Applied Energy 212 (2018) 1589–1600

Contents lists available at ScienceDirect

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

Online state-of-health estimation for lithium-ion batteries using constant- T

voltage charging current analysis
⁎
Jufeng Yanga,b, Bing Xiab,c, Wenxin Huanga, Yuhong Fub, Chris Mib,
a
Department of Electrical Engineering, Nanjing University of Aeronautics and Astronautics, 29 Jiangjun Street, Nanjing, Jiangsu 211106, China
b
Department of Electrical and Computer Engineering, San Diego State University, 5500 Campanile Drive, San Diego, CA 92182, USA
c
Department of Electrical and Computer Engineering, University of California San Diego, 9500 Gilman Dr., La Jolla, CA 92093, USA

H I G H L I G H T S

• Derived the expression of CV charging current based on ECM.

• Introduced time constant of CV charging current to estimate battery SoH.
• Established a quantitative correlation between current time constant and battery SoH.
• Discovered that current time constant is a logarithmic function of fitted data size.
• Employed uncompleted CV charging data to estimate battery SoH.

A R T I C L E I N F O A B S T R A C T

Keywords: Battery state-of-health (SoH) estimation is a critical function in a well-designed battery management system
Lithium-ion battery (BMS). In this paper, the battery SoH is detected based on the dynamic characteristic of the charging current
State-of-health (SoH) during the constant-voltage (CV) period. Firstly, according to the preliminary analysis of the battery test data,
Constant-current constant-voltage (CCCV) the time constant of CV charging current is proved to be a robust characteristic parameter related to the battery
charge
aging. Secondly, the detailed expression of the current time constant is derived based on the first order
Equivalent circuit model (ECM)
equivalent circuit model (ECM). Thirdly, the quantitative correlation between the normalized battery capacity
Current time constant
and the current time constant is established to indicate the battery SoH. Specifically, for the uncompleted CV
charging process, the logarithmic function-based current time constant prediction model and the reference
correlation curve are established to identify the battery capacity fading. At last, experimental results showed that
regardless of the adopted data size, the correlation identified from one battery can be used to indicate the SoH of
other three batteries within 2.5% error bound except a few outliers.

1. Introduction conditions (such as temperature and load profile) make it a more

challenging task [9–11].
Due to the high energy and power density, lithium-ion batteries
have proved to be a promising candidate as the energy storage system 1.1. Review of the literature
in electric vehicles (EVs) and consumer electronics [1–3]. To ensure the
safety and reliability of battery systems, the basic battery states, i.e., To describe the deterioration of the power and the energy capability
state-of-charge (SoC) [4,5], state-of-power (SoP) [6], and state-of- of the battery, the quantitative definition of the SoH is generally based
health (SoH) [7,8], should be monitored continuously by the battery on the impedance growth and/or the capacity loss. Hence, the SoH
management system (BMS). Hence, the accurate, robust and practical estimation can be converted to the identification of the present im-
battery state estimation is one of the key functions of a BMS. However, pedance and/or capacity.
batteries are sophisticated electrochemical devices with various non- Abundant research work has been conducted to seek the accurate
linear characteristics. It is difficult to estimate the above states with impedance identification, ranging from offline tests, such as active/
high accuracy and strong robustness. Especially for the SoH estimation, passive impedance spectroscopy [12,13], to online estimation algo-
the complicated aging mechanism and the uncertain external rithms, such as extended Kalman filter (EKF) based and least squares

⁎
Corresponding author.
E-mail addresses: [email protected] (J. Yang), [email protected] (B. Xia), [email protected] (W. Huang), [email protected] (Y. Fu), [email protected] (C. Mi).

https://doi.org/10.1016/j.apenergy.2018.01.010
Received 21 August 2017; Received in revised form 30 December 2017; Accepted 3 January 2018
0306-2619/ © 2018 Elsevier Ltd. All rights reserved.
J. Yang et al. Applied Energy 212 (2018) 1589–1600

based methods [14–17]. constant-voltage (CCCV) charging data. One of the commonly used
The battery capacity can generally be determined by two groups of methods is based on the incremental capacity (IC) or the differential
approaches: the model-based and the correlation-based techniques. voltage (DV) analysis [36–38]. With this type of technique, the battery
According to the employed battery model, the model-based tech- aging mechanisms (i.e., loss of lithium inventory and loss of active
niques can be further classified into two categories, namely, the black material) can be identified by analyzing the peaks on the IC or DV curve
box model-based and the equivalent circuit model (ECM)-based tech- [39]. However, since all the peaks on the IC or DV curve lie within the
niques. For the black box model-based techniques, artificial neural plateaus of the OCV curve which is vulnerable to the measurement
networks [18,19], support vector machine [20,21], sample entropy noise, it is difficult to identify the distinct peaks directly from the
[22,23] and sparse bayesian modeling [24] are generally employed to measured data set. Meanwhile, the IC and DV curves are generally
estimate the capacity fade in batteries. The advantage of these ap- derived from the pseudo or real OCV curve. It requires completely
proaches is that the exact knowledge of battery aging mechanism is not charging/discharging the battery with micro currents or recording the
considered, and only external behaviors of battery (e.g., voltage, cur- battery terminal voltage after a long-time relaxation process at SoC
rent, and temperature) are required. Therefore, they are relatively points covering the entire range. Both methods are time-consuming and
simple and straightforward. However, the accuracy of the black box thus are not suitable for the on-board application [39]. Ref. [40] es-
model is closely dependent on the quantity and quality of training data, tablished a quantitative correlation between battery capacity and IC
which requires extensive offline training investigations [25]. By con- peak value, and the support vector regression was employed to extract
trast, the ECM describes battery behaviors through a series of specific the IC peak value. Ref. [41] considered the normalized location interval
electric elements such as impedance components and voltage sources. of the DV curve as the characteristic parameter, and correlated it with
The battery capacity can be estimated alone or together with other the capacity loss. The improved center least squares method was em-
model parameters (e.g. impedance parameters, open circuit voltage ployed to extract the DV curve. These approaches utilized the battery
(OCV)) and SoC through a variety of filters or observers, such as EKF charging data directly to detect the battery capacity fading, and showed
[26,27], H-infinity filter [28], and recursive least squares algorithm robust performances against the measurement noise and data size. It
[29]. These methods can theoretically estimate the capacity and other has to be noted that better results would be expected using the OCV
parameters as precisely as the battery is modeled. However, it has to be curve to obtain IC and DV curves, instead of the charging data [39,42].
noted that three disadvantages exist concerning this type of technique. In addition, because of the extensive computational power required, the
Firstly, a sufficient time is required to ensure the estimated capacity robust parameter extraction algorithm employed in these methods is
converge to the stable value [26]. Secondly, the accuracy of the esti- generally not applicable for the on-board implementation. Besides the
mation results strongly depends on the precision of the battery model, IC or DV based methods, Ref. [43] calculated the battery pack capacity
while an accurate battery model can significantly increase the compu- by transforming the charging voltage curve during the constant-current
tational cost [30]. Thirdly, the cross interference terms among the es- (CC) charging period, and only simple mathematical calculations were
timated variables can also compromise the estimation performance in needed. However, the authors made the critical assumption that the
terms of the numerical stability and accuracy [31]. battery should experience a complete CC charging cycle, which rarely
From the perspective of correlation-based techniques, the battery happens in real applications, especially in EVs. Ref. [44] used the vol-
capacity is indirectly determined based on the specific characteristic tage-capacity rate curve to identify the battery capacity, and the genetic
parameter, and the mapping relationship between capacity degradation algorithm was carried out to find the optimum transformation para-
and the related parameter variation should be established in advance. meter. Nevertheless, the performance of the voltage-capacity rate curve
Typically, there exists several methods to derive the characteristic is sensitive to the sampling frequency and the measurement noise
parameter from the dynamic discharging process [32,33]. However, [40,45]. Moreover, the adopted genetic algorithm is still complicated
besides the aging state, the battery model parameters are generally for the on-board applications. Different from the methods described
changeable with SoC, temperature and C-rate of the load current under above, Ref. [46] recognized the battery capacity loss according to the
the operating condition [12], and the accuracy of SoH estimation will constant-voltage (CV) charging data. The simple mathematical function
be reduced without considering these influences. Ref. [32] selected the was proposed to simulate the battery current behavior during the CV
solid electrolyte interphase (SEI) resistance to predict the battery re- charging period, and an obvious linear correlation between the model
maining capacity by the appropriate correlation function. This ap- parameter and the battery capacity was exploited. Although the ver-
proach was found to be insensitive to discharge C-rate and effective in ification results show that the simulated CV current matches well with
estimating the battery remaining capacity. Specifically, the average SEI the experimental data, the parameters in the developed function had no
resistance with respect to the specific SoC range was utilized to reduce explicit physical meaning. Besides, a complete CV charging period is
the influence of SoC. However, the battery parameters are generally required to determine the accurate battery capacity loss, which is in-
identified online from dynamic operating conditions with a random SoC feasible from a practical point of view.
range. Thus, obtaining the average value covering the specific SoC
range in practice is a challenging task. Ref. [33] developed a formula to 1.2. Contributions of the paper
determine the battery capacity based on the estimated diffusion capa-
citance. The temperature dependency of diffusion capacitance was In this paper, the battery capacity fading is detected through the
considered to improve the robustness of the fitting function. None- dynamic characteristic of the charging current during the CV period.
theless, extensive laboratory investigations were required to obtain an The main contributions of this paper are:
accurate correlation. Some methods were developed based on the re-
laxation data after the current interruption. With this kind of methods, (1) The time constant of CV charging current is firstly introduced to
the battery capacity can be derived based on the change in the OCV indicate the battery aging state. Compared with the CV charging
before and after a driving event [34,35]. Since only the SoC-OCV cor- time, the current time constant is a more robust characteristic
relation was employed as a characteristic parameter, which was nearly parameter related to the battery aging. Based on the employed
unchanged by the battery aging state, these methods enabled a high ECM, the CV charging current is expressed in the recursive form to
accuracy of capacity estimation over the battery lifetime. However, the obtain the detailed expression of the current time constant.
long-time relaxation period is required for the precise OCV measure- (2) An online battery SoH estimation method is developed based on the
ment, which is not suitable for real applications. Recently, many re- quantitative correlation between the normalized battery capacity
searchers have concentrated on predicting the battery capacity based and the current time constant. Experimental results show that there
on the characteristic parameter extracted from the constant-current exist a strong linear regression between the normalized battery

