Paper 13

Uploaded by

BALASREE MURALEEDHARAN

We take content rights seriously. If you suspect this is your content, claim it here.

Available Formats

Download as PDF, TXT or read online on Scribd

0% found this document useful (0 votes)

33 views10 pages

Paper 13

Uploaded by

BALASREE MURALEEDHARAN

We take content rights seriously. If you suspect this is your content, claim it here.

Available Formats

Download as PDF, TXT or read online on Scribd

You are on page 1/ 10

This article has been accepted for publication in a future issue of this journal, but has not been

fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/TVT.2015.2427659, IEEE Transactions on Vehicular Technology
VT-2014-01076.R2 1

Robust Adaptive Sliding Mode Observer Using

RBF Neural Network for Lithium-ion Battery
State of Charge Estimation in Electric Vehicles
Xiaopeng Chen, Weixiang Shen, Member, IEEE, Mingxiang Dai, Zhenwei Cao, Member, IEEE, Jiong
Jin, Member, IEEE, and Ajay Kapoor

 SOC reflects the ratio of the utilizable capacity to its nominal

Abstract—This paper presents a robust sliding mode observer capacity [4]. It is substantial for managing battery energy
(RSMO) for state of charge (SOC) estimation of a utilization efficiently, preventing the battery from being
lithium-polymer battery (LiPB) in electric vehicles (EVs). A radial over-charged and over-discharged [5], [6]. Thus, a
basis function (RBF) neural network (NN) is employed to
well-developed method is essential for accurate SOC
adaptively learn an upper bound of system uncertainty. The
switching gain of the RSMO is adjusted based on the learned estimation.
upper bound to achieve asymptotic error convergence of the SOC The ampere-hour (Ah) counting is the most straightforward
estimation. A battery equivalent circuit model (BECM) is approach for the SOC estimation, which simply integrates the
constructed for battery modeling and its BECM are identified in charge and discharge currents over operating period. The
real time by using a forgetting factor recursive least square accuracy of this method is primarily reliant on precise current
algorithm. The experiments under the discharge current profiles
measurement. But, the noises are often embedded into the
based on EV driving cycles are conducted on the LiPB to validate
the effectiveness and accuracy of the proposed framework for the measured current, which attributes to the SOC drift. Therefore,
SOC estimation. a variety of techniques has been proposed to improve the SOC
estimation and each technique has its merits in certain aspects
Index Terms—Battery management system (BMS), battery [7], [8]. One technique is the battery internal impedance
equivalent circuit model (BECM), electric vehicles (EVs), measurement, which loads a series of small amplitude ac
lithium-ion batteries, neural networks (NNs), sliding mode signals with a wide range of frequencies into the battery to
observer (SMO), state of charge (SOC).
detect the responses for internal impedance calculation [9],
[10].
I. INTRODUCTION Another category of the SOC estimation methods is
established on machine learning strategies, which encompass
W ITH the rapid innovation of battery technologies,
lithium-ion (Li-ion) batteries have exhibited outstanding
performance in comparison with other types of batteries such as
artificial neural networks (NNs), fuzzy NNs and adaptive fuzzy
NNs [11-13]. These data-oriented approaches treat the battery
as a black-box system. Without requiring prior knowledge of
lead-acid batteries and nickel-metal hydride batteries. Li-ion battery internal structure, NNs are able to model a battery based
batteries offer high energy and power density, fast charge on a large amount of training data. But, it heightens
capability, no memory effect, longer cycle life and low computational burden in the BMSs.
self-discharge rate. These outstanding features facilitate Li-ion Some techniques based on different types of observers have
batteries as the most promising power sources for the also become more popular for the SOC estimation [14-22]. The
eco-friendly electric vehicles (EVs) [1-3]. Among existing EV Kalman filter (KF) has been adopted to estimate the SOC based
technologies, reliable battery management system (BMS) is a on linear battery models [14]. To deal with nonlinear behavior
remaining design challenge and the pivotal function of this of battery model, the enhanced versions of the KF have been
BMS is the indication of battery state of charge (SOC). The further investigated to achieve online SOC estimation, such as
extended KF (EKF) [15], [16] and unscented KF (UKF) [17],
Manuscript received June 20, 2014; revised October 30, 2014 and February
28, 2015; accepted March 27, 2015. Date of publication March 29, 2015; date [18]. The underlying principle of the EKF is to linearize the
of current version April 28, 2015. This research work is supported by nonlinear functions by means of the first-order Taylor series
Commonwealth of Australia, through the Cooperative Research Centre for expansion. This local linearization can result in large errors at
Advanced Automotive Technology (AutoCRC), under the project of Electric
Vehicle Control Systems and Power Management (C2-801). the high degree of nonlinearity in battery models. On the other
X.P. Chen is with the Faulty of Science, Engineering and Technology, hand, the UKF applies an unscented transformation to
Swinburne University of Technology, Hawthorn, Victoria, 3122, Australia approximate the probability density function of the nonlinear
(e-mail: [email protected]).
W.X. Shen is now with the Faulty of Science, Engineering and Technology, systems with a set of sample points and offers better SOC
Swinburne University of Technology, Hawthorn, Victoria, 3122, Australia estimation results. A sliding mode observer (SMO) has been
(e-mail: [email protected]). employed to accomplish the SOC estimation with the capability

