simple LEDs to indicate the state of the motorcycle and a single LCD screen displayed
digital real-time values for battery power, battery energy consumed and vehicle speed.
Interconnect
Cover
LaserWelded ~cmrsinBnCmrsinBn
BusBar
Heatsink Plates
Cell
Lthium IonPrismatic
Figure 3.2: Battery module assembly (left) and exploded view (right). Images from A123 Systems.
34
Figure 3.3: CAD was used to ensure sufficient volume for the batteries and motors.
400A Fuse
Manual Disconnect
Battery Pack
Figure 3.4: Wiring diagram of the motorcycle's power electronics.
Table 3.2: Sensors used on the motorcycle.

Subsystem Measurement Specific Location
Batteries Temperature Between battery cells
Temperature Cell interconnections
Current Around battery cable
Continuity Across main fuse to sense if fuse has blown
Continuity Attached to frame chassis to check for ground fault
Voltage Each cell group and entire pack
Motors Temperature Back side of brush holder (1 per motor)
Current Main power cable from motor controller (1 per motor)
System Accelerometer Mounted in motorcycle tail
GPS Mounted in motorcycle tail
Magnet mounted on the front disk brake, sensor on front
suspension fork to count wheel rotations
35
3.2 Mechanical Design
The entire motorcycle was designed in CAD, which enabled the various subsystems to fit
inside the motorcycle with ~0.5 millimeter accuracy.3 A structural frame was needed to support
the batteries, motor, and supporting electronics; the removal of the engine needed to be
considered in the mechanical design, since the engine was originally a structural element (Figure
3.5). Instead of a more common space-frame weldment, it was decided to design a frame that
could be made purely on the waterjet using aluminum plates. The waterjet was chosen because it
is rapid, economical, and provides significant design flexibility. The final frame design is shown
in Figure 3.6. Designing a set of 2D plates also made prototyping easy; a laser cutter was used to
cut wood and cardboard into quick prototypes (Figure 3.7). The frame was assembled like puzzle
pieces with small tabs connecting the various 2D plates together (Figure 3.8a). The frame,
including the tab-slots, was welded while bolted in place on the motorcycle chassis. A rigid steel
jig was made to preserve the proper spacing and parallelism for the motors during welding
(Figure 3.8b). It should also be noted that an integrated design was used to house both the motors
and batteries in one weldment assembly.
The motors were connected via a custom designed drive shaft (Figure 3.9). Finite
Element Analysis (FEA) was used to ensure that the shaft could handle the maximum torque
loads of the motors and chain. Typically there is a flexible element in the drive shaft to provide
compliance for misalignment, though in this case it was assumed that the flexture in the frame
mounting was sufficient.4
3 CAD of the SIOOORR chassis was provided by BMW.

4 Based on a discussion with the motor designer (Cedric Lynch).
36
(a) (b)
(c)
Figure 3.5: The structural loop needed to be preserved when the gasoline engine was replaced by
batteries and electric motors.
Figure 3.6: The final design of the battery and motor structural frame (shown with motors and
batteries). An integrated design was used to house both the motors and batteries.
Figure 3.7: The frame and battery assemblies were prototyped using cardboard and wood cut on a
laser-cutter according to the CAD files.
37
Welded
Tabs
Motorcycle
Frame
(a) (b)
Figure 3.8: The frame was designed to be fabricated using a waterjet. The tab-slots of the frame
were welded while bolted to the actual motorcycle chassis (a). A rigid steel jig was used to set the
proper motor spacing during welding (b).
Plug Welds *fD
16T Sprocket
(a) (b)
Figure 3.9: A rigid shaft was designed to connect the two motors together (a). FEA was used to
ensure that the designs were adequate under maximum motor loading (b).
3.3 Testing
A series of incremental tests was performed to ensure that each subsystem operated as
predicted. This included the following tests: bench-top, wind tunnel, chassis dynamometer, track,
and finally road. The methods, results and usefulness of these tests are discussed in this
subsection.
Bench-top tests were performed to ensure that each subsystem was functioning as
expected before it was integrated into the rest of the motorcycle. The motorcycle chassis, main
fairings, and cardboard mock-ups of the batteries and motors were assembled and tested in the
MIT wind tunnel to estimate the aerodynamic coefficient (Figure 3.10).
38
The fully assembled motorcycle was then tested extensively on a chassis dynamometer
(Figure 3.11). The main function of a dynamometer is to measure the power output as the
motorcycle is strapped on a stationary stand. The front wheel is held rigidly, while the rear wheel
is free to spin on a drum that provides resistance. A sensor and computer measures and records
the torque and speed of the drum, which can be translated to power, torque, and speed curves.
The dynamometer was used to do the following:
1. Ensure that the drivetrain, and other components were able to withstand full load
and speed: Since safety was a main consideration, it was important to ensure that the
motorcycle would not have failures - especially under heavy load and high speeds. The
heavy loading and high speeds were tested on the dynamometer while there was easy
visual and physical access.
2. Match the current going to each motor: A single motor controller was used and the
motors were wired in parallel. Thus the motors needed to be adjusted (timed) so that a
nearly equal amount of current was going to each motor. The brushes were advanced or
retarded by a very small amount depending on the conditions (Figure 3.12c).
3. Understand the thermal characteristics of the batteries and motor: Initially it was
uncertain how the temperature of the motors and batteries would change with time. The
dynamometer represented a worst-case scenario, since the actual airflow at 140+ km/h
would provide much more cooling than the dynamometer fans (Figure 3.12a and b).
4. Use the data obtained from the dynamometer to validate and tune the analytical
models: The dynamometer generated the curve shown in Figure 3.12d, which was used
to validate and tune the models discussed in Section 2.1.
5. Measure the drive-train efficiency: The dynamometer is able to measure power output
at the rear wheel (Pdyno). The current (I) and voltage (V) can be measured at the motor
leads (power terminals/connections) to determine the power going into the motor. The
drivetrain efficiency (motor shaft to rear wheel) can then be estimated as:
Pdyno
k7drivetrain n
IJmotor 3.1
It should be noted that this method also includes the rolling resistance of the rear wheel
since the wheel is rolling on the dynamometer drum.
39
6. Test the microelectronics and software: The motorcycle had multiple microprocessors
and sensors that were wirelessly streaming data to on- and off-board data loggers and
computers. The dyno enabled testing of the full microelectronic system and software.
(a) (b)
Figure 3.10: The motorcycle was put in a wind tunnel. A manikin was used to simulate the rider (a).
A custom mount secured the motorcycle to the wind tunnel test stand (b).
After multiple dynamometer tests, upgrades, and fixes, the motorcycle was tested on a track at
New Hampshire Motor Speedway where it performed as previously predicted (Table 3.3 and
Figure 3.13). The motorcycle was then sent by air from Boston to the Isle of Man inside a
wooden crate. Partial disassembly was required since import/export regulation required the
batteries to be shipped separately by the manufacturer (A123 Systems). The motorcycle was
reassembled on the Isle of Man and a locally available dynamometer was used to run the
motorcycle under full load and speed. Finally, the motorcycle was road-tested on the Isle of Man
before the races (Figure 3.14).
Cooling
fa ns
Rolling
d rum
Figure 3.11: The motorcycle on a dynamometer.
40
200A Continuous From Battery Pack
40 90
,35 80
0.
30 70
'
t25
E 60
i 20 50
S
15 40
10 30
0 2 4 6 8 10121416182022242628 0 2 4 6 8 10 12 14 16 18
Time (min) Time (min)
(a) (b)
160 35
-- Right Motor - Model
155 +*LeftMotor 30 -Dynamometer
25
S150
20
= 145
U
0
15
2 140 CL.
WU
10
135 5
cc
130 0
0 2 4 6
8 10 12 14 16 18 0 10 20 30 40 50
Time (min) Speed (m/s)
(c) (d)
Figure 3.12: Example set of data collected on the dynamometer: battery (a) and motor (b)
temperature, and motor current (c) data taken while the motorcycle was under load. The
dynamometer can also be used to validate and tune models; plot (d) shows a full-throttle run, and
thus the maximum power available at the given speed.
Table 3.3: Motorcycle performance specifications obtained from testing and simulation.
o to 90 MPH (sprocket ratio = 3.3) 25 sec
Maximum Speed (sprocket ratio = 3.3) 44 m/s
0 to 60 MPH time (sprocket ratio = 5.3)* 8 sec
0 to 60 MPH time (sprocket ratio = 5.3, max motor 43 sec
output)* 4.3_sec
*Simulation only
41
Figure 3.13: Track testing at New Hampshire Motor Speedway.
Figure 3.14: Road testing on the Isle of Man.
42
4. Motorcycle Race Results and Analysis
The motorcycle and rider competed in two qualifying races and one final race (Figure
4.1). The following subsections analyze the race data to understand the motorcycle performance
(Section 4.1) and validate the models and assumptions made during the design process (Section
4.2). The final section reflects on the racing experience to outline a set of guidelines for
designing zero emission races with the aim of promoting innovation (Section 4.3).
Figure 4.1: The motorcycle during the TT race.
4.1 Analyzing the Motorcycle Performance

The average vehicle speed and energy consumption increased with each race (Table 4.1),
which was likely caused by: (1) adjustments in the motor controller settings (2) changes in the
sprocket ratio between each race, and (3) the rider learning to better control the motorcycle with
each additional race. The speed versus distance over the entire course is shown in Figure 4.2 and
each dip in the curve represents a point in the course where the rider slowed down to enter a turn.
The speed data is shown with the corresponding altitude, which highlights the speed reduction
while traversing up the mountain. The motors were closely matched as measured by current
during the final race (Figure 4.4), though the left motor was nearly 20"C warmer than the right
43
(Table 4.1). The temperature differences likely resulted from one motor having a different gap
between the commutator and magnets. The gap differed because the motors were rebuilt between
each race and the reassembly process was modified each time. The motors were operated within
their current limits (Table 3.1) and the battery power fluctuated about a safe nominal value of
-22 kW (Figure 4.3). Finally, the battery cells, interconnects and terminals stayed at ambient
temperatures throughout the race, which indicates that sufficient cooling was achieved.
Table 4.1: Comparing the qualifying and final race results.
Parameter Units Qual. Qual. Race

1 2 _ _
Average Speed [km/hr] 122 126 127

Finishing Time [mins] 29:54 28:50 28:58
Battery Energy [kWh] 9.4 9.7 9.8
Average Battery [kW] 19 20 20
Power
Average Battery [A] 198 215 220
Current
Battery [Wh/km] 147 155 161
Energy/Distance
Left Motor Temp* ["C] 72 63 52
Right Motor Temp* [*C] 68 87 35
Average Current to [A] - - 149
Left Motor
Average Current to [A] - - 142
Right Motor
*Temperature taken at the end of the race using infrared temperature gun
160 450
140 400
. 120 350
300 a,
*44100 -E
E 250 0
80 -M
200
60 150
0
VI 40 100
20 50
0 0
0 20
10 30 40 50 60
Distance (kn)
Figure 4.2: The motorcycle averaged 127 km/h during the race.
44
40 450
35 400
~30 350
'25 300,,
250.
3 20
0 200
t15 150
010 1004
50c 50
0 0
0 10 20 30 40 50 60
Distance (km)
Figure 4.3: The battery had an average power output of-20 kW.
250
Right Motor
Left Motor
200
W 150 .N
L.
i
1.. 100
0
0
2 50
0 ---- -
-- -
1 1.5 2 2.5 3
Time (mins)
Figure 4.4: The race data shows that the motors were equally matched.
4.2 Model Validation

