Journal of Power Sources 217 (2012) 209e219

Contents lists available at SciVerse ScienceDirect

Journal of Power Sources

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

Robustness analysis of State-of-Charge estimation methods for two types

of Li-ion batteries
Xiaosong Hu a, b, *, Shengbo Li a, c, Huei Peng a, Fengchun Sun b
a
Department of Mechanical Engineering, The University of Michigan, Ann Arbor, MI 48109, USA
b
National Engineering Laboratory for Electric Vehicles, Beijing Institute of Technology, No. 5 South Zhongguancun Street, Haidian District, Beijing 100081, China
c
State Key Laboratory of Automotive Safety and Energy, Tsinghua University, Beijing 100084, China

h i g h l i g h t s

< Analytic functions describing the battery model parameters are optimized.
< EKF based on the optimal analytic model is adopted as the SOC estimator.
< The robustness of the SOC estimator against varying loading profiles is evaluated.
< The robustness of the SOC estimator against varying temperatures is analyzed.
< The robustness of the SOC estimator against varying aging levels is assessed.

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

Article history: Battery State of Charge (SOC) estimation is an important function for battery management systems and
Received 18 April 2012 critical for the reliable operations of batteries. This paper analyzes the robustness of SOC estimation
Received in revised form algorithms for two types of Li-ion batteries under varying loading conditions, temperatures and aging
29 May 2012
levels. Based on the model templates identified in an earlier research, the model parameters are
Accepted 1 June 2012
Available online 9 June 2012
determined. The Extended Kalman Filter (EKF) technique is then adopted as the SOC estimation algo-
rithm. The robustness of the estimator against varying loading profiles and temperatures is evaluated
and compared against the Coulomb counting method. We subsequently used data from cells that have
Keywords:
Battery management systems
significantly aged to assess the robustness of the SOC estimation algorithm. Finally, the need for model
SOC estimation parameter updates is analyzed.
Li-ion battery Ó 2012 Elsevier B.V. All rights reserved.
Robustness analysis

1. Introduction The Coulomb counting method was often used as a core tech-
nology for battery SOC estimation [1e3]. This method is easy to
Electrified vehicles including battery electric vehicles (BEVs), implement but has three challenges. First, the initial SOC at key-on
hybrid electric vehicles (HEVs) and plug-in hybrid electric vehicles must be estimated accurately because subsequent estimates may
(PHEVs) have been actively studied and developed. A critical be biased by the initial SOC error [4]. Secondly, the method is highly
element for their successful commercialization is technologies for dependent on the accuracy of the current sensor. The current
reliable battery operations. Battery management systems (BMS) sensor is often subject to noise, drift induced by temperature and
have been designed to provide monitoring, diagnosis, and control other uncertainties. Finally, the battery capacity reduces with age
functions to enhance the operations of battery packs. A key func- and will affect SOC calculation. To remedy these three challenges,
tion of BMS is to accurately estimate battery state of charge (SOC). periodic resets are needed. In actual vehicle operations, the open
Poor SOC estimation can result in larger SOC swing than specified, circuit voltage (OCV) is often used to calculate initial SOC. Since
and can lead to reduced cycle life or lower efficiency. there is typically a monotonic relationship between OCV and SOC.
Inverting this algebraic relation leads to a SOC estimate. However,
error can exist if the battery is not fully relaxed. Sometimes the
* Corresponding author. National Engineering Laboratory for Electric Vehicles, OCVeSOC curve is very flat in the middle region (e.g., between 30
Beijing Institute of Technology, No. 5 South Zhongguancun Street, Haidian District, and 80% SOC), particularly for lithium iron phosphate cells. A small
Beijing 100081, China. Tel./fax: þ86 10 6891 4625. OCV error can result in a large SOC error. Hence, the OCV-based
E-mail addresses: [email protected], [email protected] (X. Hu).

0378-7753/$ e see front matter Ó 2012 Elsevier B.V. All rights reserved.
http://dx.doi.org/10.1016/j.jpowsour.2012.06.005
210 X. Hu et al. / Journal of Power Sources 217 (2012) 209e219

Table 1 Table 2
Equations of the best performing models identified in Ref. [24]. Candidate OCV functions evaluated in this paper.

Model Equations No. OCV functions and description

(1) First-order RC Ukþ1 ¼ expðDt=s1 ÞUk þ R1 ½1  expðDt=s1 ÞIk (1) Voc ¼ K0  K1/z  K2z þ K3 ln (z) þ K4 ln (1  z) [13]
model for LiNMC Vk ¼ Voc ðzk Þ  R0 Ik  Uk where Voc and z are the OCV and SOC, respectively.
cell where I, V and z are current, output voltage The optimization variable vector q ¼ [K0, K1, K2, K3, K4].
and SOC, respectively. Dt is the sampling time (2) Voc ¼ K0 þ a1 z þ a2 ð1  ea3 z Þ þ a4 ð1  ea5 =1z Þ [26]
and Voc is OCV. R0 is the internal ohmic where q ¼ [K0, a1, a2, a3, a4, a5].
resistance which depends on the current (3) Voc ¼ K0  a1 =z þ a2 ea3 ð1zÞ [27]
direction. U and s1 ¼ R1C1 are the voltage and where q ¼ [K0, a1, a2, a3].
time constant of the RC network. (4) Voc ¼ a1 ea2 z þ a3 þ a4 z þ a5 z2 þ a6 z3 [28]
(
(2) First-order RC Ukþ1 ¼ expðDt=s1 ÞUk þ R1 ½1  expðDt=s1 ÞIk where q ¼ [a1, a2, a3, a4, a5, a6].
model with hkþ1 ¼ expðjkIk DtjÞhk þ ½1  expðjkIk DtjÞH (5) Voc ¼ a1z6 þ a2z5 þ a3z4 þ a4z3 þ a5z2 þ a6z þ a7 [29]
one-state Vk ¼ Voc ðzk Þ  R0 Ik  Uk þ hk where q ¼ [a1, a2, a3, a4, a5, a6, a7].
hysteresis where h is the hysteresis voltage, k is a decaying
for LiFePO4 cell factor and H is the maximum amount of
hysteresis voltage which is positive for charge
and negative for discharge.

