Journal of Power Sources 258 (2014) 228e237

Contents lists available at ScienceDirect

Journal of Power Sources

journal homepage: www.elsevier.com/locate/jpowsour

A unified open-circuit-voltage model of lithium-ion batteries for

state-of-charge estimation and state-of-health monitoringq
Caihao Weng a, *, Jing Sun a, Huei Peng b
a
Department of Naval Architecture and Marine Engineering, University of Michigan, Ann Arbor, MI 48109, USA
b
Department of Mechanical Engineering, University of Michigan, Ann Arbor, MI 48109, USA

h i g h l i g h t s

Proposed a unified OCV model for Li-ion battery SOC estimation and SOH monitoring.

Considered Li-ion staging phenomenon during intercalation in model parametrization.

The model improves SOC estimation accuracy compared to existing OCV models.

The model captures battery electrochemical properties at different aging stages.

Established relations between the model parameters and cell performance degradation.

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

Article history: Open-circuit-voltage (OCV) data is widely used for characterizing battery properties under different
Received 13 August 2013 conditions. It contains important information that can help to identify battery state-of-charge (SOC) and
Received in revised form state-of-health (SOH). While various OCV models have been developed for battery SOC estimation, few
6 January 2014
have been designed for SOH monitoring. In this paper, we propose a unified OCV model that can be
Accepted 7 February 2014
Available online 19 February 2014
applied for both SOC estimation and SOH monitoring. Improvements in SOC estimation using the new
model compared to other existing models are demonstrated. Moreover, it is shown that the proposed
OCV model can be used to perform battery SOH monitoring as it effectively captures aging information
Keywords:
Lithium-ion batteries
based on incremental capacity analysis (ICA). Parametric analysis and model complexity reduction are
Open-circuit-voltage also addressed. Experimental data is used to illustrate the effectiveness of the model and its simplified
State-of-charge version in the application context of SOC estimation and SOH monitoring.
State-of-health Ó 2014 Elsevier B.V. All rights reserved.
Incremental capacity analysis

1. Introduction applications [4]. Given the fact that batteries have been used to
power an increasingly diverse range of applications, from cell
The growing concern over oil shortage and environmental is- phones to electric vehicles, more reliable and efficient battery
sues has greatly accelerated the development of alternative power management systems (BMSs) need to be developed [5].
and energy solutions to displace fossil fuel in recent years [1e3]. As BMS has, among many functions, two important tasks, namely
an important component in the clean energy intuitive, energy the state-of-charge (SOC) estimation and state-of-health (SOH)
storage system is a critical enabler for the next generation power/ determination [6,7]. SOC is commonly defined as “the percentage
energy technologies. Among the various choices of energy storage of the maximum possible charge that is present inside a
technologies, it is well recognized that battery systems can offer a rechargeable battery”, and the estimation of SOC serves as the fuel
number of high-value opportunities given their advantages in gauge for batteries. On the other hand, SOH is “a ‘measure’ that
portability and energy efficiency, particularly for mobile reflects the general condition of a battery and its ability to deliver
the specified performance in comparison with a fresh battery” [8].
Typically, the quantitative definition of SOH is based either on the
battery capacity or the internal resistance, depending on specific
q This work is supported by USeChina Clean Energy Research Center-Clean
applications [7]. Due to the requirement for online estimation of
Vehicle Consortium (CERC-CVC).
both SOC and SOH during system operation, the development of an
* Corresponding author. Tel.: þ1 734 763 7963; fax: þ1 734 936 8820.
E-mail addresses: [email protected] (C. Weng), [email protected] (J. Sun), simple but accurate battery model is crucial for an effective and
[email protected] (H. Peng). robust BMS.

http://dx.doi.org/10.1016/j.jpowsour.2014.02.026
0378-7753/Ó 2014 Elsevier B.V. All rights reserved.
 C. Weng et al. / Journal of Power Sources 258 (2014) 228e237 229

