2018 23rd International Scientific-Professional Conference on Information Technology (IT) Žabljak, Montenegro

State-of-Charge Estimation of Lithium-ion

Batteries using Extended Kalman filter and
Unscented Kalman filter
Ivan Jokić, Student Member, IEEE, Žarko Zečević, Member, IEEE and Božo Krstajić, Member, IEEE

 considered: Extended Kalman filter and Unscented

Abstract — Lithium-ion battery stands for vital segment of Kalman filter. Their SoC estimates are compared under
the hybrid-electrical vehicles (HEV). Accurate monitoring of two simulation tests.
battery status, which is the main task of battery management In Section II adopted model for Li-ion battery is
system (BMS), ensures reliable operation and preserve
battery performance. Successful management of battery relies
discussed. Also some methods for parameter estimation are
on accurate estimation of state-of-charge (SoC) parameter. mentioned. General overview of standard Kalman filter is
Proposed methods for SoC estimation tend to consider presented in Section III. Furthermore, EKF and UKF are
nonlinear nature of battery system with time and temperature explained, respectively, together with equations and block
dependent parameters. Kalman-based filters have been diagrams. Conducted simulation tests can be found in
widely used for SoC estimation. In this paper, two Kalman- Section IV together with comments on their results, while
based filters have been used fo SoC estimation, Extended
Kalman filter (EKF) and Uscented Kalman filter (UKF). general comment on paper is given in Section V.
These two methods have been compared and simulation II.BATERRY MODEL AND SOC ESTIMATION
results are presented.
Keywords — BMS, EKF, Extended Kalman filter, Battery model used in this paper is presented on figure 1
Lithium–ion battery, SoC estimation, UKF, Unscented and has been adopted from [1]. Its parameters are
Kalman filter. estimated using measurements and implementing Least
Square approach. Parameter values are given in table 1.
I. INTRODUCTION R1 R2

E
R0
LECTRIC vehicles (EVs) have become widely accepted
alternative to commonly used technologies in C1 C2
automotive industry [1]. In that manner, battery system as - + - +
v1 v2 v0
a most expensive part of EVs have been researched
+
thoroughly. Compared with other types of batteries, E0
-
Lithium-ion battery can storage considerably amount of i
energy and has no memory effect [2], [3]. Knowing that
battery system is the most expensive part of EVs, Fig. 1. Equivalent circuit model
reliability of its functioning became an imperative. BMS A. SoC estimation
coordinates with thousands of Li-ion battery cells inside SoC parameter is a state variable in adopted model and is
EV with aim to monitor battery vital characteristics and to defined as a ratio of available and total battery capacity:
contribute safe and reliable operation of vehicles [3], [4]. t

Battery performance can be affected by many factors, such s(t ) s(t0 )  K ³ i (W )dW , (1)
as operating temperature and aging of the battery. State of t0

Charge (SoC) is recognized as a vital parameter of battery. with Ʉ=1/(3600C), where C is battery capacity and its
Its precise estimate can extend its lifecycle, increase dimension is Ah. Accurate calculating directly the SoC is
battery performance and enhance safety and reliability of not possible because of few reasons. Noise presented in
the system [1], [5]. However, it is not possible to estimate measurements of the current would be accumulated in SoC
SoC directly. Nonlinearity of the system, temperature and estimates and therefore provide inaccurate estimate. Also,
time dependent characteristics of the battery make direct another problem is to calculate initial values of SoC [1].
calculation inaccurate. From those reasons, SoC estimation As it is noted, model of battery has been adopted from
is widely investigated [6]. [1] and is shown on figure 1. It consists of voltage source
In this paper two versions of Kalman filter are E0, called battery internal open circuit voltage, while other
parameters form circuit, such as R0, R1, C1, R2 and C2 tend
to model dynamics of the battery. Voltage on battery
Ivan Jokić, Žarko Zečević and Božo Krstajić, Elektrotehnički fakultet
terminals is defined as v0. SoC parameter is in nonlinear
Podgorica, Univerzitet Crne Gore, Džordža Vašingtona bb, 81000
Podgorica, Crna Gora (e-mails: [email protected], [email protected], relation with E0 and this relation can be modeled by
[email protected]) polynomial function, that is adopted from [1]:

978-1-5386-3620-6/18/$31.00 ©2018 IEEE

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY DURGAPUR. Downloaded on December 17,2022 at 11:31:49 UTC from IEEE Xplore. Restrictions apply.
 2018 23rd International Scientific-Professional Conference on Information Technology (IT) Žabljak, Montenegro