0018-9545 (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/TVT.2015.2427659, IEEE Transactions on Vehicular Technology
VT-2014-01076.R2 2

of compensating the effects of nonlinearity and uncertainty in experimental data obtained from the pulse current discharge
battery models [19], [20]. This observer relies on the (PCD) and pulse current charge (PCC) tests (see details in
determination of the appropriate switching gain associated with section III). A resistor, Rin characterizes the electrolyte and
system uncertainty bound, which turns out the trade-off interphase resistance of the battery. Two parallel-connected
between the chattering levels in the SOC estimation and the networks consist of the electrochemical polarization resistance,
error convergence speed. To achieve better SOC estimation Rpe, capacitance, Cpe, the concentration polarization resistance,
accuracy, the SMO with an adaptive switching gain Rpc, and capacitance, Cpc, which reflect the short-term and
(ASGSMO) has been proposed to counteract the modeling long-term transient responses as the results of battery relaxation
errors and mitigate chattering effects. However, the uncertainty effect, respectively. The relaxation effect refers to slow
bound is still inevitable to be specified for the initial value of convergence of battery terminal voltage to the true OCV at its
the switching gain [21], [22]. The underestimated or equilibrium state after hours of relaxation at the end of charging
overestimated switching gain has caused the poor tracking or discharging process [23].
capability in the SOC estimation. The SOC describes the ratio of the remaining capacity to the
In this paper, a robust SMO (RSMO) is proposed for the nominal capacity in the battery, and it can be expressed as
SOC estimation of a lithium-polymer battery (LiPB) in EVs. t  I  
This robust strategy integrates a radial basis function (RBF) Z (t )  Z (t0 )  
t0 Cn
d (1)
NN, which is learning the bound of uncertain dynamics of
battery equivalent circuit model (BECM) in real time. The where Z(t0) is the initial SOC of the battery, I(τ) is the
outputs of the RBF NN are used to adaptively tune the instantaneous current and it is assumed to be positive for
switching gain of the RSMO, which ensures the estimation discharge and negative for charge, respectively. The η denotes
error asymptotically converging to zero. Moreover, online the coulomb efficiency and it is normally taken as one for
identification of the BECM parameters has been made through discharging and less than or close to one for charging LiPB in
a forgetting factor recursive least square (FFRLS) algorithm. the broad range of current and temperature.
The identified parameters of the BECM are incorporated with The time derivative of the SOC in (1) gives
the RSMO to diminish the impact of modeling errors [14-22].
With the synergy of the RBF NN learning capability and online  I 
Z    (2)
parameter identification, the RSMO demonstrates strong  Cn 
robustness with respect to time-varying dynamics and
nonlinearity of the battery. This novel method is implemented According to the circuit theory, Vt in Fig. 1 can be defined as
and tested on the LiPB with the current profiles based on EV
driving cycles. The experimental results show that the proposed
Vt  Voc (Z )  Vpe  Vpc  IRin (3)
RSMO has higher accuracy and fast convergence for the SOC
estimation in the comparison with conventional SMO and
ASGSMO. The dynamics of the voltages across the polarization
The rest of paper is organized as follows. Section II outlines capacitors are given by
the BECM for Li-ion batteries. Section III introduces the
FFRLS algorithm for online parameter identification. Section 1 1
V pe   V pe  I (4)
IV explains the RSMO design incorporating with the RBF NN R pe C pe C pe
for the SOC estimation in details. Section V shows the
1 1
experimental setups, validation results and the comparison of V pc   V pc  I (5)
the proposed method with the conventional SMO and R pc C pc C pc
ASGSMO approaches. Section VI draws the conclusions.
Despite the nonlinearity of the OCV-SOC curves as shown in
Fig. 2, Voc(Z) has a piecewise linear relationship in certain
II. BATTERY MODELING ranges of the SOC indicated by the dots in the curves. Hence,
Various types of the BECM have been proposed to capture the OCV can be expressed as a nonlinear function of the SOC
dynamic characteristics of EV batteries [23-25]. One of the such that
prevailing battery models is the BECM as shown in Fig. 1,
which is utilized to represent the dynamic behaviors of the Voc  Z    Z  g  Z  (6)
LiPB in this paper. A capacitor, Cn indicates the battery
nominal capacity in ampere hours and a current source, I and Vt where κZ and g(Z) stand for the linear and nonlinear parts of the
denote the battery current and the terminal voltage, OCV, respectively. Thus, the derivative of Voc with respect to
respectively. The voltage across Cn as the open circuit voltage SOC, Z gives
(OCV), Voc varies in the range of the SOC, Z between 0% and
100%. The current-controlled voltage source, Voc(Z) is used for dVoc
mapping the nonlinear relationship between the SOC and the    g Z  (7)
dZ
OCV as illustrated in Fig. 2, which is derived by fitting the

It is worthwhile to mention that the rate of change of the   x, u, t    t  x  R 4 , u  R1 , t  0 (13)

charging or discharging current can be negligible, namely dI/dt
≈0 due to the high sampling rate [19-22]. Therefore, the time
derivative of Vt in (3) and the substitute of (4)-(5) and (7) into it where   t  is the desired upper bound of system uncertainty.
yields According to (12), the equation (10) can be rewritten as

 I  V pe I V pc I x(t )  Ax(t )  Bu(t )  ( x, u, t ) (14)

Vt     g  Z        (8)
 Cn  R peC pe C pe R pc C pc C pc
To obtain the nonlinear relationship between the OCV and the
SOC as well as the battery model parameters, two specifically
Solving I in (3) and substituting it into (8) as well as rearranging
designed experiments and an online parameter identification
(4), (5) and (8) to form the state-space equations as
method are explained in next section.