The speed versus distance plot in Figure 4.5 shows that the models derived in Section 2.2
closely match the real world data when full-throttle was applied during the race. The parameters
shown in Table 2.2 were used in the simulations. Additionally, Figure 4.6 shows that the racing
data supports the hypothesis described in Section 2.6 that the actual battery energy required to
traverse the course is in between best and worst cases of constant speed and power, respectively.
Interestingly, an arithmetic mean of the constant speed and power curves matches the actual race
data at the finish with <5% error (Figure 4.7). It should be noted that the battery energy was
45
measured by a shunt and voltage measurement at the output of the battery and thus internal
battery losses were not included. Table 4.2 lists the comparisons between the models and racing
data. The conclusive result is that the models and assumptions were able to sufficiently predict
motorcycle performance.
180
160
140
120
100
E
80
60
0. 40
20
0
0 5 10 15 20 25 30 35
Time (sec)
Figure 4.5: The models closely match the actual full-throttle conditions.
, r~ 450
Simulated Energy (Const. Speed)
1 - Simulated Energy (Const. Power) 400
0.0 - Actual/Race Energy
----- Altitude 350
LU@1 8.0 300 E
250 '0
6.0
200
a 2.0.. 150
i5
100
2.0 .. ,.
50
0.0 0
0 10 20 30 40 50 60
Distance (km)
Figure 4.6: The race data supports the hypothesis described in Section 2.6 that the actual battery
energy required to traverse the course is in between best and worst cases of constant speed and
power, respectively.
46
12 -- Actual/Race Energy
10 0
8 -0.05
6 -0.1
m 4 -0.15 0
(2 -0.2
0 - - -0.25
0 10 20 30 40 50 60
Distance (kn)
Figure 4.7: An arithmetic mean of the constant speed and power curves matches the actual race
data at the finish to within 5% error.
Table 4.2: A comparison between the model predictions and final race results.
Parameter Units Model Race Error
Battery Energy [kWh] 9.6 9.8 2%
Battery [Wh/km] 158 161 2%
Energy/Distance
Average Speed [km/h] 127* 127 N/A *
Maximum [km/h] 148 155 5%
Speed I I
0 to 144 km/hr [sec] 24 25 4%
(z = 3.3) _ 1 '
*the average speed used in the simulation was set equal to the actual race speed
4.3 Framework for Spurring Innovation Through Racing

The Isle of Man TT Zero is an example of a new breed of "zero emission" races. The aim
of these races is to spur innovation that will reduce the environmental impact of consumer
vehicles. Racing has historically been a catalyst for innovation, particularly in the early years of
motorcycles and automobiles [15]. New concepts were tested on the track and the desire to win
drove companies to produce superior technology. Consumer demand for better performance
motivated companies to transfer the technology from the racetrack to the mass market.
The fundamental question is whether or not zero emission racing will yield the desired
outcome. With the goal of contributing to the success of zero emission racing, this subsection
outlines a set of guidelines for designing zero emission races that will yield relevant innovation.
47
Innovation in this context is defined as the act of generating a product or service that (1) reduces
the environmental impact of vehicles and (2) consumers want to purchase.
4.3.1 Consider the Historical Context

Gasoline vehicle racing has evolved dramatically over the past 100 years. Because of
this, caution should be used when copying a modem gasoline race with a zero emission
equivalent. Zero emission racing might require a different approach and lessons may be learned
from looking back into the beginnings of gasoline racing.
Patience will also be required when directly comparing modem gasoline and zero
emission racing. It is easy to forget that it took decades for gasoline engines to make dramatic
improvements. For example, it took 50 years for the first gasoline motorcycle to reach a 100 mph
average lap at the TT. The electric motorcycles reached this same milestone within 4 years from
the start of electric motorcycle racing at the TT.
4.3.2 Utilize the Power of Regulation

Regulations should be used as the fundamental tool to engineer a race for a desired
outcome. For example, assume that consumers want to refuel their vehicle quickly; if winning a
zero emission race is dependent on fast refueling, then the regulations are successfully guiding
development. A successful racing innovation platform must focus on technology relevant to the
consumer market.
4.3.3 Drive Technology

Many diverse participants, including inventors, academia, and corporate research labs
contribute to generating and developing innovative ideas. Consumer-focused companies choose
relevant developments, refine them, and promote them to the consumer market. Identifying
which ideas will succeed is a challenge facing all vehicle companies. Resources are often not
available to invest in multiple emerging technologies. For example, it is costly for an automobile
company to invest in batteries, fuel cells, and super capacitors simultaneously. Racing
competitions should be structured to accelerate the transition from ideas to mass production and
simultaneously facilitate the development of multiple technologies.
4.3.4 Provide Valued Entertainment

Any repeated event that the public finds entertaining will draw a large number of
spectators both in person and through the media (e.g. internet, TV, etc.). Spectators and media
48
drive advertising, which creates an influx of funds through team, rider and event sponsorship.
These funds help finance the teams who in turn develop the technology. Thus valued
entertainment is drawing in extra research and development funds that would otherwise not be
availablefor that purpose (Figure 4.8). For example, an energy drink manufacturer might be
indirectly funding battery research. This could translate into significant funds spent on zero
emission innovation [16].
The influx of available sponsorship also reduces the risk that the team with the most
personal wealth will win. In other words, sponsorships are typically chosen based on which team
is likely to win; if the teams generating the most innovative vehicles are more likely to win, these
teams would be rewarded through sponsorship funds to develop even better technology.
Product Companies
Goods or / / Sponsorship
Services $
Entertainment
Consumers - Zero Emission Race
Innovative $ Sponsorship $7 Technology
Vehicles / Development
Vehicle Companies
Figure 4.8: This diagram illustrates how the entertainment from racing can indirectly generate
research and development funds and drive Technology Development. A Zero Emission Race
provides entertainment to Consumers and the associated advertising at the race motivates
consumers to purchase vehicles and products from race sponsors. This money eventually flows
back into the Zero Emission Race through sponsorship. The race generates Technology
Development that goes back into Vehicle Companies, which in the long run will provide consumers
with more advanced vehicles.
4.3.5 Inspire Consumer Demand

It is critical that the races inspire consumers to purchase the technology that is found
superior on the racetrack. Otherwise, true innovation will not be achieved through racing and the
objective of reducing the environmental impact of vehicles will not be achieved. One way this
can be accomplished is through styling, and ensuring that the race vehicle has brand identity. For
example, a motorcycle company should use styling that is distinct and that connects their race
vehicle to their commercially available vehicles.
Secondly, inspiration can be found through education. The race should strive to inform
the consumer of the environmental effects and implications of the various technologies.
49
Finally, races can inspire consumer demand by building confidence in new technologies.
For example, racing could prove that rapid charging is feasible, which might convince the
skeptical consumer that the technology will satisfy their needs.
50
Part II: Estimating an Electric Vehicle's
Distance to Empty
51
52
5. Introduction to Part II
An electric vehicle's Distance to Empty (DTE) is defined as the distance the vehicle can
be driven before recharging is required. A real-time estimate of DTE is commonly displayed on
the vehicle's dashboard and is used by the driver for route planning (Figure 5.1). The objective
of Part II of this thesis is to introduce a new DTE algorithm that uses measurements from the past
along with knowledge of the future route. For example, if the driver provides their destination(s)
beforehand, a navigation system could obtain route, traffic and weather information via the
internet and this information could be used to improve the DTE estimate. Instead of physics-based
models, a multivariate linear regression model is used that adjusts a historical average of energy
consumption based on estimated changes in speeds, traffic and temperature. This approach could
be implemented as a cloud-based mobile phone application since it is computationally light and
fitted using historical driving data.
The new DTE algorithm was quantitatively compared to conventional methods. To
perform this comparison using real driving data, a large set of driving data would be needed that
contains speed, energy, traffic and temperature information. Since this dataset is not currently
available, a large variety of driving conditions were simulated using a set of physics-based
models. The models were based on the motorcycle work (Part I) and data from a fleet of electric
vehicles. The input to the models is a Markov-based stochastic speed profile generator, which
attempts to capture the stochastic nature of driving. Example simulations show that including
future driving conditions in the DTE algorithm can significantly reduce error.
Figure 5.1: The dashboard of the Nissan Leaf displays the DTE and the SOC.
53
5.1 Motivation
The maximum DTE for electric vehicles (EV), also referred to as range, is typically 100 to
400 km less than gasoline vehicles and a full recharge usually takes hours instead of minutes
(Table 5.1). Also, the energy consumption of EVs is more influenced by auxiliary loads (e.g.
heating) [17]. An undesirable scenario is one where the driver is left stranded on the side of the
road without the ability to quickly recharge. For these reasons it is important to provide an
accurate DTE estimate. Studies have shown that current DTE algorithms are insufficient and often
cause "range anxiety" among drivers [18][19]. Estimating DTE is difficult because of the
stochastic nature of driver behavior and the environment, the lack of a quantitative understanding
for how these factors affect energy use, and the fairly basic algorithms currently being used.
Recent studies have shown that the maximum DTE can vary by as much as 55% depending on the
driving conditions [20]. The usefulness and importance of the DTE estimate was confirmed by a
study of EV users, which concluded that "providing drivers with a reliable usable range
[estimate] may be more important than enhancing maximal range in an electric mobility system"
[18]. In other words, a more accurate DTE estimate may be more useful than increasing the size
of the battery pack.
Table 5.1: Comparing the energy storage, recharging times, fuel economy and range of various
vehicles. The approximate values were obtained from the manufacturer's websites.
Fuel
Energy Recharge
Vehicle Type Make/Model (km per hour of Economy Range (kn)
Storage charging) (L/100km
or EV-equivalent)
Gasoline Honda Civic 50 Liters N/A 7.4 680 (gas)
Plug-in Hybrid Toyota Prius 4.4 kWh 2.5 (EV) 18 (EV)
(PHEV) Plug-in 40 Liters 4.7 (hybrid) 870 (total)
Plug-in Hybrid GM Volt 16 kWh 2.5 (EV) 56 (EV)
(PHEV) 35 Liters 6.4 (gas) 655 (total)
Battery Electric Nissan Leaf 24 kWh 23 2.4 160 (EV)
(BEV)
Battery Electric Tesla Model S 40-85 kWh 50-480 2.6 258-483
(BEV) (EV)
5.2 Overview
Chapter 6 develops a stochastic vehicle simulation environment that was used to compare
DTE algorithms. This approach captures the random nature of driving and can be used to simulate
an unlimited number of full battery discharges. Chapter 7 introduces key concepts and equations
54
that aid in developing and evaluating DTE algorithms. Chapter 8 proposes a new approach that
uses estimates of future driving conditions to more accurately predict DTE. Instead of using
physics-based models, a multivariate linear regression-based model is derived, which adjusts a
historical average of energy consumption up or down based on estimated changes in route, traffic
and traffic. Both the conventional and new algorithms are compared in Chapter 9 using the
stochastic vehicle simulation environment.
5.3 Related Work

A previous study used a basic vehicle model combined with static assumptions about
auxiliary use to estimate the maximum DTE [21]. Though their simulations and experimentation
closely matched the expected values published by the manufacturers, they did not take into
account the true stochastic nature of the driving conditions nor did they investigate the accuracy
of DTE with distance. Some have approached the more general topic of estimating the energy
consumed between two points [22][23]; they used commercially available simulation software
with historical traffic and road data. There are similar approaches that require physics-based
models to estimate energy use and/or DTE [24]. Others describe a method where historical data is
used to predict the energy required for future trips [25][26]. Many of the techniques related to
energy estimation were applied to power-split control algorithms for plug-in hybrids or gasoline
vehicles [27]. Finally, there are many patents issued by automobile companies that describe
various averaging techniques used to estimate DTE [32][33][34].
None of the research described above shows results of a DTE simulation or
experimentation, nor do they explain how their approaches would be implemented. They also all
stop short of measuring the quality of the DTE estimates and making comparisons to other
methods. Finally, there has been little effort to understand the fundamental relationship between
energy estimation and DTE error. This chapter aims to contribute in these areas and start a more
open and in-depth discussion of DTE algorithms.
55
56
6. Stochastic Vehicle Simulations for
Evaluating DTE Algorithms
This chapter derives a parametric physics-based electric vehicle simulation, which was
used to simulate an unlimited number of full battery discharges under realistic driving
conditions. The overall simulation architecture is discussed in Section 6.1. The simulation uses a
stochastic speed profile (speed versus time) as input, which is discussed in Section 6.2. The
subsystem models (e.g. vehicle, motor, etc.) are discussed in Section 6.3. Finally, Section 6.4
summarizes how the models were used together to generate a stochastic vehicle simulation.
6.1 Simulation Architecture