reset is not suitable for batteries that are mostly half-charged. For
EVs/PHEVs, the reset can be more effectively done, since the SOC
frequently reach the two ends where the OCVeSOC slope is steep.
Many SOC estimation models have been proposed, such as
artificial neural networks based models [5e7], fuzzy logic models
[8,9] and support vector regression (SVR) based models [10,11]. The
robustness of these models strongly relies on the quantity and
quality of the training data set. A limited training data set (e.g.,
obtained from new cells) may result in limited model robustness,
thus reducing the applicability of the model.
Kalman filter and other observer-based approaches have also
been used to estimate the battery SOC. These methods use output
feedback and can have better robustness than non-feedback
methods. In Refs. [12e14], an Extended Kalman filter (EKF), based
on several nonlinear state-space models, was used to estimate the
SOC of an HEV Li-polymer cell. A central difference Kalman filter Fig. 2. OCV fitting results for LiNMC cell at 22  C.
(CDKF) using a nonlinear enhanced self-correcting battery model
was also developed to estimate SOC [15,16]. An unscented Kalman
filter (UKF) by means of a nonlinear electrochemical battery model exploring different temperatures, battery ages, or highly transient
was also used to estimate SOC of a lithium-ion cell [17]. The loadings. In other words, the robustness of these SOC estimation
performance of these filters depends on the model accuracy. The algorithms was not sufficiently assessed. For example, many SOC
sigma-point based CDKF/UKF often provides better estimates than estimation approaches mentioned above were evaluated under
EKF at the expense of higher complexity and computational cost. only one battery loading profile and one environmental tempera-
Several variants of Kalman filter were also studied with similar ture. Moreover, the performance and robustness of these SOC
results [18e20]. Besides Kalman filters, sliding-mode observers algorithms against aging were not adequately studied. A key
[21,22] and linear parameter-varying observers [23] were also contribution of this paper is that the SOC estimator performance for
applied to predict the battery SOC. two types of Li-ion batteries was evaluated under different loading
Most of the estimation methods described above were validated profiles, temperatures, and cell aging levels.
using battery data under a narrow set of scenarios, without

Fig. 1. The test schedule to collect battery cell data. Fig. 3. OCV fitting results for LiFePO4 cell at 22  C.
 X. Hu et al. / Journal of Power Sources 217 (2012) 209e219 211

Table 3
Three candidate analytic functions for RC and hysteresis parameters.

Type Functions and description

(1) Polynomial [29] f ¼ cnzn þ cn1zn1 þ/þ c0
where f and z are model parameter and SOC,
respectively. n is the degree of the polynomial.
cn is a coefficient to be optimized.
(2) Exponential [28] f ¼ c1 ec2 z þ c3
(3) Power series [30] f ¼ c1 zc2 þ c3

Based on the preferred model structures identified in our earlier

work [24], the optimal analytic functions explicitly depicting the
dependency of model parameters on SOC and temperature are
determined. These models are then used in EKF to estimate SOC.
The robustness of the estimator under varying loading profiles,
temperatures and cell age are then analyzed.
The remainder of this paper is organized as follows. In Section 2,
determination of the optimal analytic functions to explicitly
delineate the model dependence on SOC and temperature is
described. In Section 3, the EKF-based SOC estimator is introduced.
The evaluation results of the estimator robustness are discussed in
Section 4, followed by conclusions in Section 5.

Fig. 5. Comparison results of polynomials with different degrees for the LiNMC cell.
2. Optimization of model dependency on SOC and
temperature

In Ref. [24], we studied twelve equivalent circuit battery

model structures reported in the literature, and compared their first-order RC model with one-state hysteresis, respectively. The
complexity, training and validation accuracies using test data equations of the two models are shown in Table 1. The aim of Ref.
from multiple cells. The preferred model structures for the [24] was to find the optimal model structures and the parameters
lithium nickelemanganeseecobalt oxide (LiNMC) UR14650P and of all the twelve models were assumed to be constant (inde-
lithium iron phosphate (LiFePO4) APR18650M1A cells were found pendent of SOC). To explore the full potential of these two
to be the first-order resistanceecapacitance (RC) model and the preferred model structures, in this paper we will establish the
dependence of model parameters on the battery SOC and
temperature. The test schedules shown in Fig. 1 are designed to
excite and age the two types of battery cells. The datasets are
described in details in Ref. [24]. The three dynamic test cycles
(Hybrid Pulse Test DST and FUDS) are used to identify the SOC
and temperature-dependent analytic functions.

Fig. 4. Schematic diagram of the two-step optimization process. Fig. 6. Comparison results of the three types of analytic functions for the LiNMC cell.
212 X. Hu et al. / Journal of Power Sources 217 (2012) 209e219

Table 4
Optimal coefficients of the analytic functions for the LiNMC cell.

