Computers and Chemical Engineering 64 (2014) 114123

Contents lists available at ScienceDirect

Computers and Chemical Engineering
j our nal home page: www. el sevi er . com/ l ocat e/ compchemeng
Adaptive gain sliding mode observer for state of charge estimation
based on combined battery equivalent circuit model
Xiaopeng Chen, Weixiang Shen

, Zhenwei Cao, Ajay Kapoor

Faculty of Science, Engineering and Technology, Swinburne University of Technology, Hawthorn, Victoria 3122, Australia
a r t i c l e i n f o
Article history:
Received 3 February 2013
Received in revised form8 February 2014
Accepted 16 February 2014
Available online 22 February 2014
Keywords:
Adaptive gain sliding mode observer
Battery management system
Combined battery equivalent circuit model
Electric vehicle
Lithium-polymer battery
State of charge
a b s t r a c t
An adaptive gain sliding mode observer (AGSMO) for battery state of charge (SOC) estimation based on a
combined battery equivalent circuit model (CBECM) is presented. The error convergence of the AGSMO
for the SOC estimation is proved by Lyapunov stability theory. Comparing with conventional sliding
mode observers for the SOC estimation, the AGSMO can minimise chattering levels and improve the
accuracy by adaptively adjusting switching gains to compensate modelling errors. To design the AGSMO
for the SOC estimation, the state equations of the CBECM are derived to capture dynamics of a battery. A
lithium-polymer battery (LiPB) is used to conduct experiments for extracting parameters of the CBECM
and verifying the effectiveness of the proposed AGSMO for the SOC estimation.
2014 Elsevier Ltd. All rights reserved.
1. Introduction
In recent decades, the progressive increase of petrol costs and
air pollution of the exhaust fumes frompetrol-driven vehicles has
stimulated a surge of research and innovation in electric vehi-
cle (EV) technologies. Lithium-ion or lithium-polymer batteries
(LiPBs) have been adopted as primary power sources in EVs due
to their merits in high power and energy densities, high operating
voltages, extremely low self-discharge rate and long cycle life in
the comparison with other types of batteries such as lead-acid or
nickel-metal hydride batteries. For the application of the batter-
ies in EVs, the state of charge (SOC) is one of the key parameters
which corresponds to the amount of residual available capacity, its
accurate indication is crucial for optimising battery energy utilisa-
tion, informing drivers the reliable EV travelling range, preventing
batteries from over-charging or over-discharging and extend-
ing battery life cycles. Unfortunately, the SOC cannot be directly
measured by a sensor as it involves in complex electrochemical
processes of a battery. An advanced algorithm is required to esti-
mate the SOC with the aids of measurable parameters of a battery
such as terminal voltage and current.
A variety of the SOC estimation techniques has been reviewed
by Piller, Perrin, and Jossen (2001) and each method has its own

Corresponding author. Tel.: +61 3 9214 5886; fax: +61 3 9214 8264.
E-mail addresses: xchen@swin.edu.au (X. Chen), wshen@swin.edu.au (W. Shen).
advantages incertainaspects. Theampere-hour (Ah) countingis the
most applicable approach for the SOC indication in many commer-
cial battery management systems (BMSs). It simply integrates the
battery charge and discharge currents over time and accumulates
errors caused by the embedded noises in current measurements.
Furthermore, this non-model and open-loop based method has
difculty in determining the initial SOC value. An improved ver-
sion of the Ah counting has exhibited better SOC estimation results
by on-line evaluating charge and discharge efciencies with the
recalibration of the cell capacity (Ng, Moo, Chen, & Hsieh, 2009).
Battery impedance measurement technique is also used for the
SOC estimation through injecting small ac signals with a wide
range of frequencies into a battery to detect the variation of battery
internal impedances (Rodrigues, Munichandraiah, &Shukla, 2000).
However, the measured impedances cannot completely model the
dynamics of batteries in the case of large discharge current in EVs.
Furthermore, the application of impedance spectroscopy has to be
carried out in temperature-controlled environment that requires
bulky and costly auxiliary equipment since the temperature signif-
icantly affects impedance curves.
Another category of the SOC estimation methods is based on
black-box established on machine learning strategies, which
includes articial neural networks (ANNs) (Shen, 2007; Shen, Chan,
Lo, &Chau, 2002), fuzzy neural networks (Li, Wang, Su, &Lee, 2007),
adaptive fuzzy neural networks (Chau, Wu, Chan, & Shen, 2003)
and support vector machine (Hansen & Wang, 2005). These data-
oriented approaches can accurately estimate the SOC without its
http://dx.doi.org/10.1016/j.compchemeng.2014.02.015
0098-1354/ 2014 Elsevier Ltd. All rights reserved.
X. Chen et al. / Computers and Chemical Engineering 64 (2014) 114123 115
Nomenclatures
C
n
nominal capacity of LiPB (Ah)
C
p
polarisation capacitance (F)
e
Vt
, e
Z
, e
Voc
, e
Vp
estimation errors
^f
1
, ^f
2
, ^f
3
systemuncertainty terms
R
i
ohmic resistance (!)
R
p
polarisation resistance (!)
Z state of charge

