Applied Energy 134 (2014) 332341

Contents lists available at ScienceDirect

Applied Energy
journal homepage: www.elsevier.com/locate/apenergy

Optimal energy management strategy for battery powered electric

vehicles
Jiaqi Xi a, Mian Li a,b,, Min Xu b
a
b

University of MichiganShanghai Jiao Tong University Joint Institute, Shanghai Jiao Tong University, Shanghai 200240, China
National Engineering Laboratory for the Automotive Electronic Control Technology, Shanghai Jiao Tong University, Shanghai 200240, China

h i g h l i g h t s
 The power usage for battery-powered electrical vehicles with in-wheel motors is maximized.
 The battery and motor dynamics are examined emphasized on the power conversion and utilization.
 The optimal control strategy is derived and veried by simulations.
 An analytic expression of the optimal operating point is obtained.

a r t i c l e

i n f o

Article history:
Received 18 April 2014
Received in revised form 17 July 2014
Accepted 7 August 2014
Available online 29 August 2014
Keywords:
Energy management
Battery powered
EVs
Optimal control

a b s t r a c t
Due to limited energy density of batteries, energy management has always played a critical role in
improving the overall energy efciency of electric vehicles. In this paper, a key issue within the energy
management problem will be carefully tackled, i.e., maximizing the power usage of batteries for battery-powered electrical vehicles with in-wheel motors. To this end, the battery and motor dynamics will
be thoroughly examined with particular emphasis on the power conversion and power utilization. The
optimal control strategy will then be derived based on the analysis. One signicant contribution of this
work is that an analytic expression for the optimal operating point in terms of the component and environment parameters can be obtained. Owing to this nding, the derived control strategy is also rendered
a simple structure for real-time implementation. Simulation results demonstrate that the proposed
strategy works both adaptively and robustly under different driving scenarios.
2014 Elsevier Ltd. All rights reserved.

1. Introduction
In response to renewed pleadings for energy efciency and
environment protection, the electric vehicles (EVs), as a promising
substitute for the conventional ones, have received much more
attention than ever before. However, due to limited energy density
of batteries nowadays, energy management has always been the
central and critical issue in the control of EVs. Although the term
energy management may have various meanings in different contexts, all share the common goal of improving the energy efciency
and maximizing the utilization of stored energy in the batteries
equipped on the vehicle.

Corresponding author at: University of MichiganShanghai Jiao Tong University

Joint Institute, Shanghai Jiao Tong University, Shanghai 200240, China. Tel.: +86 21
34207212.
E-mail address: mianli@sjtu.edu.cn (M. Li).
http://dx.doi.org/10.1016/j.apenergy.2014.08.033
0306-2619/ 2014 Elsevier Ltd. All rights reserved.

In order to achieve the purpose of energy management, extensive research work has focused on energy control by analyzing
the component characteristics, especially for battery-powered
EVs. For example, Capasso and Veneri veried the applicability of
lithium-based batteries for EV applications [1]. Xiong et al. and
Hu et al. proposed adaptive state-of-charge estimation methods
based on real-time measurements on the battery terminal voltage
and current [2,3]. Zhong et al. developed a method for state-ofcharge estimation of the battery pack which accounted for the difference among the cells [4]. Besides the in-depth analyses of battery performance, the study of other key components of the EV
system, such as power converters and motors, also abounds in
the literature. Pahlevaninezhad et al. proposed a Control-Lyapunov-Function based approach to regulate the input power of the
inverter so that higher energy efciencies and larger stability margin could be attained [5]. Faiz et al. designed a direct torque control
law for induction motors used in EVs with the improved overall
efciency and reasonable dynamic response [6]. Although physical

