Applied Energy 207 (2017) 384–393

Contents lists available at ScienceDirect

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

A novel fractional order model based state-of-charge estimation method

for lithium-ion battery
Hao Mu a, Rui Xiong a,⇑, Hongfei Zheng a, Yuhua Chang b, Zeyu Chen c
a
Collaborative Innovation Center of Electric Vehicles in Beijing, School of Mechanical Engineering, Beijing Institute of Technology, Beijing 100081, China
b
Faculty of Automotive and Construction Machinery Engineering, Warsaw University of Technology, Narbutta 84, 02-524 Warsaw, Poland
c
School of Mechanical Engineering and Automation, Northeastern University, Shenyang 110819, China

h i g h l i g h t s

A novel fractional order impedance model for lithium-ion battery was proposed.

The new state-space equations of impedance model was presented.

Based on the GL definition and short-memory principle, the FOUKF algorithm was inferred.

The feasibility and robustness of the proposed method were verified under three various cycles.

The results show that the proposed method can promote the SoC estimation accuracy.

a r t i c l e i n f o a b s t r a c t

Article history: Accurate state of charge estimation of lithium-ion battery is directly related to the safe operation of elec-
Received 4 May 2017 tric vehicles and also an indispensable function of the battery management system. Four aspects of efforts
Received in revised form 5 July 2017 are made to improve the estimation accuracy. First, for overcoming the drawbacks of equivalent circuit
Accepted 8 July 2017
model and electrochemical model, the fractional order impedance model is built via electrochemical
Available online 22 July 2017
impedance spectroscopy data and the fractional element is used to describe the polarization effect in a
simple and meaningful way. Second, the discrete state-space equations of the impedance model are
Keywords:
inferred by Grünwald-Letnikov definition and parameters of the model including the order of the frac-
Lithium-ion battery
Electrochemical impedance spectroscopy
tional element are identified together by genetic algorithm (GA) and the experiment data of the dynamic
Battery model driving cycles. Third, the fractional order unscented Kalman filter technique is presented and the ‘short
State of charge memory’ technique is employed to improve the computation efficiency of fractional operator. Lastly,
Fractional order unscented Kalman filter experimental validation is implemented to verify the effectiveness of the proposed approach and results
show that the SoC estimation accuracy can be improved by the proposed model and estimation method.
The estimation error can be controlled within the range of 3%.
Ó 2017 Elsevier Ltd. All rights reserved.

1. Introduction
Due to the high energy and power density, long cycling and cal-
endar life-time, and low self-discharge rate, lithium-ion batteries
(LIBs) have become the predominant choice for electric vehicles
[1,2]. Despite those advantages, the practical application has
encountered many challenges, such as inconsistency of battery
pack, shortcut issues, and heat-runaway. To prevent these
problems from happening, reliable states estimation of LIBs, partic-
⇑ Corresponding author at: Department of Vehicle Engineering, School of
Mechanical Engineering, Beijing Institute of Technology, No. 5 South Zhongguancun
Street, Haidian District, Beijing 100081, China.
E-mail address: rxiong@bit.edu.cn (R. Xiong).
ularly, for the state of charge (SoC), seems rather crucial for battery
management systems (BMS). Nevertheless, the sophisticated
chemical reactions occur inside the batteries which give rise to
the strong nonlinearity and time-varying property of external
characteristic, which intensify the difficulty of SoC estimation [3].
In order to tackle this problem, some model-based estimation
methods are investigated [4,5]. Among several types of models,
the equivalent circuit model is regarded as the most appropriate
one for online estimation, and based on this type of model the
adaptive filter approaches and many observer techniques are pre-
sented to estimate the SoC of the battery effectively [6,7]. Hu et al.
[8] proposed the adaptive unscented Kalman filter (AUKF) to esti-
mate the SoC of Lithium Nickel-Manganese-Cobalt Oxide (NMC)
battery module based on the zero-state hysteresis battery model.

http://dx.doi.org/10.1016/j.apenergy.2017.07.003
0306-2619/Ó 2017 Elsevier Ltd. All rights reserved.
 H. Mu et al. / Applied Energy 207 (2017) 384–393 385