The models derived in this section form a backward-looking simulation similar to the
methods described for the motorcycle in Section 2.1.1. The main differences are the (1)
stochastic speed profile generator and (2) auxiliary model (Figure 6.1). The speed profile
generator produced a speed versus time dataset, x(t), that was used as input to the vehicle model.
The vehicle model then determined the wheel force (F,,) required to follow the prescribed profile
based on a set of design parameters. Similarly, the transmission and motor models determined
the amount of electric power needed by the motor (P,,) in order to generate the desired wheel
force. Next, the motor power combined with the Auxiliary Power (Pa) was the net electrical
energy needed from the battery (Pb). The battery model translated the total electrical need (Pm +
Pa) to battery energy (Eb) and included internal losses. The battery losses depended on the
battery current (Ib). The following subsections derive each of these models.
Motor Torque Electrical Power
Start Vehicle
Speed
WheelForce
Speed
Vehicle Motor Speed to Motor
Speed kk Vehicle Fwkkk Transmission T motor
Mkdmk Pmk
Profile0
Genera tor Model Model Model
Current from
Battery Model Battery
Battery Ib,k
Losses Battery Losses Power from
FI-- Battery
Figure 6.1: The backward-looking vehicle simulation.
57
6.2 Stochastic Speed Profile Generator
Backward-looking simulations require a speed profile (speed versus time) as input. It is
most common to use standardized Environmental Protection Agency (EPA) driving cycles as
inputs, which specify the speed profile and auxiliary use (e.g. A/C and heater use). The following
are the most commonly used EPA driving cycles:
* City driving (FTP-75)
* Highway driving (HWFET) See Figure 6.2
* Aggressive driving (SFTP US06)
* Air conditioning test (SFTP SC03)
The following are key limitations to using EPA driving cycles for DTE simulations:
* Short Length: The EPA driving cycles are short in length (~10 km) and thus would have
to be repeated multiple times in order to simulate a full battery discharge (100 to 300
km).
e Lack of Randomness: A DTE algorithm could be tuned to work well with EPA driving
cycle(s), but it might not work well under broader (more realistic) conditions.
For these reasons, it is undesirable to use standard EPA driving cycles as the input to the DTE
simulations. Instead, this section derives a method for generating stochastic speed profiles that
can be of any length with specified conditions (e.g. city or highway).
Various methods for generating stochastic speed profiles have been described in literature
[36][37][38][39][40][41]. The general approach used in this thesis is to view any time-series
speed profile as a Markov process, which is a stochastic process with no memory [42]. The
selection of the next state only depends on the current state and not on prior states. In this case
the states are sequential sets of data that represent deceleration, cruise, acceleration or idle events
- or states. These states are referred to as modes (Figure 6.3a) since they represent particular
operating modes of the vehicle. Since there are different intensities of acceleration and
deceleration at difference speeds (e.g. low acceleration at high speeds and high acceleration at
low speeds), the acceleration and deceleration modes were more finely divided for a total of 10
states in the Markov model: 4 deceleration, 1 cruise, 4 acceleration and 1 idle state.
A large set of driving data (speed versus time) was parsed, and the Markov Transition
Probability Matrix (TPM) was determined by observing the transitions between states (e.g.
58
counting the number of times the state transitioned from the cruise to deceleration mode). The
TPM along with a random number generator was used to string together a stochastic speed
profile of any length. To simulate a variety of conditions, the TPM matrix was modified to give
preference to certain modes (e.g. less aggressive driving was simulated by putting zeros in the
columns of the TPM associated with the largest acceleration and deceleration modes). This
section will describe the details of this approach.
120
100
C 80
E
- 60
-o
a40
U0~
20
0 5 10 15
Time (sec)
Figure 6.2: The above plot is a speed profile published by the EPA. The EPA speed profiles are too
short and lack variety in conditions and thus are not adequate for DTE simulations.
6.2.1 Raw Driving Data
The speed profiles are derived from a large set of driving data collected by the EPA in
Kansas City in 2004 [43]. The data contains over 400 hours of GPS-based speed data logged at 1
Hz for 480 randomly selected light duty vehicles in the Kansas City metropolitan area. An EPA
subcontractor performed extensive statistical analysis to ensure that a variety of ages, genders,
races and vehicle types were considered. However, the method derived in this section is
independent of the data collected, so any large set of speed traces could be used.
6.2.2 Categorizing and Grouping Data

A speed trace is defined as the raw speed data collected by GPS. The large database of
speed traces was converted to a readable format for Matlab, and an extensive set of checks were
performed to remove erroneous data (e.g. loss of GPS signals or general errors). Next, the speed
traces were divided into "micro-trips," which are defined as the speed data collected in between
moments when the speed was zero (Figure 6.4). Each micro-trip was further categorized as either
highway or city based on maximum speed.
59
Each micro-trip is composed of a series of deceleration, cruise, acceleration and/or idle
"modes." It is necessary to develop an algorithm that determines the sequence of modes for each
micro trip as shown in Figure 6.3a. To this end, Lin [36] derived a method based on Maximum
Likelihood Estimation (MLE). However, the MLE method was implemented and tested as part
of this research and found to yield inconsistent and unreliable results. Thus the following
approach was developed that used the fractional increase in speed between each 1 Hz speed data
point was calculated:
A = . 6.1
Where ii is the speed of data point i. Then the mode was selected based on the values outlined in
Table 6.1.
"modes" Calculate average speed and acceleration
25 25
- Decel - Decel
20 . -Cruise 20 +Cruise
Accel * * Accel
E115- Idle E15 / Idle
401006 100
a. . o.I.!
(f)()
0
20 40 60 80 0 20 40 60 80
(a) (b)
Figure 6.3: A speed profile can be thought of a series of deceleration, cruise, acceleration and/or
idle "modes." Each set of points that compose a mode is called a "snippet" (a). The average speed
and acceleration were calculated for each snippet (b).
Each sequential string of mode data is referred to as a "snippet" (Figure 6.3a). A snippet
is discarded if it contains fewer than 3 data points. The average speed and acceleration are
calculated for each snippet and the speed-acceleration pairs for each snippet are grouped by
mode (Figure 6.3b). The snippets that are acceleration or deceleration modes are further
subdivided to form a total of ten modes: Deceleration (1 through 4), Cruise (5), Acceleration (6-
9) and Idle (10). The subdivision is done with a k-means clustering algorithm (Figure 6.5) [44].
Thus each snippet is assigned to one of the ten possible modes based on the snippet's average
60
acceleration and speed. The entire categorizing and grouping method is summarized in Figure
6.6.
"micro trips"
25
20
15 -
Co
10. ..
C'-__ _ _
'0 200 400 600 800 1000

Time (sec)
Figure 6.4: The speed traces were divided into micro-trips.
Table 6.1: Criteria for assigning the driving "mode" to each data point.
Ait Mode
=<-0.01 Deceleration
>-0.01
& <0.01
>=0.01 Acceleration
0 Idle
2
C) E
E 1
Decel 1
Decel 2
C75 - Decel 3
Decel 4
Cruise 5
-1 Accel 6
Decel a) Accel 7
- Cruise Accel 8
0 -2 -2[ Accel9
Accel
Idle 10
Idle
40 5 10 15 20 25 5 10 15 20 25
Average Speed (mis) Average Speed (m/s)
(a) (b)
Figure 6.5: The speed-acceleration pairs for all of the snippets were grouped by mode (a). The
deceleration and acceleration groups were further subdivided using a k-means clustering technique
(b). The result is 10 distinct modes that were used in the Markov model.
61
Raw driving data: Raw Driving Data
"micro Speed v. Time
25-
Zero-to-Zero speed
Micro Trips
Determine max speed

200 400 60 WO 1000
Highway or City?
A single micro-trip: Form snippets

based on value of /A*i
Dee
20! Qe "modes"
Decelerabon Cruise Acceleration Idle
Modes/
'b
20 o aok-Means Clustering based
k-means clustering: on values of Snippers
(Ave Speed, Ave Accel)
2
12 3 4 5 6 7 8 10
-Z =Modes+
S 6A 0 1p
5m 02)
Figure 6.6: A summary of the method used to categorize data into snippets, and then assign each
snippet to one of ten modes.
6.2.3 Markov Model

It was assumed that speed profiles can be modeled as a Markov process, which is a
stochastic process with no memory [42]. The selection of the next state only depends on the
current state and not on prior states. In this case the states are synonymous with modes and thus
there are ten possible states.
The approach involved sequentially stepping through each snippet in the order that they
occurred and counting the mode transitions. For example, the micro trip shown in Figure 6.3a
transitions from Mode 10 (idle) to Mode 9 (accel), Mode 5 (cruise), Mode 7 (accel), Mode 5
(cruise), Mode 2 (decel) etc. The goal was to determine the probability of transitioning from one
mode to another. These are referred to as transitionprobabilitiesand were calculated using:
11 = 11j6.2
ni
62
Where ri; is the number of times a snippet transitioned from Mode i to Mode j and ni is the total
number of snippets in the Mode i group. H1 is the transitionprobabilitymatrix and by definition
the following is true for all values of i:
10
I li; = 1 6.3
j=1
An example transition probability matrix (TPM) is shown in Figure 6.7.
I.200.0 00.0004.000305.403529
r 0092
An 0xa
0l 0a000 probability 0.25ix.
0.02 0.00S 0.002 0.00 0.002 0.332 0,229 0.0006 001 00007
-- U0.010 0.20 0.175 0.031 0.359 0.000 00089 0.000 0.001 0.000
Simuatin Highway City4and0Aggrssiv Drivin Conditions.22

6.2.4 Aggrssiv
0.240Sm0ati0.0ghay 0.073 0and 0rivin 0Conditions7000
maxmumsped ssownnFigure 6.. Ale triotinprobiitymarxmsedgeae.hno
equal to 65 km/h were considered highway driving and all of the slower traces were considered
city driving. Separate TPMs were made for both the city and highway data.
Next, to create a TPM representing less aggressive driving, the values in column 4 and 8
were replaced by zeros, which effectively removed the largest deceleration and acceleration
modes, respectively. In general the TPM can be modified to yield various outcomes. Thus four
TPMs were generated: city, less aggressive city, highway and less aggressive highway. All four
matrices were used to generate a variety of mixed combinations as discussed in the next section.
63
6.2.5 Creating a Stochastic Speed Profile
The TPMs were used to join snippets together to form a speed profile of a specified
length. Below is a summary of the algorithm:
1. One of four TPMs was chosen: city, less aggressive city, highway and less aggressive
highway (see Section 6.2.4).
2. A snippet was chosen at random with an initial speed equal to zero and an average speed
greater than zero. The random initial snippet selection shown in Figure 6.8 was from
Mode 6 through 9 (accel).
3. The subsequent mode was chosen based on a distribution formed from the corresponding
row of the TPM (Figure 6.7 and Figure 6.10). In this example, Mode 9 has the transition
probabilities listed in row 9 of the TPM. A snippet of this mode type was then chosen
randomly (using a random number generator) - though it must satisfy the requirement
that the initial speed was within a specified percentage of the final speed from the
previous snippet.
4. Step 3 was repeated until the desired speed profile length was obtained (Figure 6.9).
Switching between highway and city TPMs can generate speed profiles with varying conditions.
For example, half of a speed profile could be city and the highway. The proportion of highway
versus city driving for a given speed profile was set using a random number generator. The same
concept was applied to varying aggressiveness. Thus each speed profile was unique, stochastic
and simulated a mix of highway and city driving conditions.
t 2[
isY Mode 9E
Snippet r 1
15 Dec0
Decel 2
th -.- Decel 3
Decel 4
10 Cruise 5
-1 *- *Accel 6
Accel 7
5 Accel9
Mde10
0 5 10 15 20 10 15 20 25
Time (sec) Average Speed (m/s)
Figure 6.8: An initial snippet was chosen at random that has an initial speed equal to zero and an
average speed greater than zero.
64
20 1
15 C 25
E 20
l10
(D15
(D5
W1)10
5
Mode 9 Snippet
Mode 5 Snippet
0 5 10 15 20 25 30 00 100 2000 3000 4000 5000 6000

(a) (b)
Figure 6.9: The TPM was used to select subsequent snippets (a). The process was repeated until a
specified length of time was achieved (b).
0DO
7T2ol iw
rATh
(a) (b)
Figure 6.10: The transition histogram (non-normalized) of all modes (a) and only Mode 9 (b).
6.3 Subsystem Models