Parameter Coefficients
OCV Voc ¼ a1z6 þ a2z5 þ a3z4 þ a4z3 þ a5z2 þ a6z þ a7 (0.1  z  0.9)
where
a1 ¼ f 0:1674ðTe  10Þ þ 2:9192; 10  Te < 22 ðTe : TemperatureÞ; 0:1939ðTe  22Þ þ 0:9109; 22  Te  35g
a2 ¼ f 0:3394ðTe  10Þ  5:8608; 10  Te < 22; 0:3675ðTe  22Þ  1:7875; 22  Te  35g
a3 ¼ f 0:1811ðTe  10Þ þ 1:3408; 10  Te < 22; 0:0469ðTe  22Þ  0:8330; 22  Te  35g
a4 ¼ f 0:0614ðTe  10Þ þ 5:1684; 10  Te < 22; 0:3023ðTe  22Þ þ 4:4320; 22  Te  35g
a5 ¼ f 0:0948ðTe  10Þ  4:4983; 10  Te < 22; 0:2336ðTe  22Þ  3:3606; 22  Te  35g
a6 ¼ f 0:0290ðTe  10Þ þ 1:5815; 10  Te < 22; 0:0648ðTe  22Þ þ 1:2331; 22  Te  35g
a7 ¼ f 0:0025ðTe  10Þ þ 3:5600; 10  Te < 22; 0:0063ðTe  22Þ þ 3:5901; 22  Te  35g
Charging R
0 R
0 ¼ c1 e
c2 z þ c (0.1  z  0.9)
3
where
c1 ¼ f 0:0116ðTe  10Þ þ 0:1728; 10  Te < 22; 0:0741ðTe  22Þ þ 0:0337; 22  Te  35g
c2 ¼ f 0:5632ðTe  10Þ  24:5426; 10  Te < 22; 1:0925ðTe  22Þ  17:7848; 22  Te  35g
c3 ¼ f 0:0009ðTe  10Þ þ 0:1086; 10  Te < 22; 0:0005ðTe  22Þ þ 0:0972; 22  Te  35g
Discharging Rþ
0 Rþ
0 ¼ c1 e
c2 z þ c (0.1  z  0.9)
3
where
c1 ¼ f 0:0072ðTe  10Þ þ 0:0761; 10  Te < 22; 0:0125ðTe  22Þ þ 0:1621; 22  Te  35g
c2 ¼ f 2:0975ðTe  10Þ  4:7917; 10  Te < 22; 0:6137ðTe  22Þ  29:9621; 22  Te  35g
c3 ¼ f 0:0010ðTe  10Þ þ 0:1142; 10  Te < 22; 0:0006ðTe  22Þ þ 0:1021; 22  Te  35g
R1 (RC network) R1 ¼ c1 ec2 z þ c3 (0.1  z  0.9)
where
c1 ¼ f 0; 10  Te < 22; 0; 22  Te  35g c2 ¼ f 0:1403ðTe  10Þ  21:9590; 10  Te < 22; 0:6098ðTe  22Þ  23:6430; 22  Te  35g
c3 ¼ f 0:0015ðTe  10Þ þ 0:0608; 10  Te < 22; 0:0011ðTe  22Þ þ 0:0433; 22  Te  35g
Time constant s1 s1 ¼ c1 ec2 z þ c3 (0.1  z  0.9)
(RC network) where
c1 ¼ f 2:1871ðTe  10Þ þ 19:3967; 10  Te < 22; 1:9460ðTe  22Þ þ 45:6424; 22  Te  35g
c2 ¼ f 4:1106ðTe  10Þ  0:6723; 10  Te < 22; 1:9550ðTe  22Þ  50:0000; 22  Te  35g
c3 ¼ f 0:5091ðTe  10Þ þ 32:7608; 10  Te < 22; 1:7591ðTe  22Þ þ 38:8700; 22  Te  35g

2.1. OCV function identification are shown in Figs. 2 and 3, respectively. It seems that all candidate
functions work reasonably well and the polynomial (5) matches the
Compared to lookup tables, analytic OCVeSOC functions can be OCV points a little better for both cells. The fitting results at tem-
beneficial. For example, function derivatives can be calculated more perature ¼ 10  C and 35  C are similar. It was argued in Ref. [25] that
easily and accurately. Several candidate functions proposed in the rich features in low or high SOC can be better fitted by a high-order
literature to depict the battery OCV are summarized in Table 2. We smooth polynomial. Therefore, function (5) in Table 2 is selected.
fit the five candidate functions shown in Table 2 to the 12 OCV
points obtained in [24] by nonlinear least-squares optimization.
2.2. Optimization process for the RC circuit and one-state hysteresis
The results of the LiNMC and LiFePO4 cells at temperature ¼ 22  C

In Ref. [24], the first-order RC model with one-state hysteresis

was found to fit the LiFePO4 cell results the best while the first-
order RC model was found to be the best for the LiNMC cells.
Three analytic functions commonly used to fit the RC parameters