Many studies of battery SOC estimation have been reported in based on the dV/dQ curve. The accuracy of DVA based capacity
the literature (see Ref. [9] and references therein). Most of those fading prediction has been shown in Refs. [26,27].
methods are model-based approach and require an accurate open- One major difficulty in performing ICA and DVA is the sensitivity
circuit-voltage (OCV) model which relates OCV to SOC [9e11]. The to noise in battery voltage measurement [7,21,23]. Since all the
OCVeSOC function is implemented in the BMS either as a look-up peaks on an IC curve lie within the flat region of the OCV curve,
table or an analytical expression, while the latter has several ad- computing the derivatives directly from the data could lead to
vantages including computational efficiency (since no interpolation inaccurate and undesirable results, even after careful data filtering
is needed) as well as the convenience for analysis. Because of the [7]. This issue could be addressed by implementing specially
wide flat region on the OCVeSOC curve for lithium iron phosphate designed high precision testing equipment [23,25]. However, for
batteries, a small mismatch in OCV fitting may cause a large devi- on-board application, the measurement precision is usually limited
ation in SOC estimation. Therefore, improving OCV models can and appropriate data processing is required.
contribute significantly to increasing the accuracy of SOC By constructing an OCV model that represents the underlying
estimation. physical phenomenon of lithium-ion intercalaction/dein-
In Ref. [12], several phenomenological OCV models, which are tercalation process and therefore captures the subtle transitions
built with curve fitting without considering the complex battery over the flat region, one can take advantage of the analytical
physical behavior during the lithium-ion intercalaction/dein- form of the model to extract useful information associated with
tercalation process [1,13], are summarized. Although those models battery electrochemical properties and aging status using ICA
represent part of the nonlinear characteristics of battery OCV, they [5,20]. The effect of measurement noise could also be mitigated
fail to capture the voltage plateaus and transitions (please see Fig. 1) by fitting the raw data with a specially structured parametric
that correspond to the staging phenomenon at the graphite anode model. Therefore, for both SOC estimation and SOH monitoring
[7,14e18]. through ICA, an OCV model that can capture the physical process
The modeling of OCVeSOC function is also found to be crucial over the flat region of the OCVeSOC curve and be capable of
when it comes to battery SOH monitoring, as OCV data often reflect supporting ICA is needed.
battery aging and performance degradation [19]. In particular, by In this paper, we propose a unified OCVeSOC model that is
differentiating the battery charged capacity (Q) with respect to the intended for both SOC estimation and SOH monitoring. The original
terminal voltage (V), the voltage plateaus on the OCV curve is idea was initially proposed in Ref. [37]. The study is based on a
transformed into clearly identifiable dQ/dV peaks on the incre- battery life cycle test data set collected from eight A123 APR18650
mental capacity (IC) curve (see Fig. 1) [7,14,18]. This so-called in- cells, which use LiFePO4 as the positive electrode material and
cremental capacity analysis (ICA) technique helps to detect gradual graphite as the negative electrode, over a period of 18 months
changes in cell behavior, based on OCV data collected from life cycle [7,28]. The test data used for this study are acquired through a
test, with greater sensitivity than those based on conventional battery test bench, which includes an Arbin BT2000 tester, a ther-
methods [5]. ICA is applicable to lithium ion batteries that use mal chamber for environment control, a computer for userema-
graphite as the negative electrode material since it originates from chine interface and data storage, a switch board for cable
the study of the lithium intercalation process [14e17]. The efficacy connection, and battery cells [7,28]. The data acquisition system has
of ICA has been shown with various lithium ion batteries (LiFePO4, a logging frequency of 10 Hz, and the measurement precision of
LiNMC, LNCAO, LiMn2O4, etc.) [7,18,20e23]. In Ref. [23], the aging both current and voltage is 0.02% (i.e., 1 mV for voltage
signature extracted using ICA is further amplified through the delta measurement).
differential capacity analysis technique, which enables people to The remainder of this paper is organized as follows. We first
detect small battery degradation in a shorter time interval propose the development and parametrization of the new OCV
[14,24,25]. One alternative way of extracting battery aging signa- model in Section 2. Then we compare the new OCV model with
ture and analyzing battery SOH from the voltage curve is the dif- other existing models in terms of OCV data fitting and SOC esti-
ferential voltage analysis (DVA). In contrast to the ICA, the DVA is mation accuracy in Section 3. In Section 4, we elucidate the appli-
cation of the OCV model for SOH monitoring, followed by
parametric analysis and model simplification in Section 6. The
4 conclusions are given in Section 7.
Measured Voltage (V)

3.5

2.5
0 0.2 0.4 0.6 0.8 1
SOC
Incremental Capacity (V )

30
−1

20

10

0
2.8 2.9 3 3.1 3.2 3.3 3.4 3.5
Measured Voltage (V)

Fig. 1. ICA performed on charging voltage curve at 1/20 C rate. Fig. 2. Fitting result with the proposed OCV model.
230 C. Weng et al. / Journal of Power Sources 258 (2014) 228e237

Fig. 3. OCV model fitted over SOC from 10% to 90%. Fig. 4. Comparison of OCV fitting results.

