Вы находитесь на странице: 1из 13

Journal of Power Sources 289 (2015) 50e62

Contents lists available at ScienceDirect

Journal of Power Sources


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

Adaptive unscented Kalman lter based state of energy and power


capability estimation approach for lithium-ion battery
Weige Zhang*, Wei Shi, Zeyu Ma
National Active Distribution Network Technology Research Center, School of Electrical Engineering, Beijing Jiaotong University, No.3 Shangyuancun Street,
Haidian District, Beijing 100044, China

h i g h l i g h t s

 The three-dimensional response surface based open circuit voltage model is developed.
 An adaptive unscented Kalman lter has been employed to develop a joint estimator.
 A novel battery model-based SoE and SoP joint estimation approach is developed.
 The robustness of joint estimator against uncertain operating temperatures and aging levels is systematic evaluated.

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

Article history: Accurate estimations of battery energy and available power capability are of great of importance for
Received 7 February 2015 realizing an efcient and reliable operation of electric vehicles. To improve the estimation accuracy and
Received in revised form reliability for battery state of energy and power capability, a novel model-based joint estimation
22 April 2015
approach has been proposed against uncertain external operating conditions and internal degradation
Accepted 24 April 2015
status of battery cells. Firstly, it proposes a three-dimensional response surface open circuit voltage
model to calibrate the estimation inaccuracies of battery state of energy. Secondly, the adaptive un-
scented Kalman lter (AUKF) is employed to develop a novel model-based joint state estimator for
Keywords:
Lithium-ion battery
battery state of energy and power capability. The AUKF algorithm utilizes the well-known features of the
Uncertainty Kalman lter but employs the method of unscented transform (UT) and adaptive error covariance
Model-based matching technology to improve the state estimation accuracy. Thirdly, the proposed joint estimator has
State of energy been veried by a LiFePO4 lithium-ion battery cell under different operating temperatures and aging
State of available power capability levels. The result indicates that the estimation errors of battery voltage and state-of-energy are less than
Adaptive unscented Kalman lter 2% even if given a large erroneous initial value, which makes the state of available power capability
predict more accurate and reliable for the electric vehicles application.
2015 Elsevier B.V. All rights reserved.

1. Introduction measures energy left in a battery compared with its maximum


value, and it is critical for indicating the remaining mileage of
With the global issues of energy shortage and environmental electric vehicle. SoP of battery is necessary for regulating the pro-
degradation, lithium-ion battery has become one of the most pelling power and coordinating the regenerative braking and fric-
readily available and low-cost energy storage in electric vehicles tion braking of EVs [3e5].
(EVs) [1,2]. However, due to the strong time-variable and nonlinear
characteristics, as well as inuences by such random factors as 1.1. Review of the SoE and SoP estimation approaches
driving loads and operating environment in its application, an ac-
curate and reliable estimation of state of energy (SoE) and state of With the increasingly widespread application of batteries, the
available power capability (SoP) still remains big challenges. SoE functional requirement of battery management system appears a
more sophisticated and complex trend. On the other hand, battery
state of charge (SoC) is not appropriate to predict the residual en-
* Corresponding author. ergy and remaining mileage of electric vehicle, managing the en-
E-mail address: wgzhang@bjtu.edu.cn (W. Zhang). ergy/power ows from the various energy storage devices [4]. SoE

http://dx.doi.org/10.1016/j.jpowsour.2015.04.148
0378-7753/ 2015 Elsevier B.V. All rights reserved.
W. Zhang et al. / Journal of Power Sources 289 (2015) 50e62 51