Fig. 7. Comparison results of polynomials with different degrees for the LiFePO4 cell. Fig. 8. Comparison results of the three types of analytic functions for the LiFePO4 cell.
 X. Hu et al. / Journal of Power Sources 217 (2012) 209e219 213

are summarized in Table 3. The same basis functions are applied to algorithm. The optimization is conducted under each temperature.
the hysteresis parameters as well, to reduce the complexity in Then, we fit piecewise linear functions to the optimal solutions for
optimization. Since each model parameter is expressed as an the three temperatures, thereby establishing an explicit tempera-
explicit function with unknown coefficients, all are to be found by ture dependency. The OCV coefficients (sixth-degree polynomial)
optimization routines. Furthermore, unlike the model parameters are also optimized in the two-step procedure, together with other
(e.g., internal resistance), we often have no clear knowledge of the model parameters.
bounds of the unknown coefficients. The global multi-swarm
particle swarm optimization (MPSO) method [24] is used to opti- (1) LiNMC cell
mize these coefficients. However, its computation load is very
heavy. If gradient-based methods such as sequential quadratic The results using polynomial basis functions of different degrees
programming (SQP) are used, they frequently got trapped at local are shown in Fig. 5. The maximum, minimum and average RMS
minima. Therefore, a two-step optimization procedure is adopted. errors under three temperatures are computed. The hybrid pulse
The schematic diagram of the two-step optimization is shown in test dataset is used as the training dataset while the Dynamic Stress
Fig. 4. With loose bounds for all the unknown coefficients, 4000 Test (DST) and Federal Urban Dynamic Schedule (FUDS) datasets
generations in the global MPSO are first iterated such that the RMS [24] are used as validation datasets. The third-degree polynomial is
error is less than 30 mV. Then, the 3 best solutions (particles) selected based on accuracy-complexity trade-off. Fig. 6 shows the
attained by MPSO are used as the initial solutions of the SQP comparison results of the third-degree polynomial, exponential (2)

Table 5
Optimal coefficients of the analytic functions for the LiFePO4 cell.

Parameter Coefficients
OCV Voc ¼ a1z6 þ a2z5 þ a3z4 þ a4z3 þ a5z2 þ a6z þ a7 (0.1  z  0.9)
where
a1 ¼ f 0:4629ðTe  10Þ þ 1:9754; 10  Te < 22; 0:2506ðTe  22Þ  3:5795; 22  Te  35g
a2 ¼ f 1:0943ðTe  10Þ  3:6208; 10  Te < 22; 0:7233ðTe  22Þ þ 9:5106; 22  Te  35g
a3 ¼ f 0:8341ðTe  10Þ þ 0:0094; 10  Te < 22; 0:8529ðTe  22Þ  10:0000; 22  Te  35g
a4 ¼ f 0:1023ðTe  10Þ þ 4:3001; 10  Te < 22; 0:4813ðTe  22Þ þ 5:5273; 22  Te  35g
a5 ¼ f 0:1514ðTe  10Þ  3:7275; 10  Te < 22; 0:1121ðTe  22Þ  1:9108; 22  Te  35g
a6 ¼ f 0:0631ðTe  10Þ þ 1:3896; 10  Te < 22; 0:0049ðTe  22Þ þ 0:6326; 22  Te  35g
a7 ¼ f 0:0068ðTe  10Þ þ 3:0629; 10  Te < 22; 0:0002ðTe  22Þ þ 3:1440; 22  Te  35g
Charging resistance R
0 R 3 2
0 ¼ c1 z þ c2 z þ c3 z þ c4 (0.1  z  0.9)
where
c1 ¼ f 0:0148ðTe  10Þ  0:1670; 10  Te < 22; 0:0018ðTe  22Þ þ 0:0105; 22  Te  35g
c2 ¼ f 0:0221ðTe  10Þ þ 0:2458; 10  Te < 22; 0:0027ðTe  22Þ  0:0200; 22  Te  35g
c3 ¼ f 0:0098ðTe  10Þ  0:1027; 10  Te < 22; 0:0014ðTe  22Þ þ 0:0149; 22  Te  35g
c4 ¼ f 0:0016ðTe  10Þ þ 0:0349; 10  Te < 22; 0:0001ðTe  22Þ þ 0:0151; 22  Te  35g
Discharging Rþ
0 Rþ 3 2
0 ¼ c1 z þ c2 z þ c3 z þ c4 (0.1  z  0.9)
where
c1 ¼ f 0:0012ðTe  10Þ  0:0010; 10  Te < 22; 0:0106ðTe  22Þ  0:0153; 22  Te  35g
c2 ¼ f 0:0005ðTe  10Þ þ 0:0274; 10  Te < 22; 0:0163ðTe  22Þ þ 0:0333; 22  Te  35g
c3 ¼ f 0:0009ðTe  10Þ  0:0340; 10  Te < 22; 0:0071ðTe  22Þ  0:0231; 22  Te  35g
c4 ¼ f 0:0013ðTe  10Þ þ 0:0432; 10  Te < 22; 0:0014ðTe  22Þ þ 0:0277; 22  Te  35g
R1 (RC network) R1 ¼ c1z3 þ c2z2 þ c3z þ c4 (0.1  z  0.9)
where
c1 ¼ f 0:0272ðTe  10Þ  0:3042; 10  Te < 22; 0:0058ðTe  22Þ þ 0:0220; 22  Te  35g
c2 ¼ f 0:0424ðTe  10Þ þ 0:4777; 10  Te < 22; 0:0078ðTe  22Þ  0:0309; 22  Te  35g
c3 ¼ f 0:0226ðTe  10Þ  0:2715; 10  Te < 22; 0:0022ðTe  22Þ  0:0000; 22  Te  35g
c4 ¼ f 0:0054ðTe  10Þ þ 0:0956; 10  Te < 22; 0:0007ðTe  22Þ þ 0:0311; 22  Te  35g
Time constant s1 (RC network) s1 ¼ c1z3 þ c2z2 þ c3z þ c4 (0.1  z  0.9)
where
c1 ¼ f 0:1124ðTe  10Þ  4:3311; 10  Te < 22; 0:2546ðTe  22Þ  2:9824; 22  Te  35g
c2 ¼ f 0:4379ðTe  10Þ þ 1:1861; 10  Te < 22; 0:0584ðTe  22Þ  4:0682; 22  Te  35g
c3 ¼ f 0:2447ðTe  10Þ þ 6:2657; 10  Te < 22; 1:4771ðTe  22Þ þ 9:2026; 22  Te  35g
c4 ¼ f 0:9800ðTe  10Þ þ 35:2236; 10  Te < 22; 0:5832ðTe  22Þ þ 23:4637; 22  Te  35g
Hysteresis decaying factor k k ¼ c1z3 þ c2z2 þ c3z þ c4 (0.1  z  0.9)
where
c1 ¼ f 0:0379ðTe  10Þ  0:5478; 10  Te < 22; 0:0243ðTe  22Þ  0:0925; 22  Te  35g
c2 ¼ f 0:0521ðTe  10Þ þ 0:7931; 10  Te < 22; 0:0339ðTe  22Þ þ 0:1675; 22  Te  35g
c3 ¼ f 0:0188ðTe  10Þ  0:3157; 10  Te < 22; 0:0124ðTe  22Þ  0:0903; 22  Te  35g
c4 ¼ f 0:0015ðTe  10Þ þ 0:0410; 10  Te < 22; 0:0011ðTe  22Þ þ 0:0233; 22  Te  35g
Maximum charging hysteresis H H ¼ c1z3 þ c2z2 þ c3z þ c4 (0.1  z  0.9)
where
c1 ¼ f 0:0248ðTe  10Þ þ 0:9525; 10  Te < 22; 0:0725ðTe  22Þ þ 0:6554; 22  Te  35g
c2 ¼ f 0:0306ðTe  10Þ  1:3109; 10  Te < 22; 0:1021ðTe  22Þ  0:9440; 22  Te  35g
c3 ¼ f 0:0085ðTe  10Þ þ 0:4424; 10  Te < 22; 0:0331ðTe  22Þ þ 0:3399; 22  Te  35g
c4 ¼ f 0:0005ðTe  10Þ þ 0:0060; 10  Te < 22; 0:0018ðTe  22Þ þ 0:0000; 22  Te  35g
Maximum discharging hysteresis Hþ Hþ ¼ c1z3 þ c2z2 þ c3z þ c4 (0.1  z  0.9)
where
c1 ¼ f 0:0415ðTe  10Þ  0:0848; 10  Te < 22; 0:0091ðTe  22Þ þ 0:4134; 22  Te  35g
c2 ¼ f 0:0609ðTe  10Þ  0:0169; 10  Te < 22; 0:0165ðTe  22Þ  0:7478; 22  Te  35g
c3 ¼ f 0:0243ðTe  10Þ þ 0:0944; 10  Te < 22; 0:0065ðTe  22Þ þ 0:3857; 22  Te  35g
c4 ¼ f 0:0018ðTe  10Þ  0:0392; 10  Te < 22; 0:0005ðTe  22Þ  0:0613; 22  Te  35g
214 X. Hu et al. / Journal of Power Sources 217 (2012) 209e219