He et al. [9] also applied the AUKF method to estimate the SoC, but Kalman filter (UKF) method, the feasibility and superiority of pro-
the object was changed to Lithium Iron Phosphate (LFP) battery posed approaches are revealed.
and the improved combined model was selected which considered The remaining of this paper is organized as follows: In Section 2,
the influences of various temperatures and discharging rates. Hu some basic knowledge about fractional order theory, electrochem-
[10] and Xiong et al. [11] successively employed the multi-scale ical impedance spectroscopy and impedance model is introduced.
extended Kalman filter to jointly estimate the SoC and capacity Section 3 proposes the fractional order unscented Kalman filter
of the LIBs on enhanced self-correcting model and Thevenin model, technique to estimate the SoC of the LIB. The validation process
respectively and both of them obtained desirable estimation is implemented in Section 4. Finally, some conclusions are drawn
results. Wei et al. [12] addressed the negative effect of noise con- in Section 5.
tamination on model identification and SOC estimate and the
Frisch scheme based bias compensating recursive least squares 2. Fractional order impedance model
(FBCRLS) observer was proposed to co-estimate parameters of
the enhanced Thevenin model and SoC with high accuracy and In order to conquer the shortcoming of traditional ECMs — the
robustness. Besides, Zhang et al. [13] used the RC model which is lack of physical-chemical meaning and meanwhile guarantee the
quite different with the ones mentioned above and presented the optimal trade-off between model complexity and computation
H1 observer to estimate the SoC. However, the lack of physical- efficiency, the impedance model is investigated, based on frac-
chemical explanation for the microscopic movements in the bat- tional order and electrochemical impedance spectroscopy theory.
tery is the fatal drawback of this model, and meanwhile the elec-
trochemical model [14,15], which could illustrate the charge 2.1. Fractional calculus
transfer between two electrodes and reveal the electrochemical
mechanism, is too complicated to be used for online calculation. Fractional calculus is a generalization of the traditional calculus
In recent years, the fractional order impedance model (FOIM) and can be traced from the letters of Leibnitz to L’ Hospital in 1695.
inferred from electrochemical impedance spectroscopy (EIS) has It has been widely used in various fields of science, such as control
achieved a few of attraction to be investigated. Ref. [16] elaborated theory, electric circuits, viscoelasticity, and electrochemistry
the origin of the FOIM, and deduced the fractional order Kalman fil- [24,25]. The continuous integro-differential operator is defined as
ter by Grünwald-Letnikov (GL) definition in order to estimate the 8 a
SoC of the LIB. Ref. [17] specifically expounded the electrochemical < dta ;
> d
a>0
a
mechanism of the EIS variations along with the battery aging and a Dt ¼
>
1; a¼0 ð1Þ
:Rt a
set the Randle model as the baseline to demonstrate the improve-
a
ðds Þ ; a<0
ments of the presented electrochemistry-based model. Ref. [18]
presented a sliding mode observer (SMO) to estimate the SoC of where a is the order of fractional operator, varying within real num-
the battery and a self-tuning strategy for the SMO gain was pro- ber domain, a and t are the bounds of the operation. Three most pre-
posed to improve the convergence speed and reduce the high fre- vailing definitions of the fractional derivatives and integrals are the
quency chattering. Refs. [19,20] deduced the inner relationship Grünwald-Letnikov (GL) definition, the Riemann-Liouville (RL) and
between traditional ECMs and fractional order ECMs, and used the Caputo definitions. The GL definition is inferred from integer-
observer technique to estimate the SoC of the battery which order derivatives and integrals and expresses the unique discrete
acquired satisfactory estimation accuracy and robustness. Ref. [21] form of fractional calculus, which is shown as
 
presented a simple parameters identification method, namely 1X 1
j a
‘srivcf’, which can keep high fitting precision for electrochemical Dat f ðtÞ ¼ lim a ð1Þ f ðt  jhÞ ð2Þ
h!0 h j¼0 j
impendence spectroscopy data in frequency-domain. Ref. [22] ana-
lyzed the parameters sensitivity of fractional order models and  
where the symbol h is the sampling interval and
a is the bino-
divided these parameters into three categories to distinguish the j
impacts on the terminal voltages. Ref. [23] systematically com-
mial coefficients which can be calculated as
pared the fractional order model and ordinary integer order model (
  Cðaþ1Þ
on modelling technique and SoC estimation, and the advantages of a a!
j!ðajÞ!
¼ Cðkþ1Þ Cðajþ1Þ ; j > 0
fractional order model were prominent. ¼ ð3Þ
j 1; j¼0
Although some work has been done to expand the application
of fractional order theory for lithium-ion battery modelling and where C(
) is the gamma function. Commonly, in order to imple-
states estimation, the increased nonlinearity and model complex- ment the fractional derivatives and integrals, it is unavoidable to
 
ity due to the fractional calculus cannot be neglected. Therefore,
compute the coefficients ð1Þ j
a . One recursive approach has
simplified modelling methods and the appropriate estimation j
techniques, particularly for this kind of strong nonlinear problem, been investigated for simplifying the computation process, shown
should be focused. This paper is involved to the SoC estimation below [24]
issue of the lithium-ion battery and some contributions are made  
aþ1
as follows: (i) Based on the EIS data and fractional order theory, x0ðaÞ ¼ 1; xðj aÞ ¼ 1  aÞ
xðj1 ; j ¼ 1; 2; 3    ð4Þ
j
the fractional order impedance model is established; (ii) The order
 