Z estimated state of charge

V
oc
open circuit voltage (V)
V
oc
(Z) open circuit voltage as a function of state of charge
V
p
polarisation voltage (V)

v
p
estimated polarisation voltage (V)

v
t
estimated battery terminal voltage (V)
V
t
battery terminal voltage (V)
q coulomb efciency

i
uncertainty bounds

1
1
,

1
2
,

1
3
adaptive switching gains

1
1
,

1
2
,

1
3
adaptive switching gains updating laws

1
,
2
,
3
adaptation speed adjusting values
accurate initial state, but they require a large amount of data to
train ANNs, which leads to the large computation burden in the
BMS. Moreover, the SOC estimation results would be unpredictable
in the presence of the conditions where the current proles in EVs
are different fromthose represented by the training data.
The Kalman lter (KF), as an optimal recursive estimator which
is able to estimate the states of a linear dynamic system (Ristic,
Arulampalam, &Gordon, 2004), has beendevelopedtoestimate the
SOC based on linear state space battery models (Barbarisi, Vasca, &
Glielmo, 2006). For nonlinear battery models, the enhanced ver-
sions of KF have been intensively investigated to achieve better
results for on-line SOC estimation, such as extended KF (EKF) (Dai,
Wei, Sun, Wang, & Gu, 2012; Hu, Youn, & Chung, 2012; Hu, Li, &
Peng, 2012), adaptive extended KF (AEKF) (Han, Kim, & Sunwoo,
2009), sigma-point KF (SKF) (Plett, 2006a,b) and unscented KF
(UKF) (He, Williard, Chen, & Pecht, 2013; Zhang & Xia, 2011). The
EKF utilises the rst-order Taylor series expansion to linearise
the nonlinear function. This local linearisation can give rise to
large estimation errors when the degrees of nonlinearity in battery
models are signicant and the covariance of process and measure-
ment noises is assumed to be constant. Adaptively updating the
covariance of process and measurement noises, the AEKF has been
developed to improve the online SOC estimation accuracy. Instead
of local linearisation, the SKF and the UKF use an unscented trans-
formation to approximate the probability density function of the
nonlinear systems with a set of sample points or so-called sigma
points. Essentially, all above-mentioned KF-based approaches are
based on the assumption that the covariance of measurement and
process noises described by a Gaussian probability density func-
tion has to be known a priori. Moreover, their complex matrix
operations may result in numeric instabilities.
The H

observer based approach has also been proposed to

estimate the SOC without the requirement of the exact statistical
properties of the battery model (Zhang, Liu, Fang, & Wang, 2012).
This approach minimises the errors between the outputs of the
battery and its model so that the SOC estimation error is less than
a given attenuation level. However, in order to tackle modelling
errors and external disturbances, the feedback gain of H

observer
must be obtained by solving a linear matrix inequality, which may
not provide the optimal solution for ensuring tracking error con-
vergence.
More recently, sliding mode observer (SMO) based SOC esti-
mation methods were adopted to overcome battery model
uncertainties, external disturbances and measurement noises with
sufcient large switching gains (Kim, 2006; Chen, Shen, Cao, &
Kapoor, 2012). This method relies on the exhaustive understanding
of battery dynamics for the appropriate selection of the switching
gains, which lead to the trade-off between the magnitude of chat-
tering in the SOC estimation and the convergence speed to reach
the sliding mode surface and trigger the sliding motion.
In this paper, an adaptive gain slide mode observer (AGSMO)
based on a combined battery equivalent circuit model (CBECM)
has been proposed for the SOC estimation. The main advantage
of the AGSMO is that the robust behaviour of the SOC estima-
tion is guaranteed in the presence of the modelling errors, which
are considered as the bounded uncertainties. This is achieved by
dynamically adjusting the switching gain of the SMO in response
to the tracking error while ensuring the reachability of the slid-
ing mode surface and triggering the sliding mode. Once the sliding
mode is activated, the switching gain is self-tuned to an adequate
level to counteract the modelling errors and reduce the chattering
levels, thereby improving the SOC estimation accuracy.
This rest of this paper is organised as follows. In Section 2, a
CBECM is presented to model the battery dynamic behaviour. In
Section 3, the AGSMO design methodology for estimating the SOC
is explained. Section 4 elaborates the procedures to extract battery
model parameters. Section 5 validates the proposed AGSMOfor the
SOC estimation by experimental results and Section 6 concludes.
2. Battery modelling
A suitable battery model is essential to the development of
the model-based BMS in real EVs, which requires less computa-
tion power and fast response to ever-changing road conditions.
Many types of models are developed to capture lithium-ion bat-
tey dynamics for various purposes (Ramadesigan et al., 2012). In
general, they can be categorised into two main groups, which
are electrochemical and equivalent circuit models (He, Xiong, &
Fan, 2011; Hussein & Batarseh, 2011; Hu, Youn et al., 2012; Hu,
Li et al., 2012). The electrochemical models describe the physical
phenomena which occur inside batteries such as the material and
charge transfer processes, ionic conduction, solid phase diffusion.
They utilise partial differential equations with a large number of
unknownparameters andthus a large amount of memory required,
which leads to long computation time and slowresponse. They are
usually used for battery design and simulation and hardly suitable
for the BMS design in real EVs (Smith, Rahn, & Wang, 2010).
On the other hand, the equivalent circuit models simply consist
of resistors, capacitors and voltage sources to form a circuit net-
work, which leads to short computation time and quick response.
Furthermore, they are the circuit in nature which is easily inte-
grated into the BMS and power control in real EVs. Various battery
equivalent circuit models have been proposed to reect dynamic
characteristics of the battery as a result of the trade-off between
modellingaccuracyandcomplexity(Lee, Kim, Lee, &Cho, 2008; Cho
et al., 2012; Chen, Gabriel, & Mora, 2006; Abu-Sharkh & Doerffel,
2004).
In this paper, the combined battery equivalent circuit model
(CBECM) is used to represent the dynamical behaviors of LiPB,
as shown in Fig. 1. A capacitor, C
n
represents the nominal capac-
ity of the battery on the left in the model. The current source, I
denotes the discharge or charge current of the LiPB and the corre-
sponding battery terminal voltage is expressed by V
t
. The voltage
across the C
n
as the open circuit voltage (OCV), V
oc
varies in the
range of the SOC, Z from 0% to 100% and it represents the SOC of
the battery quantitatively. A resistor, R
i
and a parallel-connected
116 X. Chen et al. / Computers and Chemical Engineering 64 (2014) 114123
Fig. 1. Schematic of combined battery equivalent circuit model.
network of the polarisation resistance, R
p
and polarisation capaci-
tance, C
p
on the right are used to characterise the ohmic resistance
and simulate the transient responses or the battery relaxation
effect, respectively. The relaxation effect is dened as the slow
convergence of the battery terminal voltage to the OCV at its equi-
librium state after hours of relaxation at the end of charging or
dischargingprocess. It is causedbythediffusioneffect andadouble-
layer chargingor dischargingeffect inthebattery(Chenet al., 2006).
The voltage-controlled voltage source, V
oc
(Z) is used to bridge the
nonlinear relationship between the SOC and the OCV as shown
in Fig. 2, which can be derived by tting the experimental data
obtained fromthe pulse current discharge (PCD) and pulse current
charge (PCC) tests (for details see Section 4). The self-discharge
resistance is ignored in this model as a LiPB has extremely low
self-discharge rate.
The SOC describes the ratio of the remaining capacity to the
present maximum available capacity of a battery, and it can be
expressed as:
Z(t) = Z(0)
_
t
0
ql(z)
C
n
dz (1)
where Z(0) is the initial SOC of the battery, I(z) is the instantaneous
current and it is assumed to be positive for discharge current and
negative for charge current. The q denotes the coulomb efciency
and it can be normally taken one for discharging and less than or
close to one for charging LiPB in the wide range of the current and
the temperature.
Fig. 2. Experimental OCVSOC curves of LiPB.
According to Fig. 1, V
t
can be written as follows
v
t
= v
cc
(Z) v
p
lR
i
(2)
The time derivatives of polarisation voltage and the SOC yield