6.3.1 Vehicle Model
An electric vehicle can be modeled as a mass on an incline with externally applied forces
(Figure 6.11). It is propelled forward by an electric motor acting through the front and/or rear
wheel with a force, F., but slowed down by the aerodynamic drag, Faero, the rolling resistance
from all four wheels, Froiing,and the horizontal component of weight, Fgry. Rewriting Equation
2.3, which was derived in Section 2.1.1:
Fw = mxz + 1/2 p CdA(i + w) 2 + mgCrrCosf3 + mgSinfl 2.3
65
Figure 6.11: A vehicle can be modeled as a mass on an incline with externally applied forces.
6.3.2 Transmission Model

The force on the wheel(s), F., originates from a motor acting through a transmission. The
transmission generates a torque on the wheels, 'Tw. Summing the torques around the wheel yields
(Figure 6.12) the same as Equation 2.12:
Tm = + F, ---- 2.12
rr,
Figure 6.12: Torques and forces on the wheel.
6.3.3 Motor Model

The same motor model derived in Section 2.1.1 was used. Rewriting Equation 2.15 with
P,c defined as the power going to the motor controller:
MC (2w F6.4
PmmccIM'ld
66
6.3.4 Battery Model
The battery model derived in Section 2.1.1 was used, though auxiliary loads must be
considered for non-racing consumer applications. The total power from the battery is the sum of
the power to the motor and the power to the accessories (Power In - Losses = Power Out):
b-IRb = Pm + Pa 6.5
Where Pa is the power going to the auxiliary loads. The power from the battery can be written as:
Pb = VbnIb 6.6
Where Vbn is the voltage on the battery pack when no current is being drawn. Combining
Equations 6.5 and 6.6 yields:
Vbnlb - Ib R = Pm + Pa 6.7
The only remaining unknown is 1b, which can be determined using the quadratic equation. Thus:
4
Vb=nl n - Rb(Pm + Pa) 6.8
b b 2 Rb
Combining Equation 6.6 and 2.44:
Vbfn Vn - 4Rb(Pm + Pa)

p_FV z 6.9
Pbn 2 Rb
2 Rb
The energy consumed from the battery can now be determined by integrating Equation 6.9:
Eb - Pbn dt 6.10
6.3.5 Auxiliary Model

The auxiliary model was meant to simulate the energy used to run the heating, air-
conditioner, radio, headlights, etc. [17]. The model is shown in Figure 6.13 and is based on
measurements taken with an electric sedan similar but not identical to the Nissan Leaf This
simple model assumes that the auxiliary energy use only depends on the outside (ambient) air
temperature. It also assumes that heating and cooling take the same amount of energy for a given
change in ambient air temperature. This increase is meant to simulate the energy use required to
keep the inside air (cabin) at a comfortable temperature. Other factors (e.g. defrosting) are
lumped into this model.
67
0 3.5.
20
Ambient Air Temperature (*C)
Figure 6.13: Auxiliary power model used for the vehicle simulations.
6.4 Generating Stochastic Vehicle Simulations

This section summarizes the process of simulating an unlimited number of full battery
discharges (see Figure 6.1). The inputs to the simulation are the speed profile, ambient air
temperature, and vehicle parameters (drag coefficient, mass, battery voltage, etc.). The speed
profile, i(t), is generated using the algorithm described in Section 6.2.5. The following
equations can be used to numerically solve for the force, torque, power and energy values for
each increment in time. The force on the wheel is:
F, = mx + 1/2 pCdA(i + w) 2 + mgCrrCosfl + mgSin 2.3
Then the corresponding torque from the motor:
(CM
k2 Fw r
zrlM
2.12
And the power going to the motor from the batteries:
(tXw+ F.) ic
6.4
Pm = 2
?lmcTmild
Where the efficiencies are obtained from experimental data. The battery power, which includes
the internal losses, can be determined using:
V2 Vbn Vb2n- 4Rb(Pm + Pa)

2R 6.9
pb =_n
b 2Rb 2R
68
The auxiliary energy, Pa, is determined using the model shown in Figure 6.13. And the total
battery energy consumed is:
Eb - Pbn dt 6.10
The vehicle parameters (e.g. CdA) are fixed and the ambient air temperature is selected
based on the specific scenario. The simulations used in this thesis assumed a generic sedan EV -
similar but not identical to the Nissan Leaf and BMW ActiveE. The output is the simulated
battery energy, speed and distance with time for a full battery discharge.
69
70
7. Fundamental DTE Concepts
This chapter first derives key equations that quantify DTE error and aid in formulating DTE
algorithms (7.1). It is shown that these equations take on an analytical form that is similar to
other applications such as predicting the time remaining until a vehicle reaches a destination
(7.2). Next, conventional DTE algorithms are derived to serve as examples of algorithms being
used in practice today (7.4). Conventional algorithms blend together a long-term and short-term
average of past energy use. It is shown that this approach works well except in cases when
significant changes in driving conditions occur for sustained periods of time (e.g. changes in
traffic or auxiliary energy use). These conventional algorithms are compared to a new algorithm
in Chapter 9.
7.1 Key DTE Equations

The energy stored in the battery, Eb, is consumed as the vehicle is driven a distance x
(Figure 7.1 a). An important metric used frequently in this discussion is the average energy use,
A, which is defined as the quantity of energy AEb consumed over a distance Ax and has units of
Wh/km:
_ AEb
P A 7.1
Ax
Thefuture average energy use, pf(t), can be determined by evaluating Equation 7.1 between the
current time, t, and final time, ty:
f(t) E (t 7.2
x(tf) - x(t)
Where t -* R E {to, tf }, to is the time when the battery is fully charged and tj is the time when
the battery is fully discharged. Eb(t) is the battery energy remaining at time t. By definition, DTE
can be written as (Figure 7.1a):
DTE M = X(tf X (t) 7.3
DTE() -M 7.4
pf(t)
71
EXt t) Estimate
Actual
(t) tDTuE
Edt) il " DTE(t)DTt)
LpDT.Egt) ~.Actual
x(t) xtxx
x (Distance Traveled) x (Distance Traveled)

(a) (b)
Figure 7.1: A schematic of the vehicle's battery energy (a) and DTE (b) with distance. The actual DTE
is a straight line with distance traveled.
Evaluating Equation 7.2 at to yields:
_t Eb(to) Eb (to)
pto DTE(to) X(t
Now solving for x(t):
x(t) = Eb
ijj(to)
(to) 7.6
Combining Equations 7.3 and 7.6 yields a different form of DTE:

Ebato)
DTE(t) = -f (to) x(t) 7.7
pf (to)
Conceptually, both Equations 7.4 and 7.7 show that DTE can be determined if the current
battery energy and the future energy consumption are known. An onboard Battery Management
System (BMS) measures Eb (t) and was assumed to be known perfectly. Thus the task of a DTE
algorithm is to estimate the future energy consumption, Pf. Also, Equation 7.7 reveals that a
perfect algorithm would predict a linear relationship for DTE with distance (Figure 7.1 b).
Errors in DTE are caused by the algorithm's inability to perfectly predict p. Thus it is
desirable to quantify how errors in estimating 15 relate to errors in DTE. Assume that a DTE
algorithm estimates Pf(t) and the associated error is defined as:
e5 (t) = p(t) - 5f(t) 7.8
Where A designates an estimate of the actual value.
72
The corresponding error in the DTE estimate is then (Figure 7.1b):
eDTE (t) E (t) - DTE ( 7.9
Substituting Equations 7.4 and 7.8 into Equation 7.9 yields:
eDTE (t) = Eb ( t) + ef (t) f 7.10
Rearranging:
eDTE W Ef(t) ef(t)1 ] 7.11

pf (t)
Combining with Equation 7.4 results in:
eDTE t) = -DTE (t) ef(tff t) 7.12

.f (t)
Equation 7.12 shows that when Equation 7.4 is used the error in estimating DTE will attenuate to
zero towards the end of the discharge. Equation 7.12 can be rearranged as:
eDTE(0 ef t).1)
DTE M eflf(t)
Pf +1 7
Now defining the fractional error, E, as:

_eDTE (t) _p M__
EDTE t - DTE(t) and E,(t) - eif(t) 7.14

[Eg,(t))
EDTE (t) 1+) 7.15
_5f()+ 11
Equation 7.15 is plotted in Figure 7.2. The plot quantifies the relationship between errors in
estimating p and the resulting DTE errors. It can be seen that the relationship is linear for small
errors (~10%) but becomes increasingly non-linear for larger errors. The plot also shows that DTE
error is greater when f is underestimated versus when it is overestimated (e.g. when pf is

underestimated by 30% the corresponding DTE error is ~43% while when Pf is overestimated by
30% the corresponding DTE error is ~23%).
73
50
40
30
20
underestimating 10 overestimating
1U 0
i- 0 -30 -20 -10 -10 20 30 40
-20
-30
-40
Error in Estimating Pf (%)

Figure 7.2: A plot of Equation 7.15 shows quantitatively how errors in estimating Pi result in DTE
errors. DTE error is larger when Pf is overestimated.
7.2 Measuring the Remaining Battery Energy

As mentioned in the previous section, it was assumed that the BMS is able to perfectly
measure the remaining battery energy, Eb(t), which has units of Watt-hrs (Wh). The value of Eb
is analogous to the battery's State of Charge (SOC), which is defined as the percentage of
battery energy remaining. There is a large amount of literature related to SOC estimation, which
describes both model- and empirical-based methods [45]. An alternative approach is to use SOC
(%) instead of Wh and thus Eb would have units of percentage and 15 would have units of SOC
(%) per kilometer. A subtle but very important distinction assumed in this thesis is that Eb and
SOC includes the energy that will be consumed by internal battery losses. For example, if 12
kWh of battery energy is remaining, some fraction of this energy will be lost as heat within the
battery.
Another common measurement of battery capacity is Amp-hours (Ah). Both Ah and Wh,
as measured at the output of the battery, depend on the rate of discharge (e.g. fast discharge
results in fewer Ah and Wh). So neither are fixed quantities in this case. However, if the control
volume is drawn around the entire battery to include internal losses then both Ah and Wh are
fixed quantities. Wh are used in this thesis because they can be directly converted to Joules (J)
which is more useful in the mechanical domain. Also, the battery losses are modeled as a resistor
and the losses have units of J (or Watts) and not Ah.
74
7.3 Other Applications with a Similar Formulation
It should be noted that there are other applications that attempt to solve a problem similar
to Distance to Empty. For example, estimating the time remaining on a battery for mobile phones
or computers takes on a similar form as Equation 7.4:
Eb ()
TTE (0) - 7.16
bf (t)
Where TTE is the time (minutes) remaining until the battery is empty, Eb is the remaining battery
energy and bf is the future average energy use (Wh/min). Similar to conventional approaches for
estimating DTE, the past work in this area has used historical averages to estimate the future [45].
Another example is the time required to reach a destination, TTD, which is estimated by
navigation and mapping software:
D(t)
TTD - 7.17
Where D is the distance (km) remaining until the destination is reached and Vf is the future
average speed (km/hour). Past work in this area uses posted speed limits and crowd-sourced
information to estimate Vf [47]. Even though each vehicle will accelerate slightly differently, the
average speed is approximately the same and thus it is possible to estimate if based on crowd-
sourced information. Crowd-sourcing p, for Distance to Empty, however, is more difficult

because each vehicle will consume energy differently over an identical path based on the vehicle
type and driver. Crowd-sourced data has been used for predicting fuel consumption in gasoline
vehicles by adjusting the crowd-sourced values based on vehicle and driver type [27]. A similar
approach might be possible for electric vehicles, though further research is required.
7.4 Conventional DTE Algorithms

