Torque Vectoring with a feedback and feed forward controller - applied to a

through the road hybrid electric vehicle

Gerd Kaiser, Frdric Holzmann, Benot Chretien and Matthias Korte
Intedis GmbH, Delpstrasse 4, 97084 Wrzburg, Germany
Herbert Werner
Technische Universitt Hamburg-Harburg, Germany
Abstract
This paper concentrates on the torque commands
for electric propulsion motors in a through the road hy-
brid electric vehicle. By using a linear quadratic gaus-
sian controller, a at feed forward controller and a lin-
ear desired value generator the lateral vehicle dynam-
ics are inuenced. Understeering, oversteering, agility
and cornering speed can be optimized by proper con-
troller design. A 14 degree of freedom vehicle model
with a Dugoff tire model is used to simulate the vehicle
behavior. The simulation results show improved vehicle
dynamics and increased handling for the driver com-
pared to a equal distributed torque command.
1. Introduction
Today more and more vehicles are produced as hy-
brid electric vehicles (HEV). On possible solution to
implement an electric drivetrain into a internal combus-
tion engine (ICE) based vehicle is the so called Through
the Road (TtR) hybrid electric vehicle. In a TtR-HEV
[1] the existing drive line will not be changed. Only
the axle with no drive will be equipped with the elec-
tric components. These components include an electric
energy storage - mostly a lithium-ion battery - and an
electric machine controlled by an inverter.
For the electric drive there exist two design possibili-
ties. In the rst case there is one electric motor which
is controlled by one inverter and mounted somewhere
in the chassis. This system applies the motor torque
to a nal drive which routes the torque to the wheels.
For the second design two electric motors with two in-
verters are necessary. These motors can be located in
the chassis or inside the driving wheels as hub motors.
This work was supported by the European Commission under
Grant agreement no. 258133
combustion
drive line
e
n
g
i
n
e
electric
drive line
Electric
Motors
F
x,FR
F
x,FL F
x,RL
F
x,RR
M
z
F
x
CoG
Figure 1. Basic design of a TtR-HEV
With two motors the size of the machines can be smaller
and the nal drive is unnecessary. With two indepen-
dent motors it is possible to apply positive and negative
torque individually to each wheel. This means that it
is possible to accelerate and brake the wheels indepen-
dently. The torque generation of electric machines is
very quick and accurate for accelerating and braking the
driving wheels [2]. With the possibility of controlling
the wheels individually the question arises: How to dis-
tribute the torque between the two motors? To answer
this question a torque vectoring controller is developed.
In this paper a control strategy for torque vectoring will
be presented and simulated for a through the road hy-
brid electric vehicle. The control structure and design
of the sub-controllers will be shown in section 2. In
section 3 the controllers will be simulated with a 14 de-
gree of freedom vehicle model. Conclusions are given
in section 5.
2. Torque Vectoring
The basic idea of torque vectoring is that given re-
quests from the driver (steering angle, brake and accel-
eration pedal signals) will be processed and distributed
as torque commands to the wheels of the vehicle. With
the individual torque distribution the vehicle perfor-
desired
value
generator
torque
distribution
unit
vehicle
Ti,wheel
Mz
Mz
Fx
Vx,des
des
des
Vx real
real
real

a,req
Fx
LQG
& PID
controller
flat
feed
forward
Figure 2. Control structure for torque vectoring
mance, agility and safety [3] can be improved. This
topic includes optimal energy management [4] between
the ICE and the electric drive and optimal distribution
[5] of power from the combustion part and the electric
drive to the wheels.
In order to distribute the torque the structure from g-
ure 2 is used. A desired value generator gets the steer-
ing angle of the wheels and the desired acceleration
request a
req
from the pedals. With these inputs the de-
sired value generator creates favoured states (velocity,
side slip angle and yaw rate) of the vehicle. A at
feed forward controller in parallel to a combination of
a PID and Linear Quadratic Gaussian (LQG) controller
are calculating the desired force F
x
in longitudinal direc-
tion and the desired yaw moment M
z
around the vertical
axis of the vehicle. The feed forward controller is used
to improve the dynamics of the system. The LQG con-
troller reduces lateral errors due to model uncertainties
and parameter variations. Disturbances like side winds
are canceled out with the LQG controller. The PID con-
troller is used to cancel disturbances in longitudinal di-
rection. In a torque distribution unit (TDU) the yaw
and longitudinal commands are transfered to individual
wheel torque requests T
wheel,i
for the four wheels.
2.1. Feedback control
To calculate the moment about the z-axis a linear-
quadratic-Gaussian (LQG) controller [6] is used. The
states x are the desired values for the side slip angle