1590
J. Yang et al. Applied Energy 212 (2018) 1589–1600

Current
Capacity test CCCV charging test
Voltage

HPPC test CC discharging test

Temperature
Tested batteries
UDDS test
Reach certain Y
aging cycles?

N N
TCP/IP Reach EoL?

Y
Host computer Arbin BT2000 End test

Fig. 1. Configuration of the battery test bench.

Characterization test Aging test
capacity and the current time constant. Besides, the correlation Fig. 2. Battery aging test procedure.
function extracted from one battery is able to evaluate the SoH of
other three batteries with less than 2.5% absolute error except a few room temperature (the ambient temperature is 25 ± 2 °C), and the
outliers. measured data sets, including current, voltage and temperature, are
(3) The uncompleted CV charging data is utilized to estimate battery recorded with the sampling frequency (fs) of 1 Hz.
SoH. Firstly, the logarithmic function-based prediction model is The aging tests are conducted to explore the aging mechanism of the
established to predict the reference current time constant. Then, the battery, e.g., loss of lithium inventory, loss of active material and
battery SoH can be estimated by substituting the reference time Ohmic resistance increase [9]. It cycles until the battery reaches the
constant into the reference correlation curve. Comparison results end-of-life (EoL) condition (i.e., the battery’s capacity reduces to 80% of
demonstrate the superiority of the proposed method in terms of its its nominal capacity [32]). During each aging cycle, all four batteries
robustness to data size. are at 1C rate for CCCV charge and at 4C rate for CC discharge with no
rest periods.
2. Experimental tests and CV charging current analysis The characterization tests are performed periodically after certain
aging cycles, which are aimed to extract battery parameters and acquire
2.1. Experimental setup the corresponding variation tendency along with aging. As illustrated in
Fig. 2, the characterization tests consist of a capacity test, a hybrid pulse
The battery test bench includes the tested batteries, an Arbin power characterization (HPPC) test and an Urban Dynamometer
BT2000 cycle-based tester, and a host computer with MITS Pro software Driving Schedule (UDDS) test. The battery capacity is derived from the
for experiment control and data storage, as shown in Fig. 1. Four li- capacity test comprising 5 charge-discharge cycles. Each cycle includes
thium iron phosphate (LiFePO4) IFR26650PC batteries from Valence a 1/2C rate CCCV charge (CC-CV transition voltage: 3.65 V, cutoff
are adopted under test. All four batteries, which are numbered from #1 current: 1/20C) and a 1/2C rate CC discharge processes. Finally, the
to #4, are cycled with the same current excitations for comparison and battery capacity is calculated as the mean value of five cycles. The
validation. The specific parameters of the tested battery are listed in purpose of the HPPC test is to discharge/charge the battery to a certain
Table 1. The Arbin BT2000 cycle-based tester is used to charge and SoC and excite the batteries dynamically. In the HPPC test, the battery
discharge the batteries and collect the test data. The voltage and current is discharged/charged at a 5% SoC interval with 1/2C rate, and a 2 h
measurement ranges are 0–5 V and ± 100 A, respectively. The current rest is applied after each SoC variation operation. As shown in Fig. 3,
measurement error in the battery tester is less than 0.05%, and the the HPPC pulses include one pair of discharge and charge pulses. The
initial SoC can be obtained precisely from the test platform. Thus, it is magnitude and duration of each current pulse are 2C and 10 s, re-
feasible to assume that the accumulative battery capacity calculated spectively. The rest period between each discharge and charge pulses
through the recorded current can be considered as the reference value (T1 in Fig. 3) is 40 s, and the rest period after current pulses (T2 in
for comparison. Fig. 3) is 60 s. The objective of the UDDS test is to validate the

2.2. Test procedure

7.5
The test procedure, as shown in Fig. 2, is designed to generate rich 5
excitations for the tested batteries. It mainly includes two parts: the
2.5
Current (A)

aging and the characterization tests. All of the tests are conducted at
0
Table 1
Specifications of the tested battery. -2.5
Type LiFePO4 -5 T1 T2
Nominal capacity 2.5 Ah -7.5
Nominal voltage 3.2 V
0 10 20 30 40 50 60 70 80 90 100 110 120
Charge cutoff voltage 3.65 V Time (s)
Discharge cutoff voltage 2.0 V
Fig. 3. The current profile of HPPC pulses.

1591
J. Yang et al. Applied Energy 212 (2018) 1589–1600

1400 1.25
Battery #1 0150 cycles
1200 Battery #2 0450 cycles
Battery #3 1 0900 cycles
Battery #4 1500 cycles
1000
CV charging time (s)

1800 cycles
2100 cycles

Current (A)
0.75
800

600 0.5

400
0.25
200
0
0 0 200 400 600 800 1000 1200
150 450 900 1500 1800 2100
Time (s)
Cycles
Fig. 6. Comparison of CV charging current after different aging cycles.
Fig. 4. Comparison of CV charging time for all four batteries.