Vt  a11Vt  a12 Z  a13V pe  a14V pc  b11I  1 (Z , I )  f1 III. BATTERY MODEL PARAMETERS IDENTIFICATION
Z  a21Vt  a22 Z  a23V pe  a24V pc   2 ( Z )  f 2 A Turnigy LiPB is used in the experiments as shown in Fig.
(9) 3. It is composed of a LiMn2O4 cathode and an artificial
V pe  a33V pe  b31 I  f3
graphite anode. It has a nominal capacity of 5.0Ah and a
V pc  a44V pc  b41I  f 4 nominal voltage of 3.7V. The maximum charging voltage is
4.2V and the cut-off voltage is 2.7V. The experiments are
where a11 = -[1/(RpeCpe)+1/(RpcCpc)], a12 = conducted on Arbin BT2000 battery testing equipment as
κ[1/(RpeCpe)+1/(RpcCpc)], a13 = -1/(RpcCpc), a14 = -1/(RpeCpe), a21 shown in Fig. 4. This battery testing platform can control
= 1/(RinCn), a22 = -κ/(RinCn), a23 = 1/(RinCn), a24 = 1/(RinCn), a33 charging and discharging processes of batteries via the Arbin
= -1/(RpeCpe), a44 = -1/(RpcCpc), b11 =-[κ/Cn + Rin/(RpeCpe) + software program.
1/Cpe + Rin/(RpcCpc)+ 1/Cpc], b31 = 1/Cpe and b41 = 1/Cpc. Instead of a hybrid pulse power characterization (HPPC) test
as a combination version of the PCD and PCC tests, two tests
Moreover, two nonlinear function terms, ξi (i=1, 2) represent have been conducted separately to obtain accurate OCV-SOC
the nonlinear parts of the model which are induced by the curves. The OCV-SOC discharge curve is obtained from the
linearization of the SOC and input current. Four modeling error PCD test, which comprises of a sequence of ten current pulses
terms, Δf j (j=1, 2, 3 and 4) are caused by model uncertainty and with six minutes discharge period and one hour rest as shown in
external disturbances. If the input and the output of the BECM Fig. 5. The discharge current of each pulse is 5.0A, which
are taken as u(t) = I and y(t) = Vt, respectively, and the system corresponds to 1Cn discharge rate of the battery. For the fully
state variables are defined as Vt, Z, Vpe and Vpc, then (9) can be charged LiPB (Z = 100%), each pulse current discharges the
concisely expressed in the form of matrices battery for six minutes which is equivalent to 10% of nominal
capacity or 10% of the SOC deduction. The one hour rest
x(t )  Ax(t )  Bu(t )  (Z , u, f ) (10) allows the battery terminal voltage to stabilize towards the
equilibrium potential (e.g. the true OCV) before running next
y(t )  Cx(t ) (11) cycle. The pulse current discharge and recovery are iterated
where until the cut-off voltage of 2.7V (Z = 0%) is reached. Similarly,
 a11 a12 a13 a14   b11  the PCC test as shown in Fig. 6 is designed to acquire the
a a22 a23 
a24  0 OCV-SOC charge curve. The fully discharged LiPB is charged
A   21 , B    , C = [1 0 0 0] and x(t) for six minutes at the recommended 1Cn charge rate, which is
 0 0 a33 0  b31  equivalent to 10% of the SOC increase, followed by one hour
   
 0 0 0 a44  b41  rest, and this process is repeated until the battery reaches the
= [Vt Z Vpe Vpc] . T fully charged state (Z =100%).
After each hour rest, the OCV is measured corresponding to
The unknown function, (Z , u, f ) represents system the SOC. These measured OCVs versus the SOCs are shown in
Fig. 2. It is observed that the OCV levels during charging are
uncertainty and it can be assumed to satisfy the matching
always higher than those during discharging at the same SOC.
condition such that
Thus, there exists a hysteresis phenomenon between two
OCV-SOC curves. For the consideration of hysteresis effect,
(Z , u, f )   ( x, u, t ) (12) the OCV as a function of SOC is defined as the average OCV
value between charge and discharge curves indicated by the
where Γ is the system uncertainty input matrix and the function dashed lines in Fig. 2.
is unknown but bounded with respect to the input current which
has a limited operation range. Thus, there exists a bound such TABLE I
that

ˆ  k   ˆ  k  1  K  k  Vt  k     k ˆ  k  1 
IDENTIFIED BECM PARAMETERS B ASED ON OFFLINE LS APPROACH IN
PCD C URRENT PROFILE  
P  k  1  T
k 
Rin Rpe Cpe Rpc Cpc
K k   (17)
(mΩ) (mΩ) (kF) (mΩ) (kF)   T
 k  P  k  1   k 
1
P k   P  k  1  K  k   T  k  P  k  1 
3.1536 2.6213 3.4105 3.8534 6.2614
 

where ˆ  k  is the estimates of the coefficients vector θ(k) , K(k)

The least square (LS) method can be applied to identify the
BECM parameters offline from the experimental data of the is the gain and P(k) is the covariance matrix. λ is the forgetting
PCD test [21]. For the same PCD profile, a small deviation factor which is used to put more weight to present data and less
between the terminal voltages obtained from the experimental weight to past data to enhance computational efficiency.
data of the LiPB and those calculated from the BECM with the Typically, the forgetting factor is selected within [0.95 1].
identified parameters in Table I is observed in Fig. 7. However, As an example, the FFRLS algorithm is used to identify the
when other profiles deviate from the profile obtained under the parameters of the BECM for the UDDS profile in real time,
PCD test, the BECM with the parameters in Table 1 leads to where the forgetting factor λ is chosen as 0.98. The identified
considerable modeling errors. For the instance of the current results of Rin, Rpe, Rpc, Cpe and Cpc are illustrated in Fig. 9. It is
profile of the urban dynamometer driving schedule (UDDS), a seen that the BECM parameters are closely related to the
large deviation between the terminal voltages from the LiPB current and terminal voltage which are fluctuating within
and those from the BECM is observed in Fig. 8. In fact, the certain ranges in response to the dynamics of the LiPB
BECM parameters vary significantly with discharge current corresponding to the UDDS profile.
rates as well as the SOC [26]. Therefore, the online system To verify the model accuracy by using the FFRLS algorithm,
identification is necessary to determine parameters of the the same UDDS current profile is reloaded to the LiPB, the
BECM for reducing the modeling errors. In this paper, a terminal voltages obtained from the LiPB and the BECM and
forgetting factor recursive least square (FFRLS) algorithm is their absolute voltage errors are shown in Fig. 10, respectively.
utilized to identify the parameters of the BECM for the SOC It is clear that the model accuracy has been greatly improved as
estimation in real time while the battery is discharging or the modeling errors bounded in the range of -0.04V to 0.04V.
charging. Since the present data reflects more accurate Also in Fig. 10, the part of the results during the time interval
information of the battery current status that the past data, the from 3000 to 4000 seconds is zoomed to show the closeness of
FFRLS algorithm gives the trusting degree to the present data the terminal voltages. As a further comparison, the statistics of
higher than the past data so as to eliminate the phenomenon of the modeling errors obtained from both offline LS and online
“data saturation” and reduce modeling errors [27]. FFRLS algorithms in various current profiles are elaborated in
The identified parameters are Rin, Rpe, Cpe, Rpc and Cpc, which Table II. It can be seen that the BECM with the parameters
are the linear part of the BECM represented by (3), (4) and (5). identified by online FFRLS algorithm outperforms the BECM
They can be discretized into a difference equation as with the parameters identified by the offline LS algorithm in
modeling accuracy.
TABLE II
Vt (k )  1Vt (k  1)   2Vt (k  2)  3 I (k )
(15)
 4 I (k  1)  5 I (k  2)  6 STATISTICS OF M ODELING E RRORS FROM OFFLINE LS AND O NLINE FFRLS
M ETHODS