v
p
=
v
p
R
p
C
p
+
l
C
p
(3)

Z =
l
C
n
(4)
where V
p
is the polarisation voltages across the C
p
.
Despite the nonlinearity of the OCVSOC curves as shown in
Fig. 2, there exists a piecewise linear relationship between the OCV
and the SOC in a certain range of the SOC indicated by the dots in
the curves. Therefore, the OCV is expressed as a function of the SOC
by using piecewise linearisation method
v
cc
(Z) = kZ +d (5)
where the values of k and d are the constants in a certain range of
SOC. In this paper, there are one pair of k and d in every 10% SOC
and totally ten pairs for 0% to 100%. Thus, the time derivative of V
oc
in each 10% SOC segment is

v
cc
(Z) = k

Z (6)
Subsitituting Eq. (4) into Eq. (6) gives

v
cc
(Z) = k
_
l
C
n
_
(7)
The rate of change of the current during charging or discharging,
I canbe negligible due to the the fast sampling interval as explained
as follows (Chenet al., 2012; Chiang, Sean, &Ke, 2011). For instance,
5Ah LiPB has been conducted discharge at the current of 1C
n
(5A), the variation of SOC in 1s sampling period with respect to
time as given in Eq. (4) is dZ/dt =5/(53600) =0.00028, namely
dZ/dt 0. It shows that the current within 1s has an insignicant
impact of theSOC, thus thecurrent is assumedtobeconstant ineach
sampling period (e.g., one second in this paper), namely dI/dt 0.
Therefore, the time derivative of V
t
in Eq. (2) with the substitutions
of Eqs. (3)(7) gives

v
t
= k
_
l
C
n
_
+
v
p
R
p
C
p

l
C
p
(8)
Solving I in Eq. (2) and substituting it into Eq. (4) as well as
rearranging Eqs. (3)(5) result in the state-space equations of the
CBECMas

v
t
= u
1
v
t
+u
1
v
cc
(Z) b
1
l

Z = u
2
v
t
u
2
v
cc
(Z) +u
2
v
p

v
p
= u
1
v
p
+b
2
l
(9)
where a
1
=1/(R
p
C
p
), a
2
=1/(R
i
C
n
), b
1
=k/C
n
+R
i
/(R
p
C
p
) +1/C
p
and
b
2
=1/C
p
.
There are two main causes for modelling errors by using the
CBECMto represent battery behaviours. Firstly, the circuit parame-
ters of the CBECMequations are taken as the constant values, but in
fact they are varying with the battery SOC. Secondly, the OCVSOC
curve is piecewise linearised. These modelling errors represented
by the uncertainty terms, ^f are added to Eq. (9) as

v
t
= u
1
v
t
+u
1
v
cc
(Z) b
1
l +zj
1

Z = u
2
v
t
u
2
v
cc
(Z) +u
2
v
p
+zj
2

v
p
= u
1
v
p
+b
2
l +zj
3
(10)
where ^f
1
, ^f
2
and ^f
3
satisfy the following bounded conditions:

^j
i

-
i
where i =1, 2 and 3. Since the proposed SOC estimation
approach is applied to EV applications, EV driving schedules are
X. Chen et al. / Computers and Chemical Engineering 64 (2014) 114123 117
Fig. 3. SMO based SOC estimation with conservative switching gains.
studied to determine average discharge rates of current proles for
battery testing, where the average EV driving speed is equivalently
converted to average current rates of 1/3C
n
1C
n
. In this paper, the
average current rate of discharge current proles up to 1.5C
n
is
adopted for the battery tests to ensure that the proposed discharge
current proles have included the maximumpossible currents for
discharging or charging. As a result, both charge and discharge cur-
rent rates as the inputs to the model and the battery in EVs are
bounded and so are the terminal voltages as the outputs. Therefore,
the uncertainty bound,
i
can be determined by using the largest
modellingerrors betweentheLiPBandtheCBECMunder thelargest
average discharge rate of 1.5C
n
.
3. Design of adaptive gain sliding mode observer for SOC
estimation
Conventional SMOs with the constant switching gains for the
SOC estimation have demonstrated the robustness to compen-
sate modelling errors and uncertainties with the properly selected
switching gains (Kim, 2006; Chenet al., 2012). However, the under-
estimated or overestimated switching gain has given rise to the
poor tracking performance or undesired chattering phenomena in
the SOC estimation. The SMO with a lower switching gain has no
unwanted chattering in the SOC estimation as shown in Fig. 3, but
its tracking performance under the randomly selected initial SOC
is very poor with the mean square errors (MSEs) in SOC estimation
higher than 10%. On the other hand, the SMO with a large switching
gain has a considerable chattering around the true SOC as shown in
Fig. 4 and its tracking performance is robust with most of the MSEs
Fig. 4. SMO based SOC estimation with large switching gains.
bounded within 5%, but the higher magnitude of chattering ripples
blurs the SOC estimation and affects the stability of observer.
In order to accurately estimate the SOC, an AGSMO based on an
equivalent control concept is proposed as follows

v
t
= u
1

v
t
+u
1
v
cc
(

Scc) b
1
l +

1
1
sgn(c
vt
)

Z = u
2

v
t
u
2
v
cc
(

Z) +u
2

v
p
+

1
2
sgn(c
vcc
)

v
p
= u
1

v
p
+b
2
l +

1
3
sgn(c
vp
)
(11)
where

v
t
,

Z and

v
p
are the estimated V
t
, Z and V
p
, respectively, the

1
1
,

1
2
and

1
3
are the adaptive switching gains which are adapted
according to the following updating laws:

1
1
=
1
|c
vt
| ,

1
2
=
2
|c
vcc
| and

1
3
=
3

c
vp

(12)
where the terms
1
,
2
and
3
are positive constants that should
be chosen suitably small so that they can ensure the adaptation
speed of the switching gains for state errrors convergence while
preventing the corresponding

1
i
from becoming too large and
guaranteeing suitable bounded magnitude of the switching gains.
Accordingly, the battery terminal voltage and state estimation
errors are dened as
c
vt
= v
t


v
t
c
vcc
= v
cc
(Z) v
cc
(

Z) = k(Z

Z) = kc
Z
c
vp
= v
p

v
p
(13)
Thus, by substracting Eq. (11) fromEq. (10), the error dynamics
of battery terminal voltage and other states are expressed as
c
vt
= u
1
c
vt
+u
1
c
vcc
+zj
1

1
1
sgn(c
vt
)
c
Z
= u
2
c
vt
u
2
kc
Z
+u
2
c
vp
+zj
2

1
2
sgn(c
Z
)
c
vp
= u
1
c
vp
+zj
3

1
3
sgn(c
vp
)
(14)
where sgn() is the signumfunction
sgn(c
vt
) =
_
+1, c
vt
:0
1, c
vt
-0
It can be seen fromEq. (14) that if the switching gain

1
1
is prop-
erly adjusted so that a sliding mode motion can be induced on the
terminal voltageerror stateinEq. (14). Theasymptotic convergence
of the terminal voltage error can be proved by Lyapunov stability
theory via choosing the candidate of Lyapunov function as follows
v
1
=
1
2
(c
2
vt
+
1
1

1
2
1
) (15)
where

1
1
=

1
1
1
1
and since the SOC operation range of the LiPB
is varying from 0 to 1, as can been seen in Fig. 2, the SOC esti-
mation error is bounded as |e
z
| <1 and the V
oc
is also bounded as
|c
vcc
| =

v
cc
(Z) v
cc
(

Z)

k(Z

Z)

-|k| |c
Z
| -|k| , and the time
derivative of the candidate of Lyapunov function V
1
results in

v
1
= c
vt
c
vt
+
1
1

1
1

1
1
= c
vt
[u
1
c
vt
+u
1
c
vcc
+^j
1

1
1
sgn(c
vt
)]
+
1
1
(

1
1
1
1
)

1
1
= [u
1
c
2
vt
+u
1
c
vt
c
vcc
+c
vt
^j
1

1
1
|c
vt
|]
+(

1
1
1
1
) |c
vt
| = u
1
c
2
vt
+u
1
c
vt
c
vcc
+c
vt
^j
1
1
1
|c
vt
| -u
1
|c
vt
| |c
vcc
| +|c
vt
|

^j
1

1
1
|c
vt
| = |c
vt
| (u
1
|c
vcc
| +

^j
1

1
1
) (16)
There exists an unknown nite non-negative switching gain
1
1
such that 1
1
:u
1
|c
vcc
| +
1
, leading to

v
1
-0, which satises
118 X. Chen et al. / Computers and Chemical Engineering 64 (2014) 114123
Fig. 5. Testing platformof LiPB.
the second method of Lyapunov stability theory. Thus, the termi-
nal voltage error as the sliding variable in Eq. (14) asymptotically
converges to zero as time tends to innity. In other words, the
sliding surface is reached during the sliding motion as the slid-
ing variable is equal to zero, where the sliding surface is dened as
S ={e
Vt
=0}. Once the sliding surface is reached, the sliding mode
would be induced to ensure c
vt
(t) = c
vt
(t) = 0 and the inuence
of the bounded uncertainty term ^f
1
is compensated. According
to the equivalent control concept (Edwards & Spurgeon, 1994), the
unmeasurable SOC error can be derived by solving Eq. (14) in terms
of I
1
sgn(e
Vt
) after inserting zeros for c
vt
and e
Vt
c
Z
=
__