may also cause the battery to finish the CV period earlier. Compared
effectiveness of identified model parameters. In the UDDS test, the in- with the complete CV charging process, the shorter TCV would be ob-
itial SoC is 90% and the cycle is repeated without rest until the SoC tained from either of the above two scenarios, which in turn causes a
reaches 20%. significant SoH estimation error. For example, TCV extracted after 2000
cycles of battery #1 is 1158 s, and the corresponding Cn should be
2.3. Preliminary CV charging current analysis 0.8782 (i.e., point A in Fig. 5). Nevertheless, if the CV charging process
is terminated when TCV equals 1000 s, the estimated Cn is 0.9195 when
Fig. 4 shows the CV charging time of all four batteries after different substituting TCV into the correlation curve, as illustrated in Fig. 5. Thus,
aging cycles. For this study, all the CV charging data is extracted from TCV is highly sensitive to the external interference, and a more robust
the 5th cycle in the capacity test. It can be seen from Fig. 4 that the CV method is needed.
charging time (TCV) generally possesses the increasing tendency with The CV charging current of battery #1 after different aging cycles
the growing aging cycles. It indicates that the battery aging state can be are plotted in Fig. 6. It can be observed from Fig. 6 that with the
directly investigated from the CV charging time. In order to quantify the growing aging cycles, the charging current reduces to the cut-off value
correlation between the battery capacity and TCV, the normalized ca- at a slower variation rate, which in turn leads to a longer duration of the
pacity (Cn) of battery #1 (battery #1 is considered as the reference CV period. Since the current variation rate is closely associated with the
battery in this paper) is plotted versus the corresponding TCV in Fig. 5. relevant time constant, the current time constant (τI) for the CV char-
The normalized capacity, as defined in (1), is adopted in this paper due ging period can thereby be adopted to investigate the battery aging
to the inconsistency among different batteries. state. Unlike TCV, which should be obtained when the whole CV char-
Cactual ging process is completed, τI can be extracted from the partial CV
Cn =
Cinitial (1) charging data, which will be discussed in great detail in Section 4.2.
Hence, τI is a more robust characteristic parameter related to the bat-
where Cactual is the battery capacity obtained after each characterization tery aging, in comparison to the CV charging time.
test, and Cinitial is the battery capacity derived after the first char-
acterization test.
It can be seen from Fig. 5 that Cn shows a monotonically decreasing 3. Current time constant identification
relationship with TCV. The solid line in Fig. 5 denotes the fitted linear
correlation between Cn and TCV. Fig. 5 demonstrates that with the 3.1. Battery model and parameter identification
complete CV charging process, the battery capacity can be obtained by
substituting TCV into the correlation established beforehand. However, A proper battery model plays an important role in a high-perfor-
it has to be noted that some batteries in the practical application cannot mance BMS. Theoretically, the dynamic characteristic of a battery can
reach the cut-off current value due to the uncompleted charging pro- be accurately described by the ECM with infinite resistor-capacitor (RC)
cess. In addition, the noise disturbance on the current measurement networks [47]. However, this is an impractical battery model because
of the heavy computational effort. Hence, considering the tradeoff be-
1.02 tween the model fidelity and computational complexity, the first order
1 Measured capacity ECM is adopted in this paper.
Normalized capacity

0.98 Fitted curve The architecture of the first order ECM is shown in Fig. 7. It com-
0.96 prises a voltage source which represents the OCV, an Ohmic resistance
0.94 (Ro) and an RC network (Rp//Cp). It should be noted that the above
0.92 model parameters are generally represented as functions of SoC and
0.9 temperature (T). Besides, the polarization resistance (Rp) and OCV also
0.88 depend on the C-rate of the load current (CI) [12] and hysteresis effect
A H [48], respectively, i.e.:
0.86
0.84 VOC = VOC (SoC ,T ,H ), Rp = Rp (SoC ,T ,CI ), Cp = Cp (SoC ,T ),
600 700 800 900 1000 1100 1200 1300
CV charging time (s) Ro = Ro (SoC ,T ) (2)
Fig. 5. Correlation between Cn and TCV of battery #1.
where Cp represents the polarization capacitance.

1592
J. Yang et al. Applied Energy 212 (2018) 1589–1600

Rp where vectors ∂t and X are defined as

Ro IL dV
∂t = ⎡ dtt
dVOC dRo dIL
⎤, X T = [1 − 1 − IL − Ro ]
Cp ⎣ dt dt dt ⎦ (12)

Since the battery terminal voltage is controlled constant during the

Vp CV period, the value of dVt/dt can be considered as 0.
OCV Vt Considering (2), the variation of Voc and Ro with respect to time can
be further expressed as
dVOC ∂VOC dSoC ∂V dT ∂VOC dH
⎧ dt
= ∂SoC · dt + ∂OC T
· dt + ∂H
· dt
Fig. 7. The first order ECM.
⎨ dRo = ∂Ro · dSoC + ∂Ro · dT
⎩ dt ∂SoC dt ∂T dt (13)
The electrical behavior of the first order ECM can be expressed as In addition, Eq. (13) can be simplified considering the following
dVp Vp assumptions:
Cp + = IL
dt Rp (3)
(1) dSoC/dt can be further express as IL/(3600Ccap) (where Ccap is the
VOC + Ro IL + Vp = Vt (4) capacity of the battery in Ah). Assuming that the battery is charged
by a 2C constant current (generally less than this value in most
where Vt represents the battery terminal voltage, Vp represents the applications [42], especially during the CV period), the value of
voltage across the RC network, VOC represents the OCV, and IL is the dSoC/dt is IL/(3600Ccap) = 2/3600 = 0.00056, which is very small.
load current, a positive value represents the charging scenario and a Hence, dSoC/dt ≈ 0 holds for normal charging conditions.
negative value represents the discharging scenario. (2) Depending on the thermal management in the BMS, the battery
Based on (3) and (4), the transfer function of the first order ECM in temperature should be controlled in a proper range and the relevant
the Laplace domain can be expressed as variation should be slow, thus dT/dt ≈ 0 can be obtained [50].
IL (s ) 1 + Rp Cp s (3) Unlike the dynamic discharging process, the charging process is
H (s ) = = relatively simple and predictable. Hence, the model fidelity can be
Vt (s )−Voc (s ) Ro + Rp + Ro Rp Cp s (5)
guaranteed by using the charging SoC-OCV relationship, and thus
By means of the z-transform, Eq. (5) yields the hysteresis effect can be neglected, i.e., dH/dt ≈ 0.

IL (z ) z−exp(−Ts / Rp Cp)
H (z ) = = Based on the aforementioned analysis, dVoc/dt ≈ 0 and dRo/dt ≈ 0
Vt (z )−Voc (z ) Ro [z−exp(−Ts / Rp Cp)] + Rp [1−exp(−Ts / Rp Cp)] hold for the charging condition, thus (11) can be rewritten as
(6)
dIL Rp + Ro V −V
where Ts is the sampling period and is equal to 1 s in this paper. + IL = t OC
dt Cp Rp Ro Cp Rp Ro (14)
This function can be rewritten as
where Vt − VOC is considered as a constant item during one sampling
IL,k = IL,k − 1 θ1 + (Vt ,k−Voc,k ) θ2 + (Vt ,k − 1−Voc,k − 1) θ3 (7)
period, and the impedance parameters (Cp, Rp and Ro) can be identified
where in advance through the algorithms mentioned above. Based on the
general solution of the first order linear differential equation, IL can be
θ1 = exp(−Ts / Rp Cp)−(Rp / Ro)[1−exp(−Ts / Rp Cp)] solved as
θ2 = 1/ Ro
θ3 = −(1/ Ro)exp(−Ts /Rp Cp) (8) ( ∫ bdt)
IL = IL,t0 exp −

The regression form of (7) can be expressed as + exp (− ∫ bdt ) ∫ ⎡ (V −V t OC )exp (∫ bdt) c⎤ dt
⎣ ⎦ (15)
IL,k = φk θ (9)
where IL,t0 is the initial load current, b = (Rp + Ro)/(CpRpRo), and
where c = CpRpRo. Defining tk = t0 and tk+1=t0 + Ts, Eq. (15) can be simpli-
fied as
φk = [ IL,k − 1 Vt ,k−Voc,k Vt ,k − 1−Voc,k − 1]
IL,k + 1 = IL,k exp(−bTs ) + [(Vt ,k−VOC ,k )/ d][1−exp(−bTs )] (16)
θ = [θ1 θ2 θ3 ]T (10)
where IL,k and IL,k+1 represent the load current at time tk and tk+1,
Among the parameters in (10), IL and Vt can be measured directly,
respectively, and d = Rp + Ro.
and VOC can be obtained through a lookup table, which is normally
In the system expressed by (16), Vt-VOC can be regarded as the
identified by the HPPC test in advance. Besides, the impedance para-
system input, and the load current IL is considered as the system re-
meters (Ro, Rp and Cp) can be obtained by solving (9) with the discrete-
sponse. If the system input is zero, the system response will decrease to
time least squares (DT LS) method, which is widely implemented in the
36.79% of its initial value when Ts = 1/b. Similarly, for an increasing
real applications. The algorithmic procedure of the DT LS method is not
step input, if the initial current equals 0, the system response will reach
detailed in this paper, but can be referred to Ref. [49].
63.21% of its stable value when Ts = 1/b. Hence, it can be concluded
that the time constant for the CV charging current can be expressed as
3.2. Expression of current time constant
1 Cp Rp Ro
τI = =
In order to analyze the kinetic characteristic of the CV charging b Rp + Ro (17)
current, the detailed expression of the load current is needed.
Since Vp cannot be measured directly, the removal of Vp term from
(3) and (4) is desired in order to obtain the current expression. Hence, 3.3. Parametric sensitivity to sampling frequency
substituting (4) into (3), the differential equation can be rewritten as
The sampling frequency is crucial for the performance of DT LS
Rp IL = Cp Rp ∂t X + (Vt −VOC−Ro IL) (11) method. Without considering the limitation of on-board storage and