where the battery current and voltage are sampled at time steps Identification Current Minimum Maximum Mean Variance
k, k-1 and k-2, and the model parameters are estimated based on Method Profile (V) (V) (V) (V2)
the coefficients in (15), which are θ1=-a11, θ2=a13 a 14, θ2=Rin,
θ4=-a13Rin+b41-a14Rin+b31, θ5=a13a14Rin-a13b31-a14b41, and Offline LS PCD -0.0723 0.0357 0.0064 1.95e-05
θ6=Voc(1-θ6 -θ2). Offline LS PCC -0.1453 0.2047 0.0154 5.25e-04
Defining θ(k)=[ θ1 θ2 θ3 θ4 θ5 θ6 ]T and σ(k)=[Vt(k-1)
Offline LS UDDS 0.3185 1.8537 0.4212 9.54e-04
Vt (k-2) I(k) I(k-1) I(k-2) 1], can be concisely expressed as
Online
PCD -0.0416 0.0352 0.0038 2.58e-05
Vt (k )    k   k 
FFRLS
(16)
Online
PCC -0.0397 0.0407 0.0029 2.83e-05
FFRLS
For the system identification of (16), the FFRLS is realized as
follows Online
UDDS -0.0382 0.0378 0.0018 2.08e-05
FFRLS

IV. ROBUST SLIDING MODE OBSERVER DESIGN  xˆ  c 2 

Since the unpredictable driving conditions of EVs give rise i  xˆ   exp   i  i  1, ..., n (24)
 i2 
to unknown system uncertainty, it is impractical to use the  
predefined bounds of system uncertainties to determine the
switching gain in SMOs. Also, inappropriate switching gains of where ci  R 41 and  i  R1 denote the center vector and
SMOs have caused either slow estimation error convergence or
width of Gaussian function for the ith neuron node, respectively.
large chattering magnitudes in the SOC estimation [19-22].
n represents the number of neuron nodes in the hidden layer.
Thus, the RSMO with the switching gain adaptive to the
The width of Gaussian function, gives the covering scope of
learned upper bounds in real time is proposed as follows
the NN inputs. The wider width of the Gaussian function is, the
greater the covering scope of the NN inputs. The center vector,
xˆ  t   Axˆ  t   Bu  t   G Cxˆ t   y t   v t  (18) ci represents the center coordination of Gaussian function for
yˆ  t   Cxˆ  t  (19) each neuron node i. The nearer the center vector is to the NN
inputs, the better the sensitivity of Gaussian function is to the
NN inputs. Moreover, the NN approximation error is also
where the feedback gain matrix, G and the observer input
closely related to the number of neuron nodes. As a rule of
function, v(t) can be designed so that the robustness and errors
thumb, the number of neurons, n can be selected as (2m + 1) in
convergence of the RSMO are guaranteed.
the hidden layer and m is the number of NN inputs so that the
The state estimation errors between estimated states and true
number of neuron nodes is 9 in this study. By using the
states are defined as e(t )  xˆ (t )  x(t ), and subtracting (14) experimental data of the PCD and PCC tests to train the RBF
from (18) yields the error dynamics NN, local training technique is applied to predetermine the
center vectors and width [32], [33]. In this study, a matrix, H
e  t   A0e  t     x, u, t   v t  (20) formed by the center vectors and the width are obtained as
 1 0.75 0.5 0.25 0 0.25 0.5 0.75 1
where A0=(A-GC).  1 0.75 0.5 0.25 0 0.25 0.5 0.75 1
Remark 1: There exists a feedback gain matrix G such that   and
 1 0.75 0.5 0.25 0 0.25 0.5 0.75 1
A0=(A-GC) is a strictly Hurwitz matrix, where G can be directly  
obtained by using either the pole placement method or linear  1 0.75 0.5 0.25 0 0.25 0.5 0.75 1
quadratic regulator (LQR) method. For further analysis, a i  0.05 , respectively. Meanwhile, the adjustable RBF NN
sliding variable, s is defined as weight vector, Ŵ appears linearly with respect to the known
nonlinear function  ( xˆ ) . The initial values of Ŵ can be
s  Me(t )  FCe(t )  F  Cxˆ(t )  Cx(t )   F Cxˆ(t )  y(t )  (21)
randomly selected as Wˆ (0) = [0.2 0.2 0.2 0.2 0.2 0.2 0.2
0.2]T.
where M=FC and it is designed such that the following matrix
According to the universal approximation properties of the
RBF NN, the two assumptions are made as follows.
1
AS  ( M T MA0  A0T M T M ) (22) Assumption 1: Given an arbitrary small positive constant,  0
2
and a continuous function  (t ) on a compact set U, there exists
satisfies λmax(As)≤0, and λmax is the maximum eigenvalue of an optimal constant weight vector W * such that the output of
matrix As. the optimal RBF NN with neuron nodes satisfies
As a typical feed-forward NN, the RBF NNs have been
applied in numerous applications due to its merits of simple W *T  ( xˆ )  (t )   f   0 (25)
structure, fast training capability and capable of convergence to
global optimization [28-31]. As shown in Fig. 11, the RBF NN
Assumption 2: The norm of the system uncertainty and its
is adopted to learn the upper bound of system uncertainty in (13)
upper bound  (t ) satisfy the following inequality on the
as
compact set U

