State of Charge Estimation for Electric Vehicle

Batteries under an Adaptive Filtering Framework

Wei He, Nicholas Williard, Chaochao Chen, Michael Pecht

Center for Advanced Life Cycle Engineering, University of Maryland, College Park, MD, 20742, USA
[email protected]

Abstract- Electric vehicles (EVs), which are powered by lithium drift in Coulomb counting. Another popular method is the
batteries, will penetrate the automobile market within the next voltage-based method that is to infer SOC by an open circuit
few years. This is mainly due to the increasing concerns of global voltage (OCV)-SOC lookup table [6]. However, OCV
warming and fossil fuel depletion. However, there are still some measurement requires a long period of rest before the tenninal
challenges for EVs that remain to be solved. The most notable voltage converges to the actual OCV. With the use of a battery
one is the state of charge (SOC) estimation and prediction for
model, it is possible to infer the battery's OCV from its
relieving EV drivers' range anxiety. To address this problem, an
terminal voltage. But this approach can generate large error if
equivalent circuit model is built to simulate battery behavior
the model employed is not accurate. A ±O.O1 V modeling error
under dynamical loading conditions. The parameters of the
model should be tuned on-line in order to handle the prediction
in the OCV could produce 10% percent error in the SOC
uncertainty arising from unit to unit variations and loading
estimation. In order to address these problems, this paper
condition changes. This paper proposed an Unscented Kalman
proposes a SOC estimation method using unscented Kalman
filtering-based method to self-adjust the model parameters and
filter (UKF) with a simple battery state-space model. This
provide the SOC estimation. The performance of the proposed
method does not require initial SOC infonnation. It can self­
method is demonstrated using data collected from LiFeP04 correct the SOC estimation by adding a correcting term which
batteries cycled with two dynamical discharge profiles. is generated based on the difference between the measurement
and model prediction. The proposed method is validated using
Keywords-State of Charge Estimation; Lithium-ion two dynamical discharge profiles.
Battery; Unscented Kalman Filter; styling; Electric Vechicles

II. BATTERY MODELING

I. INTRODUCTION A straightforward way to model the tenninal voltage V of a
With the increasing concerns on global wanning and fossil battery is to model it by the OCV minus the voltage drop of the
fuel depletion, the automobile industry is facing a landmark internal resistance R, as in Eq. (1):
transition from internal combustion engine (ICE) vehicles to
V=OCV-/xR (1)
electric vehicles (EVs). The major industrialized nations have
outlined their plans for the EV development and production. The schematic diagram of this model is shown in Fig.l, in
For example, the U.S. government set a goal to have one which the OCV and R is connected in series. OCV is the
million EVs on the road by 2015 [1]. The Chinese government battery terminal voltage when no current is put in or draw out
plans to have five million EVs on the road by 2020 [2]. of the battery. OCV is a function of SOC. As SOC decrease,
Though EVs will inevitably penneate the market, OCV will also decrease. The OCV-SOC relationship can be
challenges still exist. A notable one is the "range anxiety" determined and stored in a loop-up table under well controlled
problem, which describes the driver's fear of running out of lab experiments.
battery power on road [3]. The driving range of an EV is only
40-100 miles today, which is much less than ICE vehicles.
What make things worse is that there are very few charging ocv R
stations on road at present. To prevent EVs from being stalled
on the road, predictions of its residual range is needed. The first
step for residual range prediction is to know how much
capacity remains in the battery or state of charge (SOC). The �I
most common method for SOC estimation is Coulomb -------------------------------------------------

counting [3-5], in which the remaining charge is calculated by v

integrating the current entering or leaving the battery overtime.
Coulomb counting is simple and easy to be implemented for Figure 1. A simple battery model
on-board applications. However, it requires knowledge of the
starting SOc. In addition, Coulomb counting is an open-loop In this research, the experiment for OCV-SOC
method. Measurement noise and battery ageing can cause the determination contains the follow steps [7]:

978-1-4577-1911-0/12/$26.00 ©20121EEE MU3297 2012 Prognostics & System Health Management Conference
(PHM-2012 Beijing)
 (1) Discharge the cell at 0.1 A from its fully charged state federal driving schedules. The current profile was generated by
to its fully discharged state. combining the urban dynamometer driving schedule (UUDS)
[8] and US06 [9], which emulates both city and highway
(2) Rest for 2 hours. driving conditions experienced by an EV battery. The driving
(3) Charge the cell at 0.1 A to its fully charged state. schedule and the corresponding current profile are shown in
Fig. 3.
(4) Averaging the discharge and charge voltage provides
the OCV. The terminal voltage was measured every second. Based on
the collected data, R can be estimated by a least-square
The batteries tested have a rated capacity of 2.3 Ah. The algorithm. Fig. 4 (a) and (b) shows the fitted result and the
resultant OCV-SOC is shown in Fig. 2. error respectively. The root mean square (RMS) error of the fit
is 0.0212, and the mean error is 0.0180. The mean error
suggests there is a bias in the model, which can also be seen
3.8 ,-------.:--c--.:---,:---, from Fig. 4(b). Thus, the model can be improved by adding a
3.6
constant C to Eq. (1):
3.4 ) V=OCV-IxR +C (2)
The fitted result of Eq. (2) is shown in Fig. 5. The RMS
error of this model is decreased to 0.0064, and the mean error is
reduced to -6.6107e-016. Using Eq. (2), we can formulate a
state-space model for UKF estimation. SOC and R are selected

{
as states in the state-space model, and the measurement model
2.6 is the terminal voltage, which is a function of SOC and R.
2.4 . _ J(k) x I'lT
SOc(k +1) SOC(k) + WI
Cn
=

2.2
(3)
2 ' R
(k +l ) R(k) + W2
=

0 20 40 60 80 100
SOC (%) V(k) OCV[SOC(k) ]-J(k)· R(k) +C +6'
=

Figure 2. OCY-SOC Here Cn is the total capacity of the battery; W

I
and w2 are
process noise, and 6' is measurement noise.
(a)
100 (a)
:2
0- 80
5 3.6
"0
v 60
v
0- v
� 3.4
<;I) co
v 40
.::l
Q "0 3.2
:.c
v >
20
>
3
0
0 500 1000 1500 2000 2500 3000 3500 4000
Time (s) 500 1000 1500 2000 2500 3000 3500 4000
Time(s)
(b) (b)
2

0 0.05
"
�
-2
>
:::l ';::'
u -4 2 0
�
-6
-0.05
-8
0 500 1000 1500 2000 2500 3000 3500 4000
Time (s) 0 500 1000 1500 2000 2500 3000 3500 4000
Time(s)

Figure 3. (a) Federal driving schedule and (b) the corresponding current
profile. Figure 4. (a) The fitted result of model Eq.(I) and (b) the error of the model
fit
In order to identify the model parameter R and test the
model performance, experiments were conducted based on

978-1-4577-1911-0/12/$26.00 ©2012 IEEE MU3297 2012 Prognostics & System Health Management Conference
(PHM-2012 Beijing)
 Ai
Wo(c) /( a2
L+A )+1- + f3
=

(a)

1I1 11�1
(4)
, ) - W(c)
W(III , -- /I/,{ 2 ( L+A)}
3.6
i 1,. · · ,2L
=

" I : ' �I '1��p�m��lrnml

"
� 3.4
CIl
.::l
"0
>
3.2 "
where A a2(L+ K) - L is a scaling parameter. a determines
=

3 the spread of the sigma points around X . K is another scaling

parameter which is set to 3-L. f3 is used to incorporate prior
2.8
0 500 1000 1500 2000 2500 3000 3500 4000 knowledge of the distribution of x. For Gaussian distributions,
Time(s)
f3 2 is optimal [10].
=
(�(
L + A) P" is the i-th column of the )i
(b) matrix square root of ( L + A ) p". Then each sigma point is
0.05
propagated trough the nonlinear function y, g(X, ) , =

?::: 0
i 0,.. ,2L . The estimated mean and covariance of y are
=

�
computed by the weighted sample mean and covariance as
follows:
-0.05
2L
0 500 1000 1500 2000 2500
Time(s)
3000 3500 4000
Y
=
LW,(III)Y,
;=0