1
1
ku
1
_
sgn(c
vt
)
_
cq
(17)
Similarly, the SOC error in Eq. (14) asymptotically converges to
zero as time tends to innity. After c
Z
= 0 and e
Z
=0, the inuence
of the bounded uncertainty term ^f
2
is compensated. Finally, the
V
p
error equation can be derived fromEq. (14) as follows
c
vp
=
_
_

1
2
u
2
_
sgn
_
__

1
1
ku
1
_
sgn(c
vt
)
_
cq
__
cq
(18)
Eqs. (17) and (18) are substituted into Eq. (14), a set of the
AGSMO equations based on the equivalent control concept is
obtained

vt= u
1

vt +u
1
vcc (

Z) b
1
l +

1
1
sgn(cvt )

Z = u
2

vt u
2
vcc (

Z) +u
2

vp +

1
2
sgn
_
__

1
1
ku
1
_
sgn(cvt )
_
cq
_

vp=u
1

vp +b
2
l +

1
3
sgn
__
_

1
2
u
2
_
sgn
_
__

1
1
ku
1
_
sgn(cvt )
_
cq
__
cq
_
(19)
FromEq. (19), it is worth mentioning that the proposed AGSMO
has no requirement for the detailed knowledge of modelling errors
as long as they are bounded whereas the KF based approaches
require the covariance values of the process and measurement
noises. These covariance values are determined either by the
time-consuming trial-and-error method or by the recursive iden-
tication method. Since a priori knowledge of the noise statistical
properties is normally unknown, the former may cause the large
SOCestimationerror withslowconvergenceif inappropriatevalues
of the noise covariance were used and the latter increases compu-
tational complexity.
4. Battery model parameters determination
The component values of the CBECM in Fig. 1 are obtained
fromthe transient-response method via an experiment of the PCD
prole at room temperature. A LiPB is used in the test and it has
a nominal capacity of 5.0Ah and a nominal voltage of 3.8V. The
dimension of the cell is 135mm50mm9mm and the weight
of the cell is 130g. A battery testing platformas illustrated in Fig. 5
is constructed to perform the experiments, and it consists of pro-
grammable power supply (Sorensen DLM50-60), electronic load
(Prodigit 3320) and switches safety box. The testing platform can
control charging/discharging battery, sample experimental data
and store the data into the PC via a graphic user interface program
designed by using the LABVIEWsoftware.
The nonlinear relationship between battery OCV and SOC has
been identied by performing PCD and PCC tests on the LiPB. For
discharge test, the PCD is comprised of a sequence of pulse current
with 6-min discharge and 1-h rest to allow the battery to return
to its equilibrium state before running the next cycle as shown in
Fig. 6. The discharge current of 5.0A is used, which corresponds to
1C
n
rate. For the fully charged LiPB (Z =100%), each pulse discharge
approximates 10%of nominal capacity equivalent to 10%of the SOC
reduction, and the procedure of the pulse discharge and recovery is
repeated until the battery is fully discharged to the cut-off voltage
of 2.7V (Z =0%). For charge test, the PCC is similar to the PCD test, a
fully discharged LiPB has been charged from 0% to 10% SOC at the
Fig. 6. Pulse current discharge and corresponding terminal voltage.
X. Chen et al. / Computers and Chemical Engineering 64 (2014) 114123 119
Table 1
Parameters of CBECMand tting errors extracted fromthe PCD prole.
No. of set zp (s) R
i
(m!) Rp (m!) Cp (kF) MSEs
1 48.52 102.3 5.5 8.822 1.2321%
2 83.94 102.7 7.4 11.343 1.4641%
3 319.2 103.1 8.5 37.553 5.8081%
4 157.5 103.2 5.0 31.50 3.0625%
5 45.09 103.2 4.1 10.998 0.8281%
6 71.08 103.3 5.1 13.937 1.1664%
7 107.9 103.6 4.6 23.457 1.6384%
8 121.8 103.7 4.1 29.71 2.1316%
9 252.3 103.5 6.5 38.554 6.4009%
10 436.3 105.8 20.4 21.387 0.4624%
recommended 0.5C
n
rate, followed by 1-h rest, and this process is
repeated until the battery reaches 100% SOC. Fig. 2 shows the mea-
sured OCVat different SOCs during discharge and charge processes.
It canbe observedthat the OCVof charging process is always higher
thanthat of thedischargingprocess, whichaccounts for ahysteresis
phenomenon between two OCV-SOC curves during the discharge
and charge, respectively. In fact, the hysteresis effect is correlated
with the relaxation effect due to lithium ion diffusion inside the
LiPB and the level of hysteresis is decreasing with the longer rest
period. For the consideration of hysteresis effect, the OCVas a func-
tion of SOC is dened as the average OCV values between charging
and discharging curves as shown in the blue dashed line in Fig. 2.
The circuit model parameters are extracted based on the PCDas
shown in Fig. 6. The corresponding terminal voltage response after