of the fractional element is identified together with other parame-
where xj
ðaÞ
denotes the ð1Þ j
a .
ters in the model by GA method using dynamic driving cycles data j
rather than being identified independently in Refs. [16,19], which
could make parameters reflect more comprehensive characteristics 2.2. Electrochemical impedance spectroscopy
of the battery; (iii) Compared with Refs. [16,23], novel state-space
functions from the model are inferred by the GL definition and the Electrochemical impedance spectroscopy (EIS) is a powerful
fractional order unscented Kalman filter (FOUKF) approach is technique for separating the electrochemical reactions and track-
investigated to deal with this nonlinear issue; (iv) Through com- ing the variation of the performance under different state of health
parison with the results of traditional ECM calculated by unscented (SoH) of the battery in a nondestructive manner [26,27]. Previous
386 H. Mu et al. / Applied Energy 207 (2017) 384–393
High Middle Low
Fig. 1. Typical diagram of impedance spectra of the LIB.
studies have demonstrated that the measured impedance spectra
of the LIB consists of three main parts: (1) an inductive tail at high
frequency, (2) a semicircle at middle frequency and (3) a straight
line at low frequency, which can be seen in Fig. 1. Each part owns
its electrochemistry-based explanation. For example, the semicir-
cle is related to charge-transfer reaction and double layer effect,
and the straight line at low frequency represents the diffusion
dynamics in two electrodes. Since the EIS unveils the inner
mechanism of the electrochemical reaction in isolated frequency-
domain, and it varies with some factors, like SoC, SoH and temper-
ature, the battery model built through it will contain comprehensive
and fruitful information.
2.3. Fractional order impedance model
Combining the fractional order theory with the EIS theory, the
FOIM is merged. As mentioned above, the FOIM is superior to the
ECM because of its reflections to electrochemical mechanism,
and meanwhile it conquers the drawback of the electrochemical
model with respect to the complexity. From the architecture, it is
similar to the ECM, which is shown in Fig. 2. It is constituted with
the open circuit voltage source, the ohmic resistance. These two
elements can be found in conventional ECMs as well. Besides, the
RC-like elements here named ‘ZARC’ and ‘Warburg’ are placed in
series, which have totally different meanings and mathematical
formulas with normal paralleled RC part and capacitor.
ZARC element including a resistor and a constant phase element
(CPE) is to describe the charge-transfer between electrode/elec-
trolyte interface and double layer effect at the middle frequency
region in Fig. 1. Warburg impedance represented by a simple CPE
corresponds to simulate the diffusion process in solid phases at
low frequency band. Due to the double layer effect with dispersion
Warburg
element
ZARC
element
Fig. 2. Architecture of the fractional order impedance model.
phenomenon between solid electrodes and the electrolyte, it is
inconsistent with the ideal capacitor in frequency response [28].
Therefore, the constant phase element (CPE) is designed to mimic
this deviation and its mathematical function can be expressed as
Z CPE ðsÞ ¼ 1=ðC CPE sa Þ ð5Þ
where ZCPE is the impedance of CPE, CCPE is the capacitance-like
index, s is the Laplace operator, a is the fractional order to describe
the dispersion effect (a 2 R+, 0  a  1). Note that when a equals to
1, the CPE behaves like the ideal capacitor and when a is 0, the CPE
degenerates to the resistor.
FOIM is facilitated to precisely depict the sophisticated electro-
chemical dynamics by adding more fractional order elements. In
this paper, considering the practical applications of LIBs for electric
vehicles, the EIS data at over-high frequency and over-low fre-
quency band are ignored. Hence, the final impedance model is sim-
plified to the Thevenin-like, i.e., the Warburg impedance is
excluded in Fig. 2.
3. Parameters identification and SoC estimation methods
Compared with the same kind of the ECM, the impedance
model owns higher freedom degree to describe the nonlinear
dynamics since there exists the fractional order and simultane-
ously the model complexity is not much higher than that of the
ECM. However, the additional nonlinearity of the model due to
fractional operator asks for particular approaches to figure out.
Therefore, in this section the UKF algorithm is evolved to the
FOUKF for dealing with the fractional order system.
3.1. Model parameters identification
Before estimating the implicit states of the battery, parameters
of the battery model should be obtained ahead [29,30]. Several
methods have been proposed to identify parameters in the impe-
dance model, such as the least squares (LS) method, hybrid multi-
ple particle swarm optimization (HMPSO) method and genetic
algorithm (GA). But before applying all these approaches, there is
a prerequisite that the fractional order of CPE has been given.
Ref. [16] used the LS method to identify the remaining parameters
of impedance model other than the order. Ref. [31] tried 0.1–0.9
with 0.1 interval of the order to evaluate the root mean square
error (RMSE) of terminal voltage and eventually chose 0.7 as the
optimal order.
In this paper, the fractional order is identified concurrently with
other parameters in the model and GA method is applied. Accord-
ing to the model presented in Section 2.3 and Kirchhoff’s law, the
terminal voltage of the battery can be expressed as
U t ðtÞ ¼ U oc ðtÞ  IL ðtÞRo  U CPE ðtÞ ð6Þ
where Ut denotes the terminal voltage of the battery which can be
measured on real time, Uoc is the open circuit voltage which is
monotonic with SoC and its relation can be derived through tests,
UCPE is the voltage drop of ZARC element, Ro denotes the ohmic
resistance and IL is the charging/discharging current through the
battery which can be measured as well. In Eq. (6), terms of Uoc
and UCPE should be unequivocal further. As mentioned above, the
Uoc can be fitted by the following polynomial form
U oc ðtÞ ¼ K 0 þ K 1 z þ K 2 z2 þ K 3 z3 þ    ð7Þ
where Ki denotes the coefficients of polynomial equation and z
denotes the SoC of the battery (the same below). Moreover, incor-
porating with the Eq. (5), the UCPE can be formulated as
U CPE ðtÞ
C CPE Da U CPE ðtÞ ¼ IL ðtÞ  ð8Þ
Rct
 H. Mu et al. / Applied Energy 207 (2017) 384–393 387