Conventional DTE algorithms assume that the future energy use will be similar to the past.
In other words, the following is assumed:
Pf ~Pp 7.18
Where p, is the average energy use of past driving, which can be determined using past
(historical) driving energy data (e.g. 1 km, running, or blended averages as defined below). The
DTE can then be estimated by combining Equations 7.4 and 7.18:
75
Eb()
DTE(t = 7.19
The values of DTE will deviate from DTE according to Equation 7.12 and the amount depends on
the validity of Equation 7.18 (Figure 7.1 b). This is considered a conventional approach since it is
likely very similar to the methods being used in EVs today based on the limited amount of
related literature [32][33][34]. The average energy use over the past 1 km is defined as:
_ Eb (tlkm) - Eb (t)
PikmNt = ,( 8k)7.20
X(t) - X(tlkm)
where t is the current time and tik, is the time 1 km in the past. And the running average is
defined as:
Eb to) - Eb()
Prunning(t) = - 7.21
X(0) - Xx(to)
(to)
A "blended" average uses a long-term average energy use, fliong, during the beginning of
discharge but blends to a more recent short-term average energy use measurement, Tshort, as the
battery is discharged. This can be written as:
Pblend (t) = flong ( - (Along (W - Pshort (0) 7.22
Where b is a "blending factor," b - R E fO,1}. The value of b is typically chosen based on a
linear function that changes with State of Charge (SOC):
b(t) = 1 - SOC(t)/100 7.23
SOC is defined as the percentage of battery energy remaining:
SOC(t) = 100 7.24

Eb (to)
These historical averages work well as long as the future conditions are similar to the
past. However, there are cases when changes in energy use will cause significant errors. For
example, Figure 7.3a shows that switching on auxiliary loads could have significant impacts on
energy use. Each curve represents a constant auxiliary load and was obtained by simulating a
sedan-sized vehicle at constant driving speeds. For example, a 2 kW constant auxiliary load will
cause a 30% increase in energy consumption when driving at a constant speed of 45 km/h.
Figure 7.3b-d are meant to show how a 30% change in energy consumption would impact DTE
error. To this end, assume that a vehicle has been consuming energy at a constant rate of 210
Wh/km for the long and short-term past. When the vehicle is ~50 km into a full discharge, a
heater load is turned on, which causes a 30% increase in energy consumption (Figure 7.3a and
76
b). Figure 7.3c shows the actual and estimated DTE when the running average or blend algorithms
were used. Figure 7.3d shows that the corresponding DTE error for the blended algorithm ranges
from 0 to 30% depending on the distance traveled.
Figure 7.4a and b show a similar situation, where the energy consumption is increased
due to auxiliary loads part way through the discharge, though this time a stochastic vehicle
simulation is used instead. The stochastic vehicle simulation is described in a Chapter 6, and uses
stochastic speed profiles and physics-based models to more accurately represent a vehicle's
energy consumption. The simulation runs multiple discharges in order to generate a long-term
average energy use, Plong, which in this case is the Wh/km averaged over the past 300 km. A
final example is shown in Figure 7.4c and d, and for this case there is no change in auxiliary use
or traffic.
400 300
Constant Loads: E
I350 Aux
-kW c 250
E 300 -2 kw Aux
250 -5 kW Aux W 200
200 >% 150
150 M C 100
I"
2J W - Prun.
1001 c
50 4 50 Pblenc
05
0 0
10 35 60 85 110 135 160 0 40 80 120 160
Constant Speed (km/hr) Distance Traveled (km)
(a) (b)
160 35
-Actual
E -Running Ave I-30
1*..Blend 25
E %%E 20
w 80 21
--\0 15
40 510
5
0 0
0 40 80 120 160 0 40 80 120 160
Distance Traveled (kin) Distance Traveled (kin)
(c) (d)
Figure 7.3: A simple example showing how auxiliary use can cause errors in the DTE estimate when
the running average or blended algorithms are used.
77
EP -Actual
600
-5-a
EE
nnn
tal n E -*-Prunnino
n Plong
~--*----
S500 Z"100 ,i'"
IA)400 CI) I
It *
u 300 50
200- 6
0
1000 20 40 60 80 100 20 40 60 80 100

Distance Traveled (kin) Distance Traveled (km)
(a) (b)
700 150
E -1 P -Actual
600 Prunnin Prunnn
Plong Plon
-Plong
S500 - 100
400
S300 50
L200 Fl
0
1020 40 60 80 100 0 20 40 60 80 100
Distance Traveled (km) Distance Traveled (km)
(c) (d)
Figure 7.4: Example output from a stochastic vehicle simulation. The plots show the corresponding
DTE using Equation 7.19 with P, equal to Pikm, Prunning and Plong. For these examples iprong is
defined as the energy use averaged over the previous 300 km. Figures a and b simulate an increase
in auxiliary load part way through the discharge while c and d have a constant auxiliary load. The
above simulations show that using past data and assuming jir i works well as long as there are
P,
no sustained changes in driving conditions.
All three of these examples support the following conclusion: using averages and blends
of past data and assuming that 15 ~ p works well as long as there are no sustained changes in
driving conditions between the past average and the future average (e.g. sudden use of heater for
a sustained period of time). It is interesting to note that in Figure 7.4c the assumption that
P ~: plang is especially true at the beginning to middle of discharge since f5prog is a measure of
average energy use over a similar length scale as Pf (to). In other words, both )long and pf(to)
are aggregates of a large amount of varying and stochastic conditions (temperature, traffic, etc.)
over a similar distance.
78
The examples above reveal that DTE algorithms could be improved if future changes in
driving conditions were anticipated, which is a concept described in the next section.
7.5 Using Estimates of Future Conditions and Models to

Improve DTE Predictions
The previous section showed that DTE errors occur when the average energy use changes
significantly from the past. This section describes the concept of detecting these changes
beforehand and using this information, along with a model, to improve the DTE estimate. For
example, assume that the driver provides their destination(s) to an on-board navigation system
before departure. The navigation system could then obtain navigation, traffic and weather
information via the internet. Thus changes in driving conditions could be detected before they
occur and fed into a model that estimates pf. The physics-based modeling methods described in
Chapter 2 could be used, though there are differences between racing and consumer driving that
make that approach less feasible. For example, there are more factors that significantly influence
energy consumption of consumer vehicles and thus additional models would be required (Table
7.1).
Figure 7.5 shows that auxiliary energy has a much larger impact on the total energy use
(Wh/km) at lower speeds. Since the motorcycle race is at high speeds (~150 km/hr), the total
energy use is less impacted by auxiliary energy consumption. Consumer vehicles, on the other
hand, are driven at lower speeds and thus more impacted by auxiliary use. The conclusion is that
a fluctuation in auxiliary power has a larger impact on energy use, which causes larger errors in
DTE estimates.
Figure 7.6 attempts to capture all of the factors that affect energy use in consumer EVs.
Modeling each of these factors using physics-based models would be complex, require vehicle-
specific parameters and can be computational intensive for real-time applications. To avoid these
shortcomings, the following section will derive a method that uses a model that can be "learned"
as the vehicle is driven.
79
Table 7.1: Comparing factors that influence energy use for motorcycle racing and consumer vehicle
applications.
A single trip in a single Multiple trips in a single

Route
battery discharge battery discharge
Average Speed Steady -150 km/hr Variable <110 km/hr
Auxiliary Loads Insignificant Variable and significant
Traffic None Variable and significant
Driver Behavior Steady and repeatable Variable and significant
Cargo Mass Constant Variable and significant
500
400
300
200 Drive Energy

-Auxiliary Energy (1 kW)
-Total (1 kW Aux)
100 -Total (2kW Aux)
-Total (5 kW Aux)
0
10 35 60 85 110 135 160
Constant Speed (km/hr)
Figure 7.5: A simulation of the motorcycle showing how the total energy use (Wh/km) changes with
speed and auxiliary load. Constant speed driving on a flat road and a constant auxiliary load were
assumed. The exact shape of the curve depends on many factors specific to the vehicle's loss
parameters (e.g. drag coefficient, rolling resistance)
80
U
X-
Cycle .j
History (cell
degradation)
Snow ad Ie
Speed)Sola Flux
.... Profile
4-1 4- Capac' Fade Outs' e Temp
E
SKe: Driver Energy Energy
U--
Options Stored Lost
Figure 7.6: This diagram attempts to show all of the factors that influence a consumer vehicle's
energy consumption and specifically how various driver decisions and options (purple) lead to
energy losses (red) and storage (green).
81
82
8. A Regression-based DTE Algorithm
This chapter derives an algorithm that uses estimates of future driving conditions to more
accurately predict DTE. As discussed in Chapter 7.2, the past energy use, Plong, is often a good
estimate for future energy consumption as long as there are no significant changes in driving
conditions. An improved algorithm is one that adjusts the value oflpiong, based on estimated
changes in future driving conditions, to yield a better estimate of Pf . The approach is to multiply
5 iong by an adjustment factor, y, that makes the following true:
Pf(t) = y(t) iong(t) 8.1
Combing Equations 7.4 and 8.1 yields:
DTE t - 8.2
y(t) piong(t)
Since y cannot be determined perfectly beforehand, an estimate 9 can be made when an
unknown residual error, E, is included:
y(t) = f(t) + E (t) 8.3
Rearranging:
f(t) = y(t) - e(t) 8.4
Then an estimate for DTE can be written as:
Eb(t)
DTE W = 9(t) - ng() 8.5
Section 8.1 proposes a multivariate linear regression model for determining 9. The regression
model uses a training set to learn (fit) the relationship between explanatory variables and 9.
Section 8.2 defines the explanatory variables as changes in ambient temperature, traffic and
average speed. Section 8.3 describes how the training set is obtained using historical driving data
and Section 8.4 checks the validity of the regression model.
The advantage to this new approach is that there is no need for physics-based models. In
other words, instead of precisely predicting the future energy consumption using physics-based
models, a past measurement is simply adjusted up or down based on estimates of future
conditions and a regression model.
83
8.1 Multivariate Linear Regression Model
A regression model uses a training dataset to learn the relationship between explanatory
variables, X, and the response variable, f. Assuming a linear model of the form [29]:
f(ti) = #h + #1iXi1 + fl2Xi2 + + f#mXnm 8.6
Where /3 is a set of m unknown coefficients that are determined from historical data (training
set). The variables X must be measurable and predictable factors that cause differences in energy
use between the past and future. For the sake of this derivation, the values of X are assumed to be
known. Equation 8.6 can be written in matrix form:
y ='X# 8.7
Once f is known, Equation 8.5 can be used to determine UTE- Rewriting Equation 8.3 in matrix
form:
y = y + e8.8
Y =,X# + 8.9
Solving for the residual error:
E = Y - X# 8.10
The residual error can be minimized through a least squares estimator, which can be written as a
function S(c):
S(C) = T= (y -XI) T(Y - XP) 8.11
Expanding the right side of the equation:
S(c) = yly - yTXP - BTXTy + PTXTXP 8.12
Setting the derivative of S(c) to zero solves for the minimum residual error:
a= -2Xy + 2XTX
ac = 0 8.13
Rearranging Equation 8.13 yields the Normal Equation:
X Y =X TX# 8.14
Solving for fl:
# = (XTX) lXTy 8.15

The strategy is to "learn" values of f using known historical values ofy and X and Equation 8.15
(training set). Then the learned values of j8 and the real-time values of Xj can be used to calculate
84
f(ti) using Equation 8.6. The following subsection will explain how the explanatory variables X
are defined and calculated.
8.2 Explanatory Variables