and SoP compared to battery SoC, are two more direct indicators for voltage. Additionally, the dual/joint estimation for battery SoE and
EVs application. SoP have not been fully considered in the existing methods.
SoC denotes the state of electrical charge/capacity rather than
the energy of battery. Although there is a positive correlation be- 1.2. Innovation
tween battery SoE and SoC, they have no explicit quantitative
relationship. The SoC changes linearly with operating current, but The key contribution of this work is that a novel model-based
battery energy is the integral between the product of voltage and joint estimation approach of state of energy and available power
current with time. Battery SoE has direct link with its real-time for EVs battery has been proposed with an adaptive unscented
voltage, as a result, it is hard to be accurately calculated. Kalman lter. It has the potential to accurately project the trajectory
Compared with battery SoC estimation approaches [6e15], few of battery SoE and SoP in real-time.
researchers have carried out systematic research for battery SoE
[16e18]. 1.3. Organization of the paper
Refs. [17,18] proposed new criterion for direct evaluating the
remaining energy of a battery, where the discharge stages of bat- Section 2 describes implement owchart of the parametric
tery have been considered. Ref. [4] used the particle lter jointly modeling approach. Section 3 presents the joint estimation
estimate SoC and SoE of battery, and the developed method has implement owchart of battery SoE and SoP. To evaluate the pro-
been veried with IFP1865140-type batteries under both constant posed approach, a LiFePO4 battery is used to carry out systematic
and dynamic current conditions. In their following research pre- battery experiment under different temperatures. The experiment,
sented in Ref. [16], they proposed a black box model to estimate the simulation results and evaluation of the proposed method are re-
SoE online. In the input layer, the battery terminal voltage, the ported in Section 4 where conclusions are drawn in Section 5.
current and the temperature are taken as the input parameters, and
the output layer is the estimated SoE. 2. Battery model
The above methods for battery SoE estimation have achieved
their own advantages. However, for batteries used in EVs would 2.1. Lumped parameters battery model
inevitably experience uncertain operating conditions and degra-
dation. Therefore, the real-time performance of battery model Lithium-ion battery is a very complex electrochemical system
should be enhanced. What's more, battery available power links with physical/chemical processes and some extra side reactions,
with its available energy, therefore, estimating the SoE separately such as aging, diffusion and self-discharge effects. As a result, it is
hardly can ensure the safety and efcient operation of batteries difcult to obtain an accurate battery model to describe the rela-
used in electric vehicles. tionship between terminal voltage and varied dynamic loads. In
In addition to SoE information, the SoP information of battery most cases, researchers choose to use a data-driven model, which
is necessary for regulating the propelling power and coordinating means that the battery model is established according to the
the regenerative braking and friction braking of electric vehicles measured data, such as Refs. [5,22e24]. Ref. [13] proposed a
[5,19]. Inappropriate SoP may result in overly conservative vehicle recursive least square algorithm (RLS) based parameter identica-
energy management or yield to over-charge, over-discharge and tion for Thevenin battery model (Schematic diagram of the The-
premature failure of battery. Therefore, the available power venin battery model can be found in Ref. [13]). Thevenin model is a
capability can be dened as an additional state of the battery and typical lumped parameters battery model and has been selected for
its prediction has to be performed by the battery management this study. Its electrical behavior is expressed by Eq. (1).
system (BMS).
8
Unfortunately, the true power capability of the battery is > 1 1
< U_ D  UD I
difcult to determine for the running EVs. Thus the power capa- t CD L (1)
bility estimation is primarily carried out with the design limits of >
:
the battery itself and the EVs or powertrain, such as upper cutoff Ut Uoc  UD  IL Ro
voltage, lower cutter off voltage, maximum allowable currents and
powers, maximum and minimum SoE values et al. With the where Uoc denotes the OCV (open circuit voltage), Ro denotes the
development of EVs technology, some power capability estimation equivalent ohmic resistance. The RC network, which includes
approaches are presented [3,5,19e21]. Ref. [20] reviewed several diffusion resistance (RD) and diffusion capacitance (CD), is used to
typical kinds of existing methods and each method has their own describe the mass transport effects and dynamic voltage perfor-
advantage. mances, t RD CD . IL is the load current (assumed positive for
The most commonly used approach is hybrid pulse power discharge, negative for charge), Ut is the terminal voltage. Through
characterization (HPPC) method [5]. In considering that its esti- general solution of the linear differential equation, the UD can be
mates are over optimism owing to it determines the static peak solved.
power in laboratory environments, Ref. [19] proposed a voltage- Z Z Z
1 1 Z 1
limited method. As following researches, Refs. [5,14] proposed an  dt  dt IL  dt
UD Ct e t e t e t dt (2)
online parameter identication-based multi-parameter peak po- CD
wer capability estimation method. It is helpful for improving the
control efciency of the energy management system in EVs, espe- where Ct denotes a constant item in the equation. Assuming
cially when acceleration, regenerative braking or gradient climbing. UD;k1 Dt0 UD;k as its initial value, Eq. (2) can be simplied by the
Obviously, the on-line parameter identied based SoP prediction following equation.
method has the best performance against random operating
environment. UD;k1 UD;k expDt=t 1  expDt=tRD IL;k (3)
However, these methods have not both considered the in-
uences of operating conditions and aging of battery. As a result, whereDt denotes the sampling interval, UD,k1 and UD,k denote that
the predicted results are not reliable enough. What's worse, these the diffusion voltage at time tk1 and tk, respectively. Let us dene
methods have not considered the inuence from battery dynamic Et UtUoc, Eq. (1) can be rewritten as:
52 W. Zhang et al. / Journal of Power Sources 289 (2015) 50e62