E0 where Q and R represents process noise and measurement

a1 SoC 7  a2 SoC 6  a3 SoC 5  a4 SoC 4  a5 SoC 3 
(2) noise covariance matrices. Kalman filter achieves the
 a6 SoC 2  a7 SoC  a8 ,
minimum mean-square state error estimate [1].
where coefficients are obtained using fitting curve The estimation process consists of two steps, as can be
technique and are: seen from figure 2. First step is time update for estimating
the states, based on the model(8). This estimation is called
a1 8.4073, a2 19.892, a3 11.497, a4 4.161, a priori estimation and is denoted with “¯”. The input
(3) signals in this step are states from previous iteration, input
a5 4.5533, a6 0.34365, a7 0.64685, a8 3.5016.
signal sample and process covariance matrix. Second step
In order to create a state-space based model of the is measurement update and is implemented by feedback.
system, voltage drop v1 over capacitor C1, voltage drop v2 This step uses error in estimating output signal to correct a
over capacitor C2 and SoC have been adopted for states. priori estimation of the states and it is named as a
Circuit from figure 1 can be modeled with next equation: posteriori estimation, denoted by “^”.
v0 E0  v1  v2  iR0 , (4)

where current represents input signal for the model and is

used in relations for derivatives of two first states: Time Update Measurement Update
(a priori estimate) (a posteriori estimate)
v1 dv v2 dv
i  C1 1  C2 2 . (5)
R1 dt R2 dt

Using equations (1)-(5), state space model can be Fig. 2. Discrete Kalman filter cycle
constructed:

ª 1 R1C1 0 0º ª1 C1 º The idea behind Kalman filter is to make an optimum

« 0 tradeoff between measurements and estimates based on the
x « 1 R2 C2 0 »» x  ««1 C2 »» i, (6) model, in order to achieve optimal state estimates [1].
¬« 0 0 0 ¼» «¬ K »¼
B. Extended Kalman filter
with state vector defined as x [v1 v2 SoC] . After T
The Extended Kalman filter represents a version of
discretizing, discrete state space model is obtained: Kalman filter for nonlinear systems [10]. Its block diagram
is presented on figure 3. EKF linearizes functions around
ª 't R1C1  1 0 0º ª 't C1 º the current mean and covariance [7]. Nonlinear system can
xk « 0 't R2 C2  1 0 » xk -1  « 't C2 »» ik , (7)
» «
« be described by state-space model:
¬« 0 0 1 ¼» ¬« K't ¼» xk f (xk -1 , uk )  wk
(10)
where time interval is denoted by ȟt, and superscript k and yk h(xk )  vk ,
k-1 determines time by defining iteration.
where f(xk-1,uk-1) and h(xk) represents nonlinear functions.
III. KALMAN-BASED FILTERS These nonlinearities are approximated in EKF using
linearization through first order Taylor expansion:
A. Kalman filter
Kalman filter represents an optimal state estimator of the wf (xk -1 , uk 1 ) wh(xk , uk )
Ak -1 , Ck (11)
linear systems, under process and measurement noise [7]. wxk -1 wxk
x k-1 xˆ k-1 x k xˆ k ,
It is named after R.E. Kalman, who has invented this filter
[8], [9]. The discrete-time system can be described by where Ak-1 is defined as Jacobian matrix of partial
equations: derivatives of f with respect to xk-1 and uk-1, and C is
defined as Jacobian matrix of partial derivatives of h with
xk Axk -1  Buk 1  wk 1 respect to xk and uk [7]. A priori and a posteriori error and
(8)
zk Hxk  vk , error covariance are defined as follows:

where xԖ Ըn represents the state of the system, A is named ek- xk  xˆ k- , ek xk  xˆ k

as transition matrix, B is called input matrix, xk-1 and uk-1 (12)
P -
k E ª¬e e º¼ , Pk
- -T
k k E ª¬ek ekT º¼
are state and input signals from previous iteration, H
represents measurement matrix, while wk and vk are defined
uk
as process and measurement noise, respectively. These two y (k )   xˆ k xˆ k
random variables are considered as independent. They are Lk f Delay

also assumed as white and with normal probability  

distributions [7]: h( xˆk , uk )

p ( w) N (0, Q) h uk
(9)
p (v) N (0, P), Fig. 3. EKF estimation block diagram

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY DURGAPUR. Downloaded on December 17,2022 at 11:31:49 UTC from IEEE Xplore. Restrictions apply.
 2018 23rd International Scientific-Professional Conference on Information Technology (IT) Žabljak, Montenegro