Table 6
EKF equations for the SOC estimation.

Nonlinear battery model

xkþ1 ¼ A(zk, Te)xk þ B(zk, Te)uk þ uk
yk ¼ g(xk, Ik) þ nk
where x, I and y are state variables, current and voltage, respectively.
The first element of x is SOC z. u is the model input. A and B are SOC
and temperature-dependent matrices. u and n are assumed to be
independent, zero-mean, Gaussian noise processes
of covariance matrices Q and V.
(1) Initialization

Assign the initial state estimate b

x o , error covariance P0, Q and V.
(2) Prediction

(
b z k1 ; TeÞ b
x kjk1 ¼ Aðb x k1 þ Buk1
P kjk1 ¼ Aðb
z k1 ; TeÞP k1 Aðbz k1 ; TeÞT þ Q
(3) Correction

(G ¼ P kjk1 Cðb
z kjk1 ; TeÞT ½Cðb z kjk1 ; TeÞT þ W1
z kjk1 ; TeÞP kjk1 Cðb
k
Fig. 9. OCV function of the LiNMC cell. b
xk ¼ b
x kjk1 þ Gk ½yk  gð b x kjk1 ; Ik Þ
Pk ¼ ½E  Gk Cðb z k ; TeÞP kjk1
where
and power series (3). The results are all very similar (i.e., they all vgðxk ; Ik Þ
Cðb
z kjk1 ; TeÞ ¼ jx ¼bx
work reasonably well). The exponential function shows slightly vxk k kjk1

better validation accuracy (especially in FUDS data) and is quite

simple. Therefore, the exponential function is used to describe the
RC parameters of the first-order RC model for the LiNMC cells. The Table 7
optimal coefficients of the analytic functions are summarized in Matrices of the state-space models for the LINMC and LiFePO4 cells.
Table 4. (1) The first-order RC model of the LiNMC cell