Table 1
Et;k UD;k  IL;k Ro (4) Detailed computation process of recursive least squares algorithm.

(1) Initialization
Adding with Eq. (3), Eq. (4) can be rewritten as follows.
f0 ; q0 ; Pls;0 ; Kls;0 ; m
(2) Data collection
Et;k1 UD;k expDt=t  1  expDt=tRD IL;k  IL;k1 Ro Calculating thefk in Eq. (11) for next computation.
(5) (3) Computation for error covariance matrix Pls and algorithm gain Kls.
8
>
> Pls;k1 fTk
>
> K
With Eq. (4), we can get: >
< ls;k l fT Pls;k1 fk
n;k
(13)
  >
>
> Pls;k1  Kls;k fTn;k Pls:k1
Dt >
>
: Pls;k
Et;k1 exp  Et;k  Ro IL;k1 m
t (6)
expDt=tRo  1  expDt=tRD IL;k (4) Calculating model prediction inaccuracy and update model parameter.

k yk  fk qk
b (14)
Then we have, qk bq k1 Kls;k k

(5) Update the predicted terminal voltage of battery


Et;k1 a1 Et;k a2 IL;k1 a3 IL;k (7) Ut;k Uoc;k Et;k : (15)

where a1 , a2 and a3 are:


8
< a1 expDt=t
a Ro (8) because it depends on linearization to propagate the mean and
: 2
a3 expDt=tRo  1  expDt=tRD covariance of the state. The AUKF, it is an extension of the KF and
has the potential to decrease linearization errors of EKF. It can
Then the parameters of the battery model can be solved by:
provide considerable improvement over the EKF [25]. In consid-
8 ering that the complex chemical reaction in battery, it shows strong
>
> Ro a2
>
> time-varying nonlinear characteristic in itself. Comparing with the
>
< R a1 a2 a3
> EKF, the AUKF does not need to calculate the Jacobi matrix during
D
a1  1 (9) its linearization procedure [26e29]. On the contrary, AUKF uses
>
>
>
> 1  a1 Dt classical statistical methods based unscented transform (UT) line-
>
> CD
:
a1 a2 a3 loga1 arized solutions for nonlinear systems. Namely, it executes a
nonlinear transformation on a single point (rather than an entire
Based on the format of the ARX (auto regressive exogenous) probability density function (PDF)). It can greatly improve the
model, the electrical equation of the battery model can be rewritten estimation accuracy. From the comparative analysis of the state
as follows. estimate results in Refs. [28], we also can nd that the AUKF based
state estimators have more superior performance than the EKF or
yk fk qk (10)
UKF based state estimators. For a system of our interest described
by Eq. (12), such state and measurement equations can be written
where f and q denote data matrix and the parameter matrix
in the discrete form as follows:
respectively. denotes the model error.
   