The EKF estimatin process can be presented in five covariance of the measurement, as well as cross covariance
steps with equations (13), repectively: of the state and measurement. It is obtained by next
1. state estimate time update, equations(17):
2. error covariance time update,
3. Kalman gain matrix, yˆ ki h(xˆ ki , uk ), i 0,1, 2,..., 2n
2n
4. state estimate measurement update,
5. error covariance measurement update [1],
yˆ k ¦w
i 0
m
i yˆ ki
xˆ k- f (xˆ k -1 , uk 1 ), 2n (17)
P -
k A k -1 Pk -1 A T
k -1  Q(t ),
P k
h
¦ w ( yˆ
i 0
c
i
i
k  yˆ k )( yˆ ki  yˆ k )T  Rk

Kk Pk- CkT [Ck Pk- CkT  R(t )]1 , (13) 2n

Pkfh ¦ w (xˆ c
i
i
k  xˆ k- )( yˆ ki  yˆ k )T ,
xˆ k xˆ k-  K k [ yk  h(xˆ k- , uk )], i 0

Pk (I  K k CkT )Pk- , where yˆ ki represents measurement estimated from sigma

where I is unity matrix. point at kth iteration, yˆ k and Pkh represent mean and
covariance of yˆ ki , while Pk fh is cross covariance of the state
C. Unscented Kalman filter
xˆki and measurement yˆ ki .
Unlike EKF, Unscented Kalman filter does not linearize
state space equations. It provides sigma points for states Finally, acquired covariance matrices of measurements
instead, by nonlinear unscented transformation (UT), and states are used to define Kalman gain(18) and also
make an a posteriori estimate of the states(19) and state
which mean and error covariance are calculated and
covariance(20):
updated iteratively. These sigma points are then
Pk fh  Pkh 
1
propagated through the nonlinear model functions, by Kk (18)
which it is provided a priori estimate of the states and of
output signal. The mean and covariance of those variables xˆ k+ xˆ k-  Kk ( yk  yˆk ) (19)
are then computed based on their statistics [3]. Feedback P +
k P  Kk P K .
-
k
h
k
T
k (20)
with output measurement is used to update (a posterior
estimate) provided estimate and to provide optimal state Initialize
estimation. State update
(11, 12, 13)
In each iteration 2n+1 sigma points, where n is number
Generate sigma Measurement
of states and coefficients wc, wm are calculated by next
points (7,8) estimate (10)
equations(14)-(15):
Propagate sigma
xˆ k0 -1 xˆ k+ -1 points (9)
Fig. 4. UKF estimating cycle
xˆ ki -1 xˆ k+ -1  n  O  
Pk -1 , i 1, 2,..., n (14)
IV. SIMULATION RESULTS
xˆ ki -1 xˆ k+ -1  nO  Pk -1 ,i n  1,..., 2n Time step of 1s is used in simulation tests. Model
parameters are adopted from[1]. They are estimated by
O O Least Square approach and are given in table 1.
w0m , w0c 1 E D 2
(n  O ) (n  O ) TABLE 1: PAGE LAYOUT DESCRIPTION.
(15) Parameters Value
1
wim wic , i 1, 2,..., 2n R0 0.0013ȳ
(2(n  O ))
R1 0.0042ȳ
where ɉ=Ƚ2(n+k)-n is defined as parameter controlling the C1 17111F
spread of sigma points around their mean. In this work R2 0.0024ȳ
parameters Ƚ and Ⱦ are set to 1 and 0, respectively. C2 440.57F
Parameter k can take values 0 or 3-n. Next, ξPk-1 is
decomposed matrix Pk-1. As it is explained, these sigma For EKF, next measurement and process covariance
points are propagated through state-space equations: matrices are used:

xˆ ki f (xˆ ki -1 , uk ), i 1, 2,..., 2n Q diag  0.1 0.1 0.01 , P diag  0.1 0.1 0.01 ,
(21)
2n R 1000,
xˆ -
k ¦w m
i xˆ i
k (16)
i 0 while for UKF they have next values:
2n
P -
k ¦ w (xˆ c
i
i
k  xˆ )(xˆ  xˆ )  Qk .
-
k
i
k
- T
k Q diag  0.01 0.1 0  , P  0.1 0.1 10 6  ,
i 0 (22)
R 15.
In order to estimate output signal, sigma points of states
are propagated through measurement function. Those A. Constant discharge simulation
acquired values are also used for calculating the mean and First simulation test is battery discharging with constant

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY DURGAPUR. Downloaded on December 17,2022 at 11:31:49 UTC from IEEE Xplore. Restrictions apply.
 2018 23rd International Scientific-Professional Conference on Information Technology (IT) Žabljak, Montenegro

current of 20A, applied at the input of the model. SoC performs considerably better results under described
estimates from both filters are presented on figure 5. On conditions.
figure 6 SoC estimation error in percentage is shown. It
can be seen that EKF estimates more precisely in first part
of simulation, while over considerably bigger range of SoC
values UKF provides better estimate.