1 1
OCVðzÞ ¼ K0 þ K1 þ K2
2. A new open-circuit-voltage model 1 þ ea1 ðzb1 Þ 1 þ ea2 ðzb2 Þ
1 1
We propose a new OCV model structure that can capture the þ K3 þ K4 þ K5 z (1)
1 þ ea3 ðz1Þ 1 þ ea4 z
phenomenological characteristics associated with the lithium-ion
intercalaction/deintercalation process while fitting the OCVeSOC where z is the SOC, K0w5 are the linear parameters, and a1w4 and
data. A specific model structure is chosen to enforce the model to b1w2 are the nonlinear parameters. This parametrization shares
exhibit plateaus on the OCV curve. These plateaus are results from features with some of the electrochemical model of LiFePO4 pro-
the lithium-ion staging phenomena, and have been irrefutably posed in Ref. [1]. More specifically, our model can be interpreted as
observed from experimental data. The parameters of the proposed a simplified composition and generalization of the cathode and
model will be identified with experiment data. Instead of anode OCV functions in Ref. [1] (where hyperbolic and exponential
measuring the OCV after a long relaxation period at different SOC functions are used).
levels that span the entire range, the data used are collected by An example of a complete data set for model identification is
charging/discharging a LiFePO4 battery cell at low rate of 1/20 C shown in Fig. 2, where the collected charging/discharging data are
under room temperature. The voltage data obtained by this slow averaged to yield OCV data. The parameters in Eq. (1) are deter-
charging/discharging process reflects the OCV at a close-to- mined using the Matlab curve fitting toolbox. We focus on fitting
equilibrium status [18]. Although the close-to-equilibrium OCV the data in the middle SOC range, namely 10%e90% SOC, given that
curve shown in Fig. 2 may be affected by hysteresis and diffusion in practice only data in the middle SOC range are available. More-
voltage due to lack of relaxation, it can sufficiently represent the over, this allows us to achieve higher accuracy in the range where
generic electrochemical properties for ICA and quantifying capacity the model will be most useful. The fitted OCVeSOC curve is plotted
fading as presented in Ref. [18]. The upper and lower cut-off voltage in Fig. 3, with a reference made to the averaged measured OCV data.
limits are set to be 3.6 V and 2.1 V respectively. The total charge The rms and maximum values of fitting error is 1 mV and 2.5 mV,
stored between the upper and lower limits is defined as the battery reduced from 2.3 mV (rms) and 7 mV (maximum) that were ob-
capacity. tained when the full SOC range is used. Table. 1

2.1. Open-circuit-voltage parametric model 2.2. Model comparison and analysis

Based on the studies presented in Refs. [7,14,18], there are The new OCV model is compared in this section with the five
generally three observable voltage plateaus and two transitions, different models summarized in Ref. [12], where the polynomial
which are important in characterizing the OCV and battery capac- model is reported to be the most accurate. All the parameters in
ity, over the flat area on the OCV curve of LiFePO4 battery cells. In those five OCV models are refitted for the data presented in Fig. 2
our model, those plateaus and transitions are represented by the using Matlab curve fitting toolbox, the results with their rms and
sigmoid functions as follows, maximum errors are shown and summarized in Fig. 4 and Table 2. It

Table 1
OCV models summarized in Ref. [12] together with the new model.

# OCV models Ref.

1 OCV(z) ¼ K0  K1/z  K2z þ K3ln(z) þ K4ln(1  z) [32]

a2 
2 OCVðzÞ ¼ K0 þ K1 ð1  ea1 z Þ þ K2 1  e1z þ K3 z [33]

3 OCVðzÞ ¼ K0 þ K1 ea1 ð1zÞ  K2 =z [34]

4 OCVðzÞ ¼ K0 þ K1 ea1 z þ K2 z þ K3 z2 þ K4 z3 [35]
5 OCV(z) ¼ K0 þ K1z þ K2z2 þ K3z3 þ K4z4 þ K5z5 þ K6z6 [36]
6 OCVðzÞ ¼ K0 þ K1 1þea11 ðzb1 Þ þ K2 1þea12 ðzb2 Þ þ K3 1þea13 ðz1Þ þ K4 1þe1a4 z þ K5 z Proposed model
 C. Weng et al. / Journal of Power Sources 258 (2014) 228e237 231

Table 2 models #2 and #5, dQ/dV is not defined at 3.35 V because of a

Fitting results of OCV models from Table 1 singularity in the math expression. Consequently, spikes in both
Model# RMS error (mV) Max error (mV) directions are shown in Fig. 5 for the two IC curves, corresponding
1 5.2 12.6
to models #2 and #5.
2 3.1 8.9
3 5.6 21.3
3. State-of-charge estimation based on extended Kalman
4 4.7 12.7
5 2.1 7.3
filter
6 1.0 2.5
The extended Kalman filter (EKF) based approach discussed in
Ref. [29] is used to illustrate the implementation of the new OCV
model for SOC estimation. The battery test data for this study is
collected through the experimental set-up introduced in Ref. [28]
and the first-order RC model [30] is used for calculating the SOC
with the following algorithm:

RC Model :
zkþ1 ¼ f ðzk ; Ik Þ þ wk ;
Vk ¼ gðzk ; Ik Þ þ vk ;
EKF Algorithm :

vf ðzk1 ;Ik1 Þ
Ak1 ¼ vzk1  ;
z b
k1 ¼ z k1

k ;Ik Þ
Ck ¼ vgðz
vzk  ;
zk1 ¼b
z kjk1 (2)
z kjk1 ¼ f ðb
b z k1 ; Ik1 Þ;

Pkjk1 ¼ Ak1 Pk1 ATk1 þ Q ;

h i1
Fig. 5. Comparison of IC curves from fitted OCV.
Gk ¼ Pkjk1 CkT Ck Pkjk1 CkT þ V ;
h
i
zk ¼ b
b z kjk1 þ Gk Vk  g b z kjk1 ; Ik ;
should be noted that none of the five OCV models is suitable for IC
analysis, as their model structures do not take into account the Pk ¼ ½I  Gk Ck Pkjk1 ;
staging mechanism in intercalation/deintercalation process. One
can see that the new OCV model proposed in Eq. (1) has better where
fitting accuracy than all those five models. Consequently,


f ðzk ; Ik Þ ¼ zk  hQi Dc t Ik ;
improvement in the SOC estimation results may also be expected
when the new OCV model is incorporated.
The IC curves based on different OCV models are shown in Fig. 5. gðzk ; Ik Þ ¼ OCVðzk Þ þ Vrc;k þ R0 Ik ;
It can be observed that the new model (model #6) captures the two
 h
i
Vrc;kþ1 ¼ exp  Dt  Dt
IC peaks associated with staging, whereas all other models only R1 C Vrc;k þ R1 1  exp R1 C Ik ;
show one peak on the IC curves. Therefore, model #6 can extract
more capacity information from the OCV data compared to other zk is the SOC, hi is the charging/discharing efficiency, Dt is the
models and will be used for further analysis in SOH monitoring. For time step, Qc is the battery capacity, Ik is the input current, Vk is the
total terminal voltage, Vrc,k is voltage of the RC circuit, R0, R1 and C
are battery internal resistance and capacitance, Q is the error
covariance of process noises, and V is the error covariance of
observation noises [28]. The OCV function is represented by the
model given by Eq. (1).
In this study, all the parameters in the RC model are set to be
constants without performing iterative optimizations as discussed
in Ref. [12].
The battery data tested under the Federal Urban Driving
Schedule (FUDS) is shown in Fig. 6. As in Ref. [29], the SOC data
obtained by Coulomb counting in the Arbin battery test bench,
where high precision current sensor is used, is assumed to be the
“true” SOC and used as a reference for performance evaluation.
Therefore the SOC estimation error is defined as,

eSOC;k ¼ b
z k  zk (3)

where eSOC,k is the SOC estimation error, b

z k is the SOC estimated by
the EKF algorithm and zk is the reference SOC calculated from the
measured data. Please note that the SOC from Arbin Coulomb
Fig. 6. Battery data tested under FUDS. counting, while representing the best reference that we can get, is
232 C. Weng et al. / Journal of Power Sources 258 (2014) 228e237

Table 3
Parameters of models#5 and#6 used in SOC estimation.

Model parameters Model#5 Model#6

K0 3.0896 3.4002
K1 1.1627 0.0080
K2 2.3821 0.0785
K3 2.1870 0.2150
K4 0.5444 1.3032
K5 0.1939 0.0891
K6 0.0582 N/A
a1 N/A 14
a2 N/A 18
a3 N/A 28
a4 N/A 40

Table 4
Parameters of EKF based SOC estimator.

Parameters Values
 
Error covariance of process noises, Q 0:15 0 Fig. 8. V-SOC and IC curve from 0.5C charging data.
0 1
Error covariance of observation noises, V 0.01
The 5% error bound can be achieved with other initial error (e.g.,
20% or 30%).
only approximately accurate due to the unavoidable error and in- In summary, the new OCV model performs well in EKF based
tegral accumulation of the error in the current measurement [29]. SOC estimation. The model delivers accurate estimation results
The model parameters used in the EKF based SOC estimation are with less than 5% errors without extensive calibration and training.
shown in Table 3. The results presented in this section demonstrates the potential of
A detailed description of the EKF algorithm can be found in both implementing this new OCV model for SOC estimation problems.
Refs. [29] and [12]. The parameters of the EKF (i.e., V and Q used in
Eq. (2)) are calibrated using the data collected in our aging tests, 4. State-of-health monitoring based on incremental capacity
and the same parameters (listed in Table 4) are used in the per- analysis
formance evaluation for the new model (#6) and the polynomial
model (#5). Even though ICA was originally proposed for “close-to-equilib-
Fig. 7 displays the EKF based SOC estimation results. The two rium” conditions, it was shown in Ref. [7] that, for normal charging
plots represent two cases with different initial SOC error (þ10% data, the peaks on the IC curve can also be identified and they
and 10% respectively). One can see that the estimated SOC reveal significant information about battery SOH. Hence, the results
converge into the 5% estimation error bound when either model #5 presented in this section for ICA are all based on 0.5C battery
or #6 is used. In particular, with positive initial error, the SOC charging data (same as in Ref. [7]). It will be shown that the pro-
estimated with model #6 converges much faster than the SOC posed parametrization is applicable to both OCV data and normal
estimated with model #5. This difference in convergence rate with charging data (i.e., charged at 0.5C).
positive initial error can be explained by the curve fitting results As shown in Fig. 8, where the charging data is fitted with the
shown in Figs. 3 and 4, where model #6 have better accuracy than model given by Eq. (1) with OCV replaced by the terminal voltage
#5 in the high SOC region. On the other hand, both models have collected at normal charging rate, we can obtain the IC curve
approximately the same accuracy when SOC is below 70% and directly by differentiating the analytical V-SOC function and then
therefore have similar convergence rate with negative initial error. taking the reciprocal.