J. Xi et al. / Applied Energy 134 (2014) 332341

characteristics of a specic component were fully explored in all

those work, the analysis was only performed in a somewhat isolated manner and the coupling effects between different components were neglected.
In parallel to component level analysis, much of attention in the
energy-control area has been directed to hybrid energy system
design by integrating, for example, ultra-capacitors into the energy
storage system of EVs. The dening features of the ultra-capacitor
are its high power density and long life cycle [7]. By combining the
complementary characteristics of the battery and the ultra-capacitor, a superior dynamic behavior can thus be achieved. Along this
direction, the research has been carried out in full length. For
example, Dougal et al. analyzed the performance improvement of
the hybrid energy system when the ultra-capacitor was connected
in parallel with the battery directly [8]. Lu et al. explored a new
topology to interface the battery and ultra-capacitor, and proposed
the corresponding energy management schemes [9]. Kuperman
et al. and Garcia et al. focused on the power ow control of the
hybrid energy system. An embedded control topology was proposed to guarantee that the extra power during peak time will
be provided by the ultra-capacitor while keeping the battery current unchanged [10,11]. Lukic et al. compared different topologies
in terms of efciency and stability and showed that an active combination of the battery with ultra-capacitor was a promising
approach [12]. Under this active topology, Laldin et al. designed
an optimal power split path in real time based on the prediction
of future load demand and the energy loss model for all system
components [13]. The similar problem was approached from
another perspective in the work done by Wang et al., in which
the way of hybridization between the battery and the ultra-capacitor was determined by solving an optimization problem for the
total fuel economy of the energy system [14]. Besides, there are
also fuzzy approaches proposed in the literature. Wang et al.
designed the fuzzy logic based on the power requirement of EVs
[15]. Hannan et al. presented a multi-source model and designed
a rule-based power sharing strategy according to the energy source
states and load conditions [16].
Whether it be control methods for the individual components
or power split mechanisms for hybrid systems, it can be noted that
most of the work above were targeted at minimizing the power
loss in power sharing for a given power demand. This demand is
usually either obtained priori or estimated in real time. On the contrary, another point of view in this area has been concerned with
the maximization of the travel distance for a given amount of
stored energy [17]. Instead of determining the power sharing prole at each instant, it aimed to nd an optimal velocity prole that
would maximize the travel distance. As a rst step, this work built
a simplied yet complete EV model. Based on the terrain information and the operating efciency of the in-wheel motors, the optimal velocity prole can be found using the dynamic programming
or other optimization algorithms. The originality of this work lies
in that it offers a new perspective to treating the energy management problem.
However, the approach in Ref. [17] is plagued by two major
issues that can undermine the optimality of its solution. First, the
battery and motor are statically viewed as the energy storage
and energy consumption components, and the dynamic behaviors
associated with those two components are totally ignored, not to
mention the energy ow associated with those dynamic behaviors.
Because of this intrinsic model discrepancy, the optimal velocity
prole obtained may not produce the corresponding optimal
power ow prole as expected in a real-world test. To make things
worse, the optimal velocity prole itself might not even be reproducible due to the coupling effects and dynamic constraints.
Relevant to the rst issue, the second one is that since everything
is viewed statically, the technical details of control implementa-

333

tion, especially the control of the DCDC converter, remain

untouched. The lack of this essential part of contents makes the
proposed strategy incomplete or inconvincible to some extent.
In this paper, a more systematic and rened approach is proposed
for energy management of EVs with in-wheel motors and a single
power source, i.e., batteries. Two major issues mentioned above in
previous work will be carefully tackled. The discussion will begin
with the component analysis by examining the battery and motor
dynamics respectively. The vehicle dynamics will also be accounted
for in our work and special attention will be paid to the power and
torque couplings between motor and vehicle dynamics. The energy
management problem will then be reformulated as an optimal control problem with input constraints. With the help of justied simplications, we can nally nd an analytic solution to this energy
management problem, which can serve as an important guideline
to energy control of battery-powered EVs. After that we will move
on to the control implementation of the proposed energy management strategy. Specically, we will provide the details of the PI controller design for the DCDC converter in this work which accounts
for both the dynamic response and noise rejection.
One signicant contribution of this work is that, with justiable
simplications, we identify the dependence of the optimal operating point on the component and environment parameters in an
analytic form. This not only provides a guideline on the driving
mode in different road conditions, but also offers insights to the
physical mechanism behind the optimality for further theoretical
investigation. Owing to this nding, the derived energy management strategy is also rendered a simple and compact structure that
facilitates the real-time implementation. Simulation results for
two case studies will be demonstrated and analyzed. It can be
shown that the proposed strategy works both adaptively and
robustly under different driving scenarios.
The rest of the paper is organized as follows. A brief introduction to the vehicle architecture considered in our study is provided
in Section 2. In Section 3, the component model for the battery and
the motor are derived step by step. Based on this model, the control strategy is developed accordingly. The simulation is demonstrated in Section 4, including the initial condition settings,
simulation results, and post analysis. Finally the conclusion of
the proposed work and possible future improvement are summarized in Section 5.
2. Background
In this paper, the energy management strategy will be developed for an electric vehicle driven by two rear in-wheel motors.
The vehicle topology is shown in Fig. 1. The battery power is delivered to the motors through the DCDC converter, which is controlled by an energy management unit (EMU) on the vehicle.
Notice that in this work the ultra-capacitor is not included as a
part of the power source system like other hybrid systems.
Undoubtedly adding ultra-capacitors can improve the transient
dynamics of the vehicle at acceleration and can take care of the
energy recovery at regenerative braking much better. However,
in the authors opinion, the study of power transportation and
power usage is of more fundamental importance from the energy
management point of view. In this regard we restrict our focus to
these key aspects and temporarily set aside the discussion of the
ancillary function of the ultra-capacitor, which will be an extension
of this work and addressed in our future work.
3. Proposed energy management method
The ultimate goal of energy management is to maximize the
power usage of the battery on the vehicle. To this end, two aspects

334

J. Xi et al. / Applied Energy 134 (2014) 332341

Qc
Qc
Q AeBQ  k
ib
Qc  Q
Qc  Q

v b E0  k
Motor

where Qc is the battery capacity and Q is the actual battery charge

which is a value related to the current ib dened by:

Q_ ib
Converter

EMU

Battery

Motor

Fig. 1. Vehicle topology used in our analysis.