fk Et;k1 Et;k2 IL;k IL;k1 IL;k2 xk1 Axk Buk uk
(11) (16)
qk a1 a2 a3 a4 a5 T yk Cxk Duk yk
The state-space equation of the lumped parameters battery
where x is the n dimensional state vector, y is the m dimensional
model can be expressed as:
measurement vector and u is the r dimensional input vector. The
 modeling error, w, and measurement error, v, are zero-mean white
UD;k1 expDt=tUD;k 1  expDt=tRD IL;k
(12) noise processes with covariance matrices Q and R, respectively. A,
Ut;k1 Uoc;k1  UD;k1  Ro IL;k1
B, C, and D are system matrices. The objective is to estimate the
state xk at step k recursively based on measurements y [y1,
y2,,yk] [29].
2.2. Identication method The basic steps involve initialization, time-update and
measurement-update, detailed computation processes of the AUKF
The RLS algorithm has been employed to carry out the online algorithm are as follows [30,31].
parameter identication for the lumped parameters battery model.
The detailed computational processes for solving the model pa- (I) Initialization. For k 0, set:
rameters are listed in Table 1 [13,23].where m denotes the forgetting
factor for ignoring the inuence from the old measured data

gradually. Initial state : b
x 0 Ex0 ; (17)

3. Battery SoE and SoP joint estimator





T
3.1. Adaptive unscented Kalman lter Initial posteriori error covariance : P0 E x0  b
x0 x0  b
x0

(18)
The extend Kalman lter (EKF) is the most commonly used state
estimation algorithm for nonlinear systems. However, it often gives Initial measurement noise covariance: Q0 , Window size for
untrustworthy estimations if the system nonlinearities are serious, covariance matching: M.
W. Zhang et al. / Journal of Power Sources 289 (2015) 50e62 53

(II) Computation: time updateprior estimation. (From (k1)


to (k)). For k 1, 2,, compute 8 j
b bj
(a) Sampling for Sigma points: < y k C x k1 Duk yk
>

X2n
(24)
:b
> yk wm bj
j yk
j0
8 j
>
>
> x k1 b
b x k1 ~ xj j 1; 2; :::; 2n
>
>
>
>  q
 (b) Update the covariance: covariance of the predicted mea-
>
< ~j
T
surement Py and cross covariance Pxy between b

x k and b
yk .
x nP j 1; 2; :::; n
k1 (19)
>
>
i
>
> qT 8

>
> > X
2n
j  j  T
>
>~xnj  nP j 1; 2; :::; n >
> Py;k wcj byk  b
yk yk  b
b yk Rk
: k1 >
<
j j0


(25)
>
> X
2n X
2n
 T
>
> wcj wcj b
j
yk  b
 j
yk  b
b
: Pxy;k yk yk
xj and ~
where ~ xnj denote the jth and (n j)th sigma point, j0 j0
j
respectively. b x k1 denotes the system state x at step k1 consid-
q
ering the sigma point. nP (c) Update the Kalman gain.
k1 denotes the matrix square root of
q q q


nPk1 such that nP k1 T
nP
k1 nPk1 , and nP k1 j is the jth
1
K Pxy;k *Py;k (26)
q
row of nP k1 .
(d) Update the state estimate.
(b) Calculate the weights w.

 
xk b
b x k K yk  b
yk (27)
8 wm l=n l
>
<
0

wc0 l=n l 1  a2 b (e) Update the error covariance.


(20)
>
: m
wj wcj 1=2n l; i 1; 2; :::; 2n 
Px;k Px;k  KPy;k  K T (28)

where l a2 n k  n is a scaling parameter, a determines the (f) Adaptive lawcovariance matching.


spread of the sigma points around b x and is usually set to a small
positive value (e.g., 1e3). k is a secondary scaling parameter which 8
is usually set to 0, and b is used to incorporate prior knowledge of >
> 1 Xk