1593
J. Yang et al. Applied Energy 212 (2018) 1589–1600

computation capabilities, the rapid sampling is normally desired so that 1.3

the information on the fast dynamics can be captured [51]. However, Experiment
1.1
the high sampling frequency will result in the eigenvalue being close to Estimate
the unit circle in the z-domain, which can affect the stability for para- 0.9

Current (A)
meter identification [52]. Hence, the sensitivity of τI to fs is worth in-
vestigating. 0.7
Generally, the location of the eigenvalue in the z-domain can be 0.5
expressed as
0.3
z = exp(λTs ) (18)
0.1
where λ is the eigenvalue of the system, and |λ| = 1/τI. 0 100 200 300 400 500 600 700
Based on (18), the sensitivity of τI with respect to eigenvalue loca-
Time (s)
tion in the z-domain is expressed as [49]:
(a) 150 cycles
z ∂τI 1
SzτI = = = τI fs
τI ∂z |λ|Ts (19) 1.3
It can be inferred from (19) that the higher sampling frequency Experiment
1.1
yields higher sensitivity of identified τI. In a practical BMS, the sampling Estimate
frequency can be 100 Hz, and τI identified after 2300 cycles is 233.8 s. 0.9

Current (A)
Based on (19), the sensitivity can be as high as 23,380, which indicates
that τI estimation is extremely sensitive to the eigenvalue location. 0.7
Hence, to guarantee a good numerical stability, the sampling frequency 0.5
in real application is recommended to limit to lower than a certain
level. 0.3
0.1
3.4. Battery model verification 0 200 400 600 800 1000
Time (s)
To verify the accuracy of the identified model parameters and the
derived current expression during the CV charging period, the com-
(b) 1500 cycles
parison results of battery #1 after 150, 1500 and 2200 cycles are shown 1.3
in Fig. 8. Two evaluation criterions, including the Root Mean-Square
Error (RMSE) and the R-square (R2), are employed to evaluate the 1.1 Experiment
waveform similarity, as listed in Table 2. The definitions of RMSE and Estimate
0.9
Current (A)

R2 are described as
1 n
0.7
RMSE = ∑k =1 (yexp,k −yest,k )2 (20)
n 0.5
n
∑k = 1 (yexp,k −yest ,k )2 0.3
R2 = 1− n
∑k = 1 (yexp,k −yexp )2 (21) 0.1
0 200 400 600 800 1000 1200
where yexp,k and yest,k denote the experimental and the estimated values,
respectively, yexp is the mean value of the experimental data, and n is the Time (s)
data size. (c) 2200 cycles
It can be inferred from (20) and (21) that a lower RMSE (closer to 0)
Fig. 8. Comparison between measured and estimated CV charging current after different
and a larger R2 (closer to 1) indicate a better fitting performance. It can cycles.
be concluded from Fig. 8 and Table 2 that the estimated current mat-
ches well with the measured charging current, which confirms that the
extracted parameter set and the derived current expression can describe Table 2
Evaluation criterions of estimation performance after different cycles.
the transient behavior of the CV charging current with high fidelity.
Cycles RMSE R2
4. SoH estimation based on the correlation between Cn and τI
0150 0.02812 0.9855
1500 0.02812 0.9870
4.1. Correlation identified from the complete CV charging data 2200 0.02455 0.9905

Based on (16) and (17), τI can be identified from the CV charging

data. Taking the data of battery #1 as the benchmark, Cn is plotted represent the mean value of Cn and τI, respectively. The range of rxy is
versus the corresponding τI in Fig. 9. It can be seen from Fig. 9 that Cn [−1, 1], where close to +1 indicates a strong positive correlation, close
shows a monotonically decreasing relationship with τI. To quantita- to −1 indicates a strong negative correlation, and close to 0 indicates a
tively measure the degree of the linear relationship between Cn and τI, weak or totally missing correlation.
the correlation coefficient, or the Pearson product-moment correlation The correlation coefficient between Cn and τI of battery #1 is
coefficient, is adopted in this paper. It can be described by (22) [38,53]. −0.9880, thus the linear function, as shown in (23), is employed to fit
n the obvious linear relationship between these two values.
∑k = 1 (xk −x )(yk −y )
rxy = n n
∑k = 1 (xk −x )2 ∑k = 1 (yk −y )2 (22) Cn = a1 τI + b1 (23)

where x and y correspond to Cn and τI, respectively, and x and y where a1 and b1 are the fitting function coefficients which can be

1594
J. Yang et al. Applied Energy 212 (2018) 1589–1600

1.02 1
1 Measured capacity 0.98 ld =300
Normalized capacity

0.98 Fitted curve ld =400

Normalized capacity
0.96 0.96 ld =600
whole
0.94 0.94
0.92 0.92
0.9
0.9
0.88 B
0.86 0.88
0.84 0.86
120 140 160 180 200 220 240 260 90 110 130 150 170 190 210 230 250
Current time constant (s) Current time constant (s)
Fig. 9. Correlation between Cn and τI of battery #1. Fig. 11. Comparison of correlations between Cn and τI with respect to different ld.

Table 3 Table 4
Curve fitting results of battery #1. Curve fitting results with respect to different ld.

a1 b1 RMSE R2 Data size a1 b1 RMSE R2 Correlation coefficient

−0.001063 1.143 0.0113 0.9761 ld = 300 −0.002348 1.218 0.01496 0.8912 −0.9440
ld = 400 −0.00222 1.239 0.008814 0.9622 −0.9809
ld = 600 −0.00196 1.252 0.007512 0.9731 −0.9865
determined through curve-fitting method. Whole −0.001163 1.155 0.00701 0.9761 −0.9880

The fitted correlation curve is plotted as a solid line in Fig. 9. The

detailed fitting function coefficients and two fitting quality criterions,
battery #1, Cn is plotted versus the corresponding τI identified from the
including RMSE and R2, are listed in Table 3.
partial (the former 300, 400 and 600 sampling data) and the whole CV
It can be observed from Table 3 that the value of RMSE is small and
charging data in Fig. 11. In addition, the relevant correlation curves and
the value of R2 is close to 1, which indicates a good fitting performance.
the corresponding fitting results are shown in Fig. 11 and Table 4.
Cn and τI of all four batteries are plotted in Fig. 10. The solid line in the
It can be observed from Fig. 11 that even in the same aging state, τI
figure represents the correlation curve fitted through the data of battery
identified from the data with different lengths shows diverse values,
#1, and the dotted lines give the error bound of 2.5%. It can be ob-
which in turn leads to the various fitted correlation curves. Hence,
served that most of the points are within the error bonds except one
applying τI corresponds to one specific data length (ld) to the correlation
outlier, which confirms that the relationship between Cn and τI can be
curve with respect to other ld can cause a significant SoH estimation
employed to indicate the battery SoH. In addition, it can be concluded
error. For example, τI identified from the former 400 sampling data is
that the linear relationship identified from the reference battery (bat-
159.1 s, and the corresponding Cn should be 0.8912 (i.e., point B in
tery #1) can also be used to evaluate the capacity degradation of the
Fig. 11). However, the estimated Cn is 0.9404 and 0.9701 when sub-
other three batteries. Hence, for the batteries from the same manu-
stituting τI into the correlation curve fitted from the former 600 sam-
facturer, only the linear relationship of the reference battery should be
pling data and the whole data size, respectively, as illustrated in Fig. 11.
established in advance, which can effectively reduce the testing effort.
The straightforward way to overcome this problem is to identify the
correlation curve covering a wide range of ld with a particular data
4.2. Correlation identified from the partial CV charging data length interval (Δld) in advance, then store them in the on-board mi-
crocontroller as a database. When τI of a specific ld is identified, it can
4.2.1. Reference correlation curve selection be put into the “Cn-τI database” to find a correlation curve with respect
In practical applications, some batteries cannot reach the cut-off to a closest ld. The accuracy of the estimated Cn depends on the selected
current value due to the uncompleted charging process or the mea- Δld. The smaller Δld yields more details and thus the higher Cn esti-
surement noise of the current sensors. Thus, only part of the CV char- mation accuracy, whereas it requires more storage resources and offline
ging current curve can be obtained in this case. Based on the existing identification efforts, especially for the battery with long CV charging
research work [54,55], the identified time constant of the battery period. Besides, it can be concluded from Table 4 that the correlation
terminal voltage is not a constant value, but a variable as a function of curve identified from the data with less ld shows a worse fitting per-
open-circuit time. Similarly, the value of τI should also be a variable formance, which in turn leads to a larger SoH estimation error. An al-
with respect to the length of CV charging data. Based on the data of ternative method is to predict the specific τI (τI,ref) corresponding to a
predetermined longer ld (ld,ref), based on the uncompleted CV charging
1.02 data. Then, the battery SoH can be estimated by substituting the pre-
1 dicted τI,ref into the reference correlation curve, which is identified from
Normalized capacity