Here the battery constant voltage E0 and polarization voltage k, as

well as the exponential zone coefcients A and B, can be identied
using the experimental data. The detailed discussion on this model
can be found in the literature [18]. Our interest is to analyze the
impact of the current intensity on the total power loss both under
charging and discharging conditions. The instant power of the battery can be calculated as: p = vb  ib, where vb is the no-load voltage
of the battery. By integrating the instant power over a period of
time, we can obtain the corresponding energy ow of the battery
as below:


E0  k

tf

t0

Qc
Q AeBQ
Qc  Q


 dQ 

tf

t0

Qc
2
i dt
Qc  Q b

W1 W2
of the matter need to be considered concurrently. The rst issue is
to minimize the power loss during the energy transportation on
the vehicle. Another issue, of course closely coupled with the rst
one, is to extend the driving range as much as possible for a given
amount of energy. These two issues occur at two ends of the same
problem and are associated with two key components of the vehicle: the battery and motor. In the following discussion, we will
start with the component-level analysis.
In the analysis of battery dynamics, we will thoroughly investigate the impact of the current intensity on the power loss during
the energy transportation. The analysis will be based on a wellknown battery model which is generally used to characterize properties of different types of batteries. In this regard, our objective for
the battery analysis is to nd a discharge prole to minimize the
power loss during the energy transportation.
In the analysis of motor dynamics, we will study the coupling
relationship between the electrical and mechanical behavior of
in-wheel motors to clearly demonstrate their dynamic characteristics (which were usually neglected in the previous work). Since inwheel motors are applied in this work, the mechanical coupling
between motor and vehicle dynamics, which depends on the way
of connection between the wheel and the vehicle, will also be
investigated and claried. Given that, our objective of study on
motor dynamics is to nd an optimal operating point which minimizes the total energy usage for a given driving range.
The study of these two components goes in parallel to each
other and gets merged under the system-level analysis by combining results from them together. The overall control hierarchy will
be proposed based on results from the component-level analysis;
and then we will discuss the real time implementation of our control strategy on the DCDC converter. At the end of this section,
several practical considerations of the control strategy will be provided including the choice of lter parameters and control
frequency.
3.1. Component analysis: battery
A generic battery model proper for dynamic characterization
has been proposed by Tremblay et al. [18], who assume that the
battery state-of-charge (SOC) is the only state variable and other
characteristics of the battery can be derived from it. In their work,
the battery is modeled as a controlled voltage source in series connection with a resistance. The expression for the voltage source is
given by

Two terms can be recognized from Eq. (3): W1 and W2. Since W1 is a
function solely depending on Q, its value is independent of any specic power ow path. In other words, W1 is a state function. W2, on
the contrary, is a process function. In view of this particular energy
feature, we only need to focus on the term W2 when analyzing the
power loss since W1 will not be affected by any specic power ow
path. A closer look at Eq. (3) reveals that the term Qc/(Qc  Q) is
strictly positive, which means the entire integrand of W2 is strictly
negative contributing to the energy loss during the irreversible process. The question following is whether there exist such a current
prole (power ow path) that can minimize this energy loss, that
is, to minimize W2. Reformulating this problem mathematically will
give us a minimization problem by changing the variable Q:

min 

tf

k
t0

Qc _ 2
Q dt;
Qc  Q

with Q t 0 Q 0 and Q t f Q f

where Q0 and Qf correspond to the initial and nal battery charge

respectively. Both of them can be easily obtained by on-line SOC
estimation of the battery. The charging and discharging process
can be differentiated by the relative magnitude of Q0 and Qf. Let
us denote the integrand in Eq. (4) as L. By invoking the Euler
Lagrangian condition [19], the following condition has to be satised for an extremal:

LQ

d
L_
dt Q

where LQ and LQ_ denote the partial derivative with respect to Q and
Q_ . Expressing the Eq. (5) explicitly yields

Q_ 2 2  Q  Q c Q

which is a second order nonlinear differential equation and the feasible solution to it takes the form as:
2

Q 2c b
 t 2 bQ c  t Q 0 ; where
4Q 0  Q c
s! 

Qf  Qc
1
Q0
1
1

b2
tf
Q0  Qc
Qc

Q

Notice that the term b in Eq. (7) will be positive in the discharging
process (Q0 < Qf) and negative in the charging process. The current
intensity can then be readily calculated as
2

ib

Q 2c b
 t bQ c
2Q 0  Q c

335

J. Xi et al. / Applied Energy 134 (2014) 332341

When the time span tf is long enough, b2 will be relatively small


compared to b. So the optimal current intensity ib will stay on a constant value b for a long charging or discharging duration.

Crr

Cd
MV

3.2. Component analysis: motor

The motor dynamics can be fully characterized in terms of its

electrical and mechanical behaviors [20]. The governing equations
of these two behaviors are derived from Kirchhoffs Voltage Law
and Newtons Second Law, respectively, as below:

dim
v m  k b  x  Rm  i m
dt
dx
Jm
kb  im  C f  x  sL
dt

Lm

9
10

In these two equations, vm and im are the input voltage and current
to the motor, and x is the angular velocity of the motor. Rm and Lm
represent the resistance and inductance of the rotor loop, while Jm
and Cf are the moment of inertia of the rotor and coefcient of viscous friction, respectively. The electrical and mechanical parts are
linked with each other through the electromagnetic conversion factor kb, which measures how the electrical energy is transformed
into the mechanical one. In Eq. (10), sL represents the external load
on the motor, which depends on the way of the connection between
the motor and the vehicle system as well as on the vehicle dynamics. In order to link these two aspects together, here we consider a
simplied vehicle dynamics model shown in Fig. 2, with road friction, rolling resistance, and drag coefcients indicated as l, Crr,
and Cd, respectively.
The vehicle here is powered by two rear in-wheel motors.
Before the analysis two basic assumptions are made: (1) the mass
center of the vehicle body is low, so the lifting effect can be
neglected and only the forces in the horizontal direction are considered; (2) the variation of the vehicle speed is within a certain
range and thus the air resistance can be considered proportional
to the vehicle velocity v [21]. The free body diagram of the entire
system including the front wheels, the rear wheels and the main
body is shown in Fig. 3. The wheels are pin-connected with the
vehicle body.
Applying the force and torque balance to the system gives:

dx
2sL  2f 1  rw 2f 2  r w  C rr  M v g
dt
dv
Mv
2f 1  2f 2  qa C d AF  r w x
dt

4J w

11
12

where x is the angular velocity of the wheel (which is the same as

that of the motor), rw is the wheel radius, and v is the corresponding
linear velocity (v = x  rw). f1 and f2 represent the static road frictions on the front and rear wheels respectively. The air resistance
has been written out as a multiplication of the air density qa, the
drag coefcient Cd, the cross-sectional area AF, and the velocity v.
sL is driving torque provided by the in-wheel motor. It is assumed
that the vehicle mass Mv is evenly distributed among the four
wheels and no slip occurs. Multiplying Eq. (12) with rw and adding

Jw

f1

f2

Fig. 3. Free body diagram of the vehicle.

to Eq. (11), we can ultimately obtain the relationship between sL

and the external load to the vehicle.

2sL M v r 2 4J w 

dx
qa C d AF  r w x C rr M v
dt

13

Substituting the expression for sL into Eq. (10) will give us

 dx
Mv r 2w 4J w 2J m 
2kb  im  2C f qa C d AF r w  x  C rr M v g
dt
14

Since in general Mv  rw2 is much larger than Jw and Jm, and Cd is

much larger than Cf, Eq. (14) can be simplied as Eq. (15) with only
dominant terms reserved

Jv

dx
2kb  im  C v  x  sv
dt

15

where Jv = Mv  rw2, Cv = qaCdAF  rw and sv = CrrMvg. Then the augmented motor model can be written as

dim
v m  kb  x  Rm  im
dt
dx
2kb  im  C v  x  sv
Jv
dt

Lm

16
17

3.2.1. Solving for the optimal operating point

With the augmented motor model discussed above, the next
goal is to nd an optimal operating prole for the motor in the
sense that the energy consumption for a given driving range is
minimized. In view of its complicated nature, we will break down
the analysis into two stages. Namely we will rst study an unconstrained case, i.e., there is no constraint on the system states and
control variable. Then we will come to tackle with the constrained
one. It will become clear later that the approach we take in solving
the unconstrained case will shed meaningful lights on the derivation of controller for the constrained one. Both of these cases will
be studied in the optimal control context. As a preliminary step,
let us specify the control objective mathematically. The control
system under consideration takes the form as:

dim
1

v m  kb  x  Rm  im
Lm
dt
dx 1
2kb  im  C v  x  sv
Jv
dt
dh
x
dt

18
19
20

where h is the traversed angle of the wheel. Dene the state vector
x = (im, x, h) and the control variable u = vm, we can rewrite Eqs.
(18)(20) in a compact form:

Cd

C rr

Jw

MV

x_ f x; u; t

21

The control objective is to drive the state h from h0 to hf with the

minimal total electrical energy. So the cost functional is

Fig. 2. Schematic of an electric vehicle on the road.

tf

t0

v m  im dt;

where

vm 2 U

22

336

J. Xi et al. / Applied Energy 134 (2014) 332341

where U is the feasible control set, which will be discussed in details

in Section 3.2.3. Overall Eq. (22) is a free time xed endpoint problem. The Hamiltonian associated with the system is

we mean that all the dynamics of the state variables tend to zero,
which means two derivatives in Eqs. (34) and (35) become zero.
Thus we have:

Hx; p; u; t pT  f x; u; t  v s  im

24


Cv
kb C v
k sv
 im
x b
 p3 0
Jv
Rm J v
Rm J v 2
2kb
Cv
1
 im 
 x   sv 0
Jv
Jv
Jv

25

Combining them with the condition in Eq. (28), we can nd a

unique solution which is:

26

23

and the corresponding co-state p is given by

dp1 Rm
2kb

p 
 p2 v s
dt
Lm 1
Jv
dp2 kb
Cv

p
 p2  p3
dt
Lm 1 J v
dp3
0
dt

By invoking the Pontryagins Minimum Principle [19], we have the

following two conditions:


Hx ; p ; u ; t 6 Hx ; p ; u; t;

for all u 2 U

Hx ; p ; u ; t 0

27
28

Eq. (28) follows the fact that H does not explicitly depend on time t.
Given the discussion above, now we can discuss the unconstrained
and constrained cases separately one by one.
3.2.2. Optimum of the unconstrained case
In the unconstrained case, the condition in Eq. (27) will indicate
that u* is a stationery point of H.