des
and the yaw rate
des
. A second pair of states are
the measured values for the side slip angle
meas
and
the yaw rate
meas
. It is possible to minimize the side
slip angle error e

=
des

meas
and the yaw rate er-
ror e

=
des

meas
while reducing the required yaw
moment M
z
[7]. Noise models, observer characteristics
and integral behavior can be implemented in the con-
troller design [8]. As plant basis for the LQG controller
a single track model is used. The lateral vehicle dynam-
ics [9] are reduced to the following two equations:

a
f
C
f
+a
r
C
r
mv
2
x
1

C
f
+C
r
mv
x
+
C
f
mv
x

(1)
=
a
2
f
C
f
+a
2
r
C
r
I
z
v
x
+
a
r
C
r
a
f
C
f
I
z

+
a
f
C
f
I
z
+
1
I
z
M
z
(2)
The model states are the side slip angle of the vehicle
and the yaw rate . Fixed parameters are the distance
from the center of gravity (CoG) to the front axle a
f
and
the distance from the CoG to the rear axle a
r
. C
f
repre-
sents the front cornering stiffness, C
r
the rear cornering
stiffness, m represents the mass of the vehicle and I
z
the
moment of inertia around the vertical axis.
For the LQG controller the model is linearised around
the steering angle = 0 rad. The longitudinal velocity
is labeled v
x
and will be set to v
x
= 80
km
/h in order to
linearise the model. M
z
is the torque which is generated
to turn the vehicle around the vertical axis and serves as
vehicle model input.
2.2. Feed forward control
To improve the dynamics of the system a at feed
forward controller is developed. The concept of at-
ness [10] is useful to analyse and design linear and non-
linear systems. Basically a system is called at if all in-
puts u
i
and states x
i
can be completely described by the
dened, at outputs y
i
and a number of its differentials
y
(k)
i
with respect to time. The at outputs do not have
to be the desired outputs and can be ctitious outputs
without physical relevance. To determine if a system is
at it has to fulll three conditions [11]:
1. It is possible to dene the ctitious outputs y
i
with
i = 1, ..., m as functions of the states x
i
, i = 1, ..., n,
inputs u
i
, i = 1, ..., m and a nite number of the
input derivatives u
(k)
i
, k = 1, ...,
i
. This can be ex-
pressed as followed:
y = F(x, u
1
, ..., u
(
1
)
1
, ..., u
m
, ..., u
(
m
)
m
)
= F(x, u, u, ..., u
()
)
(3)
2. The states x
i
and the inputs u
i
can be expressed as a
function of the at outputs y
i
and a nite number
of its derivatives y
(k)
i
, k = 1, ...,
i
with
m
i=1

i
n.
x =
x
(y
1
, ..., y

1
1
1
, ..., y
m
, ...., y
(1)
m
)
=
x
(y, y, ..., y
1
)
(4)
u =
u
(y
1
, ..., y

1
1
, ..., y
m
, ..., y

m
m
)
=
u
(y, y, ..., y

)
(5)
3. The number of inputs has to be the same for inputs
and outputs which means that dim y = dim u.
If a system fullls the three conditions it is called a (dif-
ferential) at system. The at system is used to build a
at feed forward controller. Advantages of a at feed
forward controller are:
The stability of the closed-loop system is not inu-
enced.
The feedback and feed forward controller can be
designed independently.
The feed forward controller is an additional tuning
possibility to change the system behavior.
The system dynamics are improved with a at feed
forward controller.
The rst step to create a at system is to use equations
1 and 2. However v
x
is a known but changing parame-
ter. By setting v
x
as a state it is possible to deal with the
change of this variable. So the vehicle model has to be
extended by the longitudinal dynamics. With the equa-
tions of the single track vehicle on the plain (x,y) it is
possible to fulll the requirements for at systems [11].
The nonlinear but at model has the following equa-
tions:
v
x
=
F
x
m
+ v
x
(6)