>   T
the distribution of x (for Gaussian distribution, b 2 is optimal). < Hk yk  by k yk  b yk
M (29)
ikM1
>
>
>
: R H  C P CT ; Q K H KT
(c) Update the prior state of 2n sigma point: Use the known
k k k x;k k k k k k
nonlinear system equation to transform the sigma points
j
into b
x k vectors. Through the detailed computation processes, we can project the
system state if the random initial information is set. The AUKF will be
j j employed to develop SoE and SoP estimator in next section. It is
x k Ab
b x k1 Buk uk (21)
noted that the drawback of the EKF/UKF algorithm, which relies on
the good estimations of the covariance matrices of measurement
j noise and process noise, has been efciently solved by the error
(d) Combine prior state estimate: Combine bx k vectors to obtain
the a priori state estimate at time k. matrix matching technology of AUKF algorithm. It provides a further
innovation in the algorithm using the lter's innovation sequence.
The innovation allows the parameters of above covariance matrices

X
2n
j to be estimated and updated iteratively through Eq. (29).
b
xk wm b
j xk (22)
j0
3.2. Model-based SoE estimator

(e) Update the error covariance: Estimate the a priori error SoE denotes the state of remaining electrical energy stored in a
covariance. We should add Qk1 to the end of the equation to battery. It can be dened as follows.
take the process noise into account.
zk zk1  DEk =Ea zk1  pk Dt=Ea (30)
X
2n

T
 c j  j 
Px;k wj xk  b
b xk xk  b
b xk Qk1 (23) where z denotes the SoE of battery, zk and zk1 denote the SoE of
j0 battery at time tk and at time tk1, respectively. DEk denotes the
consumed energy at time tk, pk denotes the input/output power
during battery operating at time tk, and p Ut IL. Ea denotes the
(III) Measurement update posteriori estimation (From (k) to maximum available energy of battery.
(k)). Combined with Eq. (12), the system state-space equation can be
(a) Update the output. achieved by:
54 W. Zhang et al. / Journal of Power Sources 289 (2015) 50e62

8      
< zk1 1 0 zk Ut;k Dt Ea
I
UD;k1 0 expDt=t UD;k 1  expDt=tRD L;k (31)
:
Ut;k1 Uoc;k1  UD;k1  Ro IL;k1

upper cut-off voltage Ut,max and lower cut-off voltage Ut,min, we can
d;v
obtain the maximum available peak discharge current Imax and
c;v
To begin with the AUKF algorithm, we should have the system minimum available peak charge current Imax .
matrices presented in Eq. (16). Through Eqs. (3) and (12), the output equation of battery voltage
at time tk1 can be obtained.
! !
8 1 0 Ut;k1 Dt Ea
>
<A
> ; B b
U t;k1 Uoc;k1  expDt=tUD;k
0
expDt=t 1  expDt=tRD (33)
>    Ro 1  expDt=tRD IL;k
>
:
C vUoc;k vzk  1 ; D Ro
Then the peak current predictions should meet the following
(32) equation.

( c;v
Ut;max Uoc;k1  expDt=tUD;k  Ro 1  expDt=tRD Imin
d;v (34)
Ut;min Uoc;k1  expDt=tUD;k  Ro 1  expDt=tRD Imax

With the detailed computational steps and the system matrices


of the state-space equation, the inner inhabited state of the battery However, the Eq. (34) cannot solve the current directly owing to
can be accurately tracked. that battery OCV is the function of battery SoE itself. Therefore,
before the solving of peak current we should expand the OCV
3.3. Model-based SoP estimator function.
Secondly, expand the OCV function to solve battery current.
Ref. [5] proposed a joint estimation approach for battery SoC and The Taylor-series expansion is used to linearize the equation and
SoP using an adaptive extended Kalman lter. The calculation of to solve the approximated values for the peak current prediction.
SoP depends on the estimation accuracy of battery SoC. In this The Taylor-series expansion equation for OCV is presented as
study, the estimation of SoC has been replaced by battery SoE follows.
owing to that the SoE is more important for predicting the
remaining mileage of an electric vehicle. What's more, the control
limits of battery SoE has been used to limit the peak power capa-  
Ut;k IL;k Dt
bility in this study. It gives a quantized relationship between bat- Uoc;k1 Uoc zk 
Ea
tery power and its renaming energy. Consequently, the calculation  
of SoP is different from the process presented in Refs. [3,5].
DtUt;k vUoc;k  DtUt;k
Uoc zk  IL;k  R1 zk ; IL;k
Through the on-line parameter identication presented in Sec- Ea vz zzk Ea
tion 2.2 and AUKF based SoE estimator presented in Section 3.2, we (35)
have the real-time parameter and SoE of battery. Then, we can carry
out the prediction for battery SoP. In considering that the rst-order residual R1  being too small
Firstly, calculate the model-based peak current. The predicted to affect the OCV at the next sampling time Dt, therefore R1  can
voltage of battery model should be accurately controlled to within be viewed as R1 z0. Then:
the limits of battery cutoff voltage. Based on the design limits of the