2 3
(2) LiFePO4 cell " # hDt
1 0  6 7
z Dt 3600Cn
xk ¼ ½ k , Aðzk ; TeÞ ¼ , Bðzk ; TeÞ ¼ 6
4
 7,
Uk 0 exp Dt 5
Polynomials of different degrees for the LiFePO4 cell are shown s1 R1  R1 exp
s1
in Fig. 7. Fig. 8 shows the performance of the third-degree poly-
nomial, exponential, and power series models. It was found that the
third-degree polynomial function shows the best balance between gðxk ; Ik Þ ¼ Voc;k  Uk  Ikþ Rþ  
0  Ik R0
where U is the voltage across the RC network at time k. Dt is the sampling time,
training/validation accuracies and complexity. The optimized
h is Coulombic efficiency, and Cn is the cell capacity. Iþ and I are discharging
coefficients of the analytic functions for the LiFePO4 cell (the first- and charging currents, respectively. When I  0 (charging), Iþ ¼ 0, I ¼ I;
order RC model with one-state hysteresis) are summarized in when I > 0 (discharging), I ¼ 0, Iþ ¼ I.
Table 5. The OCV functions of the two types of Li-ion batteries are (2) The first-order RC model with hysteresis for the LiFePO4 cell
shown in Figs. 9 and 10. It is obvious that the OCV of the LiFePO4 cell 2 3
1 0  0
between 25% and 85% SOC is much flatter than that of the LiNMC zk
6 0 exp Dt 7
xk ¼ ½ Uk , Aðzk ; TeÞ ¼ 4 0 5,
cell. The exact dependence of the model parameters on SOC and s1
hk 0 0 expðjkIk DtjÞ
temperature are unclear due to the complex reactions and
2 3
dynamics inside the cells [26]. Furthermore, since we use lumped hDt
6 0 7
models to approximate the currentevoltage relationships, the 6 3600Cn  7
Bðzk ; TeÞ ¼ 66 R1  R1 exp Dt
7,
7
model parameters are just estimates with no clear physical 4 0 5
s 1
0 1  expðjkIk DtjÞ

   
I I
u ¼ ; I  0; ; I > 0g, gðxk ; Ik Þ ¼ Voc;k  Uk þ hk  Ikþ Rþ  
0  Ik R0 .
H Hþ

meanings. Nevertheless, the models were able to represent the

lumped behavior of the battery well. With the RMS error around
merely 5 mV, the achieved analytic functions can be used to predict
battery voltage with adequate accuracy. It should be emphasized,
however, that the model error in the validation dataset is much
higher. This is perhaps because the hysteresis model was trained

Table 8
Assumed initial SOC error of Coulomb counting for the
LiNMC cell.

10  C 22  C 35  C
3.74% 3.09% 3.28%
Fig. 10. OCV function of the LiFePO4 cell.
 X. Hu et al. / Journal of Power Sources 217 (2012) 209e219 215

Table 9
Assumed initial SOC error of Coulomb counting for the LiFePO4
cell.

10  C 22  C 35  C
6.30% 5.41% 9.65%

using less-transient dynamic dataset and during the more signifi-

cant transient the model error increases.

3. EKF-based SOC estimator

EKF is a widely used state estimation method for nonlinear

dynamical systems. Based on the optimal models presented in
Section 2, EKF is adopted to estimate the battery SOC. The EKF
equations for the SOC estimation are summarized in Table 6. The
matrices for the state-space models for the LINMC and LiFePO4 cells
are summarized in Table 7.

4. Robustness of the SOC estimator

To better evaluate the robustness of the SOC estimator in an

environment close to realistic automotive environment, artificial
Gaussian noise is added to the current collected by the Arbin tester.
The mean and standard deviation of the noise are 0.01 A and
0.06 A (around 2% of the full range), respectively.

4.1. Robustness against different loading profiles and temperatures

The hybrid pulse test, DST and FUDS datasets collected under
Fig. 11. SOC estimation errors of the LiNMC cell under 22  C (before aging), (a) hybrid
pulse test; (b) DST test; (c) FUDS test.
three different temperatures for new cells are used in this subsec-
tion. The data below 20% SOC is discarded, since the battery SOC is
seldom permitted to be less than 20% in actual driving. To better

Fig. 12. RMS SOC errors under three temperatures for the LiNMC cell (new cell).
216 X. Hu et al. / Journal of Power Sources 217 (2012) 209e219

evaluate the SOC estimator, its performance is compared against

the Coulomb counting method.

(1) LiNMC cell

The OCV is often used to determine the initial SOC of Coulomb

counting using the OCVeSOC relationship. The OCV error thus leads
to the initialization error for Coulomb counting. Based on the fact
that the practically available cell-level precision is around 0.1% [31],
the OCV error incurred by voltage sensor is thus around 4 mV.
Furthermore, in practice OCV measurement is subject to electro-
magnetic interference, switching of high-current loads in nearby
wiring and insufficient relaxation. Sometimes insufficient relaxa-
tion can induce a voltage measurement error of 20 mV [32]. Based
on these considerations, we assume that the absolute measure-
ment error of the cell OCV is 8 mV. The maximum initial SOC error
of the Coulomb counting method can be algebraically calculated
based on this assumed OCV error and the SOCeOCV curve. The
results are summarized in Table 8.
The initial error of the EKF is assumed to be 10%. The initial
values of the EKF prediction and correction gain matrices are fixed
for all cases. Note that both the EKF and Coulomb counting use the
corrupted current measurement in their SOC estimation. Fig. 11
shows the SOC estimation results for 22  C. It can be seen that
the EKF-based estimator efficiently compensates for the initial
error and has a better accuracy than Coulomb counting for both the
training and validation datasets. The biased current noise has
a larger negative impact to the Coulomb counting method than to
the EKF. The results under 10  C and 35  C are similar. To more
Fig. 13. SOC estimation errors of the LiFePO4 cell under 22  C (before aging), (a) hybrid clearly compare the overall performance of the EKF and Coulomb
pulse test; (b) DST test; (c) FUDS test. counting, the RMS SOC estimation errors under three temperatures
are shown in Fig. 12. It is obvious that the EKF SOC estimator has
better accuracy. Since the model has an inferior accuracy in the
validation data (DST/FUDS dataset excluding the data below 20%
SOC) to that in the training data (hybrid pulse test dataset excluding