Hu x ; p ; u; t juu 0

29

By solving it we will have the relationship between im and p1 as:

1
p im 0
Lm 1

30

This equation is supposed to hold for all time, and so does its time
derivative. Substituting p1 with im in Eq. (24) and collecting the relative terms, we will obtain



dim
1
2kb

Rm  im 
 p2 v m
Lm
dt
Jv

31

2Rm  im

2kb
 p2 kb  x 0
Jv

32

Substituting p2 with im and x, and repeating the same procedure as

for p1 will give us


dim C v
kb C v
k sv

 im
x b
 p3
dt
Jv
Rm J v
Rm J v 2

33

34
35

and a quick calculation can show that the eigenvalues of this linear
system are:

k1;2

v
u 2
2
uC
k Cv
t 2v 2b
J v J v Rm

39

tf

v m  im

dt v m  im 

t0

gh

40

Here the physical meaning of the term g can be interpreted as the

energy consumption per unit distance. If im and x are the corresponding optimal states, then g should attain its minimum value.
In order to verify this point, we need to nd out the minimum value
of g. By expressing im and x in terms of vm, the term g can be
rewritten as

C v v m k b sv
vm
2kb v m  Rm sv

41

Taking derivatives with respect to vm and setting the equation to be

zero gives us:

2kb C v  v 2m  2Rm C v sv  v m  Rm kb s2v 0

42

After solving it we have


m

0
s1
2
Rm  sv @
2kb A

1 1
kb
Rm  C v

43

which coincides with the one obtained by solving the PMP condition. A further check of the second derivative of g at the value vm*
shows:
2

Finally we have a set of equations that im and x need to satisfy:


dim C v
kb C v
k sv

 im
x b
 p3
dt
Jv
Rm J v
Rm J v 2
dx 2kb
Cv
1

 im 
 x   sv
Jv
dt
Jv
Jv

0
s1
2
Rm  sv @
2kb A

1 1
kb
Rm  C v

38

Since in general Pontryagins Minimum Principle (PMP) is a necessary condition of the optimality, it is still necessary to verify the
solution to be a local minimum. Notice that our goal is to minimize
the cost function G. In the stationery case where im and x are all
constants, the expression for G can be written as:

By comparing Eq. (31) with Eq. (18), we will have


m

37

36

Since there is a positive eigenvalue, for a given initial condition, the

state variable im and x will eventually increase exponentially. By
examining Eq. (18), this implies that the control variable vm will
increase exponentially, which is impractical for a real controller.
So we need to nd a stationery solution with appropriate boundary
conditions that satisfy Eqs. (28) and (29). By a stationary solution,

d g
>0
dv 2m

44

So it is indeed the local minimum. Since g is a convex function of vm,

the local minimum is also a global one and any deviation from vm*
will result in an increase of g. Next we will discuss the constrained
case.
3.2.3. Optimum of the constrained case
In the constrained case, we will study how the optimal point
may shift in accordance with the constraints on the system states
and control variable. For the current control system, the constraint
mainly arises from safety considerations. In order to avoid slip, the
maximum angular acceleration of the wheel needs to be carefully
specied, which is closely related to the road friction coefcient l.

Jw 

dx
Mv
6l
dt
4

45

Since it is usually not easy to deal with state constraints in the optimal control problem, we try to translate the constraint on dx/dt
into the restriction on vm. This can be done by examining the transfer function of dx/dt with respect to vm,

337

J. Xi et al. / Applied Energy 134 (2014) 332341

Hs

x_ s

46

v m s

If we further assume the control variable vm to be of some particular

forms, e.g., piecewise constant functions, the inverse Laplace Transform of H(s)  vm(s) can then be written out explicitly and the limit
on the magnitude of vm(t) can be calculated accordingly,

L1 H  v m  6

lMv

47

4J w

After obtaining the limit umax(l), the feasible control set U will then
be the functional space of all the piecewise constant functions with
additional norm constraint: max |u(t)| 6 umax(l) for all t e [t0, tf].
It is worthwhile to note from the derivation in the unconstrained case that as long as vm* stays within the control set U,
we can always attain the global optimum. Therefore the problem
we need to deal with here is the case when vm* falls out of the
set U, i.e., when vm* > umax(l). Motivated by the fact vm* is a global
minimum of the convex function g, we expect the optimum to
occur at the value of umax(l) in the constrained case. In order to
verify this case, we follow the same line of reasoning in the unconstrained case. Again, we need to nd a stationery solution in the
sense that the dynamics of the state variables along with the costate variables tend to become zero. A slight modication for the
current case is that, in addition to the requirements on im and x,
we also require that the time derivatives of p1 and p2 should be
zero too. After solving the set of equations explicitly, we then need
to check the inequality condition of the PMP. It turns out that the
inequality always holds for any vm* > umax(l). This also conrms
with the conclusion made previously in the unconstrained case.
To sum up, we are ended with the following result for the optimal
control problem.