The explanatory variables X must be (1) measurable and predictable factors that (2) cause
differences in energy use between the past and future. The past is defined by plong For example,
if Thong is the energy use over the past 300 miles, then the past is defined as the past 300 miles
worth of driving. The value of fliong compresses a large amount of microstructure driving data
into a single measurement that quantifies how energy has been used historically. When using
5 iong to estimate the future, the value needs to be adjusted up or down if the future is different
from the past. The future is the data contained between x(t,) and x(t). (1) The Explanatory
Variables must be factors that indicate a difference in average energy use between the past
(lpiong) and future (f). For example, assume that the energy and distance values used to
determine Plong were collected while the average ambient temperature was 15'C, while it is
predicted that the future will contain a average ambient temperature of 5*C. Since auxiliary
heater use at 15'C ambient will be less than that at 5'C, it can be assumed that the change in
ambient temperature may cause a change in future energy use. Thus the change in the ambient
temperature's arithmetic mean can be one of the explanatory variables. (2) The Explanatory
Variables should be measuredfrom past data and estimated (predicted)into the future: The
ambient temperature serves as a good example of a factor that can be easily measured using a
temperature sensor. The future temperature can be predicted using weather forecasts and/or by
assuming that the temperature at to will be equal to the arithmetic mean of the future temperature.
There are other factors that might cause a difference in energy use but few can be directly
measured and predicted. For example, the driver's mood might cause an increase in future
energy use though it cannot be easily measured and predicted. Given these constraints, the
following explanatory variables were used:
(1) Change in Ambient Temperature: As described previously, the change in the arithmetic mean
of the ambient temperature is defined as:
ALTa(ti) Ia,f(ti) - 201 - |Ta,p(ti) - 20| 8.16
85
Where Ta,p(ti) is the arithmetic mean of the past ambient temperature and Taf (ti) is the
estimated arithmetic mean of the future ambient temperature. The difference from 20*C is taken
since it is assumed that temperatures above and below 20*C cause an increase auxiliary energy.
(2) Change in Traffic Conditions: Traffic conditions have a significant influence on energy use
and recent advances in traffic sensing techniques make traffic measurements and predictions
possible. For example, Google Maps is able to provide a quantitative measure of the upcoming
(future) traffic delay and corresponding estimates of average speeds over distance segments [35].
This information is accessible in real time via their internet-based Application Programming
Interface (API). If the future route is specified, an estimate for future traffic conditions can be
made. However, to reduce the scope of this project, it was assumed that the traffic conditions can
be captured through the percentage of time spent at idle conditions (i.e. zero speed):
Atile(ti) -i) tdle,p(ti) 8.17
Where Fidl,,(ti) is the arithmetic mean of the percentage of time spent at idle in the past and
tidle,f (ti) is the estimated arithmetic mean of the percentage of time spent at idle in the future.
(3) Change in Average Speed: It is well known that high speeds require more energy [30]. It was
assumed that the arithmetic mean of the future speed can be estimated using route information
from Google Maps (or equivalent). Thus a difference in mean speed was used as an explanatory
variable:
Avave(ti) Vf (ti) - i ,(ti) 8.18
Where iU,(ti) is the arithmetic mean of the past speed and if (ti) is the estimated arithmetic
mean of the future speed.
8.3 Creating a Training Dataset

Assume that multiple historical discharges are being analyzed with each discharge
starting at 100% SOC and data is available to calculate both iong and pYf. The actual values of
y(ti) can be determined by rewriting Equation 8.1:
y(ti) = Pf(ti) piang (ti) 8.19
86
Since there are n discrete values of ti, the values of y(ti) can be written as a nx] vector:
Y 8.20
The values of ATa, At ide and AVave for each time ti form the explanatory matrix:
[1 ATa(ti) Atidle(ti) Avave(ti)1
[ 8.21
1 ATattn) Atidge (tn) Avave(tn)-
The following section describes how this process is repeated for multiple discharges to build a
training set with a sufficient regression fit and to avoid extrapolation [29].
8.4 Algorithm Simulation

It was necessary to first train the multivariate regression model by simulating -300 km of
vehicle driving (Chapter 6) to generate an initial historical dataset (e.g. plong, Ta,p, aidle,p, and
1y,) based on the specified vehicle and environmental conditions. A constant ambient temperature
of 200C was assumed though other scenarios could be easily used.
Next, multiple full-discharge simulations with varying conditions were used as a
"training dataset." For example, a random speed profile was generated with an ambient
temperature of 10 C. This process was repeated for different temperatures (e.g. 15'C, 20'C, etc.)
to generate a training dataset large enough that extrapolation does not occur when the regression
is used. The temperatures were held constant for each discharge though the speed profile was
generated stochastically with a variety of city, highway and aggressiveness conditions. The
values of ATa, Atidle and Avave and y were determined for each 1% SOC increment in the
discharge. The full training dataset was used with Equation 8.15 to determine the values of/pi.
Once values of Pl
3
ong and p, were calculated, an actual discharge dataset was obtained
and used to simulate DTE algorithms. A new stochastic speed profile and a random constant
ambient temperature were used to simulate the actual discharge dataset. The DTE algorithm then
used all of the collected datasets. This process can be repeated multiple times using different
parameters and conditions.
In summary, there were three datasets used:
Historical Dataset: Multiple full battery discharges used to determine the past driving energy
consumption data.
87
Training Dataset: Multiple full battery discharges with varying conditions used to train the
regression model and thus determine pi values.
Actual Discharge Dataset: A single discharge that is meant to be the observed discharge for
simulating DTE algorithms.
In practice, the values of Taf (t1 ), tidle,f (t1 ), and ir(ti) will need to be estimated using Google
Maps, weather information, etc. However, the simulations shown in this thesis had perfect
knowledge of the future.
Finally, it is best to discretize the problem based on State of Charge (SOC), since the
range of values is constant {0,100} between the various training sets (unlike time and distance,
which have a range of values that change based on driving conditions). In discretized form: SOC
-> Z E {0, 100}:
SOC(tj) = SOCi 8.22
For example, tso is the time when SOC = 50%.
8.5 Validating the Assumptions and Fit of the Regression

Model
The assumptions and fit of the regression model were tested using the output from the
simulations described in Section 8.4. The assumptions of the multivariate linear regression model
can be summarized as follows [29]:
" Assumption 1: The linear model adequately describes the behavior of the data.
" Assumption 2: The residual error, , (Equation 8.3), is an independent and normally
distributed random variable, with a zero mean and variance < .
There are various checks commonly used to evaluate these underlying assumptions for a given
training dataset. An observation with Cook's distance larger than three times the mean Cook's
distance might be an outlier [31] and the fit might be improved if the outliers are removed.
Figure 8.1a shows that < 3% of the data points for this training dataset are considered outliers.
The second is to perform a significance test on each of the coefficients, p, (Equation 8.6), to
ensure that the model is not over-specified. This is done by testing the null hypothesis that each
coefficient is zero, which yields a pValue (Table 8.2). The small pValues (<< 0.05) and large R2
indicates that the model is valid (Table 8.3). To verify the conditions in Assumption 2, a
88
frequency distribution of the residual error was used to verify normality. It can be seen from
Figure 8.1b that the distribution has a normal shape.
Given the results from the above analysis, there is sufficient evidence to support the
assumptions made in the multivariate linear regression model. The following sections will
investigate whether or not the regression model improves DTE estimates over the conventional
blend approach.
Table 8.2: Metrics used to determine significance of coefficients.
Standard
Estimate of fi tStat pValue
Error
Intercept 1.08 0.00381 283 0
ATa 0.00868 0.000161 53.9 3.82E-296
Atidle 0.0306 0.000746 41.1 1.24E-215
Avave 0.0291 0.000629 46.34 9.05E-250
Table 8.3: Metrics used to determine significance of coefficients and fit of regression.
Metric Value
Root Mean Squared
0.035
Error
R2 0.86
F-Statistic (v. constant 2.0e+03
model)
pValue 0
U.1 2001
0.09
(1) 0.08
0 150-
C 0.07
4-0
CO)0.06
0.05 100-
Cr
0.04
0.03 LL
0 50-
0.02
S 1 3
0.01
0 200 40o 600 800 1000 U 0.1

Row Number Residuals
(a) (b)
Figure 8.1: The Cook Distance (a) and a histogram of residuals (b) are used to evaluate the
assumptions of the regression analysis.
89
90
9. Comparing DTE Algorithms
This chapter uses the stochastic vehicle simulation (Chapter 6) to quantitatively compare
the conventional (Section 7.2) and new regression-based DTE algorithms (Chapter 8). Section 9.1
describes the metrics used for comparing the quality of DTE algorithms and the results are shown
in Section 9.2.
9.1 Measuring the Performance of DTE Algorithms

The error in estimating DTE was defined in Section 7.1 as:
eDTE ( - TE (0 - DTE (t 7.9
Figure 7.1 illustrates this same concept schematically. The performance of the DTE algorithm can
be determined by visually inspecting plots such as Figure 7.1 or more quantitatively by
determining the arithmetic mean (average) error over sections of the discharge. It is useful to
discretize the problem so that each ith value of error is written as eDTEj. The algorithm can both
over-estimate (eDTE,i > 0) and under-estimate (eDTE, <0). Thus the average over-estimation
error is:
i1
eDTE,i eDTE, > 0 9.1

nj =1
Where n is the number of increments that have eDTE,i > 0. And the average under-estimation
error is:
n2
IeDT
eDTE,i <0 9.2
n2 =1
Where n2 is the number of increments that have eDTEj < 0.

The accuracy of the DTE algorithm changes as the battery is discharged. For example, the
DTE estimate may be less accurate at the beginning of the trip but more accurate as the battery
pack is drained. Thus the discharge was divided into three sections: start, middle and end, and
the average error was calculated within these sections (Figure 9.1 b).
91
- Overestimating
Underestimating
Overestimating
nderestimating
Distance Traveled Start Mid. End

Distance Traveled (by sections)
(a) (b)
Figure 9.1: The over- and under-estimation error is shown for the entire interval (a) or averaged by
the start, middle and end sections of the full discharge (b).
9.2 Results
The stochastic vehicle simulation (Chapter 6) was used to quantitatively compare the
conventional (Section 7.2) and new regression-based DTE algorithms (Chapter 8). The training
set was first obtained as described in Section 8.3. Figure 9.2 shows one example full discharge,
though more scenarios are simulated later in this section. The speed and battery energy profiles
are shown in Figure 9.2a and b and the resulting DTE plots are shown in Figure 9.2c and d. This
particular simulation assumed that the discharge dataset occurred with an ambient temperature of
100C as to simulate a drop in temperature from the past (which had an ambient temperature of
20 0 C as described in Section 7.2). The graphical representation of DTE error, as discussed in
Section 9.1, provides a quantitative comparison of the algorithms. Figure 9.2c shows that, for
this example simulation, the regression-based approach reduced the error at the beginning of
discharge by -15%. It is important to note that the ability to accurately estimate DTE at the start
of a full charge (key-on) is likely the most critical task of an algorithm since the driver uses this
estimate to plan their route.
To better understand how the algorithms perform for a wider variety of conditions, 1000
full battery discharges were simulated using stochastic speed profiles and randomly generated
ambient temperatures (0 to 22 0 C) for the actual discharge dataset. For each case the same
historicaland training datasets described in Section 8.3 were used. To simplify the analysis,
only the DTE error at the beginning of the discharge (key-on) was used for comparison (Figure
9.3). Figure 9.4a shows that the error for the regression-based approach has a smaller arithmetic
92
mean. A t-test showed that there is statistical significance between the two means [29]. The
probability density plots were estimated using the simulation data and built-in Matlab functions
[45][49].
Another approach to comparing the two algorithms is to measure the reduction in DTE
error at the beginning of discharge, which is defined as:
Reduction in DTE Error(key - on) = IDTE,rt ~~DTE,b 9.3

|eDTE,b (t0
Where eDTE,r(t 0 ) and eDTE,b(to) are the DTE error at key-on for the regression and blended
algorithm, respectively. This measures the percentage reduction that the regression algorithm has
over the blended algorithm. For example, a -100% reduction in error corresponds to a situation
where the regression-based algorithm eliminated all of the error. The error reduction was
calculated for the 1000 full battery discharges and a frequency distribution of the results is
shown in Figure 9.4b. The regression-based algorithm performed better than the blend algorithm
for cases when the error reduction is in between -100% and zero, and worse when the reduction
is greater or equal to zero. Overall it can be seen that the regression-based algorithm reduced the
error, and thus improved the DTE estimate ~90% of the time.
Next, it was of interest to plot eDTE(tO) versus the explanatory variables (ATa, Atidle and
Avave) to explore how changes in the explanatory variables affect errors at the very beginning of
a full discharge (i.e. eDTE(tO)). For example, assume the vehicle was driven for 300 km in low
traffic with an ambient temperature of 20 0 C. Now imagine that the vehicle is being driven,
starting with a full battery, in higher traffic conditions with a hotter ambient temperature of 27 C.
This would result in a ATa = 7'C. It is expected that eDTE(tO) would increase for the blended
algorithm since it is not able to anticipate these changes. The results from the 1000 simulations
are shown in Figure 9.5. Analysis of Covariance (ANCOVA) was performed for each dataset,
which showed statistical significance between the least-squares slopes [50]. In other words, the
DTE errorfor the regression-basedalgorithm is less impacted by changes in driving conditions.
93
140
120
10
E W0 ~II' 0)
a)
C
-li 60 w
a,
40 Cu
20
0
0
1. ~!I h
I000
L
2000 3000 4000 5000
~ I;0-
6000
Time (sec Distance (km)
(a) (b)
20
tn
1
10
E
w w 0-
0
-0
-10
Blend
Regression
-20
40 60 80 100 140 Start Mid End
Distance Traveled (kin) Distance Traveled (by sections)
(c) (d)
Figure 9.2: Example output from a DTE simulation. The regression-based approach reduced the
error at the beginning of discharge by -15%.
40 60 80 100 120 140