4.1. Incremental capacity analysis at different aging stages

The IC curves of battery charging data collected at different

aging cycles under 35  C environment condition are shown in Fig. 9.
The numerical values in the plot legends represent the aging cycle
number. The IC peaks highlighted by the circle in Fig. 9 clearly
shows a monotonic decreasing trend as battery ages, which implies
a battery degradation that is mainly related to the loss of active
material at the graphite anode [14,18].
The IC curves from four different cells are plotted in Fig. 10,
where the plots focus on the second peak on the IC curve and depict
a consistent trend in capacity fading as reflected by the decreasing
IC peaks.
To obtain the correlation between the cell capacity fading and IC
peak, the peak values of eight cells at different aging stages are
identified. The normalized peak values of one of the cells (red dots
in the web version in Fig. 11) is correlated with the faded cell ca-
pacities, and used to construct the battery capacity estimation
Fig. 7. SOC estimation with OCV model #6. curve (red dashed curve in Fig. 11) as done in Ref. [7]. The data from
 C. Weng et al. / Journal of Power Sources 258 (2014) 228e237 233

Fig. 9. IC curves of data at different aging stages at 35  C. Fig. 11. Correlation between faded battery capacity and IC peak values at 35  C.

other cells are also plotted to verify the prediction capability of the second peak shows the aging status. It suggests that the test data
estimation curve. The result shows the new parametric model in corresponding to higher temperature will be more reliable for ICA
the form of Eq. (1) can be used for capacity based battery SOH and capacity fading identification. Therefore, even though the
monitoring with good accuracy. temperature dependence is not parametrized explicitly, the pro-
posed model is capable of capturing the influence of temperature
4.2. Incremental capacity analysis at different temperature difference by adapting all parameters.

Test data were also collected at different temperature 5. Other signatures on the IC curve and their correlation with
throughout the battery aging process. ICA was performed with capacity fading
respect to different temperatures to evaluate and confirm the
sensitivity to temperature and results are presented in Fig. 12. The Other than using the IC peak value decreasing as the indication
intensities of the IC peaks reduce as temperature drops, which of battery capacity fading, we also explore a different approach that
could be attributed to the slow lithium-ion diffusion or lithium uses the IC peak locations to evaluate the degradation. Since the IC
plating phenomenon at low temperature as elucidated in Ref. [31]. peaks represent the phase transitions during the intercalaction/
When the temperature is too low, say at 10  C, the second peak deintercalation process and this electrochemical reaction is
could disappear from the IC curve. However, since the operating affected by battery capacity loss, the locations of the peaks might
temperature of battery systems are usually maintained through also have a correlation with the faded capacity and can give a good
integrated thermal control in the BMS, the disappearance of IC estimation of the loss.
peaks at low temperature should not be a concern. Consider the IC as a function of SOC (rather than measured
Fig. 13 presents the IC curves at different temperature at voltage), as shown in Fig. 14. Note that the first IC peak stays at
different aging stages. A temperature dependence of the IC peaks approximately the same location whereas the second peak is
and the model parameters may be evaluated. The first peak on moving towards the left as the battery ages and capacity decreases.
these four plots shows consistent temperature sensitivity while the By normalizing both the second IC peak location and battery ca-
pacity, we obtain the correlation between the two values in Fig. 15

Fig. 10. IC curves comparison from four different cells at 35  C. Fig. 12. IC curves of data at different temperature.
234 C. Weng et al. / Journal of Power Sources 258 (2014) 228e237

Fig. 15. Correlation between IC peak locations and battery faded capacities.

Fig. 13. IC curves of data at different temperature and ages.