2L
Figure 5. (a) The fitted result of model Eq.(2) and (b) the error of the model
Py =
LW,<C) (y, - y)(y, - yt
;=0
fit. (5)
The UKF is a straightforward application of the UT for state
III. UNSCENTED KALMAN FTLTER estimation. The main steps of UKF are summarized as follows:
The SOC estimation is a nonlinear problem in this study.
The non-linearity can be seen from the observation model,
(l) Initialize with x.o=E [ xol and Po=E ( Xo- �)( Xo-x [ SJ
where OCV(SOC[kD is an non-linear function as shown in Fig. (2) For k E {l, ...,oo}, calculate sigma points:
2. For non-linear state estimation problem, extended Kalman
filter (EKF) is a standard approach. However, EKF uses fIrst
order terms of the Taylor series expansion of the nonlinear "'1..'-1 - [j$;.'-I
functions. Large error can be produced if the model is highly (3) State prediction:
nonlinear. In this study, we adopted the unscented Kalman • Propagate the sigma points through the state model:
filter (UKF) which can be accurate to the 3rd order in the sense
of Taylor series expansion for any nonlinearity [10]. UKF is a
direct application of unscented transform (UT). In UT, a
Gaussian distribution is represented by a set of carefully chosen • Calculate the propagated mean
sample points called sigma points. These sigma points can
2L
capture the mean and covariance of the Gaussian random
variables (GRV) when propagated through a nonlinear function. X-k - "W<m) "'I.,. .klk-I
�
;=0
,

Assume a random variable x (dimension L) has mean • Calculate the propagated covariance
x and covariance p. , Consider propagating x through a
2L
nonlinear function y g(x). To calculate the statistics of y,
L w,<C) ["'I..'.klk-I - iifJ ["'I..',klk-I - x� J
=

P,- =

we fust fmd a matrix X of 2L+1 sigma vectors X, with i=O

corresponding weights W, , according to the following (4) Measurement update:

equations[10]: • Propagate sigma points through measurement function
Xo =
x i 0
= Yklk-l =
H [Xklk-l ]
Xi =
X + (�(L+A) P' )i i 1,.··,L
=
• Calculate the propagated mean
2L
Xi =
X- (�( L A)
+ P. )i-L i L+l,··· ,2L
= --
Yk -- L. i )Yi,klk-l
�W(III
;=0

o ) -- /(iL+A
W(III A i 0 =
• Calculate the estimated covariance
)

978-1-4577-1911-0/12/$26.00 ©2012 IEEE MU3297 2012 Prognostics & System Health Management Conference
(PHM-2012 Beijing)
 2L

"W(c)
Py,y, - L...
;=0 i [ yi,kl k -I - -
Yk-] [ Yi,klk-I - Yk ]
-- T
3.5

2L

"W(c) [ Xi,klk-I - Xk
P -- L...
x,y,
i=O I
--] [Xi,klk-I - --]
Yk
T 3

2.5

�
• Calculate the Kalman gain K and update the state 2 ::j)
:::
�

estimation and covariance UJ (

r/) Q Ci
�
1.5
()
:j) :j) Ci

Xk X� + K ( Yk -y� )
=
Cf

Pk Pk- - KP K
=
YkYk
____ T
0.5

o
o 20 40 60 80 100
IV. STATE OF CHARGE ESTIMATION RESULT SOC(%)

The proposed method is first tested on the data collected by Figure 7. RMS error of SOC estimation when starting from different SOC
the federal driving schedule as shown in Fig. 3. The true initial levels.
SOC is 80%. To simulate actual field conditions where the
exact SOC at the beginning of usage is unknown, the initial To see whether the proposed method can handle the unit to
SOC guess for UKF estimation is intentionally set to be 35%. unit variability, it was tested on data collected from a second
The estimation result is shown in Fig. 6. Though the estimation battery of the same type. The parameter settings are the same as
deviates from the true value during the first 10% of the those used in the previous case. As shown in Fig. 8, the
discharge cycle, it fmally converges to the true SOC as more maximum RMS error of the SOC estimation is within 4%.
measurements became available. The RMS error of the
estimation is 3.035%. To examine the applicability of the proposed method to
other profIles, we collected battery discharge data using the
100·
dynamical stress testing (DST) [11]. The current load profIle
--- Estimation and the corresponding terminal voltage response are shown in
Fig. 9. The parameter settings of the UKF including the initial
80 --- True SOC guess, process and measurement noise also remains the same in
this case. The RMS error of the SOC estimation is shown in
60
-;g Fig. 10. The maximum RMS error is also within 4%.
&
0
r/) 40

20 4

3.5
0
0 0.5 1.5 2 2.5
:j)
4 3
x 10 ::j) Ci

J
� 2.5
:::
�

Ci Ci
UJ 2
r/) ::j)
� 1.5
o 0.5 1.5 2 2.5 ::j)
Time(sec) 4
x 10
Ci Q
0.5
Figure 6. UKF-based SOC estimation.
o
The proposed method was further tested by varying the o 20 40 60 80 100
beginning SOC of the battery and allowing the UKF to adapt to SOC(%)
the true value. In each case, the initial SOC guesses were set to
be 35%. Fig. 7 shows the result. It can be seen that the RMS Figure 8. RMS error of UKF-based SOC estimation for another EY battery.
errors of all case are within 3.1 %. The maximum RMS error
occurs when starting from 80% SOC. One possible reason for
this can be the relative flat OCV slop at 80% SOC as shown in
Fig.2, which means a tiny error in the OCV can cause large
error in the SOC estimation.