Distance Traveled (km)
Figure 9.3: The DTE error at key-on is defined as the error at the beginning of discharge.
94
0.1
a)
-0
0)
L-
U-
-25 0 25 50 30 -150 -100 -50 0 50 100 150 200

DTE Error (key-on) [%] Reduction in DTE Error (key-on) [%J
(a) (b)
Figure 9.4: The arithmetic mean error is significantly less for the regression-based algorithm (a). A
frequency plot of the reduction in error, as defined in Equation 9.3, shows that the regression-based
algorithm reduced the error -90% of the time (b).
40
* Blend
30 0 Linear Reg.
20
4-'
1
10
-10
-20
0) 5 10 15 20
IaI [*]
(a)
0 2 0
4-'
U
LU
-1d
-206
5 10 15 4
|Atidle [%] I|av, I [km/h]
(b) (c)
Figure 9.5: Multiple simulations show how the key-on error in DTE is affected by changes in driving
conditions (explanatory variables). These plots show that the linear regression algorithm is less
sensitive to changes in driving conditions.
95
96
Part III: Summary and Contributions
97
98
10. Conclusions
This chapter contains a summary of the entire thesis, a list of the original contributions
made to the field of electric vehicles and a description of ongoing and future work.
10.1 Summary of Thesis

Part I described the process of designing, building, testing and racing a high performance
electric motorcycle for the 2011 Isle of Man Tourist Trophy (TT) Zero race. Systems
engineering was used to size the batteries and motors, predict vehicle speeds and predict the time
required to finish the race. Two types of simulations were considered. The first estimated the
motor and battery power and vehicle speed when full-throttle is applied. The second simulation
type estimated upper and lower bounds for the battery energy required to traverse the entire
course. The lower bound, or best case, was determined by assuming the motorcycle maintained a
constant vehicle speed throughout the course, which was shown to be the most energy efficient
way to drive. The upper bound, or worst case, was determined by assuming the motorcycle
maintained a constant motor power throughout the course, which would yield large speed
fluctuations.
Two motors were combined with a rigid shaft to provide a maximum of 32 kWh of
continuous power. The energy storage consisted of 106 kg of lithium-ion batteries totaling 11.9
kWh. A structural frame was needed to support the batteries, motor, and supporting electronics.
The removal of the engine needed to be considered in the mechanical design, since it was
originally a structural element. Instead of a more common space-frame weldment, it was decided
to design a frame that could be made on the waterjet using aluminum plates. The waterjet was
chosen because it is rapid, economical, and provides significant design flexibility. Designing a
set of 2D plates also made prototyping easy; prototypes were made quickly with a laser cutter
using wood and cardboard. The frame was assembled like a puzzle with small tabs connecting
the various 2D plates. The final aluminum frame, including the tab-slots, was welded while
bolted in place on the motorcycle chassis.
The motorcycle was instrumented with sensors to measure acceleration, speed, distance,
current, voltage, and temperature. A battery management system (BMS) was used to measure the
99
voltage of each cell and perform cell balancing. A safety monitoring system from A123 Systems
checked for ground faults, the condition of the fuse, and over current. The data was transmitted
wirelessly to an off-board real-time display and logged on-board. The rider display consisted of
LEDs to indicate the state of the motorcycle and a single LCD screen. The screen displayed
digital real-time values for battery power, battery energy consumption and vehicle speed.
A series of incremental tests were performed to ensure that each subsystem operated as
predicted. This included the following tests: bench-top, wind tunnel, chassis dynamometer, track,
and finally road. In June 2011 the motorcycle design and simulations were tested in three races
on the Isle of Man. Post-race analysis showed that the predictions had less than 10% error.
The energy estimation methods that were developed for the motorcycle were
subsequently modified and applied to non-racing electric vehicles (EV). Instead of predicting the
energy required to traverse a known route (or race course), it is more useful for non-racing
applications to consider the reverse scenario, which is the distance the vehicle can travel before
charging is required. This is referred to as the Distance to Empty (DTE) and an estimate is
displayed in real-time in the vehicle's dashboard. It was shown that a DTE algorithm must predict
the future energy consumption of a vehicle. Future energy can be predicted reliably if either (i)
future energy consumption is sufficiently similar to the past or (ii) applicable information about
the future is known beforehand. A conventional DTE algorithm assumes (i) by "blending" both a
long-term and short-term average of past energy use. However, significant changes in driving
conditions (e.g. traffic or auxiliary energy use) for sustained periods of time can cause large
errors in DTE estimates. This thesis showed that DTE error can be reduced if future changes are
detected beforehand and used by the algorithm. Instead of using a complex parametric physics-
based model, a multivariate linear regression-based model was derived that adjusts the value of
P,ng up or down based on estimated changes in driving conditions (temperature, traffic and
speeds). In practice the driver would specify their destination(s) and a service such as Google
Maps would provide an estimate of the future driving conditions to the algorithm. When
compared to the conventional blend algorithm, it was shown that the DTE error for the new
regression-based algorithm is less sensitive to changes in driving conditions.
There are two additional advantages to the proposed regression-based algorithm. The first
is that it is computational light and thus can be run in real-time with a variety of processor
speeds. The second is that it does not require vehicle-specific calibration and validation. In other
100
words, the algorithm can be learned (fitted) over time by simply capturing data that is already
available on the CAN-bus of most EVs. These unique advantages make the approach conducive
to mobile phone and cloud-based computing services currently being developed [49].
The DTE algorithms were compared using a stochastic, parametric and physics-based
electric vehicle simulation. Though real driving data could be used, the simulation approach
yields an unlimited number of full battery discharges under a wide variety of realistic driving
conditions. Instead of using EPA speed profiles as input, a method for generating stochastic
speed profiles using a mode-based Markov model was derived. A large set of driving data (speed
versus time) was parsed into sequential Markov states, which were defined by deceleration,
cruise, acceleration or idle events. The Transition Probability Matrix (TPM) was determined by
observing the transitions between states (e.g. counting the number of times the state transitioned
from cruise to deceleration). The TPM along with a random number generator was used to string
together a stochastic speed profile of any length. To simulate a variety of conditions, the TPM
matrix was modified to give preference to certain modes (e.g. less aggressive driving was
simulated by putting zeros in the columns of the TPM associated with the largest acceleration
and deceleration modes). The algorithm is capable of providing speed versus time profiles that
are stochastic, variable in length (e.g. full discharge), and based on actual driving data.
10.2 Summary of Contributions

Below is a summary of the original research and contributions presented in this thesis.
(1) Derived and validated a method for estimating the battery energy required for electric
vehicle racing.
Predicting the exact amount of energy required to traverse the Isle of Man course is
difficult since the speed profile is unknown beforehand. So lower- and upper-bound scenarios
were estimated with the hypothesis that the actual required energy lay somewhere in between
these values. The lower bound, or best case, was determined by assuming the motorcycle
maintained a constant vehicle speed throughout the course, which was proven to be the most
energy efficient way to drive. The upper bound, or worst case, was determined by assuming the
motorcycle maintained a constant motor power throughout the course, which would yield large
speed fluctuations and is thus inefficient. These scenarios were simulated using first-order
101
physics-based models of the motorcycle system. It was shown that the racing data supports this
hypothesis.
(2) Developed fundamental concepts critical to understanding DTE algorithms.

Equations were derived to aid in the understanding of Distance to Empty (DTE)
algorithms. It was shown that the main objective of a DTE algorithm is to predict the future
average energy use of the vehicle (f). Future energy can be predicted reliably if either (i) future
energy consumption is sufficiently similar to the past or (ii) applicable information about the
future is known beforehand. Equations were derived to show that DTE error is larger when fif is
underestimated versus when it is overestimated by the same percentage amount.
(3) Formulated an improved DTE algorithm that uses a past energy consumption
measurement and future route information.
It was shown that DTE error can be reduced if the future route is known beforehand and
thus changes in temperature, traffic and speed are used by the algorithm. Specifically, a
multivariate linear regression-based model was derived that adjusts the past average energy use
based on estimated changes in driving conditions (temperature, traffic and speed).
(4) Created a stochastic vehicle simulation environment that was used to compare DTE
algorithms.
In order to compare DTE algorithms it was necessary to develop a stochastic, parametric
and physics-based electric vehicle simulation. The simulation approach provided an unlimited
number of full battery discharges under a wide variety of realistic driving conditions. Instead of
using EPA speed profiles as input, a method for generating stochastic speed profiles using a
mode-based Markov model was derived.
10.3 Ongoing Work

Below is a summary of the ongoing work related to the research presented in this thesis.
102
10.3.1 Increasing Average Speeds at the Isle of Man TT Zero
The speeds of the Isle of Man TT Zero race are limited by the energy density of currently
available batteries. In other words, the race is a battery energy-limited design problem where the
motors are able to output larger amounts of energy and power than the batteries can actually
store given the volumetric limitations of the motorcycle. Given this situation, it is useful for
teams to (i) increase vehicle efficiency and (ii) ensure that all of the available battery energy is
consumed during the race (i.e. no battery energy remains when the motorcycle crosses the finish
line). The best approach to improving efficiency is to reduce aerodynamic and rolling drag. But
it is also possible to increase efficiency by following the most efficient speed profile given the
terrain. For example, the overall energy efficiency might be improved by driving more quickly
during certain portions of the course than others. This has been discussed in the literature for
other applications and could be applied and tested in the Isle of Man race [8][9].
The second (ii) could be accomplished by performing real-time estimates of DTE, which is
currently not being done with any level of sophistication. It was shown in this thesis that the fast
and consistent speeds and low auxiliary loads of the motorcycle yield fairly constant energy use.
This means that a fairly simple past-averaging DTE algorithm would likely yield accurate results.
This is supported by the race data shown in Figure 10.1, which shows the energy use averaged
over 1 km (P1) and 10 km (510). It can be seen that the average energy use does a fairly good
job at predicting the future energy use, 1f, especially at the beginning of the race. Thus a basic
algorithm could use the past 10 km average of energy use to predict the future energy use and
thus DTE (see Equation 7.4).
600
500
- 400
CO,
300
>, 200
-
0
LU
0 10 20 30 40 50 60 70
Distance Traveled (kin)
Figure 10.1: Average energy use of the motorcycle during the Isle of Man race. This figures shows
that the fast and consistent speeds and low auxiliary loads of the motorcycle yielded fairly constant
energy use. A basic algorithm could use a past 10 km average of energy use to predict the future
energy use and thus DTE.
103
10.3.2 Advancing DTE Algorithms
When simulating the regression-based algorithm, it was assumed that the future driving
conditions (temperature, traffic and speed) were known perfectly. It would be useful to add noise
(errors) to these values to more accurately simulate the uncertainty that will exist in real-world
applications. It is also important to better understand the accuracy and utility of real traffic and
route data (e.g. provided by Google Maps). For example, traffic was quantified in this thesis by
measuring the average time at idle, fidle. In practice there are other metrics used to measure
traffic (e.g. a number ranging from 0 to 10 depending on the amount of traffic).
There are other regression methods that could be used to adjust pong based on estimates
of future conditions. For example, historical data could be used to track energy use of repeated
routes and/or conditions. Trip types could be categorized and phong could be adjusted
accordingly.
10.3.3 Comparing DTE Algorithms