Fig. 14. RMS SOC errors under three temperatures for the LiFePO4 cell (new cell).
 X. Hu et al. / Journal of Power Sources 217 (2012) 209e219 217

the LiNMC cell, due to the flatter OCVeSOC curves. The SOC esti-
mation errors under 22  C are shown in Fig. 13. The RMS SOC errors
under three temperatures are shown in Fig. 14. The results are
similar to those of the LiNMC cells. Further, we can see that there is
an obvious difference in estimation errors between the training and
validation datasets. Due to the flatter OCV, the SOC estimation of
the LiFePO4 cell is more susceptible to model error. The validation
errors are around 2.5%, higher than the 1% accuracy achieved on the
LiNMC cells.

4.2. Robustness against cell aging

The battery cells always degrade with time in real operations.

The degradation is manifested as capacity or power losses. The
capacity loss is partially due to loss of cyclable Li Ions caused by
multiple factors, such as cathode structure degradation, side reac-
tions, passivation form and lithium plating at the anode [33]. The
Fig. 15. Normalized capacities of the LiNMC cell in the aging process. power loss is mainly due to the increase in cell resistance. It is
important to study the robustness of the SOC estimator since
robustness against cell aging is of great significance for BMS of
electrified vehicles.
the data below 20% SOC), the error of SOC estimation in the vali-
dation data is always higher. The estimation errors in the validation (1) LiNMC cell
datasets under three temperatures are around 1%, indicating a good
robustness against varying loading profiles and temperatures. Fig. 15 shows the calibrated capacities under three temperatures
in the degradation process of the LiNMC cell at 22  C. Here, we used
(2) LiFePO4 cell data up to cell age of 342 cycles at 22  C to analyze the robustness of
the EKF estimator. The capacity reduced by about 10% at 10  C when
The initial SOC error of Coulomb counting for the LiFePO4 cell is the cell is 342 cycles old at 22  Cdwhich is roughly half way
determined using the same method as the LiNMC cell. Again through the typical life expectancy accepted for automotive appli-
assuming a voltage measurement error of 8 mV, the results for the cations (a battery pack is determined to be “at the end of its life”
LiFePO4 cell are shown in Table 9. The errors are larger than those of when its capacity loss reaches 20%). The three characterization test

Fig. 16. Comparison results of a sequence of estimators for the LiNMC cell (after aging) with different level of model parameter update.
218 X. Hu et al. / Journal of Power Sources 217 (2012) 209e219

Fig. 17. Comparison results of a sequence of estimators for the LiFePO4 cell (after aging) with different level of model parameter update.

datasets right after 342 cycles are used to evaluate the SOC esti- significantly, its dynamic behavior changes noticeably. The RMS
mator. The initial SOC error and tuning parameters of the EKF are errors in the training datasets (hybrid pulse test) are worse than the
kept unchanged when we compare a sequence of estimators with LiNMC cell, due to the very flat OCVeSOC curve of the LiFePO4 cell.
increasing level of parameter updates. Fig. 16 shows the compar- The RMS errors in the validation datasets (DST and FUDS) are
ison results. It can be seen that if no parameters, including the similar to the LiNMC cell. The nominal model of the LiFePO4 cell is
capacity Cn, are updated, the estimator performance deteriorates less accurate after degradation. Also, updating internal resistance
significantly. The RMS errors of the SOC estimation are clearly R0 improves the robustness of the SOC estimator noticeably.
larger than those shown in Fig. 12. Updating internal ohmic resis- Updating the RC network and capacity does not help much, since
tance R0 can improve the SOC estimation in most cases. Updating capacity loss is small. If all the model parameters are updated, the
parameters of the RC network in addition to R0 does not help much. SOC estimator has the best robustness based on the overall evalu-
If we update all model parameters, including the cell capacity Cn, ation of all the 9 datasets. In practice, however, it is challenging to
the RMS error can be reduced to below 1% in most cases. update all the model parameters. Considering both the model
It should be noted that updating some of the model parameters accuracy and requirement for parameter update, we concluded that
is quite easy to do. For example, the internal ohmic resistance R0 updating the internal resistance is the most cost-effective update
can be easily obtained even under simple pulse loads. This simple one can use to improve the robustness of SOC estimation method
update reduces model prediction error by more than half. Updating against cell aging.
RC parameters or the cell capacity is harder onboard. It may be
possible to obtain those model parameters during charging process, 5. Conclusions
when the charging current can be manipulated and more compu-
tation power is available. The robustness of SOC estimation algorithms of two types of Li-
ion batteries is studied against varying loading profiles, tempera-
(2) LiFePO4 cell tures and aging levels. The preferred model structures determined
in our previous work [24] for the LiNMC and LiFePO4 cells are used.
The LiFePO4 cells we tested show much slower degradation. We The optimal functions of SOC and temperature explicitly describing
have finished 2600 cycles and the capacity loss at 22  C is only the model parameters are selected from commonly used generic
around 3%. The three characterization test datasets collected right functions for battery modeling. The sixth-degree polynomial is
after 2587 cycles are used. The RMS errors of the estimators with found to depict OCVs of both cells well. The exponential function is
different levels of model parameter updates are shown in Fig. 17. determined to be the best for the RC parameters of the first-order
Again, the estimator without parameter update shows poor per- RC model for the LiNMC cell; the third-degree polynomial is
formancedthis is because while the cell capacity does not reduce preferred for the RC and hysteresis parameters of the first-order RC
 X. Hu et al. / Journal of Power Sources 217 (2012) 209e219 219