ˆ ( xˆ,Wˆ )  Wˆ T  ( xˆ) (23)

 (t )    x, u, t    0   f (26)

where the estimated state vector, x̂ is the input of the RBF NN,
In practice, the bound estimation error, εf cannot be exactly
Wˆ  R n is the weight vector of the RBF NN and the vector
confined to zero due to the changes of system uncertainty and
 ( xˆ )  R n is Gaussian type of functions and individual Gaussian
dynamics of the LiPB. But it can be constrained to a small value
function of each neuron node i, is expressed as to fulfil (25) and (26) so that optimal estimation of uncertainty
bound can be attained. Thus, ε0 = 0.002 and εf = 0.001 are used
in this study.

For the adjustment of the weight vector and the analysis of V  sT s   1W T W  sT Me(t )   1W T W
the estimation error convergence in the proposed RSMO, the
input function of the RSMO is designed as  eT M T MA0 e  eT M T M 
 eT M T Mv   1W T W

 
T
T
 s(t ) M

v(t )  
s(t )T M
2  
ˆ s(t )T M    ,
if s(t )T M   0
(27)
 0.5eT ( M T MA0  A0T M T M )e
 sT M   sT Mv   1W T W
(30)

 otherwise  eT AS e  sT M   sT Mv   1W T W
 0,
 max ( AS ) e  sT M   sT Mv   1W T W
2
 2
where   2  s(t ) , these adaptation parameters, γ and β
  sT M   sT Mv   1W T W
are positive scalars satisfying β > 0 and 0 < γ < 1, respectively.
Faster convergence is achieved by increasing adaptation
parameter β at the expense of more chattering in the SOC Substituting (27) and (28) into (30) gives
estimation when the initial estimation errors are significant.
The moderate values of β = 2 and γ = 0.5 are chosen to balance V   sT M 
the convergence speed and chattering level. The weight vector,  
T T
Ŵ is adjusted by applying the following adaptive law  (s M ) ˆ T 2 
 s M 
T
2
( s M   2   s )
T
 s M 
 
Wˆ   s(t )T M   ( xˆ )
 
(28)
  1W T Wˆ
where    0   f  M 0 2
  sT M   (ˆ sT M   2   s ) (31)
The output of the RBF NN can be instantaneously updated
by using the estimated states of the BECM with the simple
adaptive law in (28). The experimental data of the PCD and
1

   

  W *T  Wˆ T    sT M   ( xˆ )  

PCC tests are only used to predetermine the parameters of the 
 sT M    (ˆ sT M   2   s )
2
RBF NN in the offline training process, such as the center
vectors and width in the initial training step. Also, it should be  sT M  W *T  ( xˆ )  Wˆ T  ( xˆ ) 
noted that if the input function, v(t) is properly adapted by using  
the estimated upper bound from the RBF NN, then a sliding
mode motion can be induced on the system error states such Substituting (23), (25) and (26) into (31) yields
that the state estimation errors can be asymptotically converged
2
to zero. A candidate of Lyapunov function is considered as V  sT M   ˆ sT M   2   s  sT M  [W *T  ( xˆ )  Wˆ T  ( xˆ )]
follows 2
 sT M  (  ˆ    )  2   s  sT M  [W *T  ( xˆ )  Wˆ T  ( xˆ )]
(32)
 sT M  (   )  sT M  Wˆ T  ( xˆ )  sT M  [W *T  ( xˆ )   f ]
V ( s, W ) 
2
 
1 T 1
s s   1W T W
2
(29)  sT M  [W *T  ( xˆ )  Wˆ T  ( xˆ )]  2   s
2

2
  sT M   0  sT M   f  2    s
where W  W *  Wˆ is the adaptation error between the optimal   sT M  ( 0   f )  2   s
2

weight vector and the estimated weight vector of the RBF NN

and also W  W *  Wˆ  Wˆ . Since and γ and β are positive scalars with β > 0, 0< γ < 1, then
Taking the time derivative of the candidate of Lyapunov
2
function V in (29) gives V  2   s 0 (33)

According to Lyapunov stability theory [34], the negative

semi-definite V implies that the trajectory of sliding variable
defined in (21) reaches the sliding surface in a finite time and
remains on the sliding surface so that the state estimation error
(20) can asymptotically converge to zero. As the bound of
system uncertainty is adaptively updated in response to the
learned output of the RBF NN, the RSMO can provide robust
tracking capability against system uncertainty and significantly