a
f
C
f
+a
r
C
r
mv
2
x
1

C
f
+C
r
mv
x
+
C
f
mv
x

(7)
=
a
2
f
C
f
+a
2
r
C
r
I
z
v
x
+
a
r
C
r
a
f
C
f
I
z

+
a
f
C
f
I
z
+
1
I
z
M
z
(8)
Flat systems are not unique and depend on the choice of
the at outputs. One solution to get a at system is the
following declaration: The at outputs y
i
of the system
are the velocity in longitudinal direction (v
x
), the side
slip angle of the vehicle () and the yaw rate ( ). By
applying the three atness criteria [11] the at inputs u
i
are the sum of the forces in longitudinal direction (F
x
),
the steering angle of the front wheels () and the yaw
torque (M
z
). To fulll the rst condition the following
equations can be derived:
y
1
= x
1
= v
x
(9)
y
2
= x
2
= (10)
y
3
= x
3
= (11)
The rst atness condition y = F(x) is fullled. To sat-
isfy the second condition the functions
x
and
u
have
to be found.
x
is solved with equation 9-11.
u
is
described with the following equations:
u
1
= ( y
1
y
1
y
2
y
3
)m (12)
u
2
=
1
C
f

my
1
y
2
+

a
f
C
f
a
r
C
r
y
1
+my
1

y
3
1
1
+(C
f
+C
r
)y
2
(13)
u
3
= I
z
y
3
+(a
2
f
C
f
+a
2
r
C
r
)
y
3
y
1
+(a
f
C
f
a
r
C
r
)y
2
a
f
C
f
u
2
(14)
The rst two conditions for atness are satised. The
third condition is fullled because the number of inputs
is three which is the same for the outputs. So the nonlin-
ear model from equation 6-8 becomes at. Using equa-
tions 12, 13 and 14 it is possible to calculate the desired
inputs for the vehicle model. The rst input F
x
from
equation 12 considers the longitudinal dynamics of the
vehicle. Until now this was a feed forward controller,
which is not robust against disturbances and uncertain-
ties. To improve the longitudinal dynamics a PID con-
troller is used. The input for the PID is the difference
between the desired v
x
and the measured v
x
. The output
of the controller is the force in longitudinal direction
F
x
.
The second input is the desired angular position of
the front wheels. To inuence the steering angle an ad-
ditional actuator like electric power steering (EPS) is
mandatory. Not every TtR-HEV is equipped with EPS
and to realize the torque vectoring controller at every
TtR-HEV some modications have to be done. The
at input will be neglected and equation 14 has to
be modied. The calculated input u
2
will be replaced
with the real, actual angular position of the front wheels
in order to calculate u
3
.
2.3. Torque distribution
From the at feed forward, the LQG and the PID
controller a force in x direction F
x
and a yaw torque
about the vertical axis M
z
is calculated. These two sig-
nals have to be modied to get the torque requests of
the individual wheels. For this purpose the plain dy-
namics of a dual track vehicle model can be used. The
important relations are the following two equations:
F
x
= F
x,FL
+F
x,FR
+F
x,RL
+F
x,RR
(15)
M
z
=
w
F
2
(F
x,FR
F
x,FL
)
+
w
R
2
(F
x,RR
F
x,RL
)
(16)
The wheel forces F
x,ii
can be rewritten as followed:
F
x,FL
= F
x,F
F
x,F
(17)
F
x,FR
= F
x,F
+F
x,F
(18)
F
x,RL
= F
x,R
F
x,R
(19)
F
x,RR
= F
x,R
+F
x,R
(20)
Two variables and are included to relate the front
and rear forces respectively. These two variables are
design parameters and can be changed during the oper-
ation of the vehicle.
F
x,R
= F
x,F
(21)
F
x,R
= F
x,F
(22)
Rearranging equations 15 - 22 into the following form:
F
x
= 2F
x,R
1+