0.98 +2.5% the data with size of ld,ref. Although the correlation curve identified
0.96 from the whole data size exhibits the best fitting performance, as illu-
0.94 strated in Table 4, it is impractical to employ it as the reference cor-
0.92 relation curve, because the complete CV charging time cannot be ob-
Battery #1
0.9 Battery #2 tained in advance. On the other hand, it can be concluded from Fig. 4
0.88 Battery #3 that the shortest complete CV charging time is slightly more than 600 s.
Battery #4 -2.5% Hence, the correlation curve identified from the former 600 sampling
0.86
data is selected as the reference correlation curve in this paper.
0.84
120 140 160 180 200 220 240 260
Current time constant (s)
4.2.2. Current time constant prediction
Fig. 10. Correlation between Cn and τI of all four batteries.
In order to predict τI,ref, the functional relationship between τI and ld

1595
J. Yang et al. Applied Energy 212 (2018) 1589–1600

250 200
Current time constant (s)
ld,ref = 600

Current time constant (s)

200 160
ld=400
150 120 ld=300

100 0600 cycles 80 IJI

1050 cycles fitted curve (ld=300)
50 2300 cycles 40 fitted curve (ld=400)

0 0
0 200 400 600 800 1000 1200 0 100 200 300 400 500 600 700 800
Data length Data length
Fig. 12. Relationship between τI and ld after different cycles. (a) 600 cycles

should be determined firstly. Scatter plots of τI versus ld after different 200 ld,ref = 600

Current time constant (s)

cycles (600, 1050 and 2300 cycles) are shown in Fig. 12 to illustrate the
relationship between these two sets of data. 160 ld=400
It can be observed from Fig. 12 that τI increases with the increase of ld=300
ld. Considering the tendency of the related data, the fitting function as 120
shown in (24) is adopted to quantitatively describe the relationship
between τI and ld [32]. 80 IJI
fitted curve (ld=300)
τI = a2ln(ld ) + b2 (24) 40 fitted curve (ld=400)
where a2 and b2 are the fitting function coefficients. The fitting results, 0
as shown in Fig. 12 (the dotted lines) and Table 5, demonstrate that the 0 100 200 300 400 500 600 700 800
employed fitting function can accurately describe the increasing ten- Data length
dency of τI. (b) 1050 cycles
To evaluate the accuracy of the predicted τI,ref estimated from par-
tial CV charging data, the identified fitting curves based on the first 300
240 ld,ref = 600
and 400 sampling data after different cycles are given in Fig. 13. The
Current time constant (s)

actual τI points covering the former 800 sampling data are also plotted 200 ld =400
in Fig. 13. It shows that the identified curves match well with the actual
160 ld =300
points. This can also be observed in Table 6, where the relative errors
between the predicted and the actual τI,ref are all less than 10% in six 120
cases. In addition, compared with the predicted τI,ref based on the IJI
former 300 sampling data, the values obtained based on the former 400 80 fitted curve (ld=300)
sampling data have a lower overall error. This is because more fitted 40 fitted curve (ld=400)
data points yield more information to predict the variation tendency of
0
τI. Considering the accuracy of the predicted τI,ref, at least 300 s of the 0 100 200 300 400 500 600 700 800
CV charging period is required for the tested batteries. Data length
(c) 2300 cycles
4.3. Developed SoH estimation scheme Fig. 13. Comparison between prediction curves and actual τI points after different cycles.

With the established reference quantitative relationship between Cn

Table 6
and τI, an online SoH (mainly on the battery capacity fading) estimation
τI,ref prediction results based on the different ld.
scheme is developed, as shown in Fig. 14. The measured battery voltage
and current are recorded in the on-board storage component as soon as Cycles τI,ref_actual τI,ref_predict Relative error τI,ref_predict Relative error
the charging process enters the CV charging mode. The recorded data (ld = 300) (ld = 300) (ld = 400) (ld = 400)
sets are processed when the charging process is finished. If the recorded
0600 140.0 s 132.2 s −5.571% 135.3 s −3.357%
data size is no less than the selected ld,ref (e.g., 600 selected in this 1050 167.7 s 159.6 s −4.830% 164.2 s −2.087%
paper), τI,ref can be identified directly from the former sampling data 2300 183.3 s 179.5 s −2.073% 183.4 s 0.05456%
with size of ld,ref. Otherwise, the values of τI covering the former data
sets with different lengths (e.g., 10, 20, … , ld) identified firstly. Sec- Note: τI,ref_actual is the actual τI,ref identified based on the former 600 sampling data, and
τI,ref_predict is the predicted τI,ref derived based on the prediction model.
ondly, the quantitative correlation between τI and ld (expressed as (24))
can be obtained through curve fitting method. Afterwards, τI,ref can be
predicted by substituting ld,ref into the identified fitting function. At last, the battery SoH can be derived using the correlation between Cn and
τI,ref, and the present battery capacity can be updated.
Table 5
Curve fitting results after different cycles.
5. Verification and discussion
Cycles a2 b2 RMSE R2
An online SoH estimation method is proposed in this paper, which is
0600 48.33 −163.4 6.1466 0.9722
based on the analysis of load current during CV charging period. The
1050 61.27 −216.7 6.1371 0.9820
2300 71.07 −272.8 5.6040 0.9890 developed method can be applied to the uncompleted CV charging
process.

1596
J. Yang et al. Applied Energy 212 (2018) 1589–1600

Measure and record battery Table 7

SoH estimation results with ld = 600 of all four batteries.
voltage and current during
the CV charging period Battery #1 Battery #2

SoHref SoHest SoHerr SoHref SoHest SoHerr

No 98.15% 97.76% −0.3981% 98.07% 96.71% −1.359%

End of charge
95.87% 96.39% 0.5234% 95.99% 96.34% 0.3485%
94.74% 93.92% −0.8211% 94.91% 92.08% −2.828%
Yes 93.77% 92.33% −1.439% 93.01% 90.64% −2.369%
92.78% 93.07% 0.2842% 90.31% 89.77% −0.5444%
No 90.12% 89.68% −0.4426% 89.23% 87.81% −1.421%
ld < ld,ref 89.12% 89.13% 0.01295% 88.57% 89.03% 0.4585%
87.82% 88.07% 0.2462% 87.02% 88.96% 1.940%
Yes Battery #3 Battery #4