Using the GL definition of fractional order derivatives, Eq. (8) can be where k is the sampling time. Due to the fractional derivative, accu-
transformed into mulative terms of UCPE incur the nonlinearity of the system. As a
  result, the matrix form of Eq. (14) should be expressed like
1X N
i a IL ðtÞ U CPE ðtÞ
Da U CPE ðtÞ ¼ a ð1Þ U CPE ðt  ihÞ ¼  ð9Þ 
h i¼0 i C CPE Rct C CPE xkþ1 ¼ Fðxk ; ukþ1 Þ þ xk
ð15Þ
yk ¼ Gðxk ; uk Þ þ v k
From Eq. (9), the discrete form of UCPE can be obtained as follows
  !
where xk+1 = [UCPE,k+1, zk+1]T denotes the system states, yk = Ut,k
Rct C CPE IL ðtÞ XN
i a
U CPE ðtÞ ¼  ð1Þ U CPE ðt  ihÞ ð10Þ denotes the terminal voltage, F(
) and G(
) are the nonlinear func-
Rct C CPE þ 1 C CPE i¼1 i
tions of the system.
where h is the sampling interval, N = (t  0)/h is the data length of Although the extended Kalman filter (EKF) is commonly feasible
historical UCPE and i is the counting number. Thus, setting the choice to estimate implicit states for nonlinear systems, it still
parameters vector to be identified h = [Ro, Rct, CCPE, a]T, if the SoC might be invalid if the nonlinearity of the systems is grave, because
of the battery is certain at every second, the terminal voltage can it depends on linearization to propagate the mean and covariance
be predicted by Eqs. (6) and (10). Given the appropriate cost func- of the states. On the contrary, UKF method using classical statisti-
tion, the optimal h will be searched via optimization method. cal method based on the unscented transform (UT) linearizes solu-
Considering the evolution algorithm is more suitable for the tions for nonlinear system, which is more beneficial to handle the
nonlinear and multiple parametric problem, so the GA method is nonlinear problem [32,33]. Comparing with the traditional UKF,
selected. And meanwhile model accuracy strongly relates to the there is no essential difference between UKF and FOUKF except
prediction accuracy of terminal voltage, hence fitness function is the prior estimation process of states. D. Sierociuk [34] did many
to minimize the RMSE of it, which is shown below investigations on the application of Kalman filter and its derivative
algorithms for fractional order systems and gained many cogent
8v ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi9
<u
results. In this paper, FOUKF approach is presented and the corre-
u1 X Nm =
min t ðU t ðkÞ  U ^ t ðk; ^hÞÞ2 ð11Þ sponding algorithm procedures are given in Table 1.
: Nm k¼1 ; Note that the most remarkable difference between UKF and
FOUKF is that all the sigma points of x ^sig
i;k at every sampling time
where Nm is the data length of measured Ut for identification, terms
Table 1
with ‘^’ on them manifest the estimated values (Similarly here- Algorithm procedures of fractional order unscented Kalman filter.
inafter). Note that on the right of Eq. (10), it refers to the pervious 
data of UCPE, and along with the time elapsing, the number of accu- xkþ1 ¼ Fðxk ; ukþ1 Þ þ xk
The nonlinear state-space system: , Q and R is
yk ¼ Gðxk ; uk Þ þ v k
mulative terms will be tremendous, which probably gives rise to the variance of process and measurement noise, respectively.
slow down the computing efficiency of BMS. Hence, the ‘short-
memory’ principle [24] should be introduced to just take into account Step 1 Initialization: For k = 0, set initial state and state error
the ‘recent past’ data of UCPE in case the ‘data blast’. This technique covariance
T
^
xþ ^þ ^þ þ ^þ ^þ T
seems to be the same as adding the sliding window with the fixed 0 ¼ ½U CPE;0 ; z0  and P 0 ¼ E½ðx0  x0 Þðx0  x0 Þ 

data length, and after using it the Eq. (10) can be rewritten as Step 2 Computation: For k = 1, 2. . .. . ., compute:
! j Prior estimation update
 
Rct C CPE IL ðtÞ XL
a (a) Creating sigma points:
8 sig
U CPE ðtÞ ¼  ð1Þi U CPE ðt  ihÞ ð12Þ > ^
x ¼^ xþ þ~ xi i ¼ 1; 2; . . . ; 2n (16)
Rct C CPE þ 1 C CPE i¼1 i >
> i;k1qffiffiffiffiffiffiffiffiffiffiffiffi
>
<
k1
ffiT
~i ¼
x nP þ
k1 i ¼ 1; 2; . . . ; n , where n is the
>
> i
qffiffiffiffiffiffiffiffiffiffiffiffiffiT
where L is the memory length. >
>
:~xnþi ¼  þ
nP k1 i ¼ 1; 2; . . . ; n
i
quantity of state variables.
3.2. SoC estimation based on fractional order model
(b) Calculating the weights:
8 m
< w0 ¼ k=ðn þ kÞ (17)
Similarly with the traditional model-based estimation method, wc0 ¼ k=ðn þ kÞ þ ð1  A2 þ IÞ
: m
the state-space equations of fractional order system and its obser- wi ¼ wci ¼ 1=ð2ðn þ kÞÞ
vation function can be built as follows where k ¼ A2 ðn þ jÞ  n is a coefficient describing width of
point expansion during the
( transformation, j is a scaling coefficient, and I is a coefficient
Da U CPE ¼ C ICPE
L
 RctUCCPECPE þ x1 corresponding with type of noise.