8
q
< Rm sv 1 1 2k2b

v m :

kb

Rm C v

48

umax l

will be maintained at the optimal value. In this regard, we will rst

analyze the entire power system in terms of the power circuit,
including the battery, the booster, and the motor connected in cascade as shown in Fig. 5.
Based on the analysis result of the battery, the battery model
has been simplied as an equivalent voltage source in series connection with a resistance as shown in Fig. 5. ib represents the outow current from the battery. The motor model, as discussed
before, is represented as an inductance in series connection with
a resistance together with a back EMF. im and vm represent the
input current and voltage to the motor, respectively. The booster
model is composed of two switches and a low pass lter. Here
the control variable in the converter is q(t), the on/off state of the
switch. The choice of L and C should meet the specications given
by:

CP

ib 1  d  T
;
Dv b

v m 1  d  T
Dim

Recalling the expression for Jv, Cv, and sv, it can be noted that the
optimal control variable in Eq. (48) is a function of road friction,
rolling resistance and air resistance parameters. All these parameters can be categorized as external factors depending solely on the
environment. When a vehicle is cruising on the road, these parameters may vary with time due to the changing environment, the
control variable is therefore supposed to adjust to these external
variations so that the vehicle is always operating optimally in the
current situation. On the other hand, the voltage input to the motor
is controlled by the DCDC converter. So the adaptive optimal voltage level serves as a reference signal fed to the DCDC converter
for the implementation in real time. The overall control hierarchy
is shown in the Fig. 4.
The discussion below will focus on the control of the DCDC
converter. The primary goal is to determine the appropriate
switching frequency so that the output voltage of the converter

tT

qs  ds

50

Dvb and Dim are the ripple components of the voltage and current at
the steady state, and they are specied based on design requirements. The state space representation of the circuit model is shown
below, in which ib, vm, im and xm are the corresponding system
states.

dib 1
v b  Rb ib  q v b  v m  1  q
L
dt
dv m 1
im  q ib  im  1  q
C
dt
dim
1

v m  Rm  im  kb  x
Lm
dt

51
52
53
54

Intrinsically this is a nonlinear time-varying system with the control in the switched state. In dealing with the system which involves
switching, a common approach is to perform the state-space averaging [22]. Instead of focusing on the full dynamics of the waveform,
we pay special attention to the low frequency averaged component,
which is of greater importance from the control perspective. The
averaging operator is dened as

Xt

1
T

xs  ds

55

tT

When acting on both sides of Eqs. (51) and (52), we will have

dib 1
v b  Rbib  d  v m  1  d
L
dt
dv m 1 
im ib  1  d
C
dt

Optimal Voltage Level

Optimizer

49

In these inequalities, T is the switching period and d is the duty

ratio, which is related to q(t) in the following way.

dx 1
2kb  im  C v  x  sv 
Jv
dt

3.3. Proposed control strategy

Rolling Resistance Crr

Air Resistance Cd

LP

PWM

DC/DC
Controller

Road Friction
Fig. 4. The overall control hierarchy.

DC/DC

Motor

56
57

338

J. Xi et al. / Applied Energy 134 (2014) 332341

ib

im

Lm

Rb

Rm
q(t)

vm
Em

vb

Fig. 5. Power circuit for electric vehicle.

verr
vreff

d
Compensator

Booster

vm

tion in a city. All the road information as well as the trafc situation is incorporated into the speed of the vehicle moving ahead.
In contrast to the previous scenario, for the most cases in the second scenario, there tends to be a continuous change in the speed.
So it sets a higher demand on the controllers adaptability. Apart
from those external variations, in both scenarios some random
noises have been purposely introduced into the model parameters,
which in some sense accounts for a discrepancy between the
model and the actual system. We would like to test whether the
controller can work robustly in the uncertain environment, i.e., in
the presence of the process noise. In the following part, quantitative descriptions of the road conditions will be given for each scenario together with the initial conditions of the EV system. Then
the simulation results will be shown followed by a discussion
focusing on the vehicle behavior and controller performance. As
an illustration of the controllers optimality, vehicle behavior at a
non-optimal operating point will also be shown and compared
with that of the optimal solution. It can be noted that any variation
in the optimal vehicle speed will ultimately contribute to a total
increase in the energy dissipation. The vehicle parameters used
in both of the simulations are listed in Table 1.
4.1. Scenario I

Fig. 6. DCDC level control.

It can be noted that in Eqs. (56) and (57), all the state variables have
been replaced with their averaged ones over the period T, and q(t)
automatically becomes d. As long as the switching period T is much
smaller than the time constant at the system input side given by L/
Rb and time constant at the system output side given by RmC, the
simplied model will sufce to capture the averaged behavior of
the system. Since it is a regulatory problem, we further linearize
the system about its operating point Xss, which yields

d~ib 1
~
Rb  d  ~ib  1 d  v~ m v b  Rbib  d
L
dt

58

dv~ m 1
~
1  d  ~ib  ~im  ib  d
dt
C

59

d~im
1
~

v~ m  Rm  ~im  kb  x
Lm
dt

60

~
dx
1
~
2kb  ~im  C v  x
Jv
dt

61

After applying the Laplace Transform and obtaining the transfer

function for the duty ratio d and output voltage vm, we can then
analyze its pole location and design the feedback PI compensator.
The block diagram of the closed-loop control system is shown in
Fig. 6. Detailed numerical simulations and corresponding results
will be given in the next section.