restrain the chattering magnitudes in the SOC estimation. The incapability of adaptively adjusting the switching gain as the
next section will present the experimental results of the LiPB to estimation errors are converging. Particularly, it produces
validate the proposed method for the SOC estimation. considerable SOC estimation errors at high or low SOC ranges
with the maximum errors of 10% after initial SOC error
V. EXPERIMENTAL RESULTS AND DISCUSSIONS convergence. Compared with the conventional SMO, the
The effectiveness and feasibility of the RSMO using the RBF RSMO is able to track the true SOC accurately without
NN for the SOC estimation have been verified by conducting chattering effects due to its capability of identifying the
experiments on the LiPB at room temperature. An overall parameters of the BECM in real time and adaptively adjusting
framework of the RSMO for the SOC estimation is outlined in the switching gain with the learned upper bound as the
Fig. 12, where the predefined current profiles have been estimation errors are converging. The overall estimation errors
programmed in the Arbin BT2000 to test the LiPB. of the proposed method are bounded within the range of 3%
As the battery is fully charged, the true initial SOC of the after initial SOC error convergence regardless of high or low
LiPB is set to 100% at the starting of each test. The currents and SOC ranges.
terminal voltages of the LiPB at the sampling rate of one second The second set of the data is obtained by performing the
are sent to the BECM and the RSMO modules programmed in highway fuel economy driving schedule (HWFET). The
the host computer for online parameter identification and the corresponding current profile and terminal voltages are shown
SOC estimation, respectively. According to (15), the currents in Fig. 15. The HWFET drive cycle represents highway driving
and terminal voltages at t, (t - 1) and (t - 2) are used for online conditions with minimal stops as approximately average speed
identificaiton of the model parameters. Thus, the identified of 100 km/h. Similarly, the initial SOC is also set to 20% away
BECM parameters can only be loaded to the RSMO module from the true value. The SOC estimation results of the proposed
after three second delay. Once the RSMO has all the parameters, RSMO, the conventional SMO and their comparison with the
it estimates system states which are used as the inputs of the true SOC and absolute estimation errors are illustrated in Figs
RBF NN to learn the upper bound of system uncertainty. There 16 and 17, respectively.
exists two second delay between the RSMO and the RBF NN It can be seen that the RSMO has the strong capability to
modules because the states have to be estimated first for accurately track the true SOC regardless of an incorrect initial
updating the inputs of the RBF NN and then the output of the SOC. The SOC errors are bounded in the range of 2% in the
RBF NN is feedback to the RSMO for adjusting the switching whole range of SOCs without chattering ripples. The
gain. As a result, the RSMO with the switching gain adaptive to conventional SMO with the BECM of the constant parameters
the upper bound learnt by using the RBF NN is used to estimate and switching gain gives significant chattering ripples and large
the SOC while overcoming the impacts of system uncertainty. SOC errors, as shown in Figs. 14 and 17, respectively.
Then, the estimated SOC, the output of the RSMO module, is To further verify the effectiveness of the RSMO, the
directly compared with the true SOC calculated by an Ah ASGSMO based on the BECM with the constant parameters
counting module to demonstrate its accuracy. and the adaptive switching gain in [21], [22] has also been used
The current profiles based on two conventional EV driving for the SOC estimation. Due to the unknown upper bound, the
cycles are used to test the LiPB, where the discharge and initial switching gain for the ASGSMO which is either
regenerative current levels with respect to the capacity of the overestimated or underestimated can cause slow SOC
selected LiPB are similar to those with respect to the capacity of convergence. Figs. 18 and 19 show the estimated results from
the battery pack in EVs. The first set of the data is obtained by the ASGSMO. It can be seen that the overall SOC errors of the
running the current profile of the UDDS as shown in Fig. 8. The ASGSMO are larger than those of the RSMO approach.
UDDS characterizes the congested traffic in city with repetitive Moreover, the standard deviation of the SOC estimation
cycles of acceleration and braking. The deceleration and errors can be represented by using root mean square errors
braking of EVs produce regenerative energy as the negative (RMSEs), which is defined as
current to charge the LiPB. However, the overall current profile
 Z 
2
is tending to discharge the LiPB during this EV driving cycles. N
 Zˆi
i 1 i
This current profile is consecutively loaded to the fully RMSE  (34)
charged LiPB (SOC=100%) until the terminal voltage reached N
its cut-off voltage of 2.7V corresponding to SOC=0%. The
SOC estimation results of the proposed RSMO and where N is the total number of the true and estimated SOC data
conventional SMO approaches and their comparisons with the in each sampled interval. Table III summarizes the RMSEs of
true SOC as well as their absolute estimation errors are shown the SOC estimation for the proposed RSMO and conventional
in Fig. 13 and Fig. 14, respectively. It can be observed that both SMO and ASGSMO approaches under the UDDS and the
approaches are robust against the incorrect initial SOC value, HWFET. It can be seen that the proposed RSMO is superior to
which has been set to 20% away from the true value. For the the conventional SMO and ASGSMO approaches. This is due
conventional SMO based on the BECM with the constant to the fact that the proposed RSMO, where the parameters of
parameters, it requires a relatively large switching gain to the BECM are updated in real time responding to the dynamics
compensate modeling errors, thus the significant chattering of the LiPB and the switching gain are adjusted adaptively to
ripples are generated on the SOC estimation results due to its adequate levels by applying the upper bound estimated from the