D1 z ¼  gCIaL þ x2 ð13Þ (c) Updating the prior states estimation:

P
2n (18)
^
xsig ¼ Fð^
xsig ; uk Þ þ xk and then ^ ^sig
U t ¼ U oc  IL Ro  U CPE þ v 1 ik i;k1
x
k ¼ wm
i xik
i¼0
(d) Updating the states error covariance
P   T
where g is the columbic efficiency (for LIB, it approximates to 1), Ca (19)
2n
P
x;k ¼ wci ^xsig
ik
^
x
k
^xsig
ik
^
x
k þQ
is the maximum available capacity of the battery under certain tem- i¼0
j Measurement update
perature, x = [x1, x2]T is the process noise and v1 is the measure- (a) Calculating the output
ment noise. In order to employ the filter methods, the discrete P
2n (20)
^sig
y ¼ Gð^
xsig ; uk Þ þ v k and then y
^k ¼ wm ^sig
i yik
form of Eq. (13) should be given. Combining Eq. (12), the discrete ik ik
i¼0
form of Eq. (13) can be inferred as (b) Calculating the covariance and Kalman gain
8
8   ! >
>
>
P2n
^sig ^ ^sig ^
T (21)
X < P y;k ¼ wci ðy yk Þðyik  yk Þ þ R
> i a
L
>
< U CPE;kþ1 ¼ R RctC CCPEþ1 IL;kþ1
 ð1Þ U CPE;kþ1j þ x1;k i¼0
ik
and Kgs = Pxy, k
P 1
C CPE >
> P
2n T y;k
ct CPE
j¼1
j >
: P xy;k ¼ wci ð^
xsig ^x ^sig ^
k Þðyik  yk Þ
>
ik
>
: gI
i¼0

zkþ1 ¼ zk  L;kþ1
Ca
þ x2;k j Posterior estimation update
Updating the states and covariance
U t;k ¼ U oc;k  IL;k Ro  U CPE;k þ v 1;k xþ
^k
¼^
x ^ þ 
k þ K gs ðyk  yk Þ and P k ¼ P x;k  K gs P y;k K gs
(22)

ð14Þ Step 3 Jump to step 2 and k = k + 1

388 H. Mu et al. / Applied Energy 207 (2017) 384–393

should be recorded within certain memory length, because the Table 2

function F(
) involves the historical information of UCPE. Besides, Parameters of the selected battery.

the memory length not only impacts on the estimation accuracy Nominal capacity 25 A h
since it determines the model precision to some extent, but it will Nominal voltage 3.65 V
also influence the computation efficiency. Reasonable value should Charging/discharging cut-off voltage 4.2 V/2.5 V
be chose to balance the estimation accuracy and calculation speed. Operation temperature 10 °C to 55 °C (charging)
The flowchart of the algorithm is shown in Fig. 3. 20 °C to 55 °C (discharging)
Weight 0.52 kg
Genetic algorithm is a global optimization method and its cross-
over, mutation and selection mechanisms guarantee the random-
ness and selective preference. In this paper, the dynamic driving
cycles data is employed to train the impedance model in order to
battery test system-Arbin BT-5HC, a thermal chamber, an electro-
make the parameters reflect more comprehensive characteristics
chemical workstation-Zahner IM6, a host computer, and several
of the battery in wide SoC working range. Once the identification
test cells. The battery test device is responsible for charging/dis-
process is done, the parameters vector which represents the bat-
charging the battery as plans and transmitting the data to the host
tery will be transferred in the states estimation module to accom-
computer via TCP/IP protocol. The programmable thermal chamber
plish the SoC estimation step.
provides stable ambient temperature in case the fluctuation of the
temperature leads to the variation of the battery performance. The
4. Validation and discussion electrochemical workstation is used for measuring the EIS of the
battery and communicating with the host computer as well.
Based on the test bench, several experiments are carried out to
verify the effectiveness of the presented algorithms. Through the 4.2. OCV fitting and model parameters identification
comparison of SoC estimation results between Thevenin model
and FOIM mentioned in Section 2.3 for three different cycles, the OCV is an important index to indicate the stationary character-
advantages of FOIM will be revealed. istic of the battery, and its intimate and unique relation with SoC
can be employed for calibrating the SoC. Normally, this relation-
ship is acquired via OCV test under fixed SoC intervals and the
4.1. Test bench
mathematical function is derived by fitting test data with the poly-
nomial equation. In this paper, the eighth-order polynomial equa-
In this paper, the NMC type lithium-ion battery with nominal
tion is selected and the fitting function and results are shown in
capacity of 25 A h is selected and its parameters are listed in
Eq. (23) and Fig. 5, respectively.
Table 2. The test bench is established to execute the pre-set exper-
iments under stable ambient environment, such as capacity, OCV,
dynamic driving cycles and EIS, meanwhile logging the informa- U oc ðtÞ ¼ K 0 þ K 1 z þ K 2 z2 þ K 3 z3 þ K 4 z4
ð23Þ
tion like current, voltage and temperature of the battery. The con- þK 5 z5 þ K 6 z6 þ K 7 z7 þ K 8 z8
figuration of this test bench is drawn in Fig. 4, which consists of a

Fractional oder Paramters

States Estimation
Impedance model identification

Selection
Identified parameters
Dynamic driving cycles data Ut, IL, Uoc

Mutation States initiation x0, P0

Cells

Sigma points calculation

Crossover
Eq. (16)
N
Y Prior estimation
Evaluation criteria Eq. (18) and (19)
k=k+1

EIS data Measurement update

Population initiation Eq. (20) and (21)

Posterior estimation
Cost function Eq. (22)

N
End?

Model parameters Y

=[Ro Rct CCPE ]T Estimated states

Fig. 3. Flowchart of the entire algorithm.