8  
> DtUt;k vUoc;k 
> c;v
< Ut;max Uoc zk  IL;k Ea
> 
vz zzk
 expDt=tUD;k  Ro 1  expDt=tRD Imin
  (36)
>
>
> DtUt;k vUoc;k  d;v
: Ut;min Uoc zk  IL;k   expDt=tUD;k  Ro 1  expDt=tRD Imax
Ea vz zzk
W. Zhang et al. / Journal of Power Sources 289 (2015) 50e62 55

Through Eq. (41), we can obtain the SoP predictions in real-time.


Then, the model based available peak current capability can be
got. 3.4. Model-based SoE and SoP joint estimator

8 Based on the proposed model-based battery SoE and SoP esti-


>
>
d;v
Imax
Uoc zk  expDt=tUD;k  Ut;min
>
< DtUt;k vUoc;k 

 1  expDt=tRD Ro
mator, we can build a joint estimation framework for these two
states, where SoE is an input state in the SoP estimation. The
Ea vz zzk
  detailed computation process is presented in Fig. 1. It can be divided
Dt (37)
>
> Uoc zk  exp UD;k  Ut;max
into three parts, including battery modeling, AUKF-based SoE
>
: Imin
c;v

DtUt;k vUoc;k 
RD CD estimation and SoP estimation processes.
Battery modeling part contains data input, on-line parameter
 1  expDt=tRD Ro identication, battery model update and estimated voltage update.
Ea vz zzk
The main calculation is carried by the RLS algorithm, which con-
Thirdly, obtain the available peak current. In considering that tains ve steps and has been listed in Table 1. It is worth noting that
the real current should be operated within the design current, prior voltage estimate can be obtained through the prior state
therefore the peak current capability of battery can be expressed by estimated by AUKF-based SoE estimation part, and the prior voltage
the following equation. estimate will be transmitted to Eq. (24) for calculating the voltage
prediction error.
8
AUKF-based SoE estimation part contains three steps. Where the
< Imax
dis d;v
min Imax ; Imax

initialization processes mainly work for battery initial state and its
(38)
: I chg max I ; I c;v error covariance, noise covariance caused by measurement and
min min min
state process. The prior estimation process implements the state
where Imax and Imin are design limits for the maximum discharge estimate and voltage prediction before the real-time measured
chg dis are
current and minimum charge current respectively. Imin and Imax voltage arrives. The posteriori estimation process carries out the
the minimum allowable charge current and maximum discharge measurement update process. With the real-time measured battery
current respectively with all limits enforced. terminal voltage, we can calculate the voltage prediction error
Fourthly, calculate the SoE-based battery available power. The based on the prior estimated battery state and then update the
allowable operated current of battery should satisfy the specied Kalman gain, project the real-time trajectory of battery's inner
SoE operation range. Therefore, with the higher cutoff SoE limit- state. It is worth noting that the initial noise information in AUKF is
zmax and upper cutoff SoE limitzmin, we will have the maximum set randomly, the adaptive covariance matching technique pre-
discharge powerPmax d;z
and minimum charge powerPmin c;z
through sented in Eq. (29) has the potential to adaptively adjust the noise
the following equation. covariance matrix and achieve accurate state estimations.
The model-based SoP estimation contains available peak cur-
( rent capability estimation with designed voltage and multiple
d;z
Pmax zk  zmin Ea =Dt constraints available peak power capability estimation. It is worth
c;z (39)
Pmin zk  zmax Ea =Dt noting that the real-time model technique and accurate SoE esti-
mation are the basis for battery SoP estimation. Beneting from the
It is very different from SoC limited method owing to that the
on-line estimated SoE, the prediction of battery SoP for next time
dynamic voltage has been coupled with the battery power. It can be
has become a possibility.
used to evaluate the power capability of a battery with the known
available energy. It is worth noting that the SoE-based battery peak
power predictions are mainly used for avoiding battery experience 4. Verication and discussion
over-charge and over discharge hazards.
Lastly, calculate battery available peak power. The state of This section mainly works for verifying the proposed SoE and
available power capability estimation can be calculated by: SoP joint estimation approach.