Identify the relationship SoHref SoHest SoHerr SoHref SoHest SoHerr

between IJI and ld
98.86% 99.20% 0.3417% 99.46% 99.35% −0.1130%
97.81% 97.95% 0.1451% 96.74% 96.58% −0.1603%
Identify IJI,ref corresponds to ld,ref
95.27% 96.59% 1.323% 95.80% 95.03% −0.7699%
92.53% 91.56% −0.9770% 94.74% 94.90% 0.1566%
Predict IJI,ref 90.51% 89.49% −1.017% 92.94% 90.92% −2.018%
89.91% 88.97% −0.9395% 90.51% 91.07% 0.5538%
89.33% 90.25% 0.9164% 88.19% 89.70% 1.516%
87.94% 86.12% −1.821% 87.72% 86.45% −1.270%
Estimate the battery SoH
Note: SoHref are obtained based on the capacity measurement, and SoHest and SoHerr are
according to the relationship
obtained based on the reference correlation curve.
between Cn and τ I
prediction model in terms of its robustness to the data size. With
respect to method 2, better results will be obtained if the stored
Update the present capacity correlation curves are identified with shorter Δld. However, more
storage resources and offline identification efforts are required,
Fig. 14. The proposed online SoH estimation scheme. especially for batteries with the long-time CV charging period (e.g.
more than 30 min for batteries in [46,56,57]).
(2) For both methods, the SoH estimated upon the longer ld generally
Based on the reference correlation curve established upon the data
has a better performance, compared with the estimation results
from battery #1 and the logarithmic function-based τI,ref prediction
based on the shorter ld. Based on the analysis in Section 4.2.2, the
model, all four batteries’ SoH can be derived online from the CV
improved performance in method 1 is supposed to be caused by the
charging data. Specifically, the definition of SoH is expressed as [33]:
more accurate τI,ref prediction results. For method 2, the improve-
SoH =
Cactual
× 100% = Cn × 100% ment is supposed to be brought by the better fitting performance of
Cinitial (25) the correlation curve, as has been discussed in Section 4.2.1.
(3) The SoH can be estimated by method 1 with less than 2.5% absolute
Firstly, to assess the performance of the reference correlation curve
established upon the data of battery #1, the SoH estimation results error except a few outliers. Hence, based on the τI,ref prediction
model, the battery SoH can be monitored by using the uncompleted
obtained from the former 600 sampling data (ld,ref = 600) of all four
batteries are listed in Table 7. It can be concluded from Table 7 that by CV charging data.
using the reference correlation curve identified through the data of
battery #1, the aging states of all four batteries can be estimated within 6. Future work
2.5% error bond except a few outliers. Hence, the battery SoH can be
estimated by the reference correlation curve successfully. Furthermore, In this paper, the simple first order ECM is adopted to characterize
for the batteries from the same manufacturer, the correlation curve can the battery behavior during the CV charging period. Besides, only
be established by a reference battery in advance. Afterwards, the SoH of LiFePO4 batteries are tested and the charging condition is nearly con-
all the other batteries can be estimated when τI,ref is derived from the stant through the aging test. Therefore, for the future work, some ex-
CV charging data. tensions to the employed model (e.g., the higher order ECM) will be
Secondly, to evaluate the performance of the proposed method considered to enhance the model performance. In addition, more ex-
when ld is less than ld,ref, the SoH estimations based on the τI,ref pre- periments and further study are required to investigate the generality of
diction model (method 1) and “Cn-τI database” (method 2) are both the proposed method, including its potential to be used for different
conducted to make a comparison. The stored Cn-τI correlations are battery chemistries, charging protocols and operating temperatures.
identified from the CV charging data ranging from the former 300 s
(ld = 300) to former 600 s (ld = 600) with a resolution of 50 s
7. Conclusion
(Δld = 50). Specifically, the former 320 s (ld = 320) and former 420 s
(ld = 420) are employed to test the robustness of two methods in terms
An online SoH estimation method is proposed in this paper, which is
of data size. The comparison of SoH estimation results obtained from
based on the analysis of load current during CV charging period.
the former 320 and 420 sampling data of all four batteries are sum-
According to the preliminary analysis of the battery test data, the time
marized in Tables 8 and 9, respectively. The results show that
constant of CV charging current, which is derived based on the ECM, is
considered as a robust characteristic parameter related to the battery
(1) Compared with method 2, method 1 yields the lower overall SoH
capacity fading. The quantitative correlation between Cn and τI is es-
estimation error in both cases. It indicates the superiority of the τI,ref
tablished to indicate the battery SoH. Specifically, for the uncompleted

1597
J. Yang et al. Applied Energy 212 (2018) 1589–1600

Table 8
SoH estimation results with ld = 320 of all four batteries.

Battery #1 Battery #2

SoHref SoHest_1 SoHerr_1 SoHest_2 SoHerr_2 SoHref SoHest_1 SoHerr_1 SoHest_2 SoHerr_2

98.15% 99.24% 1.082% 96.33% −1.824% 98.07% 98.24% 0.1745% 95.41% −2.655%
95.87% 98.28% 2.419% 95.07% −0.7923% 95.99% 96.91% 0.9141% 94.62% −1.376%
94.74% 95.56% 0.8257% 92.36% −2.375% 94.91% 95.94% 1.031% 92.69% −2.214%
93.77% 93.64% −0.1323% 91.49% −2.284% 93.01% 92.73% −0.2856% 89.47% −3.546%
92.78% 95.26% 2.478% 92.81% 0.02934% 90.31% 91.15% 0.8382% 88.96% −1.350%
90.12% 90.04% −0.07813% 87.93% −2.189% 89.23% 89.90% 0.6685% 86.43% −2.802%
89.12% 90.20% 1.085% 87.75% −1.370% 88.57% 90.32% 1.747% 88.35% −0.2163%
87.82% 88.93% 1.110% 86.73% −1.091% 87.02% 89.75% 2.729% 87.29% 0.2675%

Battery #3 Battery #4

SoHref SoHest_1 SoHerr_1 SoHest_2 SoHerr_2 SoHref SoHest_1 SoHerr_1 SoHest_2 SoHerr_2

98.86% 100.5% 1.641% 97.33% −1.529% 99.46% 99.31% −0.1487% 97.13% −2.333%
97.81% 99.57% 1.767% 96.20% −1.603% 96.74% 98.05% 1.308% 95.15% −1.597%
95.27% 97.45% 2.188% 94.67% −0.5999% 95.80% 96.56% 0.7582% 93.74% −2.055%
92.53% 93.51% 0.9769% 91.23% −1.299% 94.74% 95.80% 1.063% 92.58% −2.158%
90.51% 91.52% 1.012% 88.57% −1.938% 92.94% 92.29% −0.6482% 90.08% −2.857%
89.91% 87.78% −2.127% 87.56% −2.346% 90.51% 90.38% −0.1354% 89.15% −1.363%
89.33% 90.34% 1.007% 88.05% −1.285% 88.19% 90.65% 2.462% 91.03% 2.846%
87.94% 85.81% −2.132% 84.20% −3.740% 87.72% 85.62% −2.099% 82.38% −5.340%

Note: SoHest_1 and SoHerr_1 are obtained based on the τI,ref prediction model, and SoHest_2 and SoHerr_2 are obtained based on the Cn-τI correlation curve with respect to ld = 300.

Table 9
SoH estimation results with ld = 420 of all four batteries.