The Distance to Empty algorithms were compared in this thesis using a stochastic vehicle
simulation, though it is desirable to begin testing the algorithms using real driving data. This
would require an integrated cloud-based approach that simultaneously measures vehicle speed,
energy use, ambient temperature and traffic. This information could be obtained by recording
data from the vehicle's CAN-bus while also obtaining real-time information from Google Maps
(traffic). Though it would be a significant research effort, recent advances in cloud-based
automotive hardware make this task feasible [51] [52].
104
References
[1] Honda Isle of Man TT Race Declaration, Retrieved May 4, 2013, from
http://world.honda.com/MotoGP/history/Man-TT-Declaration/
[2] Frank, A., 2003, Honda Motorcycles, MBI Publishing.
[3] Stuecke, P., 2011, "Development of a Racing Motorbike with Electric Power Train,"
SAE International.
[4] Dorrell, D., 2012, "Design of brushless permanent-magnet DC motors for racing
motorcycles," IEEE International Symposium.
[5] Cavaiuolo, D., 2012, "Design of an Electric Motor Controller with Embedded Dynamic
Thermal Control Logic for Motorbike Racing Application," Electric Vehicle Conference
(IEVC).
[6] Pudney, P., 2000, "Optimal energy management for solar-powered cars," PhD Thesis,
University of South Australia.
[7] Pudney, P., 2002, "Critical Speed Control of a Solar Car," Optimization and Engineering.
[8] Dib, W., 2011, "Optimal Control to Minimize Trip Time and Energy Consumption in
Electric Vehicles," IEEE Vehicle Power and Propulsion Conference (VPPC).
[9] Petit, N., 2011, "Optimal drive of electric vehicles using an inversion-based trajectory
generation," 8th IFAC World Congress Milano.
[10] Hutcheson, R., 2010, "Conceptual Design of a Formula Hybrid Powertrain System
Utilizing Functionality-based Modeling Tools," ASME IDETC/CIE.
[11] Same, A., 2010, "A study on optimization of hybrid drive train using Advanced Vehicle
Simulator (ADVISOR)," Journal of Power Sources.
[12] Markel, T., 2002, "ADVISOR: a systems analysis tool for advanced vehicle modeling,"
Journal of Power Sources.
[13] Guzzella, L., 2007, "Vehicle propulsion systems: introduction to modeling and
optimization," Springer.
[14] Ehsani, M., 2010, "Modern electric, hybrid electric, and fuel cell vehicles: fundamentals,
theory, and design," CRC Press.
105
[15] Wipke, K.B., 1999, "ADVISOR 2.1: a user-friendly advanced powertrain simulation
using a combined backward/forward approach," IEEE Transactions on Vehicular
Technology.
[16] Corbett, S., 2003, "NASCAR Sponsorship: Putting Your Company in the Driver's Seat,"
Journal of Integrated Communications.
[17] Larminie, J., 2012, Electric Vehicle Technology Explained 2 "dEdition, Whiley and Sons
Ltd.
[18] Franke, T., 2011, "Experiencing Range in an Electric Vehicle - Understanding

Psychological Barriers." Applied Psychology: An International Review.
[19] Bakker, J., 2011, "Contesting range anxiety: The role of electric vehicle charging
infrastructure in the transportation transition." M.S. Thesis, Eindhoven University of
Technology.
[20] Nissan Motor Company, "How Conditions Affect Range." Retrieved May 4, 2013, from
http://www.nissanusa.com/leaf-electric-car/range
[21] Hayes, J., 2011, "Simplified Electric Vehicle Power Train Models and Range
Estimation." Vehicle Power and Propulsion Conference (VPPC) (IEEE).
[22] Karbowski, D., 2012, "Energy Consumption Prediction of a Vehicle along a User-
Specified Real-World Trip." Proceedings from the Electric Vehicle Symposium
(EVS26).
[23] Minett, C., 2011, "Eco-routing: comparing the fuel consumption of different routes
between an origin and destination using field test speed profiles and synthetic speed
profiles." IEEE Forum on Integrated and Sustainable Transportation Systems.
[24] Zhang, Y., 2011, "Remaining Driving Range Estimation of Electric Vehicle." Hitachi
(China) R&D Corporation.
[25] Ferreira, J., 2012, "Data Mining Approach for Range Prediction of Electric Vehicle."
Conference on Future Automotive Technology - Focus Electromobility.
[26] Yu, H., 2012, "Driving Pattern Identification for EV Range Estimation." Electric Vehicle
Conference (IEVC).
[27] Oehlerking, A., 2011, "StreetSmart: Modeling Vehicle Fuel Consumption with Mobile
Phone Sensor Data through a Participatory Sensing Framework." M.S. Thesis,
Massachusetts Institute of Technology.
106
[28] Rodgers, L., 2012, "Designing an electric motorcycle for the Isle of Man TT Zero race,
and how electric vehicle racing could be used to spur innovation." Proceedings for the
Electric Vehicle Symposium.
[29] Freund, R., 2006, Regression Analysis, Elsevier Inc.
[30] Guzzella, L., 2007, Vehicle propulsion systems: introduction to modeling and
optimization, Springer.
[31] MathWorks Inc., Matlab Documentation, R2012b.
[32] Viet, Q, 1994, "Electronic System and Method for Calculating Distance to Empty for
Motorized Vehicles." Ford Motor Company, Patent No. US 5301113.
[33] Cho, I, 2012, "Methods for Estimating Remaining Travel Distance of Electric Vehicle."
Kia Motor Corporation, Patent No. US 2012/0143435 Al.
[34] Siy, T., 2012, "Electric Vehicle Range Prediction." General Motors Corp., Patent No.
US20120245750 Al.
[35] Google Maps API. Retrieved May 4, 2013, from https://developers.google.com/maps/
[36] Lin, J., 2002, "A Markov Process Approach to Driving Cycle Development," University
of California Davis, PhD Thesis.
[37] Gong, Q., 2011, "An Iterative Markov Chain Approach for Generating Vehicle Driving
Cycles," SAE International.
[38] Carlson, T., 1997, "Development of Speed Correction Cycles," Sierra Research, Inc.
[39] Dai, Z., 2008, "Driving Cycles: A New Cycle-Building Method that Better Represents
Real-World Emissions," University of California Davis.
[40] Nydahl, D., 2009, "Generation of Stochastic Driving Cycles," Chalmers University of
Technology, Masters Thesis.
[41] Lin, C., 2004, "A Stochastic Control Strategy for Hybrid Electric Vehicles." Proceedings
from the American Control Conference.
[42] Stroock, D., 2005, "An Introduction to Markov Processes," Springer Publication.
[43] Environmental Protection Agenency, 2008, "Kansas City PM Characterization Study

Final Report."
[44] Mathworks Inc., K-means Clustering Documentation. Retrieved May 4, 2013, from
http://www.mathworks.com/help/stats/kmeans.html
107
[45] Samadani, S., 2012, "A Review Study of Methods for Lithium-ion Battery Health and
Remaining Life Estimation in Hybrid Electric Vehicles." SAE International.
[46] Onorato, J, 2011, "Estimating Remaining Use Time of a Mobile Device." Google Inc.,
Patent No. US 7,960,945 Bl.
[47] Tashiro, E, 2003, "Traffic routing method and apparatus for navigation system to predict
travel time and departure time." Patent No. US 2005/0096842 Al.
[48] Mathworks Inc., Kernel Smoothing Function Estimate Documentation. Retrieved May 4,
2013, from http://www.mathworks.com/help/stats/ksdensity.html
[49] Bowman, A. W., 1997, "Applied Smoothing Techniques for Data Analysis," Oxford
University Press Inc.
[50] Mathworks Inc., Interactive Analysis of Covariance Documentation. Retrieved May 4,

2013, from http://www.mathworks.com/help/stats/aoctool.html
[51] Wilhelm, E., 2012, "CloudCar: The Vehicular Internet of Things." Verizon's Innovation
Center, CloudCar.mobi
[52] Automatic Labs Inc., "Automatic, Your Car and Smartphone, Connected." Retrieved
May 4, 2013, from http://www.automatic.com/
108

Electric Vehicle Design

Uploaded by

Copyright:

Available Formats

Electric Vehicle Design

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Electric Vehicle Design

Uploaded by

Copyright:

Available Formats

Electric Vehicle Design, Racing and Distance to Empty Algorithms

Lennon Patrick Rodgers

M.S. Mechanical Engineering AJ 25

Doctor of Philosophy in Mechanical Engineering

MASSACHUSETTS INSTITUTE OF TECHNOLOGY

0 2013 Massachusetts Institute of Technology. All rights reserved.

A uth o r............................................................................ . -............ .-:................................................

C ertified by ............................................................. ..................... ..........................

Submitted to the Department of Mechanical Engineering

Thesis Supervisor: Daniel D. Frey

1. INTRODUCTION TO PART I................................................13

1 .2 O V ERV IEW ...................................................................................................................................... .. ------------------........ 13

1.3 RELATED W ORK..............................................................................................................................--.. 14

2. M OTORCYCLE SYSTEMS ENGINEERING ..................................................................................... 15

2.1 SUBSYSTEM MODELS...........................................................................................................................----------...... 16

2.1.1 Backward-lookingSimulations .................................................................................................................. 16

2.2 FULL-THROTTLE SIMULATIONS .........................................................................................................---................... 24

2.3 CONSTANT VEHICLE SPEED SIMULATIONS ............................................................................................................ 26

2.3.1 Proofthat ConstantSpeed Minimizes Energy Consumption ..................................................... 26

2.4 CONSTANT POWER SIMULATIONS ..........................................................................................................----......-----. 29

2 .5 B RA KIN G LO SSES.........................................................................................................- - - -.... --....------------............. 3 0

3. M OTORCYCLE DESIGN AND TESTING.............................................................................................. 33

3.1 ELECTRICAL DESIGN....................................................................................................-----.....---------..................... 34

3.2 M ECHANICAL DESGNIGN ..............................................................................................................--.......--------............. 36

3 .3 T ESTIN G.................................................................................................................... .. --------......--------.................... 38

4. M OTORCYCLE RACE RESULTS AND ANALYSIS.............................................................................. 43

4.1 ANALYZING THE MOTORCYCLE PERFORMANCE................................................................................................ 43

4.2 MODEL VALIDATION .........................................................................................--......--........-----------.......................... 45

4.3 FRAMEWORK FOR SPURRING INNOVATION THROUGH RACING ..................................................................... 47

4.3.1 Consider the HistoricalContext.................................................................................................................... 48

4.3.2 Utilize the Powerof Regulation....................................................................................................................48

4.3.4 Provide Valued Entertainment..................................................................................................................... 48

4.3.5 Inspire Consumer Demand.......................................................................................................................------- 49

PART II: ESTIMATING AN ELECTRIC VEHICLE'S DISTANCE TO EMPTY....................................... 51

5.2 OVERVIEW .................................................................................................................................................................... 54

5.3 RELATED W ORK .......................................................................................................................................................... 55

6. STOCHASTIC VEHICLE SIMULATIONS FOR EVALUATING DTE ALGORITHMS.....................57

6.2 STOCHASTIC SPEED PROFILE GENERATOR ....................................................................................................... 5 8

6.3.1 Vehicle M odel........................................................................................................................................................65

7. FU N DA M ENTA L DTE CO N CEPTS......................................................................................................... 71

7.2 M EASURING THE REMAINING BATTERY ENERGY ........................................................................................... 74

7.3 OTHER A PPLICATIONS W ITH A SIM ILAR FORMULATION................................................................................ 75

7.4 CONVENTIONAL DTE A LGORITHMS..........................................................................................................................75

8. A REGRESSIO N -BA SED DTE A LGO RITH M ....................................................................................... 83

8.2 EXPLANATORY VARIABLES ......................................................................................................................................... 85

8.3 CREATING A TRAINING DATASET ............................................................................................................................ 86

8.4 A LGORITHM SIMULATION ......................................................................................................................................... 87

9. CO M PA RIN G D TE A LGORITH M S........................................................................................................ 91

9.2 RESULTS ....................................................................................................................................................................... 92

PART III: SUMMARY AND CONTRIBUTIONS ......................................................................................... 97

10. CO NCLU SIO NS .......................................................................................................................................... 99

10.2 SUM MARY OF CONTRIBUTIONS........................................................................................................................... 101

10.3 O NGOING WORK.................................................................................................................................................... 102

10.3.1 IncreasingA verage Speeds at the Isle of M an TT Zero................................................................. 103

10.3.2 A dvancing D TE A lgorithms......................................................................................................................... 104

10.3.3 Com paring DTE A lgorithms....................................................................................................................... 104

R EFEREN CES ................................................................................................................................................. 105

1.3 Related Work

Emin = mi Cj *2 + 72 + w2 + 21 + 2 w + 2w)dx} 2.49