< P chg max Pmin ; Pmax


d;z chg
; Ut;kN Imin 4.1. Experiment
min

(40)
: P dis min Pmax ; P c;z ; U I dis
max min t;kN max The battery test bench contains ve components: (1) Arbin
BT2000 cycle: it is used to make battery operate in the designed the
where Pmax, Pmin are the battery's power design limits, Pmax denotes loading proles and measure the real-time operation data. (2)
the discharge power capability and Pmin denotes the charge power Chamber: it is used to control the battery operated under the
capability of its design limits respectively. designed temperatures. (3) LiFePO4 lithium-ion battery: it is used to

8           
> chg d;z Dt chg DtUt;k vUoc;k  Dt chg
>
> P zmax Pmin ; Pmax ; U oc zk  exp UD;k  I  Ro 1  exp RD I
< min t min Ea vz zzk t min

          (41)
>
> DtUt;k vUoc;k 
>
: P dis zmin P c;z Dt dis Dt dis
max max ; Pmin ; Uoc zk  exp UD;k  Imax  Ro 1  exp RD Imax
t Ea vz zzk t
56 W. Zhang et al. / Journal of Power Sources 289 (2015) 50e62

Fig. 1. Framework of the model-based SoE and SoP joint estimator.

verify the proposed approach. (4) Battery management system nd that the relationship between battery OCV and SoE is more
(BMS): it is used to ensure the safety of battery cell and transmit the sensitive to its aging levels than operating temperatures. What's
real-time data to the host computer for computation. (5) Host more, in terms of the same aging level, the mapping relationship
computer: it is used to control the Arbin BT2000 through the MITS between battery OCV and SoE under different temperatures are
Pro software and record the measurement data. similar and the variations of dSoE/dOCV are also close to the same.
The basic parameters of the used LiFePO4 lithium-ion battery Therefore, for the mapping relationship between battery OCV and
cell are listed in Table 2. It uses carbon (C) as its negative electrode SoE, the inuence of the operating temperatures could be ignored
and lithium iron phosphate LiFePO4 as its positive electrode. The for reducing the dimension of battery OCV. The OCV functionf z; Ea
LiFePO4 lithium-ion cell was placed in the thermal chamber. The is used to construct the three dimensions space responsible model
test schedules are designed to generate rich excitations. It contains for describing the mapping relationship among SoE, energy and
static energy test, open circuit voltage test and the federal urban OCV.
driving schedule (FUDS). In each procedure, it begins with a char-
acterization test at temperature T 10  C, followed by two identical f z; Ea M0 M1 z M2 z2 M3 z3 M4 lnz M5 ln1  z
tests conducted at T 25  C and T 40  C. After the series tests, the (42)
cell is loaded with 200 times constant current test for aging process
with 2C rate current and 25  C. where Mi i 0; 1; /; 5 denotes the coefcients of the OCV
The purpose of the static available energy test is to measure the function, they are the functions of battery energy. For the OCV data
cell maximum available energy. Table 3 lists the cell maximum plotted in Fig. 2(c), the parameters of Mi can be solved and the
available energy under different temperatures and aging cycles. results are listed in Table 4.
The OCV test is used to calibrate the relationship between bat- The FUDS test is used to verify the proposed algorithm. Fig. 2(e)
tery OCV and its SoE under different temperatures and aging levels. and (f) plot the FUDS proles for cell with 1000 aging cycles under
Its test method is similar with Ref. [13]. The experimental results for different temperatures.
battery OCV and FUDS cycle test are plotted in Fig. 2. Fig. 2(a)(d)
plot the measured battery OCV under different SoE, temperatures 4.2. Verication
and aging cycles. Where the OCV curve plotted in Fig. 2(a) and (b)
are for the cell with its fresh status, Fig. 2(c) and (d) are for the cell In considering that the OCV data with cycles of 0, 600 and 1500
of different aging cycles but operated at 25  C. From Fig. 2(a) and is used for developing an OCV model, herein we use the data with
(c), we can nd that LiFePO4 battery cell has typical at OCV plat- 1000 cycles for the following verication cases. Before the calcu-
form characteristic. It can be validated by Fig. 2(b) and (d), which lation for battery SoP, we should provide the design limits for
shows that per 10 mV OCV variation may cause more than 40% SoE
inaccuracies when its SoE closes to 0.6. Additionally, we also can
Table 3
Measured cell capacity under different temperatures and aging levels.