each 10% SOC discharge is also illustrated in Fig. 6. It can be seen
that totally ten sets of transient response in terminal voltage have
been generated to determine circuit parameters corresponding to
each discharge pulse and thus the ten sets of circuit parameters
are identied. Table 1 summarises those parameters and the cor-
responding tting errors represented by MSEs. Since the ninth set
of circuit parameters causes the largest model error, it is used to
determine the parameters in Eq. (9).
The transient response of the terminal voltage at the ninth pulse
current discharge indicated by a red circle in Fig. 6 is used to extract
the circuit parameters, where the circled section is magnied in
Fig. 7. It canbe seenthat whenthe battery stops discharging the ter-
minal voltage has a steep rise as the voltage drop across the ohmic
Fig. 7. Transient response of LiPB and circuit model at the circled PCD.
resistance R
i
disappears immediately, so the ohmic resistance can
be calculated by
R
i
=
zv
t
l
(20)
where ^V
t
is the change of the voltage across R
i
at the instant when
the discharge current I disappears.
During the time interval (t
0
t t
1
), the terminal voltage
increases exponentially as it slowly converges to the OCV, namely
V
oc
(t
1
). This battery terminal voltage is driven by the dynamic char-
acteristics of the battery and can be found by setting discharge
current to zero in Eqs. (2) and (3), then solving the differential
equations gives
v
t
(t) = v
cc
(t
1
) v
p
exp
_
t
z
p
_
(21)
where z
p
is the time constant for the polarisation voltage during
transient response, V
oc
(t
1
) is the OCV after a full relaxation and V
p
is the voltage of the polarisation capacitor and its value equals to
V
oc
(t
0
).
In order to identify model parameters, a curve tting technique
as a nonlinear least square algorithm is applied to search for the
best tting values which lead to the least tting errors between
the measured voltages and the voltages tted by the exponential
function, f(t) =V
1
V
2
exp(t). The coefcients V
1
, V
2
and after
curving tting can be used to determine the CBECM parameters
such as V
oc
(t
1
) =V
1
, V
p
=V
oc
(t
0
) =V
2
and z
p
=1/.
By using the time interval ^t =(t
1
t
0
) and rearranging Eq. (21),
the following equations are derived to calcualte the circuit param-
eters
R
p
=
v
p
(1 exp(zt]z
p
)l
(22)
C
p
=
z
p
R
p
(23)
From Eq. (21), it can be seen that as time t increases or tends
to innity the terminal voltage would be equal to the open cir-
cuit voltage and the battery relaxation effect disappears. With
the parameters calculated by Eqs. (21)(23), the voltage transient
response represented by the tting function is replotted in Fig. 7
in the red dash line and it can approximately match the measured
voltage.
To verify the accuracy of the extracted parameters, the PCDpro-
leis appliedtotheCBECMwiththeparameters calculatedfromthe
ninth pulse current discharge again. Fig. 8 shows the terminal vol-
tages of the battery and the CBECMas well as their corresponding
modelling errors represented by MSEs within 0.4%.
5. Verication of AGSMO for SOC estimation
The conguration of the AGSMO for the SOC estimation is illus-
trated in Fig. 9, where the properly dened current proles are
simultaneously applied to the LiPB module, the Ah counting mod-
ule and the AGSMO module. The LiPB terminal voltage is sampled
and fed into the AGSMO to generate the voltage tracking error,
which can be used to update the corresponding swiching gains of
the AGSMOto compenstate the modelling errors. The output of the
AGSMO module is the estimated SOC, which is concurrently com-
paredtothe true SOCdirectly generatedby the Ahcounting module
to demostrate the accuracy of the SOC estimation.
To validate the effectiveness and robustness of the AGSMO for
the SOC estimation, three types of current proles with different
discharge rates have been conducted on the LiPB at roomtemper-
ature. As the battery is fully charged, the initial SOC of the LiPB is
set to 100% at the beginning of discharging process.
120 X. Chen et al. / Computers and Chemical Engineering 64 (2014) 114123
Fig. 8. Terminal voltage of LiPB and CBECMand MSE in the PCD prole of 1Cn.
The testing data are obtained by using the constant current dis-
charge (CCD) prole with discharge rates of 1/3C
n
, 1C
n
and 1.5C
n
,
namely 1.67A, 5A, and 7.5A. The positive contants are selected
to satisfy the adaptation speed for the switching gains, they are