Battery #1 Battery #2

SoHref SoHest_1 SoHerr_1 SoHest_2 SoHerr_2 SoHref SoHest_1 SoHerr_1 SoHest_2 SoHerr_2

98.15% 98.56% 0.4072% 96.54% −1.616% 98.07% 97.97% −0.09717% 96.06% −2.009%
95.87% 97.05% 1.184% 95.10% −0.7621% 95.99% 96.25% 0.2590% 95.08% −0.9134%
94.74% 95.11% 0.3717% 93.20% −1.535% 94.91% 94.78% −0.1376% 92.92% −1.992%
93.77% 92.98% −0.7898% 91.82% −1.954% 93.01% 91.61% −1.405% 89.58% −3.432%
92.78% 94.27% 1.491% 92.81% 0.02712% 90.31% 90.40% 0.08593% 88.80% −1.513%
90.12% 89.93% −0.1905% 88.60% −1.519% 89.23% 89.45% 0.2200% 86.66% −2.572%
89.12% 89.51% 0.3883% 88.19% −0.9307% 88.57% 88.52% −0.0469% 86.78% −1.790%
87.82% 89.24% 1.422% 87.15% −0.6690% 87.02% 89.43% 2.407% 88.16% 1.145%

Battery #3 Battery #4

SoHref SoHest_1 SoHerr_1 SoHest_2 SoHerr_2 SoHref SoHest_1 SoHerr_1 SoHest_2 SoHerr_2

98.86% 100.0% 1.141% 97.97% −0.8868% 99.46% 99.49% 0.02794% 98.20% −1.264%
97.81% 98.96% 1.158% 97.31% −0.4918% 96.74% 97.14% 0.4014% 95.35% −1.390%
95.27% 97.09% 1.820% 95.28% 0.01479% 95.80% 96.37% 0.5701% 94.21% −1.588%
92.53% 92.37% −0.1679% 90.56% −1.971% 94.74% 95.35% 0.6095% 93.63% −1.110%
90.51% 90.75% 0.2417% 88.57% −1.943% 92.94% 91.80% −1.143% 90.42% −2.522%
89.91% 88.99% −0.9155% 86.14% −3.775% 90.51% 90.34% −0.1729% 89.51% −1.002%
89.33% 89.25% −0.07653% 87.85% −1.477% 88.19% 90.91% 2.724% 90.37% 2.183%
87.94% 86.86% −1.081% 85.25% −2.693% 87.72% 85.27% −2.449% 83.93% −3.796%

Note: SoHest_2 and SoHerr_2 are obtained based on the Cn-τI correlation curve with respect to ld = 400.

CV charging process, the reference correlation curve with respect to the the adopted data size.
specific data length is employed to estimate the battery SoH. Besides,
the logarithmic function-based prediction model is identified from the
partial CV charging data to predict τI,ref. Four LiFePO4 batteries are Acknowledgments
employed under test to verify the feasibility of the proposed method.
The results demonstrate that the correlation coefficient between Cn and The authors would like to acknowledge the funding support from
τI can reach −0.9880. Moreover, the correlation function extracted the China Scholarship Council (CSC) – China; the US DOE Graduate
from one battery is able to evaluate the SoH of other three batteries Automotive Technology Education (GATE) Center of Excellence; the
with less than 2.5% absolute error except a few outliers, regardless of National Science Foundation; and Nanjing Golden Dragon Bus Co., Ltd.

1598
J. Yang et al. Applied Energy 212 (2018) 1589–1600

Appendix A.

Abbreviations and nomenclature

BMS battery management system

CC constant-current
CCCV constant-current constant-voltage
CV constant-voltage
DT LS discrete-time least squares
DV differential voltage
ECM equivalent circuit model
EKF extended Kalman filter
EoL end-of-life
EV electric vehicle
HPPC Hybrid pulse power characterization
IC incremental capacity
LiFePO4 lithium iron phosphate
OCV and VOC open circuit voltage
RMSE Root Mean-Square Error
SEI solid electrolyte interphase
SoC state-of-charge
SoH state-of-health
SoP state-of-power
UDDS Urban Dynamometer Driving Schedule
a1 and b1 fitting function coefficients of the linear relationship
a2 and b2 fitting function coefficients of the logarithmic relationship
Cactual Battery capacity obtained after each characterization test
Ccap battery capacity
Cinitial Battery capacity obtained after the first characterization test
CI C-rate of the load current
Cp, Rp Polarization capacitance and resistance
Cn normalized capacity
H Hysteresis effect
ld data length
Δld data length interval
ld,ref reference data length
IL load current
rxy correlation coefficient between x and y
R2 R-square
Ro Ohmic resistance
T temperature
T1 rest period between each discharge and charge pulses in the HPPC test
T2 rest period after current pulses in the HPPC test
Ts and fs sampling period and frequency
TCV CV charging time
Vp voltage across the RC network
Vt battery terminal voltage
λ Eigenvalue of the system
τI current time constant
τI,ref reference current time constant
τI,ref_actual actual τI,ref identified based on the former 600 sampling data
τI,ref_predict predicted τI,ref derived based on the prediction model
SoHest_1 and SoHerr_1 SoH estimation results based on the τI,ref prediction model
SoHest_2 and SoHerr_2 SoH estimation results based on “Cn-τI database”
SoHref SoH estimation results based on the capacity measurement
SoHest and SoHerr SoH estimation results based on the reference correlation curve

References [4] Zhang C, Wang LY, Li X, Chen W, Yin GG, Jiang J. Robust and adaptive estimation
of state of charge for lithium-ion batteries. IEEE Trans Indust Electron
2015;62:4948–57.
[1] Mi C, Masrur MA, Gao DW. Hybrid electric vehicles: principles and applications [5] Lin C, Mu H, Xiong R, Shen W. A novel multi-model probability battery state of
with practical perspectives. John Wiley & Sons; 2011. charge estimation approach for electric vehicles using H-infinity algorithm. Appl
[2] Lu L, Han X, Li J, Hua J, Ouyang M. A review on the key issues for lithium-ion Energy 2016;166:76–83. [2016/03/15/].
battery management in electric vehicles. J Power Sour 2013;226:272–88. [6] Sun F, Xiong R, He H, Li W, Aussems JEE. Model-based dynamic multi-parameter
[3] Hosio S, Ferreira D, Goncalves J, van Berkel N, Luo C, Ahmed M, et al. Monetary method for peak power estimation of lithium–ion batteries. Appl Energy
assessment of battery life on smartphones. In: Proceedings of the 2016 CHI con- 2012;96:378–86. [2012/08/01/].
ference on human factors in computing systems; 2016. p. 1869–80. [7] Shen WX, C CC, Lo EWC, Chau KT. Adaptive neuro-fuzzy modeling of battery