 H. Mu et al. / Applied Energy 207 (2017) 384–393 389

Zahner IM6
electrochemical workstation

AC Power

Cells

Host computer

Power cables Thermal chamber

Data
transmission
Arbin BT-5HC

Fig. 4. The configuration of the test bench.

In this paper, parameters of models are identified by data of

Unban Dynamometer Driving Schedule (UDDS) test and the results
are listed in Table 3.

4.3. SoC estimation under DST cycles

To testify the effectiveness of the algorithms, the dynamic stress

test (DST) is executed on the LIB and the corresponding current and
SoC profiles are shown in Fig. 6.
Using FOUKF and UKF, the estimated results and errors with
respect to SoC and terminal voltages are plotted in Fig. 7. To
demonstrate the robustness of the algorithm to initial SoC error,
the initial SoC is set 50% offset to the true value. As for the FOUKF
method, the memory length is set to 500. In Fig. 7(a), it can be seen
Fig. 5. OCV curve and fitting results.
that the estimated SoCs via two approaches are convergent to true
values with similar speed. The maximum estimated errors of SoCs
Table 3 of two methods are located in the same region around 15,000 s
Parameters list of Thevenin model and FOIM. from Fig. 7(b). Fig. 7(c) illustrates that the predicted terminal volt-
Ro (mX) Rp/Rct (mX) Cp/CCPE (F) a ages by two methods can well track the measured values and at the
same time errors are controlled within 50 mV even if the SoC
Thevenin model 2.5 3.3 2.70e4 –
FOIM 2.4 4.6 1.05e4 0.75 decreases below 10% which can be found in Fig. 7(d). The relative

(a) (b)

Fig. 6. Current and SoC profiles of DST cycles, (a) current profiles and (b) SoC profiles.
390 H. Mu et al. / Applied Energy 207 (2017) 384–393

Fig. 7. Comparison of estimated SoC and predicted terminal voltages under DST cycles, (a) SoC estimation results, (b) SoC error profiles, (c) terminal voltages profiles and (d)
errors profiles of terminal voltages.

statistical information of SoC and terminal voltages are more per- 4.4. SoC estimation under UDDS cycles
suasive which are revealed in Table 4.
From Table 4, the accuracy of SoC estimation is obviously ele- The algorithms are verified under UDDS test as well which is an
vated by the FOUKF based on the FOIM since the max absolute another commonly used driving mode to simulate the practical
error (MAE) is reduced from 3.02% to 2.56% and simultaneously operations of batteries to prove its universality and the current
the stability index of the estimated SoCs, namely RMSE, is further and SoC profiles of this test are displayed in Fig. 8.
decreased from 1.06% to 0.94%, indicating that the estimation Similarly, the memory length is set to 500 for FOUKF approach
results are more reliable. Besides, the last column of Table 4 lists in this case. The estimation results revealed in Fig. 9(a)–(d).
the RMSE of predicted terminal voltages of two models and the Fig. 9(a) shows the robustness and astringency of the algorithms
depressed result also manifests the enhancement of the accuracy both of the FOUKF and traditional UKF, because the estimated SoCs
of the SoC estimation by FOUKF on impedance model. are convergent to true values even if with 50% initial errors. In
Fig. 9(b), the errors of estimated SoCs can be guaranteed within
Table 4 around 3% and the maximum estimation errors emerge in the
Statistical data of estimated SoC and terminal voltages under DST cycles. low SoC region about 18,470 s. From Fig. 9(c) and (d), the predicted
Statistical indexes State of charge Terminal voltage terminal voltages via these two methods can precisely describe the
MAE (%) RMSE (%) RMSE (mV)
measured values in broad SoC operation zone except that when the
SoC deceases below 8% since the nonlinearity of the battery is
Thevenin model 3.02 1.06 3.31
Impedance model 2.56 0.94 2.40
beyond the describing capability of the model incurring large
model errors.

(a) (b)

Fig. 8. Current and SoC profiles of UDDS cycles, (a) current profiles and (b) SoC profiles.
 H. Mu et al. / Applied Energy 207 (2017) 384–393 391

Fig. 9. Comparison of estimated SoC and predicted terminal voltages under UDDS cycles, (a) SoC estimation results, (b) SoC error profiles, (c) terminal voltages profiles and (d)
errors profiles of terminal voltages.

Table 5 4.5. SoC estimation under the charging process of OCV test
Statistical data of estimated SoC and terminal voltages under UDDS cycles.

Statistical indexes State of charge Terminal voltage Eventually, to ensure the comprehensiveness of the validation
MAE (%) RMSE (%) RMSE (mV)
process, i.e. guarantee the feasibility of this algorithm in simple
pulse cycles, data of the charging procedure of OCV test (DST) is
Thevenin model 3.01 0.95 4.61
Impedance model 2.89 0.92 3.23
employed and the concrete information of its current and SoC pro-
files are shown in Fig. 10. Due to the nominal capacity of the bat-
tery being 25 A h, the amplitude of pulse current is 1/3 C (7.5 A)
and during each 10% SoC interval, the battery is rest for 2 h.
The corresponding estimation results and their errors about SoC
From Table 5, it is much easier to understand the superiority of and terminal voltages are plotted in Fig. 11. For the identical aim,
the FOUKF by the statistical data. Similarly as the results of the DST the initial SoC is set under certain offset as well (Here the offset
test, no matter the indexes of the estimated SoCs or that of termi- is set to 20% for just deviation with former cases and 50% is also
nal voltages, all three terms are reduced comparing with the rela- demonstrated to be effective). The memory length is chose to
tive results obtained by Thevenin model and UKF technique. The 500. Fig. 11(a) shows that estimated SoCs via two approaches could
remarkable changes happen on the MAE of the estimated SoC quickly approach to true values and the maximum estimated
and RMSE of the predicted terminal voltages, varying from 3.01% errors of SoCs of two methods emerged in middle region of the
to 2.89% and from 4.61 mV to 3.23 mV, respectively, which mani- cycles which can be seen from Fig. 11(b). The inferred reason
fest the promotion of the accuracy of the SoC estimation why the estimated SoC errors fluctuate above the zero error line
comprehensively. is the applied OCV-SoC relation is the average results of charging