978-1-4577-1911-0/12/$26.00 ©20121EEE MU3297 2012 Prognostics & System Health Management Conference
(PHM-2012 Beijing)
 the SOC of an EV changes due to self-discharge or varying
5 - environmental conditions, this method will be able to self­
correct the SOC automatically The RMS errors of the SOC
� estimation in the validation tests were smaller than 4%. In
c 0
�::s addition, the proposed method can handle unit to unit
U difference and loading condition variations without changing
the setting of the UKF. The temperature is another influencing
50 100 150 200 250 300 350 factor for SOC estimation. Future work will investigate the
Time(s) SOC estimation under simultaneous dynamical temperature
4
and discharge conditions.

:> 3.5 ACKNOWLEDGMENT

OIl
"

.9 The authors would like to thank the more than 100

-0 3
> companies and organizations that support research activities at
the Center for Advanced Life Cycle Engineering at the
2.5
0 50 100 150 200 250 300 350 University of Maryland annually. The authors would also like
Time(s) to thank the members of the Prognostics and Health
Management Consortium at CALCE for their support of this
Figure 9. (a) Current profile of the DST,(b) the terminal voltage response of
work.
the DST.
REFERENCES
[I] One Million Electric Vehicles by 2015: February 2011 Status Report.
4 The US Department of Energy,2011.
�
[2] F. Van and J. Wong. China electric vehicles to hit I million by 2020:
3.5
(� report. Oct.l6 2010, http://www.reuters.com/article/20IO/IO/I6/retire­
us-autos-china-idUSTRE69FOJ820IOIOI6,cited on Feb. IS 2012.
3 Ci [3] Electric Vehicles: The Future of Driving, the Consumer Electronics
Ci :j) Association, 2010.
;t. 2.5
� [4] K.S. Ng, c.-S. Moo, Y.-P. Chen, and Y.-c. Hsieh, Enhanced coulomb
::: () counting method for estimating state-of-charge and state-of-health of
� 2
Ci
Vl :j) lithium-ion batteries. Applied Energy, 2009. 86(9): p. 1506-1511.
(�
� 1.5 [5] J.Y. Van, G.Q. Xu, H.H. Qian, and Y.S. Xu, Robust State of Charge
Estimation for Hybrid Electric Vehicles: Framework and Algorithms.
Energies,2010. 3(10): p. 1654-1672.
[6] V. Pop, H.J. Bergveld, P.H.L. Notten, and P.P.L. Regtien, State-of-the­
0.5 art of battery state-of-charge determination. Measurement Science &
Technology,2005. 16(12): p. R93-RIIO.
o [7] G.L. Plett, Extended Kalman filtering for battery management systems
o 20 40 60 80 100 of LiPB-based HEV battery packs: Part 2. Modeling and identification.
SOC(%) Journal of Power Sources,2004. 134(2): p. 262-276.
[8] The EPA Urban Dynamometer Driving Schedule (UDDS),
http://www.epa.gov/otaq/standards/light-duty/udds.htm ,cited on Feb 15,
Figure 10. RMS error of UKF-based SOC estimation for DST profile.
2012
[9] EPA US06 or Supplemental Federal Test Procedure (SFTP), accessed
CONCLUSIONS http://www.epa.gov/otaq/standards/light-duty/sc06-sftp.htm. cited on
Feb 15,2012
State of charge (SOC) provides critical infonnation for [10] E.A. Wan and R. Van Der Merwe. The unscented Kalman filter for
residual range prediction and for battery control unit to prevent nonlinear estimation. in The IEEE Adaptive Systems for Signal
the battery from over-charge and over-discharge. This paper Processing,Communications,and Control Symposium 2000.
proposed a SOC estimation method based on unscented [II] United States Advanced Battery Consortium, Electric Vehicle Battery
Test Procedures Manual,Revision 2,Southfield,MI,1996.
Kalman filter and an empirical battery model. The proposed
method does not require initial SOC information. It is able to
converge to true SOC as more measurements available. This
property is meaningful for real applications, like EVs. When

978-1-4577-1911-0/12/$26.00 ©2012 IEEE MU3297 2012 Prognostics & System Health Management Conference
(PHM-2012 Beijing)