(23)
M
z
=F
x,R
w
F
+w
R

(24)
So the force and moment requests are routed to force
requests of the axes F
x,F
, F
x,R
and axis-force differences
F
x,F
, F
x,R
. Combining these forces results in a force
request for every wheel.
F
x,FL
= F
x,F
F
x,F
=
F
x
2(1+)

M
z
w
F
+w
R
(25)
F
x,FR
= F
x,F
+F
x,F
=
F
x
2(1+)
+
M
z
w
F
+w
R
(26)
F
x,RL
= F
x,R
F
x,R
=
F
x
2

(1+)

M
z
w
F
+w
R
(27)
F
x,RR
= F
x,R
+F
x,R
=
F
x
2

(1+)
+
M
z
w
F
+w
R
(28)
With equations 25 - 28 it is possible to determine the
torque request to every single wheel. The force requests
have to be multiplied with the wheel radius to get the
torque request for every wheel. For the ICE driven axis
it is complicated and expensive to generate a torque dif-
ference at the two wheels. This can be done with dif-
ferential braking [12] which reduces the speed of the
vehicle or with an active differential [13]. But both op-
tions are adding costs and need additional implementa-
tion effort. When setting no torque difference
at the front, ICE driven wheels is required. The yaw
moment will be generated with the electric driven rear
wheels. The torque request can be processed very fast
with the electric motors and no additional devices are
necessary.
The acceleration force F
x
is shifted via . The vehicle
energy management system has to determine the value
of to shift between the ICE and electric propulsion
motor. With varying it is possible to drive purely ICE
based ( = 0) or purely electrical ( =). A combina-
tion of both like boosting (0 < < ) or recuperating
(1 < < 0) is also possible. For the simulation of the
torque vectoring control in section 3 is set to 0 and
is set to 10
6
. This implies that the acceleration force
F
x
is purely generated by the ICE driven wheels and the
yaw moment M
z
is generated by the electric driven rear
wheels.
2.4. Desired value generation
Desired values are necessary for the LQG and PID
controller. Additionally, the at feed forward controller
requires the desired at outputs v
x
, and . The veloc-
ity in longitudinal direction depends on the drivers wish,
and accordingly on the acceleration and the brake pedal
state. The transformation from pedal signal to a desired
velocity is very complex and involves knowledge of the
drivers feelings and intentions. So the transformation
part will be skipped in this paper and a requested veloc-
ity will be given.
The desired side slip angle and the yaw rate depend
on the lateral vehicle dynamics. The driver inuences
these states with the steering wheel angle. To include
the vehicle dynamics equations 1 and 2 are used (setting
M
z
= 0). With a given steering angle and longitudinal
velocity, the side slip angle and yaw rate can be calcu-
lated. With this implementation there are two possible
tuning directions:
1. Use torque vectoring to extend the region of lin-
ear vehicle behavior [14]. So the driver has an
extended, known vehicle operation region and can
drive the vehicle more safely to the limits of wheel
adhesion. With this strategy torque vectoring is not
active in the linear driving operation - where the
average driver manoeuvres the vehicle for most of
the time.