(a) (b)

Fig. 10. Current and SoC profiles of half OCV test, (a) current profiles and (b) SoC profiles.
392 H. Mu et al. / Applied Energy 207 (2017) 384–393
(a)
(c)
Fig. 11. Comparison of estimated SoC and predicted terminal voltages under the charging procedure of the OCV test, (a) SoC estimation results, (b) SoC error profiles, (c)
terminal voltages profiles and (d) errors profiles of terminal voltages.
Table 6
Statistical data of estimated SoC and terminal voltages under half OCV test.
Statistical indexes State of charge Terminal voltage
MAE (%) RMSE (%) RMSE (mV)
Thevenin model 3.44 1.76 1.23
Impedance model 2.98 1.61 1.06
and discharging processes. Therefore, for single charging process, it
has errors in the same direction and that incurs the deviations of
SoC estimation. Fig. 11(c) manifests that the desirable predicted
terminal voltages by two methods and their errors wave within
20 mV found in Fig. 11(d). Table 6 reveals the statistical informa-
tion of SoC and terminal voltages to prove the advantage of FOUKF
method.
As in the former cases, the FOUKF method based on the FOIM
improves the accuracy of SoC estimation because the MAE is
remarkably decreased from 3.44% to 2.98% and simultaneously
the RMSE is changed from 1.76% to 1.61%, indicating that the over-
all performance of estimation results is promoted. This point can
be also proved by the variation of the RMSE with respect to termi-
nal voltage in the last column of Table 6.
5. Conclusions
The state of charge of the lithium-ion battery is a significant
indicator for energy storage systems. It provides the essential basis
of the energy deployment, the load balancing and the fault progno-
sis for sophisticated energy systems. To overcome the drawbacks
of the equivalent circuit model and electrochemical model, the
impedance model based on the fractional order theory is investi-
gated and its parameters are identified by the evolutionary opti-
mization method. Concerning the enhanced nonlinearity of the
impedance model due to the fractional order of ZARC element,
the FOUKF approach is presented which is more suitable to deal
with the strong nonlinear problem. The simulation results under
different cycles sufficiently shows the effectiveness and superiority
of the proposed model and algorithm.
(b)
(d)
In the future, the modelling procedures of impedance model
will be further refined, and we will pay more attention to the vari-
ations of the fractional orders of nonlinear elements, such as ZARC
and Warburg, with SoC, temperature and aging state. The online
parameters identification method considering the variable orders
will be emphatically studied.
Acknowledgements
This work was supported in part by the National Natural
Science Foundation of China (Grant No. 51507012), Beijing
Municipal Science and Technology Project (Grant No.
Z171100000917013), Sino-polish Collaborative research in
e-mobility public transportation (Grant No. 2015DFG81930), the
Application and Demonstration of Innovative Methods in New
Energy Vehicle Industry (Grant No. 2015IM030100). The systemic
experiments of the lithium-ion batteries were performed at the
Advanced Energy Storage and Application (AESA) Group, Beijing
Institute of Technology.
References
[1] Li Z, Huang J, Liaw BY, Zhang J. On state-of-charge determination for lithium-
ion battery. J Power Sources 2017;348:281–301.
[2] Xiong R, Tian JP, Mu H, Wang C. A systematic model-based degradation
behavior recognition and health monitor method of lithium-ion batteries. Appl
Energy 2017;207:372–83.
[3] Tian Yong, Li Dong, Tian J, Xia B. State of charge estimation of lithium-ion
batteries using an optimal adaptive gain nonlinear observer. Electrochim Acta
2017;225:225–34.
[4] Xiong R, Yu QQ, Wang LY, Lin C. A novel method to obtain the open circuit
voltage for the state of charge of lithium ion batteries in electric vehicles by
using H infinity filter. Appl Energy 2017;207:346–53.
[5] Dai H, Xu T, Zhu L, Wei X. Z Sun. Adaptive model parameter identification for
large capacity Li-ion batteries on separated time scales, Appl Energy
2016;184:119–31.
[6] Sun F, Xiong R, He H. A systematic state-of-charge estimation framework for
multi-cell battery pack in electric vehicles using bias correction technique.
Appl Energy 2016;162:1399–409.
[7] Kalawoun J, Biletska K, Suard F, Montaru M. From a novel classification of the
battery state of charge estimators toward a conception of an ideal one. J Power
Sources 2015;279:694–706.
 H. Mu et al. / Applied Energy 207 (2017) 384–393
[8] Sun F, Hu X, Zou Y, Li S. Adaptive unscented Kalman filtering for state of charge
estimation of a lithium-ion battery for electric vehicles. Energy
2011;36:3531–40.