RBF NN, can compensate system uncertainty and provide [4] G. L. Plett, "Extended Kalman filtering for battery management systems
of LiPB-based HEV battery packs: Part 1. Background," Journal of Power
robust and accurate SOC estimation for EV applications. Sources, vol. 134, pp. 252-261, 2004.
[5] K. W. E. Cheng, B. P. Divakar, W. Hongjie, K. Ding, and H. Ho Fai,
TABLE III "Battery-management system (BMS) and SOC development for electrical
vehicles," IEEE Transactions on Vehicular Technology, vol. 60, pp.
RMSE S OF SOC E STIMATION R ESULTS OF THREE M ETHODS IN DIFFERENT 76-88, 2011.
C URRENT PROFILES [6] M. Zandi, A. Payman, J. P. Martin, S. Pierfederici, B. Davat, and F.
Meibody-Tabar, "Energy management of a fuel
Current RSMO SMO ASGSMO cell/supercapacitor/battery power source for electric vehicular
applications," IEEE Transactions on Vehicular Technology, vol. 60, pp.
Profiles RMSEs RMSEs RMSEs
433-443, 2011.
(%) (%) (%) [7] V. Pop, H. J. Bergveld, P. H. L. Notten, and P. P. L. Regtien,
"State-of-the-art of battery state-of-charge determination," Measurement
UDDS 2.3214 5.8123 3.4321 Science and Technology, vol. 16, p. R93, 2005.
HWFET 2.3318 5.7145 3.5134 [8] L. Lu, X. Han, J. Li, J. Hua, and M. Ouyang, "A review on the key issues
for lithium-ion battery management in electric vehicles," Journal of
Power Sources, vol. 226, pp. 272-288, 2013.
[9] S. Rodrigues, N. Munichandraiah, and A. K. Shukla, "A review of
state-of-charge indication of batteries by means of a.c. impedance
VI. CONCLUSIONS measurements," Journal of Power Sources, vol. 87, pp. 12-20, 2000.
In this paper, a robust sliding mode observer (RSMO) has [10] H. Blanke, O. Bohlen, S. Buller, R. W. De Doncker, B. Fricke, A.
Hammouche, et al., "Impedance measurements on lead–acid batteries for
been developed to estimate state of charge (SOC) for a state-of-charge, state-of-health and cranking capability prognosis in
lithium-polymer battery (LiPB) in electric vehicles (EVs). This electric and hybrid electric vehicles," Journal of Power Sources, vol. 144,
methodology integrates the RSMO with the online parameter pp. 418-425, 2005.
identification for a battery equivalent circuit model (BECM) by [11] W. X. Shen, "State of available capacity estimation for lead-acid batteries
in electric vehicles using neural network," Energy Conversion and
applying the forgetting factor recursive least square (FFRLS) Management, vol. 48, pp. 433-442, 2007.
algorithm and the learning capability of a RBF neural network [12] W. X. Shen, K. T. Chau, C. C. Chan, and E. W. C. Lo, "Neural
for the approximation of the upper bound of system network-based residual capacity indicator for nickel-metal hydride
uncertainty. This integration allows the RSMO to start with the batteries in electric vehicles," IEEE Transactions on Vehicular
Technology, vol. 54, pp. 1705-1712, 2005.
appropriate initial switching gain and then adjust the switching [13] W. X. Shen, C. C. Chan, E. W. C. Lo, and K. T. Chau, "Adaptive
gain adaptively to ensure the convergence of the SOC neuro-fuzzy modeling of battery residual capacity for electric vehicles,"
estimation errors. IEEE Transactions on Industrial Electronics, vol. 49, pp. 677-684, 2002.
The experimental results of the LiPB obtained from the [14] O. Barbarisi, F. Vasca, and L. Glielmo, "State of charge Kalman filter
estimator for automotive batteries," Control Engineering Practice, vol. 14,
current profiles based on two conventional EV drving cycles pp. 267-275, 2006.
illustrate the robust performance and highlight the effectiveness [15] M. Charkhgard and M. Farrokhi, "State-of-Charge estimation for
of the proposed observer for the SOC estimation. It shows that Lithium-ion batteries using neural networks and EKF," IEEE
the proposed RSMO is superior to conventional SMO and Transactions on Industrial Electronics, vol. 57, pp. 4178-4187, 2010.
[16] Z. Chen, Y. H. Fu, and C. C. Mi, "State of charge estimation of
ASGSMO for the SOC estimation in accuracy and tracking Lithium-ion batteries in electric drive vehicles using extended Kalman
capability. filtering," IEEE Transactions on Vehicular Technology, vol. 62, pp.
For the future work, cell balancing and thermal consistency as 1020-1030, 2013.
well as aging effect in terms of battery state of health will be [17] W. He, N. Williard, C. C. Chen, and M. Pecht, "State of charge estimation
for electric vehicle batteries using unscented kalman filtering,"
considered so that the proposed approach for the SOC Microelectronics Reliability, vol. 53, pp. 840-847, 2013.
estimation can be extended from a cell to a battery pack, [18] J. Zhang and C. Xia, "State-of-charge estimation of valve regulated lead
making it practically useful and reliable in EV applications. acid battery based on multi-state unscented Kalman filter," International
Journal of Electrical Power & Energy Systems, vol. 33, pp. 472-476,
2011.
ACKNOWLEDGMENT [19] I. S. Kim, "The novel state of charge estimation method for lithium
This research work is supported by Commonwealth of battery using sliding mode observer," Journal of Power Sources, vol. 163,
pp. 584-590, 2006.
Australia, through the Cooperative Research Centre for [20] X. P. Chen, W. X. Shen, Z. W. Cao, and A. Kapoor, "Sliding mode
Advanced Automotive Technology (AutoCRC), under the observer for state of charge estimation based on battery equivalent circuit
project of Electric Vehicle Control Systems and Power in electric vehicles," Australian Journal of Electrical & Electronic
Engineering, vol. 9, 2012.
Management (C2-801). [21] X. P. Chen, W. X. Shen, Z. W. Cao, and A. Kapoor, "A novel approach
for state of charge estimation based on adaptive switching gain sliding
REFERENCES mode observer in electric vehicles," Journal of Power Sources, vol. 246,
pp. 667-678, 2014.
[1] B. Scrosati and J. Garche, "Lithium batteries: status, prospects and [22] X. P. Chen, W. X. Shen, Z. W. Cao, and A. Kapoor, "Adaptive gain
future," Journal of Power Sources, vol. 195, pp. 2419-2430, 2010. sliding mode observer for state of charge estimation based on combined
[2] A. Affanni, A. Bellini, G. Franceschini, P. Guglielmi, and C. Tassoni, battery equivalent circuit model," Computers & Chemical Engineering,
"Battery choice and management for new-generation electric vehicles," vol. 64, pp. 114-123, 2014.
IEEE Transactions on Industrial Electronics, vol. 52, pp. 1343-1349, [23] M. Chen and G. A. Rincon-Mora, "Accurate electrical battery model
2005. capable of predicting runtime and I-V performance," IEEE Transactions
[3] X. P. Chen, W. X. Shen, T. T. Vo, Z. W. Cao, and A. Kapoor, "An on Energy Conversion, vol. 21, pp. 504-511, 2006.
overview of lithium-ion batteries for electric vehicles," in Proceedings [24] X. S. Hu, S. B. Li, and H. Peng, "A comparative study of equivalent
IPEC on Power & Energy, 2012, pp. 230-235. circuit models for Li-ion batteries," Journal of Power Sources, vol. 198,
pp. 359-367, 2012.