model with hysteresis for the LiFePO4 cell. The coefficients of these [5] C.C. Chan, E.W.C. Lo, W. Shen, J. Power Sources 87 (2000) 201e204.
[6] W. Shen, K.T. Chau, C.C. Chan, IEEE Trans. Veh. Technol. 54 (2005)
generic functions are optimized by a combined MPSO and SQP
1705e1712.
procedure. Based on the optimal model parameters, the Extended [7] B. Cheng, Z. Bai, B. Cao, Energy Convers. Manage. 49 (2008) 2788e2794.
Kalman Filter technique is used to estimate the battery SOC. Given [8] A.J. Salkind, C. Fennie, P. Singh, T. Atwater, D.E. Reisne, J. Power Sources 80
the training datasets from new cells, the robustness of the SOC (1999) 293e300.
[9] P. Singh, R. Vinjamu, X. Wang, D. Reisne, J. Power Sources 162 (2006)
estimator against varying loading profiles and temperatures is 829e836.
evaluated and compared against the Coulomb counting method. [10] T. Hansen, C. Wang, J. Power Sources 141 (2005) 351e358.
The EKF-based estimator performs better than the Coulomb [11] Q. Shi, C. Zhang, N. Cui, Int. J. Automot. Technol. 9 (2008) 759e764.
[12] G.L. Plett, J. Power Sources 134 (2004) 252e261.
counting method, and the RMS estimation errors under validation [13] G.L. Plett, J. Power Sources 134 (2004) 262e276.
datasets under three temperatures are around 1% and 2.5% for the [14] G.L. Plett, J. Power Sources 134 (2004) 277e292.
LiNMC and LiFePO4 cells, respectively. Due to a flatter OCVeSOC [15] G.L. Plett, J. Power Sources 161 (2006) 1356e1368.
[16] G.L. Plett, J. Power Sources 161 (2006) 1369e1384.
surface, the SOC estimation of the LiFePO4 cell is more vulnerable to [17] S. Santhanagopalan, R.E. White, Int. J. Energy Res. 34 (2010) 152e163.
cell aging and if the model is not updated, its performance was [18] J.Y. Han, D.C. Kim, M. Sunwoo, J. Power Sources 188 (2009) 606e612.
found to be unacceptable. Updating the internal ohmic resistance of [19] J. Wang, J. Guo, L. Ding, Energy Convers. Manage. 50 (2009) 3182e3186.
[20] F. Sun, X. Hu, Y. Zou, S. Li, Energy 36 (2011) 3531e3540.
the nominal model improves the performance of the estimator [21] I.S. Kim, J. Power Sources 163 (2006) 584e590.
significantly (by about half). Updating all the model parameters [22] I.S. Kim, IEEE Trans. Power Electron 23 (2008) 2027e2034.
achieves even better performance however this may not be easy to [23] Y. Hu, S. Yurkovich, J. Power Sources 198 (2012) 338e350.
[24] X. Hu, S. Li, H. Peng, J. Power Sources 198 (2012) 359e367.
achieve in real-time driving.
[25] Y. Hu, S. Yurkovich, Y. Guezennec, B.J. Yurkovich, Control Eng. Pract. 17 (2009)
1190e1201.
Acknowledgment [26] Y. Hu, S. Yurkovich, Y. Guezennec, B.J. Yurkovich, J. Power Sources 196 (2011)
449e457.
[27] D.E. Neumann, S. Lichte, A Multi-Dimensional Battery Discharge Model with
This work was supported by the Automotive Research Center Thermal Feedback Applied to a Lithium-ion Battery Pack, NDIA Ground
(ARC), a U.S. Army center of excellence in modeling and simulation Vehicle Systems Engineering and Technology Symposium-Modeling & Simu-
of ground vehicles. lation, Testing and Validation (MSTV) Mini-Symposium, 2011.
[28] M. Chen, G.A. Rincón-Mora, IEEE Trans. Energy Convers. 21 (2006)
504e511.
References [29] A. Szumanowski, Y. Chang, IEEE Trans. Veh. Technol. 57 (2008) 1425e1432.
[30] MATLAB, Curve Fitting ToolboxÔ 2, User’s Guide, The MathWorks Inc., 2010.
[1] J.H. Aylor, A. Thieme, B.W. Johnson, IEEE Trans. Ind. Electron 39 (1992) [31] http://www.maxim-ic.com/MAX17830.
398e409. [32] J. Qian, Rechargeable Battery Characteristics, Safety, Charging and Fuel
[2] T. Liu, D. Chen, C. Fang, Int. J. Electron 87 (2000) 211e226. Gauges, Technical Presentation, Texas Instruments Inc., 2009.http://focus.ti.
[3] Y. Çadırcı, Y. Özkazanç, J. Power Sources 129 (2004) 330e342. com/asia/download/TechDay09cn_Analog_Power_1_Battery_chemistry_fuel_
[4] Y. Hu, Identification and state estimation for linear parameter varying systems gauge.pdf.
with application to battery management system design, Ph.D. dissertation, [33] M. Alamgir, A.M. Sastry, Efficient batteries for transportation applications, SAE
Ohio State University (2010). paper (08CNVG-0036), 2008.