The analysis suggests that the signatures on the IC curves,

(data from different cells are color coded differently (in the web namely the second IC peak and the distance between two peaks,
version)), which confirms that the location of the second IC peak can be used as indicators for battery aging. The two correlations
could also be a indication of battery capacity fading. based on IC peak locations are not as strong and robust compared to
However, in order to accurately measure this peak location in the results based on peak values shown in Fig. 11. Nevertheless, we
terms of charged capacity, it is required to start the battery charging believe that those correlations can provide useful information for
from zero SOC, which is not always practical in real operation. This battery SOH monitoring from different perspectives. It is also
problem can be addressed by using the SOC span between the two possible to improve the peak location based results when more
consecutive IC peaks, accurate current sensors are available.

dIC ¼ P2  P1 (4) 6. Parametric analysis and model simplification

where dIC is the SOC interval between two consecutive peaks, P1 Although the identification results using the new OCV model
and P2 are the locations of the two IC peaks respectively. The value has shown its effectiveness, the high number of parameters and
of dIC is more easily obtainable as long as the charging starts below nonlinearity of the model may present a challenge for on-board
the SOC value that corresponds to the first IC peak. Typically, the implementation. A parametric analysis is performed to charac-
first peak on the IC curve appears around SOC ¼ 0.45. Therefore, for terize the correlations between model parameters and battery ag-
all the data we have, this condition is satisfied as long as the ing, evaluate the sensitivity and robustness of the model structure,
charging covers 40%e85% SOC. The plot with normalized dIC and and better understand the utility of the proposed model. Moreover,
battery capacity is shown in Fig. 16. This correlation obtained using the analysis results may help to identify the significance of each
dIC is looser than using P2 only, as the values of both P1 and P2 are parameter and simplify the model, so that the computational effi-
affected by measurement noises and the value of dIC is more sus- ciency in estimation can be improved.
ceptible to noise and modeling error. Consider the following function, which is derived by differen-
tiating OCV with respect to the battery SOC,

Fig. 14. IC curve plotted versus charged capacity data at 35  C. Fig. 16. Correlation between IC peak location difference and battery faded capacities.
 C. Weng et al. / Journal of Power Sources 258 (2014) 228e237 235

d
GðzÞ ¼ OCVðzÞ
dz
a ea1 ðzb1 Þ a2 ea2 ðzb2 Þ
¼ K1  1 2  K2  2
1 þ ea1 ðzb1 Þ 1 þ ea2 ðzb2 Þ
a ea3 ðzb3 Þ a4 ea4 ðzb4 Þ
 K3  3 2  K4 þ K5 : (5)
1 þ ea3 ðz1Þ ð1 þ ea4 z Þ2

Note that the IC curve is the inverse of G(z), the locations and
values of IC peaks may be calculated by letting the derivative of G(z)
equal to zero and solving the resulting algebraic equation. The
resulting equation, however, is highly nonlinear and cannot be
easily solved analytically in general. Here, an alternative approach
that separates and analyzes the terms in the Eq. (5) individually is
used to correlate the model parameters and the solution of G(z).
To understand the contribution of each term of (5), we first
define the following,
Fig. 17. IC curves obtained using simplified OCV model at 35  C.
a1 ðzb1 Þ
f1 ðzÞ ¼ K1  a1 ea b 2 ;
1þe 1 ðz 1 Þ
Table 5
a2 ðzb2 Þ
f2 ðzÞ ¼ K2  a2 ea zb 2 ; Comparison of fitting accuracy between full and simplified model using RMS error.
1þe 2 ð 2Þ
(6) Cycle# Full model (mV) Simplified model (mV)
a3 ðz1Þ
f3 ðzÞ ¼ K3 a3 ea ðz1Þ 2 ; 1000 0.760 0.805
ð1þe 3 Þ 1400 0.925 0.980
a4 ðzÞ
K4 a4 e a4 z 2 :
1900 0.946 1.247
f4 ðzÞ ¼
ð1þe Þ 2300 1.024 1.559

Through analytical and numerical analysis, we have the

following observations:
parametric analysis. Moreover, it suggests that the values of the
deteriorating peaks in the IC curves might be estimated directly

f2 and f4 decay to zero at the high SOC range, where the IC peak of
from the parameters with the following relationship,
interest occurs. Therefore, they do not contribute to the IC peaks
and their locations. Our analysis will be focused on f1 and f3;
q ¼ q0 þ q1 a1 þ q2 K1 þ q3 a3 ; (8)

The parameter K5 is a constant and therefore it does not need to
be considered here; where q, defined as q ¼ p  p0/p0, represents the amount of

Numerical results show that K1 and K3 do not vary much degradation, p represents the IC peak values and p0 is the initial
compared to a1 and a3; peak value. q0w3 are the identification parameters in this equation.