1
=0.5,
2
=0.3 and
3
=0.1. The initial SOC is also set to the ran-
domvalue away fromthe true value. The proposed AGSMO is used
to estimate the terminal voltage andthe SOCfor different discharge
rates, as an example, the results of the CCD with 1C discharge rate
are shown in Figs. 10 and 11, respectively. It can be seen that they
can track the true batery terminal voltage and the true SOC with no
chattering ripples. The SOC tracking MSEs are all within the small
range of 5% after a few seconds, which shows that the proposed
SOC observer is capable of tracking the true SOC accurately in the
presence of the incorrect initial SOC. This is due to the fact that
the switching gains are adjusted to the appropriate levels as the
corresponding errors decrease.
The testing data are obtained by using the variable current dis-
charge (VCD) prole withthe same average discharge rates as those
of the CCDprole. The proposed AGSMOis used to estimate the ter-
minal voltage and the SOC for the VCDprole in different discharge
rates. As an illustration, the results of terminal voltage and the SOC
of the VCD with the average discharge rate of 1C
n
are shown in
Figs. 12 and 13, respectively. It can be seen that the AGSMO has
Fig. 10. Terminal voltage of LiPB and AGSMO and MSE in the CCD prole of 1Cn.
peformed great robustness and capability to track the battery ter-
minal voltageandthetrueSOCregardless of anincorrect initial SOC.
The SOCestimationMSEs are boundedinthe range of 5%for 1090%
of SOC with minor chattering ripples. Again, this is due to the fact
that the switching gains are adaptively adjusted to adequate levels
against the tracking errors calculated by Eq. (12).
Inorder todemonstratetheSOCestimationbasedontheAGSMO
superior to the conventional SMO with the consideration of hys-
teresis effect, the urbandynamometer drivingschedule (UDDS) test
with corresponding current prole in Fig. 14 has been conducted
on the LiPB. The UDDS test is a typical dynamic driving cycle, which
is usually used to evaluate the vehicle performance. As shown in
Fig. 14, the UDDS cycle as speed versus time has been converted
to current versus time with respect to the capacity of 5Ah LiPB
Fig. 9. Conguration of proposed AGSMO system.
X. Chen et al. / Computers and Chemical Engineering 64 (2014) 114123 121
Fig. 11. True and AGSMO estimated SOC and SOC MSE in the CCD prole of 1Cn.
through the EV simulation. The negative current during the decel-
eration and braking of the EVs represents regenerative energy to
charge LiPB. The current prole of the UDDS is loaded to the LiPB
until the battery reaches its cut-off voltage of 2.7V and the results
are shown in Fig. 15. The same current prole is also loaded to the
proposed AGSMO is used to estimate the battery terminal voltage
andthe SOC. As showninFigs. 15 and16, the AGSMOis able to track
the battery terminal voltage and true SOC with incorrect initial
Fig. 12. Terminal voltage of LiPB and AGSMO and MSE in the VCD prole of 1Cn.
Fig. 13. True and AGSMO estimated SOC and SOC MSE in the VCD prole of 1Cn.
values. It can also achieve fast SOC convergence with minor chat-
tering ripples. The AGSMO has compared with the conventional
SMO with the constant switching gains for the SOC estimation
(Kim, 2006; Chen et al., 2012), the MSEs of the SOC estimation
for each type of current prole with different discharge rates
are summarised in Table 2. It can be seen that the MSEs of the
SOC estimation based on the AGSMO is always lower than those
based on the conventional SMO approach. Therefore, the proposed
AGSMOcan provide more robust and accurate SOC estimation than
Fig. 14. One UDDS cycle with corresponding current prole.
Table 2
MSEs of SOC estimation based SMO and AGSMO in different current proles.
Current prole SMO SOC estimation
MSEs (%)
AGSMO SOC estimation
MSEs (%)
CCD 1.5Cn 8.364 3.676
CCD 1Cn 7.523 2.583
CCD 1/3Cn 6.362 2.204
VCD 1.5Cn 9.364 3.432
VCD 1Cn 8.438 2.318
VCD 1/3Cn 7.238 2.069
UDDS 8.152 2.573
122 X. Chen et al. / Computers and Chemical Engineering 64 (2014) 114123
Fig. 15. Terminal voltage of LiPB and AGSMO and MSE in the UDDS prole.
conventional SMO-based approach. Furthermore, the proposed
AGSMO has compared with the EKF, which is one of popular KF
based approaches. Fig. 17 illustrates the EKF based SOC estimation
results in the UDDS current prole. It can be observed that the EKF
based approach has large estimation errors with slowconvergence
inthe comparisonwiththe proposedAGSMOapproachas shownin
Fig. 16, which demonstrates that the proposed AGSMOcan provide
robust tracking capability against modelling errors and incorrect
initial states.
Fig. 16. True and AGSMO estimated SOC and SOC MSE in the UDDS prole.
Fig. 17. True and EKF estimated SOC and SOC MSE in the UDDS prole.
6. Conclusions
The adaptive gain sliding mode observer (AGSMO) for the SOC
estimationbased onthe combined battery equivalent circuit model
(CBEDM) has been presented. The system state equations for the
AGSMO are derived from the CBEDM using equivalent control
concepts. The LiPB is utilised to conduct the experiments. The
parameters of the CBEDMare extractedfromthe experimental data
under a sequence of pulse current discharge. The experimental data
of the LiPB in constant current and variable current discharges are
used to verify the performance and effectiveness of the proposed
observer for the SOCestimation. It shows that the proposedAGSMO
has outperformed conventional SMO for SOC estimation of the LiPB
in terms of robust tracking capability with less chattering ripples
and high estimation accuracy.
Acknowledgment
This research work is supported by Commonwealth of Australia,
throughthe Cooperative ResearchCentre for AdvancedAutomotive
Technology (AutoCRC), under the project of Electric Vehicle Control
Systems and Power Management (C2-801).
References
Abu-Sharkh, S., & Doerffel, D. (2004). Rapid test and nonlinear model characteri-