2. Tune the vehicle parameters (m, I
z
, a
f
, a
r
,C
f
and
C
r
) in the desired value generation model. It is
possible to set an understeering, neutral or over-
steering behavior. Or with proper parameter selec-
tion the vehicle can behave like a different vehicle.
For example it is possible to have a sedan style ve-
hicle behaving like a small, compact vehicle with
reducing the mass, the moment of inertia and the
distances to the front and rear axle of the desired
value generator.
Besides the linear calculation of the desired side slip
angle and the desired yaw rate, borders [14] for these
values are included to operate the vehicle more safely.
If the vehicle leaves the regions for and the tire
forces will start to saturate which can cause a loss of
vehicle control. This dangerous situation will be sup-
pressed with limiting
des
and
des
.
3. Simulation results
For the simulation a 14 degree of freedom (DOF)
vehicle model is used. The tire characteristics are calcu-
lated with a modied Dugoff model [15]. For the simu-
lation two vehicle congurations are used. The rst one
is a vehicle with no electric motors, the second congu-
ration is a vehicle with two independent electric motors
at the rear wheels. These motors are controlled with the
proposed torque vectoring strategy.
For the desired value generation the parameter variation
is empirically developed. The mass of the vehicle mwill
be multiplied with a factor of 1.1 and the moment of in-
ertia about the vertical axis I
z
will be multiplied with a
factor of 0.5. With these changes the desired values for
the side slip angle
des
and the yaw rate
des
will have
the same amplitude with less time delay.
As driving scenario a sinusoidal movement with de-
lay [16] is used. The advantage of this test is the given
steering trajectory and better reproducibility compared
to a double lane change test. The initial driving veloc-
ity of the vehicle is set to 80
km
/h and the vehicle will
operate at a dry road. A sinusoidal movement with am-
plitude 120 degree and frequency 0.7 Hz will be applied
to the steering wheel. After reaching the second ex-
tremum the steering angle will be held for 0.5 seconds
before nishing the sinusoidal period. As outputs to the
steering angle trajectory, the yaw rate, the side slip an-
gle, the torque at the rear wheels and the yaw moment
request from the LQG and at feed forward controller
will be displayed.
The plateaus in the desired yaw rate (while steering) are
related to the limitations of safe operation and should
not be (heavily) violated. The vehicle with torque vec-
toring is following the desired yaw rate faster and the
maximum yaw rate is lower what improves the safety
of the vehicle. Looking at the side slip angle the normal
vehicle has a strong understeering behavior between 2.5
and 3.5 seconds. This unintended behavior is reduced
with torque vectoring.
The torque requests at the rear wheels are symmetric to
the 0-torque axis. This is founded by the settings for
and . The accelerating force F
x,R
is set to zero with
= 0 from equation 22. So only the force difference
0 1 2 3 4
150
100
50
0
50
100
150
steering input with 80 km/h
s
t
e
e
r
i
n
g