Our analysis show that the IC peak has about equal sensitivity to K3 is not included in this equation as it is insensitive to the variation
a1 and a3. The sensitivity analysis is performed by numerically of IC peak values. Same as our analysis in the Subsection 4.2, the
evaluating the partial derivative of the IC function, defined by 1/ values of parameters q0w3 depend on the operating temperature.
G(z) with G(z) given by (5), with respect to the parameters a1 For on-board battery SOH diagnosis, thermal control should be
and a3 at the z value where the second IC peak is located. integrated as part of the BMS. Therefore, consistent temperature
could be expected when the SOH monitoring is performed.
With the above observations, we identified a simplified model Given a set of n data points, pi, a1i, K1i, a3i, i ¼ 1,.,n, from one
in the following form: battery cell at different aging stages, the values of q0w3 can be easily
calculated from the least squares method,
1 1
OCVðzÞ ¼ K1   þ K3 þ K0
1
a1 zb1 1 þ ea3 ðz1Þ
1þe q ¼ FT F FT q;
1 1 where
þ K2   þ K4 þ K 5 z: (7)
1þe a2 zb2 1 þ ea4 z q ¼ ½q0 ; q1 ; q2 ; q3 ; (9)
F ¼ ½f1 ; f2 ; .; fn T ;
where the model structure contains only 4 changing parameters q ¼ ½q1 ; q2 ; .; qn T ;
(i.e., K1, K3, a1 and a3), and other parameters are kept constants as fi ¼ ½1; a1i ; K1i ; a3i T :
batteries age.
In Fig. 17, one can see that the IC curves obtained using the
simplified OCV model show qualitatively the same trend as the
Table 6
ones obtained by the full model (see Fig. 9) when the battery ages.
Parameters of the simplified OCV model.
Table 5 shows the comparison of fitting accuracy between the full
and the simplified model. In the worst case, the rms error is less Cycle# a1 K1 a3 K3
than 2 mV. Therefore, the simplification preserves the high fitting 1000 12.57 0.0446 25.9352 0.2459
accuracy provided by the full model. 1400 12.29 0.0467 23.2749 0.2435
The parameters for the simplified model are listed in Table 6, 1900 11.57 0.0497 20.4778 0.2413
2300 11.4 0.0531 18.5991 0.2447
where one can see the variations in a1 and a3 agree with our
236 C. Weng et al. / Journal of Power Sources 258 (2014) 228e237

Fig. 18. Identification results of (8) using battery testing data from one single cell. Fig. 20. Correlation between capacity fading and IC peak reduction.

The identification results of (8) are very accurate as shown in 7. Conclusions

Fig. 18, where the battery testing data of one single cell (denoted as
cell A) at all different ages are included. In this paper, we propose a new OCV parametric model. The
To use the model (8) identified from cell A to 7 other cells, one special parametrization considers the staging phenomenon during
can use q1w3 directly while adjusting q0 such that the value of q the lithium intercalation/deintercalation process for lithium-ion
given by (8) at the initial stage is zero. Fig. 19 show that the data batteries, therefore leading to a much better fitting accuracy
from two other cells, as an example, can also be fitted accurately by when applied to experimental battery OCV data. The new para-
applying q1w3 without further identification effort. In other words, metric model is applied, together with an EKF, to SOC estimation
the IC peak values of different cells can be estimated/predicted and its effectiveness is demonstrated on FUDS cycle data. We also
based on the parameters q1w3 identified from one reference cell. show that the parametric model can be applied for ICA based SOH
Therefore it is possible to perform battery SOH monitoring directly monitoring. Through comparisons of IC curves, the model is shown
using (8) without numerically deriving the IC curves. to be capable of reflecting battery cell’s electrochemical properties
As shown in Fig. 20, the results from Figs. 11 and 18 can be at different operating temperature conditions and aging stages.
combined so that the battery capacity fading can be directly pre- Moreover, through parametric analysis, we are able to simplify the
dicted using the parameters of the simplified OCV model. One can model structure and develop a function that characterizes the re-
see that 50% reduction in IC peak value to equivalent to 10% battery lations between the model parameters and the degradation of
capacity loss. battery performance, which enables us to monitor battery SOH and
To summarize, the simplified OCV model is capable of capturing estimate capacity fading only based on the model parameters.
the battery aging features with fewer parameters, thereby Future research will focus on developing more efficient algorithms
improving the identification efficiency and making the model for the determination of the OCV model parameters.
implementable in on-board BMS. With the established linear
relationship between the IC peak values and the OCV model pa-
Acknowledgments
rameters in (8), the battery aging information can be revealed by
only looking at the parameters.
This work is supported by the Department of Energy under
Award Number DE-PI0000012.

References

[1] M. Safari, C. Delacourt, J. Electrochem. Soc. 158 (2011) A562eA571.