The road conditions for the rst scenario and the initial settings
of the EV system are specied in Table 2, and the corresponding
results are shown in Fig. 7.
In this scenario, the vehicle is supposed to be initially cruising at
the optimal speed of 105 rad/s on the country road when a sudden
increase in the drag coefcient occurs at t = 150 s. With this variation in the drag coefcient, the optimal operating point shifts
accordingly. The energy efciency diagrams for the change of environmental conditions are shown in Fig. 7(I), where the efciency is
dened as x/(vm  im), a measure of the energy consumption per
unit angle. It can be noted that as the drag coefcient Cd increases
from 0.20 to 0.28, the optimal voltage level shifts from 66.35 V to
62.45 V and at same time the overall energy efciency drops from
0.0197 to 0.0160 (from the black curve to the grey curve). This ultimate decline indicates that more energy has to be used for overcoming the air resistance.

Table 1
Vehicle parameters used in the simulations.
Battery pack

No-load voltage vb
Internal resistance Rb

48 V
0.15 ohm

Motor

Resistance of the rotor loop Rm

Inductance of the rotor loop Lm
Conversion factor kb

0.6 ohm
12 mH
0.25 V/rad

Vehicle dynamics

Vehicle mass Mv
Wheel radius rw
Cross-sectional area AF

250 kg
0.3 m
2 m2

4. Simulation and case studies

In this section, the proposed controller performance will be
tested with respect to its adaptability and robustness under two
different scenarios using simulations. The rst scenario is driving
on a country road with variations in road conditions including
the road friction and air resistance at some particular locations
along the way, assuming no vehicle ahead of the test EV. This scenario is designed to test the controllers adaptability to discrete
abrupt changes from the external environment. The second scenario is driving behind another vehicle which has a varying speed.
This scenario is specically designed to simulate the driving situa-

Table 2
Road conditions and initial settings of the EV, Scenario I.
Road conditions

Rolling resistance coefcient Crr

Air density qa
Drag coefcient Cd

0.01
1.2 kg/m3
0.200 t<150s
0.278 t>150s

Initial settings

Battery current Ib
Input voltage to the rotor motor vm
Input current to the rotor motor Im
Angular velocity of the wheel x

132.6 A
66.4 V
80.2 A
105 rad/s

339

J. Xi et al. / Applied Energy 134 (2014) 332341

Fig. 7. (I) The efciency map under two conditions, (II) the response of the motor voltage, (III) the angular velocity of the wheel, and (IV) the corresponding instantaneous
energy efciency.

In order to adapt to this change in the environment, the input

voltage to the motor has to be altered accordingly. Under the proposed control strategy, it can be seen in Fig. 7(II) that the motor
voltage (black line) experiences a smooth transition to the new
optimal level within a few seconds. In accordance with this variation, the angular velocity of the wheel in Fig. 7(III) also transitions
to a new value, even though it takes a relatively longer time. This is
mainly because of a longer response time of the mechanical system. As the voltage and angular velocity transition to their corresponding new values, the energy efciency rst increases and
then declines rapidly and converges to the new level (as shown
in Fig. 7(IV)). The initial increase is mainly attributed to the difference in response time between the mechanical and electrical system, as mentioned before.
The motor voltage in the comparison group (grey line) is maintained at 50 V all the time and the corresponding angular velocity
prole is shown in Fig. 7(III). As the drag coefcient increases, the
angular velocity drops accordingly. It can be noted that the energy
efciency in the comparison group always lies below that of the
controlled one, which conrms the observation in Fig. 7(I). In both
conditions, the voltage level 50 V is well below the optimal one.

In this scenario there is no change in the road conditions but

there is a vehicle ahead with varying speeds as indicated in
Fig. 8(I) by the black dashed line. As long as its speed is greater
or equal to the optimal cruising speed (105 rad/s in this case),
the controlled system will just stay on the optimal one. However
when the speed of the vehicle ahead falls below, the optimal operating point cannot be attained anymore and we have to follow that
speed. It can be noted in Fig. 8(I) that the tracking performance of
the angular velocity (black solid line) is still quite good, except for a
minor undershoot around the corner point at t = 650 s. After a short
period of the transient stage, the angular velocity nally converges
to 85 rad/s, the same value we chose for the angular velocity in the
comparison group. The motor voltage which produces this velocity
prole is shown in Fig. 8(II). The corresponding energy efciency
associated with this process is shown in Fig. 8(III). As the angular
velocity deviates from the optimal value, the energy efciency
nally converges to a value below the optimal one.

Table 3
Road conditions and initial settings of the EV, Scenario II.
Road conditions

Rolling resistance coefcient Crr

Air density qa
Drag coefcient Cd

0.01
1.2 kg/m3
0.200

Initial settings

Battery current Ib
Input voltage to the rotor motor vm
Input current to the rotor motor Im
Angular velocity of the wheel x

132.6 A
66.4 V
80.2 A
105 rad/s

4.2. Scenario II
The road conditions for the second scenario and the initial settings of the EV system are specied in Table 3, and the corresponding results are shown in Fig. 8.

340

J. Xi et al. / Applied Energy 134 (2014) 332341

Fig. 8. (I) The angular velocity of the wheel, (II) the response of the motor voltage; and (III) the corresponding instantaneous energy efciency.