1599
J. Yang et al. Applied Energy 212 (2018) 1589–1600

residual capacity for electric vehicles. IEEE Trans Indust Electron 2002;49:677–84. 2013;240:184–92. [2013/10/15/].
[8] Ouyang M, Feng X, Han X, Lu L, Li Z, He X. A dynamic capacity degradation model [34] Kessels JTBA, Rosca B, Bergveld HJ, v d Bosch PPJ. On-line battery identification
and its applications considering varying load for a large format Li-ion battery. Appl for electric driving range prediction. In: 2011 IEEE vehicle power and propulsion
Energy 2016;165:48–59. [2016/03/01/]. conference; 2011. p. 1–6.
[9] Vetter J, Novák P, Wagner MR, Veit C, Möller KC, Besenhard JO, et al. Ageing [35] Einhorn M, Conte FV, Kral C, Fleig J. A method for online capacity estimation of
mechanisms in lithium-ion batteries. J Power Sour 2005;147:269–81. [2005/09/ lithium ion battery cells using the state of charge and the transferred charge. IEEE
09/]. Trans Indust Appl 2012;48:736–41.
[10] Berecibar M, Gandiaga I, Villarreal I, Omar N, Van Mierlo J, Van den Bossche P. [36] Weng C, Feng X, Sun J, Peng H. State-of-health monitoring of lithium-ion battery
Critical review of state of health estimation methods of Li-ion batteries for real modules and packs via incremental capacity peak tracking. Appl Energy
applications. Renew Sustain Energy Rev 2016;56:572–87. [2016/04/01/]. 2016;180:360–8. [2016/10/15/].
[11] Jaguemont J, Boulon L, Dubé Y. A comprehensive review of lithium-ion batteries [37] Li X, Jiang J, Wang LY, Chen D, Zhang Y, Zhang C. A capacity model based on
used in hybrid and electric vehicles at cold temperatures. Appl Energy charging process for state of health estimation of lithium ion batteries. Appl Energy
2016;164:99–114. [2016/02/15/]. 2016;177:537–43. [2016/09/01/].
[12] Waag W, Käbitz S, Sauer DU. Experimental investigation of the lithium-ion battery [38] Berecibar M, Devriendt F, Dubarry M, Villarreal I, Omar N, Verbeke W, et al. Online
impedance characteristic at various conditions and aging states and its influence on state of health estimation on NMC cells based on predictive analytics. J Power Sour
the application. Appl Energy 2013;102:885–97. [2013/02/01/]. 2016;320:239–50. [2016/07/15/].
[13] Howey DA, Mitcheson PD, Yufit V, Offer GJ, Brandon NP. Online measurement of [39] Han X, Ouyang M, Lu L, Li J, Zheng Y, Li Z. A comparative study of commercial
battery impedance using motor controller excitation. IEEE Trans Veh Technol lithium ion battery cycle life in electrical vehicle: aging mechanism identification. J
2014;63:2557–66. Power Sour 2014;251:38–54. [2014/04/01/].
[14] Plett GL. Extended Kalman filtering for battery management systems of LiPB-based [40] Weng C, Cui Y, Sun J, Peng H. On-board state of health monitoring of lithium-ion
HEV battery packs. J Power Sour 2004;134:277–92. [2004/08/12/]. batteries using incremental capacity analysis with support vector regression. J
[15] Wei Z, Meng S, Xiong B, Ji D, Tseng KJ. Enhanced online model identification and Power Sour 2013;235:36–44. [2013/08/01/].
state of charge estimation for lithium-ion battery with a FBCRLS based observer. [41] Wang L, Pan C, Liu L, Cheng Y, Zhao X. On-board state of health estimation of
Appl Energy 2016;181:332–41. [2016/11/01/]. LiFePO4 battery pack through differential voltage analysis. Appl Energy
[16] He H, Xiong R, Guo H. Online estimation of model parameters and state-of-charge 2016;168:465–72. [2016/04/15/].
of LiFePO4 batteries in electric vehicles. Appl Energy 2012;89:413–20. [2012/01/ [42] Gong X, Xiong R, Mi CC. A data-driven bias correction method based lithium-ion
01/]. battery modeling approach for electric vehicles application. In: 2014 IEEE trans-
[17] Wei Z, Lim TM, Skyllas-Kazacos M, Wai N, Tseng KJ. Online state of charge and portation electrification conference and expo (ITEC); 2014. p. 1–6.
model parameter co-estimation based on a novel multi-timescale estimator for va- [43] Wang L, Cheng Y, Zhao X. A LiFePO4 battery pack capacity estimation approach
nadium redox flow battery. Appl Energy 2016;172:169–79. [2016/06/15/]. considering in-parallel cell safety in electric vehicles. Appl Energy
[18] Lin HT, Liang TJ, Chen SM. Estimation of battery state of health using probabilistic 2015;142:293–302. [2015/03/15/].
neural network. IEEE Trans Indust Inf 2013;9:679–85. [44] Zheng Y, Lu L, Han X, Li J, Ouyang M. LiFePO4 battery pack capacity estimation for
[19] Bai G, Wang P, Hu C, Pecht M. A generic model-free approach for lithium-ion electric vehicles based on charging cell voltage curve transformation. J Power Sour
battery health management. Appl Energy 2014;135:247–60. [2014/12/15/]. 2013;226:33–41. [2013/03/15/].
[20] Klass V, Behm M, Lindbergh G. A support vector machine-based state-of-health [45] Petzl M, Danzer MA. Advancements in OCV measurement and analysis for lithium-
estimation method for lithium-ion batteries under electric vehicle operation. J ion batteries. IEEE Trans Energy Convers 2013;28:675–81.
Power Sour 2014;270:262–72. [2014/12/15/]. [46] Eddahech A, Briat O, Vinassa J-M. Determination of lithium-ion battery state-of-
[21] You G-W, Park S, Oh D. Real-time state-of-health estimation for electric vehicle health based on constant-voltage charge phase. J Power Sour 2014;258:218–27.
batteries: a data-driven approach. Appl Energy 2016;176:92–103. [2016/08/15/]. [2014/07/15/].
[22] Widodo A, Shim M-C, Caesarendra W, Yang B-S. Intelligent prognostics for battery [47] Lee J, Nam O, Cho BH. Li-ion battery SOC estimation method based on the reduced
health monitoring based on sample entropy. Expert Syst Appl 2011;38:11763–9. order extended Kalman filtering. J Power Sour 2007;174:9–15. [2007/11/22/].
[23] Hu X, Li SE, Jia Z, Egardt B. Enhanced sample entropy-based health management of [48] Pérez G, Garmendia M, Reynaud JF, Crego J, Viscarret U. Enhanced closed loop
Li-ion battery for electrified vehicles. Energy 2014;64:953–60. 2014/01/01/. state of charge estimator for lithium-ion batteries based on extended Kalman filter.
[24] Hu X, Jiang J, Cao D, Egardt B. Battery health prognosis for electric vehicles using Appl Energy 2015;155:834–45. [2015/10/01/].
sample entropy and sparse Bayesian predictive modeling. IEEE Trans Indust [49] Xia B, Zhao X, de Callafon R, Garnier H, Nguyen T, Mi C. Accurate Lithium-ion
Electron 2016;63:2645–56. battery parameter estimation with continuous-time system identification methods.
[25] Waag W, Fleischer C, Sauer DU. Critical review of the methods for monitoring of Appl Energy 2016;179:426–36. [10/1/].
lithium-ion batteries in electric and hybrid vehicles. J Power Sour [50] Chiang Y-H, Sean W-Y, Ke J-C. Online estimation of internal resistance and open-
2014;258:321–39. [2014/07/15/]. circuit voltage of lithium-ion batteries in electric vehicles. J Power Sour
[26] Wei Z, Tseng KJ, Wai N, Lim TM, Skyllas-Kazacos M. Adaptive estimation of state of 2011;196:3921–32. [2011/04/15/].
charge and capacity with online identified battery model for vanadium redox flow [51] Zheng Y, Ouyang M, Li X, Lu L, Li J, Zhou L, et al. Recording frequency optimization
battery. J Power Sour 2016;332:389–98. [2016/11/15/]. for massive battery data storage in battery management systems. Appl Energy
[27] Zou Y, Hu X, Ma H, Li SE. Combined state of charge and state of health estimation 2016;183:380–9. [2016/12/01/].
over lithium-ion battery cell cycle lifespan for electric vehicles. J Power Sour [52] Garnier H, Mensler M, Richard A. Continuous-time model identification from
2015;273:793–803. [2015/01/01/]. sampled data: implementation issues and performance evaluation. Int J Control
[28] Chen C, Xiong R, Shen W. A lithium-ion battery-in-the-loop approach to test and 2003;76:1337–57.
validate multi-scale dual H infinity filters for state of charge and capacity estima- [53] Xia B, Shang Y, Nguyen T, Mi C. A correlation based fault detection method for
tion. IEEE Trans Power Electron 2017. [pp. 1-1]. short circuits in battery packs. J Power Sour 2017;337:1–10. [2017/01/01/].
[29] Plett GL. Recursive approximate weighted total least squares estimation of battery [54] Pei L, Wang T, Lu R, Zhu C. Development of a voltage relaxation model for rapid
cell total capacity. J Power Sour 2011;196:2319–31. [2011/02/15/]. open-circuit voltage prediction in lithium-ion batteries. J Power Sour
[30] Hu X, Li S, Peng H. A comparative study of equivalent circuit models for Li-ion 2014;253:412–8. [2014/05/01/].
batteries. J Power Sour 2012;198:359–67. [2012/01/15/]. [55] Yang J, Xia B, Shang Y, Huang W, Mi C. Improved battery parameter estimation
[31] Yang J, Xia B, Shang Y, Huang W, Mi CC. Adaptive state-of-charge estimation based method considering operating scenarios for HEV/EV applications. Energies
on a split battery model for electric vehicle applications. IEEE Trans Veh Technol 2017;10:5.
2017. [pp. 1-1]. [56] Williard N, He W, Osterman M, Pecht M. Comparative analysis of features for de-
[32] Xiong R, Tian J, Mu H, Wang C. A systematic model-based degradation behavior termining state of health in lithium-ion batteries. Int J Prognost Health Manage
recognition and health monitoring method for lithium-ion batteries. Appl Energy 2013;4.
2017;207:372–83. [2017/12/01/]. [57] Ramadass P, Haran B, Gomadam PM, White R, Popov BN. Development of first
[33] Chen Z, Mi CC, Fu Y, Xu J, Gong X. Online battery state of health estimation based principles capacity fade model for Li-ion cells. J Electrochem Soc
on Genetic Algorithm for electric and hybrid vehicle applications. J Power Sour 2004;151:A196–203.

1600