[25] C. R. Gould, C. M. Bingham, D. A. Stone, and P. Bentley, "New battery Mingxiang Dai is currently a PhD
model and state-of-health determination through subspace parameter
candidate with Science and Technology on
estimation and state-observer techniques," IEEE Transactions on
Vehicular Technology, vol. 58, pp. 3905-3916, 2009. Transient Physics Laboratory, Nanjing
[26] Y. H. Chiang, W. Y. Sean, and J. C. Ke, "Online estimation of internal University of Science and Technology, and
resistance and open-circuit voltage of lithium-ion batteries in electric Nanjing, China and a PhD visitor in
vehicles," Journal of Power Sources, vol. 196, pp. 3921-3932, 2011.
Electrical and Computer Systems
[27] H. He, X. Zhang, R. Xiong, Y. Xu, and H. Guo, "Online model-based
estimation of state-of-charge and open-circuit voltage of lithium-ion Engineering, Monash University,
batteries in electric vehicles," Energy, vol. 39, pp. 310-318, 2012. Australia, Australia. His main research
[28] H. Han, Q. Chen, and J. Qiao, "Research on an online self-organizing interests include system identification,
radial basis function neural network," Neural Computing and
optimal control theory and Kalman filter.Science, Engineering
Applications, vol. 19, pp. 667-676, 2010.
[29] H. G. Han and J. F. Qiao, "Adaptive computation algorithm for RBF and Technology, Swinburne University of Technology,
neural network," IEEE Transactions on Neural Networks and Learning Australia. His research interests focus on electrical vehicles,
Systems, vol. 23, pp. 342-347, 2012. renewable energy and power system.
[30] W. E. Wong, V. Debroy, R. Golden, X. Xu, and B. Thuraisingham,
"Effective software fault localization using an RBF neural network,"
Zhenwei Cao received the B.S. and M.E.
IEEE Transactions on Reliability, vol. 61, pp. 149-169, 2012.
[31] L. Zhang, K. Li, and E.-W. Bai, "A new extension of newton algorithm degrees from the Southeast University,
for nonlinear system modelling using RBF neural networks," IEEE Nanjing, China in 1985 and 1988,
Transactions on Automatic Control, vol. 58, pp. 2929-2933, 2013. respectively, and the Ph.D. degree from the
[32] M. A. Shah and P. H. Meckl, "On-line control of a nonlinear system using
University of Newcastle, Newcastle,
radial basis function neural networks," in Proceedings of American
Control Conference, 1995, pp. 4265-4269 vol.6. Australia, in 2001, all in electrical
[33] Z. H. Man, H. R. Wu, and M. Palaniswami, "An adaptive tracking engineering.
controller using neural networks for a class of nonlinear systems," IEEE She is currently an Associate Professor in the
Transactions on Neural Networks, vol. 9, pp. 947-955, 1998.
Faculty of Science, Engineering and
[34] J. J. Slotine and W. Li, Applied Nonlinear Control: Prentice Hall, 1991.G.
O. Young, “Synthetic structure of industrial plastics,” in Plastics, 2nd Technology, Swinburne University of Technology, Australia.
ed., vol. 3, J. Peters, Ed. New York: McGraw-Hill, 1964, pp. 15– Her current research interests include robotics, mechatronics,
64. control and automation.
Xiaopeng Chen received the B.S. degree Jiong Jin (IEEE M’11) received the B.E.
in electrical and electronics engineering degree with First Class Honours in Computer
from Nottingham University, Malaysia Engineering from Nanyang Technological
Campus, in 2006 and M.S. degree with University, Singapore, in 2006, and Ph.D.
microelectronics engineering from degree from The University of Melbourne,
Swinburne University of Technology, Australia, in 2011. He is currently a Lecturer
Australia in 2011. He is currently a Ph.D in School of Software and Electrical
student at Swinburne University of Engineering, Faculty of Science,
Technology, Australia. His main research Engineering and Technology, Swinburne
interests include battery management system for electric University of Technology, Australia. Prior to it, he was a
vehicles, state estimation of battery systems such as battery Research Fellow in Department of Electrical and Electronic
modeling, battery state estimation and battery state of health. Engineering, The University of Melbourne from 2011 to 2013.
Weixiang Shen (S’00–M’02) received His research interests include network design and optimization,
Ph.D. degree from the University of Hong nonlinear systems and sliding mode control, networked
Kong, Hong Kong, China, in 2002. From robotics, wireless sensor networks and Internet of things,
2002 to 2003, He was a Lecturer in Ngee cyber-physical systems and applications in smart grids and
Ann Polytechnic, Singapore. From 2003 to smart cities.
2008, he was a Lecturer and then a Senior
Lecturer in School of Engineering, Monash Ajay Kapoor obtained his BTech and
University Sunway Campus, Malaysia. He MTech at IIT, BHU, Varanasi, and his PhD
then worked as a Research Fellow for one at Cambridge University, UK (1987).
year in the School of Electrical and Electronics Engineering, Subsequently, he lectured at IIT BHU
Nanyang Technological University, Singapore. Currently, he is before moving back to Cambridge in 1990.
an Associate Professor in Electrical Engineering in Faculty of He held academic positions at Leicester
Science, Engineering and Technology, Swinburne University University (1994 – 1995), Sheffield
of Technology, Australia. His research interests focus on University (1995 – 2004), Newcastle
electrical vehicles, renewable energy and power system. University (2004 – 2007) and moved to Swinburne University
of Technology, Australia, in 2007. Currently he is Dean, School
of Engineering, and Vice President Academic (Research
Engagement, India). Ajay has been a member of the Swinburne
Council. His research interests include electric vehicles,
mechanics of materials, wear and rolling contact fatigue. Ajay
has contributed to one book and has over 220 publications,

including several which won prestigious prizes. He has

obtained research funding of over $5M while in Australia and
of over £28M while in the UK. He sits on the editorial panels of
IMechE Journal of Rail and Rapid Transit, and several other
journals.