5. Conclusion
In this paper the energy management for battery powered electric vehicles has been approached from a comprehensive and systematic perspective. By examining the battery and motor dynamics
respectively, we have acquired an analytic expression that relates
the optimal operating point of components to the vehicle and environmental parameters. Based on this analysis, we have proposed a
controller hierarchy suitable for real-time implementation. From
the simulation results it can be seen that the controller can adapt
to different road conditions and work robustly under uncertainty.
To achieve the performance of the proposed integrated approach,
the dynamics and couplings of the battery and motor have been
focused in this work with other minor effects being neglected for
the time being, so that the focus of this work will not be diluted.
Those minor effects will be considered in our future work, for
example, as uncertain factors or additional estimated terms in
the models. Another natural extension of the current work would
be to integrate the ultra-capacitor into the power source system.
References
[1] Capasso C, Veneri O. Experimental analysis on the performance of lithium
based batteries for road full electric and hybrid vehicles. Appl Energy 2014.
[2] Xiong R, Sun F, Gong X, Gao C. A data-driven based adaptive state of charge
estimator of lithium-ion polymer battery used in electric vehicles. Appl Energy
2014;113:142133.

[3] Hu C, Youn BD, Chung J. A multiscale framework with extended kalman lter
for lithium-ion battery soc and capacity estimation. Appl Energy
2012;92:694704.
[4] Zhong L, Zhang C, He Y, Chen Z. A method for the estimation of the battery pack
state of charge based on in-pack cells uniformity analysis. Appl Energy
2014;113:55864.
[5] Pahlevaninezhad M, Das P, Drobnik J, Moschopoulos G, Jain PK, Bakhshai A. A
nonlinear optimal control approach based on the control-lyapunov function
for an ac/dc converter used in electric vehicles. IEEE Trans Ind Inform
2012;8(3):596614.
[6] Faiz J, Sharian MBB, Keyhani A, Proca AB. Sensorless direct torque control of
induction motors used in electric vehicle. IEEE Trans Energy Convers
2003;18(1):110.
[7] Faggioli E, Rena P, Danel V, Andrieu X, Mallant R, Kahlen H. Supercapacitors for
the energy management of electric vehicles. J Power Sources
1999;84(2):2619.
[8] Dougal RA, Liu S, White RE. Power and life extension of battery-ultracapacitor
hybrids. IEEE Trans Compon Pack Technol 2002;25(1):12031.
[9] Lu S, Corzine KA, Ferdowsi M. A new battery/ultracapacitor energy storage
system design and its motor drive integration for hybrid electric vehicles. IEEE
Trans Veh Technol 2007;56(4):151623.
[10] Kuperman A, Aharon I, Malki S, Kara A. Design of a semiactive batteryultracapacitor hybrid energy source. IEEE Trans Power Electron
2013;28(2):80615.
[11] Garcia F, Ferreira A, Pomilio J. Control strategy for battery-ultracapacitor
hybrid energy storage system. In: Proc applied power electronics conference
and exposition, twenty-fourth annual IEEE. IEEE; 2009. p. 82632.
[12] Lukic SM, Wirasingha SG, Rodriguez F, Cao J, Emadi A. Power management of
an ultracapacitor/battery hybrid energy storage system in an hev. In: Proc
vehicle power and propulsion conference, IEEE. IEEE; 2006. p. 16.
[13] Laldin O, Moshirvaziri M, Trescases O. Predictive algorithm for optimizing
power ow in hybrid ultracapacitor/battery storage systems for light electric
vehicles. IEEE Trans Power Electron 2013;28(8):388295.

J. Xi et al. / Applied Energy 134 (2014) 332341

[14] Wang L, Collins Jr EG, Li H. Optimal design and real-time control for energy
management in electric vehicles. IEEE Trans Veh Technol 2011;60(4):141929.
[15] Wang G, Yang P, Zhang J. Fuzzy optimal control and simulation of batteryultracapacitor dual-energy source storage system for pure electric vehicle. In:
Proc intelligent control and information processing international conference
on, IEEE; 2010. p. 55560.
[16] Hannan M, Azidin F, Mohamed A. Multi-sources model and control algorithm
of an energy management system for light electric vehicles. Energy Convers
Manage 2012;62:12330.
[17] Li X, Chen Y, Wang J. In-wheel motor electric ground vehicle energy
management strategy for maximizing the travel distance. In: Proc american
control conference, IEEE; 2012. p. 49938.

341

[18] Tremblay O, Dessaint L-A, Dekkiche A-I. A generic battery model for the
dynamic simulation of hybrid electric vehicles. In: Proc vehicle power and
propulsion conference, IEEE. IEEE; 2007. p. 2849.
[19] Liberzon D. Calculus of variations and optimal control theory: a concise
introduction. Princeton University Press; 2011.
[20] Chiasson J. Modeling and high performance control of electric
machines. Wiley-IEEE Press; 2005.
[21] Gillespie TD. Fundamentals of vehicle dynamics. Society of Automotive
Engineers; 1992.
[22] Kassakian JG, Schlecht MF, Verghese GC. Principles of power
electronics. Prentice Hall; 1991.