Fig. 8. SoC estimate in percentage under noisy constant

charge/discharge

V.CONCLUSION
Fig. 5. SoC estimate under constant current discharge Two versions of Kalman filter, Extended Kalman filter
and Unscented Kalman filter have been used in this paper
for SoC estimation of Lithium-ion battery. Durign battery
discharging under constant current it is shown that in most
part of simulation UKF reaches better performance. In
second simulation test with alternating charging and
discharging of battery by constant current and noisy
process, UKF has performed considerably better results.

REFERENCES
[1] Z. Chen, Y. Fu, and C. C. Mi, “State of Charge Estimation of
Lithium-Ion Batteries in Electric Drive Vehicles Using Extended
Kalman Filtering,” IEEE Trans. Veh. Technol., vol. 62, no. 3, pp.
1020–1030, Mar. 2013.
[2] X. Zeng and J. Wang, “A Parallel Hybrid Electric Vehicle Energy
Management Strategy Using Stochastic Model Predictive Control
With Road Grade Preview,” IEEE Trans. Control Syst. Technol.,
Fig. 6. SoC estimation error under constant current vol. 23, no. 6, pp. 2416–2423, Nov. 2015.
discharge [3] Q. Yu, R. Xiong, C. Lin, W. Shen, and J. Deng, “Lithium-Ion
Battery Parameters and State-of-Charge Joint Estimation Based on
B. Noisy charge/discharge simulation H-Infinity and Unscented Kalman Filters,” IEEE Trans. Veh.
Technol., vol. 66, no. 10, pp. 8693–8701, Oct. 2017.
Second simulation test is under constant [4] J. T. B. A. Kessels, M. Koot, B. de Jager, P. P. J. van den Bosch,
charge/discharge of battery, with noise presented in N. P. I. Aneke, and D. B. Kok, “Energy Management for the
process. A white Gaussian noise, with covariance of 0.6 is Electric Powernet in Vehicles With a Conventional Drivetrain,”
added to constant current of ±20A. Current signal is of IEEE Trans. Control Syst. Technol., vol. 15, no. 3, pp. 494–505,
May 2007.
square wave, with period of 150s and pulse width of 50%. [5] V. Agarwal, K. Uthaichana, R. A. DeCarlo, and L. H. Tsoukalas,
“Development and Validation of a Battery Model Useful for
Discharging and Charging Power Control and Lifetime
Estimation,” IEEE Trans. Energy Convers., vol. 25, no. 3, pp.
821–835, Sep. 2010.
[6] Y. Wang, H. Fang, L. Zhou, and T. Wada, “Revisiting the State-of-
Charge Estimation for Lithium-Ion Batteries: A Methodical
Investigation of the Extended Kalman Filter Approach,” IEEE
Control Syst., vol. 37, no. 4, pp. 73–96, Aug. 2017.
[7] G. Welch and G. Bishop, An Introduction to the Kalman Filter.
1995.
[8] P. Khargonekar, “Prof. R. E. Kalman - A Deeply Inspiring Mentor
[Historical Perspectives],” IEEE Control Syst., vol. 30, no. 2, pp.
98–99, Apr. 2010.
[9] A. Antoulas, Ed., Mathematical System Theory: The Influence of
R. E. Kalman. Berlin Heidelberg: Springer-Verlag, 1991.
[10] “Extended Kalman Filter and System Identification,” in Kalman
Fig. 7. SoC estimate in percentage under noisy constant Filtering, Springer, Berlin, Heidelberg, 2009, pp. 108–130.
charge/discharge
SoC estimation error in percentage is presented on figure
7, while SoC stimate, also in percentage, is presented on
figure 8. From both figures, it can be concluded that UKF

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY DURGAPUR. Downloaded on December 17,2022 at 11:31:49 UTC from IEEE Xplore. Restrictions apply.