sation of solid-state lithium-ion batteries. Journal of Power Sources, 130(12),
266274.
Barbarisi, O., Vasca, F., & Glielmo, L. (2006). State of charge Kalman lter estimator
for automotive batteries. Control Engineering Practice, 14(3), 267275.
Chau, K. T., Wu, K. C., Chan, C. C., & Shen, W. X. (2003). A new battery capacity
indicator for nickel-metal hydride battery powered electric vehicles using adap-
tive neuro-fuzzy inference system. Energy Conversion and Management, 44(13),
20592071.
Chen, M., Gabriel, A., & Mora, R. (2006). Accurate electrical battery model capable of
predicting runtime andIVperformance. IEEE Transactions onEnergy Conversion,
21(2), 504511.
Chen, X. P., Shen, W. X., Cao, Z. W., & Kapoor, A. (2012). Sliding mode observer for
state of charge estimationbasedonbatteryequivalent circuit inelectric vehicles.
Australian Journal of Electrical and Electronics Engineering, 9(3), 225234.
Chiang, Y. H., Sean, W. Y., & Ke, J. C. (2011). Online estimation of internal resistance
and open-circuit voltage of lithium-ion batteries in electric vehicles. Journal of
Power Sources, 196(8), 39213932.
Cho, S. W., Jeong, H. S., Han, C. H., Jin, S. S., Lim, J. H., & Oh, J. K. (2012). State-of-
charge estimation for lithium-ion batteries under various operating conditions
using an equivalent circuit model. Computers and Chemical Engineering, 41(11),
19.
X. Chen et al. / Computers and Chemical Engineering 64 (2014) 114123 123
Dai, H. F., Wei, X. Z., Sun, Z. C., Wang, J. Y., & Gu, W. J. (2012). Online cell SOC esti-
mation of Li-ion battery packs using a dual time-scale Kalman ltering for EV
applications. Applied Energy, 95, 227237.
Edwards, C., & Spurgeon, S. (1994). On the development of discontinuous observers.
International Journal of Control, 59(5), 12111229.
Han, J. Y., Kim, D. C., &Sunwoo, M. H. (2009). State-of-charge estimation of lead-acid
batteries using an adaptive extended Kalman lter. Journal of Power Sources,
188(2), 606612.
Hansen, T., & Wang, C. J. (2005). Support vector based battery state of charge esti-
mator. Journal of Power Sources, 141(2), 351358.
He, W., Williard, N., Chen, C. C., & Pecht, M. (2013). State of charge estimation
for electric vehicle batteries using unscented kalman ltering. Microelectronics
Reliability, 53(6), 840847.
He, H. W., Xiong, R., & Fan, J. X. (2011). Evaluation of lithium-ion battery equiva-
lent circuit models for state of charge estimation by an experimental approach.
Energies, 4(4), 582598.
Hu, D., Youn, B. D., &Chung, J. (2012). Amultiscale frameworkwithextendedKalman
lter for lithium-ion battery SOC and capacity estimation. Applied Energy, 92,
694704.
Hussein, H. A. H., & Batarseh, I. (2011). An overviewof generic battery models. In In
Proceedings of IEEE power and energy society general meeting Detroit, MI, USA.
Hu, X. S., Li, S. B., &Peng, H. (2012). Acomparative study of equivalent circuit models
for Li-ion batteries. Journal of Power Sources, 198(15), 359367.
Kim, I. S. (2006). The novel state of charge estimation method for lithium battery
using sliding mode observer. Journal of Power Sources, 163(1), 584590.
Lee, S. J., Kim, J. H., Lee, J., &Cho, B. H. (2008). State-of-charge andcapacity estimation
of lithium-ion battery using a new open-circuit voltage versus state-of-charge.
Journal of Power Source, 185(2), 13671373.
Li, I. H., Wang, W. Y., Su, S. F., & Lee, Y. S. (2007). A merged fuzzy neural network
and its applications in battery state-of-charge estimation. IEEE Transactions on
Energy Conversion, 22(3), 697708.
Ng, K. S., Moo, C. S., Chen, Y. P., & Hsieh, Y. C. (2009). Enhanced coulomb count-
ing method for estimating state-of-charge and state-of-health of lithium-ion
batteries. Applied Energy, 86(9), 15061511.
Plett, G. L. (2006a). Sigma-point Kalmanltering for battery management systems of
LiPB-based HEVbattery packs: Part 1: Introduction and state estimation. Journal
of Power Sources, 161(2), 3561368.
Plett, G. L. (2006b). Sigma-point Kalman ltering for battery management systems
of LiPB-based HEV battery packs: Part 2: Simultaneous state and parameter
estimation. Journal of Power Sources, 161(2), 13691384.
Piller, S., Perrin, M., & Jossen, A. (2001). Methods for state-of-charge determination
and their applications. Journal of Power Sources, 96(1), 113120.
Ramadesigan, V., Northrop, P. W. C., De, S., Santhanagopalan, S., Braatz, R. D., & Sub-
ramanian, V. R. (2012). Modeling and simulation of lithium-ion batteries from
a systems engineering perspective. Journal of the Electrochemical Society, 159(3),
R31R45.
Ristic, B., Arulampalam, S., & Gordon, N. (2004). Beyond the Kalman lter: Particle
lters for tracking applications. Artech House: Boston.
Rodrigues, S., Munichandraiah, N., &Shukla, A. K. (2000). Areviewof state-of-charge
indication of batteries by means of a.c. impedance measurements. Journal of
Power Sources, 87(12), 1220.
Shen, W. X. (2007). State of available capacity estimation for lead-acid batteries
in electric vehicles using neural network. Energy Conversion and Management,
48(2), 433442.
Shen, W. X., Chan, C. C., Lo, E. W. C., & Chau, K. T. (2002). A new battery available
capacity indicator for electric vehicles using neural network. Energy Conversion
and Management, 43(6), 817826.
Smith, K. A., Rahn, C. D., & Wang, C. Y. (2010). Model-based electrochemical esti-
mation and constraint management for pulse operation of lithiumion batteries.
IEEE Transactions on Control Systems Technology, 18(3), 654663.
Zhang, J. L., & Xia, C. Y. (2011). State-of-charge estimation of valve regulated lead
acid battery based on multi-state unscented Kalman lter. International Journal
of Electrical Power and Energy Systems, 33(3), 472476.
Zhang, F., Liu, G. J., Fang, L. J., & Wang, H. G. (2012). Estimation of battery state of
charge with H observer: Applied to a robot for inspecting power transmission
lines. IEEE Transactions on Industrial Electronics, 59(2), 10861095.