w
h
e
e
l

a
n
g
l
e

i
n

d
e
g
r
e
e
time in s
0 1 2 3 4
1
0.5
0
0.5
1
compare yaw rate
y
a
w

r
a
t
e

i
n

r
a
d
/
s
time in s


desired
no eMot
torVec
0 1 2 3 4
0.2
0
0.2
0.4
0.6
compare side slip angle
s
i
d
e

s
l
i
p

a
n
g
l
e

i
n

r
a
d
time in s


desired
no eMot
torVec
0 1 2 3 4
500
0
500
torque at rear wheels
t
o
r
q
u
e

i
n

N
m


left
right
0 1 2 3 4
1
0.5
0
0.5
1
x 10
4
yaw torque request
t
o
r
q
u
e

i
n

N
m
time in s


Flat
LQG
Figure 3. Simulation results for a sinus dwell
steering test with torque vectoring in a TtR-
HEV
Table 1. Parameters of the simulation model
a
f
1.126 distance front axle to CoG in m
a
r
1.341 distance rear axle to CoG in m
C
f
80,000 Cornering stiffness of the front
axle in N
C
r
70,000 Cornering stiffness of the rear
axle in N
w
F
1.458 width of the front axle in m
w
R
1.466 width of the rear axle in m
m 1340 mass of the vehicle in kg
I
z
2454 moment of inertia around verti-
cal axis in kg
T
eMot
600 maximal torque of one electric
motor in Nm
P
eMot
30 maximal power of one electric
motor in kW
Road
1 adhesion coefcient between
road and wheel
F
x,R
from equation 21 is applied to the rear wheels.
The last plot shows the inuence of the at feed forward
and the LQG controller to the desired yaw moment. If
the parameters would be perfect estimated and if all re-
lations were included in the at design the LQG con-
troller would be obsolete. But the feed forward is still
important to improve the dynamics and relieve the LQG
controller. For comparison The most important param-
eters of the simulation are written in table 1.
4. Conclusion
The control structure for torque vectoring in a
through the road hybrid electric vehicle has been de-
veloped. The concept of a at feed forward controller
is explained and improves the dynamic of the vehicle.
Disturbances and modeling errors are rejected with a
PID controller for the longitudinal dynamics and with a
LQG controller for the lateral vehicle dynamics. Con-
cept for desired values and torque distribution to the
wheels has been shown. Simulation results showed
improvements of the vehicle behaviour if the car is
equipped with torque vectoring. The next step is to in-
clude a anti skid algorithm and test the control structure
for torque vector within a real vehicle.
References
[1] B. Chretien, F. Holzmann, G. Kaiser, S. Glaser, and
S. Mammar, Concept of Through the Road Hybrid
vehicle, in Advanced Vehicle Control (AVEC, ed.),
(Loughborough), 2010.
[2] M. Milehins, C. Cheng, T.-W. Chu, and P. Jones, Han-
dling Behaviour of a TTR Hybrid Electric Vehicle with
Independent Rear Wheel Torque Control, in Advanced
Vehicle Control, pp. 556561, AVEC, 2010.
[3] L. Pinto, S. Aldworth, M. Watkinson, P. Jeary, and
M. Franco-Jorge, Advanced Yaw Motion Control of a
Hybrid Vehicle using Twin Rear Electric Motors, in Ad-
vanced Vehicle Control, pp. 640645, AVEC, 2010.
[4] J. Kim, N. Kim, S. Hwang, Y. Hori, and H. Kim, Motor
Control of Input-Split Hybrid Electric Vehicle, Interna-
tional Journal of Atuomotive Technology, vol. 10, no. 6,
pp. 733742, 2009.
[5] K. Li, T. Chen, Y. Luo, and J. Wang, Intelligent Hybrid
Electric Vehicle and Its Implementation, in Advanced
Vehicle Control (AVEC, ed.), pp. 460465, 2010.
[6] E. Hendricks, O. Jannerup, P. H. Srensen, E. Hen-
dricks, O. Jannerup, and P. H. Srensen, Optimal Ob-
servers: Kalman Filters, in Linear Systems Control,
pp. 431491, Springer Berlin Heidelberg, 2008.
[7] P. He and Y. Hori, Optimum Traction Force Distribu-
tion for Stability Improvement of 4WD EV - In Crit-
ical Driving Condition, in IEEE - AMC06-Istaanbul,
Turkey, IEEE Xplore, 2006.
[8] www.mathworks.com, Control System Toolbox - Users
Guide. The MathWorks, Inc., 9 ed., 2010.
[9] S. Anwar, Yaw Stability Control of an Automotive Ve-
hicle via Generalized Predictive Algorithm, in Ameri-
can Control Conference, pp. 435440, AACC, 2005.
[10] M. Fliess, J. Lvine, P. Martin, and P. Rouchon, Flat-
ness and defect of non-linear systems - introductory the-
ory and examples. 1995.
[11] R. Rothfuss, J. Rudolph, and M. Zeitz, Ein neuer Zu-
gang zur Steuerung und Regelung nichtlinearer Sys-
teme, Automatisierungstechnik 45, pp. 517525, 1997.
[12] R. N. a. Jazar, Vehicle Dynamics - Theory and Appli-
cation, in Vehicle Dynamics, Springer US, 1 ed., 2008.
[13] R. Denzler, C. Granzow, R. Peter, and S. Martin, Das
Hinterachsgetriebe Vector Drive, ATZ, vol. 12, no. 109,
pp. 11061115, 2007.
[14] R. Rajamani, Vehicle Dynamics and Control. Springer
eBook Collection, Engineering [Dig. Serial], Springer-
11647 [Dig. Serial], Boston, MA: Rajesh Rajamani,
2006.
[15] Z. Shiller, Optimization Tools for Automated Vehicle
Systems. 1995.
[16] G. J. Forkenbrock, An Overview of NHTSAs 2005
ESC Research Program. 2005.