[2] M. Armand, J.M. Tarascon, Nature 451 (2008) 652e657.
[3] M.A. Roscher, D.U. Sauer, J. Power Sources 196 (2011) 331e336.
[4] B. Dunn, H. Kamath, J.-M. Tarascon, Science 6058 (2011) 928e935.
[5] B.Y. Liaw, M. Dubarry, in: G. Pistoia (Ed.), Electric and Hybrid Vehicles: Power
Sources, Models, Sustainability, Infrastructure and the Market, Elsevier, 2010,
pp. 375e403.
[6] E. Meissner, G. Richter, J. Power Sources 116 (2003) 79e98.
[7] C. Weng, Y. Cui, J. Sun, H. Peng, J. Power Sources 235 (2013) 36e44.
[8] V. Pop, H.J. Bergveld, D. Danilov, P.P.L. Regtien, P.H.L. Notten, Battery Man-
agement Systems: Accurate State-of-Charge Indication for Battery-Powered
Applications, first ed., Springer, 2008.
[9] S. Piller, M. Perrin, A. Jossen, J. Power Sources 96 (2001) 113e120.
[10] S. Santhanagopalan, R.E. White, J. Power Sources 161 (2006) 1346e1355.
[11] M. Verbrugge, E. Tate, J. Power Sources 126 (2004) 236e249.
[12] X. Hu, S. Li, H. Peng, F. Sun, J. Power Sources 217 (2012) 209e219.
[13] M. Safari, C. Delacourt, J. Electrochem. Soc. 158 (2011) A63eA73.
[14] J. Groot, State-of-Health Estimation of Li-ion Batteries: Cycle Life Test
Methods, Master’s thesis, Chalmers University of Technology, 2012.
Fig. 19. Verification of (8) using parameters q1w3 identified from Fig. 18 with data from [15] R. Yazami, P. Touzain, J. Power Sources 9 (1983) 365e371.
two other cells. [16] J.R. Dahn, Phys. Rev. B 44 (1991) 9170e9177.
 C. Weng et al. / Journal of Power Sources 258 (2014) 228e237 237

[17] R. Yazami, Y. Reynier, J. Power Sources 153 (2006) 312e318. [28] X. Hu, S. Li, H. Peng, J. Power Sources 198 (2012) 359e367.
[18] M. Dubarry, V. Svoboda, R. Hwu, B.Y. Liaw, Electrochem. Solid St. 9 (2006) [29] G.L. Plett, J. Power Sources 134 (2004) 277e292.
A454eA457. [30] M. Dubarry, N. Vuillaume, B.Y. Liaw, J. Power Sources 186 (2009) 500e507.
[19] M.A. Roscher, J. Assfalg, O.S. Bohlen, IEEE Trans. Veh. Technol. 60 (2011) 98e [31] J. Vetter, P. Novák, M. Wagner, C. Veit, K.-C. Möller, J. Besenhard, M. Winter,
103. M. Wohlfahrt-Mehrens, C. Vogler, A. Hammouched, J. Power Sources 147
[20] M. Dubarry, B.Y. Liaw, J. Power Sources 194 (2009) 541e549. (2005) 269e281.
[21] X. Feng, J. Li, M. Ouyang, L. Lu, J. Li, X. He, J. Power Sources 232 (2013) 209e218. [32] G.L. Plett, J. Power Sources 134 (2004) 262e276.
[22] X. Han, M. Ouyang, L. Lu, J. Li, Y. Zheng, Z. Li, J. Power Sources 251 (2014) 38e54. [33] Y. Hu, S. Yurkovich, Y. Guezennec, B. Yurkovich, J. Power Sources 196 (2011)
[23] A.J. Smith, J.R. Dahn, J. Electrochem. Soc. 159 (2012) A290eA293. 449e457.
[24] I. Bloom, A.N. Jansen, D. Abraham, J. Knuth, S.A. Jones, V.S. Battaglia, [34] D.E. Neumann, S. Lichte, in: NDIA Ground Vehicle Systems Engineering and
G.L. Henriksen, J. Power Sources 139 (2005) 295e303. Technology Symposium.
[25] A. Smith, A High Precision Study of Li-ion Batteries, Ph.D. thesis, Dalhousie [35] M. Chen, G.A. Rincón-Mora, IEEE Trans. Veh. Technol. 21 (2008) 504e511.
University, 2012. [36] A. Szumanowski, Y. Chang, IEEE Trans. Veh. Technol. 57 (2008) 1425e1432.
[26] K. Honkuraa, H. Honboa, Y. Koishikawab, T. Horibab, ECS Trans. 13 (2008) 61e73. [37] C. Weng, J. Sun, H. Peng, in: Proceedings of the ASME 2013 Dynamic Systems
[27] K. Honkuraa, K. Takahashia, T. Horibab, J. Power Sources 196 (2011) 10141e and Control Conference, 2013, Palo Alto, CA.
10147.