[9] He Z, Gao M, Wang C, Wang L, Liu Y. Adaptive state of charge estimation for Li-
ion batteries based on an unscented Kalman filter with an enhanced battery
model. Energies 2013;6:4134–51.
[10] Chao Hu, Youn Byeng D, Chung Jaesik. A multiscale framework with extended
Kalman filter for lithium-ion battery SOC and capacity estimation. Appl Energy
2012;92:694–704.
[11] Xiong R, Sun FC, Chen Z, He HW. A data-driven multi-scale extended
Kalman filtering based parameter and state estimation approach
of lithium-ion polymer battery in electric vehicles. Appl Energy
2014;113:463–76.
[12] Wei Z, Meng S, Xiong B, Ji D, Tseng KJ. Enhanced online model identification
and state of charge estimation for lithium-ion battery with a FBCRLS based
observer. Appl Energy 2016;181:332–41.
[13] Zhang F, Liu G, Fang L, Wang H. Estimation of battery state of charge with H1
observer: applied to a robot for inspecting power transmission lines. IEEE
Trans Ind Electron 2012;5:1086–95.
[14] Sabatier J et al. Lithium-ion batteries modeling: a simple fractional
differentiation based model and its associated parameters estimation
method. Sign Process 2015;107:290–301.
[15] Tulsyan A. Yiting Tsai, R. Bhushan Gopaluni, Richard D. Braatz, State-of-charge
estimation in lithium-ion batteries: a particle filter approach. J Power Sources
2016;331:208–23.
[16] Xu J, Mi C, Cao BG, Cao J. A new method to estimate the state of charge of
lithium-ion batteries based on the battery impedance model. J Power Sources
2013;233:277–84.
[17] Li S, Wang B, Peng H, Hu X. An electrochemistry-based impedance model for
lithium-ion batteries. J Power Sources 2014;258:9–18.
[18] Zhong F, Li H, Zhong S, Zhong Q, Yin C. An SOC estimation approach based on
adaptive sliding mode observer and fractional order equivalent circuit model
for lithium-ion batteries. Commun Nonlinear Sci Numer Simulat
2015;24:127–44.
[19] Wang BJ, Li SE, Peng Huei, Liu ZY. Fractional-order modeling and parameter
identification for lithium-ion batteries. J Power Sources 2015;293:151–61.
[20] Wang B, Liu Z, Li SE, Moura SJ, Peng H. State-of-charge estimation for lithium-
ion batteries based on a nonlinear fractional model. IEEE Trans Contr Syst
Technol 2017;25(1):3–11.
393
[21] Alavi SMM, Birkl CR, Howey DA. Time-domain fitting of battery
electrochemical impedance models. J Power Sources 2015;288:345–52.
[22] Zhou D, Zhang K, Ravey A, Gao F, Miraoui A. Parameter sensitivity analysis for
fractional order modeling of lithium-ion batteries. Energies 2016;9(23).
[23] Xiao R, Shen J, Li X, Yan W, Pan E, Chen Z. Comparisons of modelling and state
of charge estimation for lithium-ion battery based on fractional order and
integral order methods. Energies 2016;9(3):184.
[24] Podlubny I. Fractional differential equations: an introduction to fractional
derivatives, fractional differential equations, to methods of their solution and
some of their applications. Academic; 1998.
[25] Petráš I. Fractional-order nonlinear systems: modeling. Analysis and
Simulation: Springer; 2011.
[26] Galeotti M, Cinà L, Giammanco C, Cordiner S, Carlo AD. Performance analysis
and SOH (state of health) evaluation of lithium polymer batteries through
electrochemical impedance spectroscopy. Energy 2015;89:678–86.
[27] Gordon IAJ, Grugeon S, Takenouti H, Tribollet B, Armand M, Gordon IAJ, et al.
Electrochemical impedance spectroscopy response study of a commercial
graphite-based negative electrode for Li-ion batteries as function of the cell
state of charge and aging. Electrochim Acta 2017;223:63–73.
[28] Barsoukov E, Macdonald JR. Impedance spectroscopy: theory, experiment, and
applications. Wiley; 2005.
[29] Plett GL. Extended Kalman filtering for battery management systems of LiPB-
based HEV battery packs Part 2. Modeling and identification. J Power Sources
2004;134:262–76.
[30] Fleischer C, Waag W, Heyn HM, Sauer DU. On-line adaptive battery impedance
parameter and state estimation considering physical principles in reduced
order equivalent circuit battery models: Part 2. Parameter and state
estimation. J Power Sources 2014;260:276–91.
[31] Zou Y, Li SE, Shao B, Wang B. State-space model with non-integer order
derivatives for lithium-ion battery. Appl Energy 2016;161:330–6.
[32] Chen C, Xiong R, Shen W. A lithium-ion battery-in-the-loop approach to test
and validate multi-scale dual H infinity filters for state of charge and capacity
estimation. IEEE T Power Electr 2017. http://dx.doi.org/10.1109/
TPEL.2017.2670081.
[33] Zhang W, Shi W, Ma Z. Adaptive unscented Kalman filter based state of energy
and power capability estimation approach for lithium-ion battery. J Power
Sources 2015;289:50–62.
[34] Sierociuk D. Uzycie ulamkowego filtru kalmana do estymacji parametrow
ukladu ulamkowego rzedu. Poland: Warsaw; 2005.