Table 2 Energy (Wh) Cycle numbers


Basic parameters of LiFePO4 lithium-ion cell.
0 600 1000 1500
Nominal Nominal Upper cutoff Lower cutoff 
Temperature 10 C 192.5 179.0 145.1 125.1
energy voltage voltage voltage
25  C 202.8 190.9 161.9 136.6
200 Wh 3.2 V 3.6 V 2.5 V 40  C 203.2 192.0 162.2 137.3
W. Zhang et al. / Journal of Power Sources 289 (2015) 50e62 57

Fig. 2. Experimental results for battery OCV and FUDS cycle test: (a) OCV; (b) dSoE/dOCV; (c) OCV; (d) dSoE/dOCV; (e) Current proles of FUDS cycle; (f) SoE proles of FUDS cycle.

battery current and power rstly. The design limits are listed in in Fig. 3, where the initial SoE are incorrectly set to 1.0 and 0.8
Table 5. respectively for analyzing the robust performance of the proposed
Addition to the design limits of battery current and power, the approach. It is noted that the sampling interval is one second in this
optimal operating range for battery SoE is from 0.1 to 1.0. For paper.
different applications, the operating ranges maybe different. Fig. 3(a) shows that the maximum errors of battery voltage are
less than 0.072 V (2% of its upper cutoff voltage). It indicates that
the proposed on-line parameter estimation method is effective for
4.2.1. Verifying the SoE estimator
capturing the real-time characteristic of a battery. This also has
been validated by Fig. 3(c) which initializes the SoE with 1.0. On the
(1) Case 01: SoE estimation under temperature 10  C

Through the detailed operating process of the model-based SoE-


SoP joint estimator and the initial parameter matrices for the AUKF Table 5
algorithm presented in Table 6, we can obtain the on-line estima- Design limits for the energy oriented LiFePO4 battery cell.

tion for battery terminal voltage and SoE. Their results are plotted Discharge current Charge current Discharge power Charge current

360A 240A 1000 W 500 W

Table 4
Parameters of the OCV function f z; Ea .
Table 6
Parameters Value
Parameters of the AUKF algorithm.
3
M0 2.836 9.821  Ea  10
x0 P0 Q0 R0 M
M1 1.491  2.781  Ea  102
M2 1.736 2.916  Ea  102 8  0:8  
> ; when z0 0:8
1000 0 105 0 10 30
0.7324  1.213  Ea  102 <
M3
 0 0 0:1 0 105
M4 0.0730 2.647  Ea  103 >
: 1:0 ; when z 1:0
M5 2.112  1035.622  Ea  104 0 0
58 W. Zhang et al. / Journal of Power Sources 289 (2015) 50e62

Fig. 3. SoE estimation result with initial SoE of 0.8 and 1.0 at 10  C (Case01): (a) Voltage estimation; (b) Voltage estimation error; (c)SoE estimation; (d)SoE estimation error.

other hand, Fig. 3 (b) and (d) show that the AUKF based battery SoE response surface model and which is helpful for reliable battery SoE
estimator can realize the SoE estimation errors of less than 2% estimation. What's more, the AUKF algorithm, which uses UT
regardless positive deviation or negative deviation in its initial SoE. technique to execute a nonlinear transformation on a single point,
Namely, the AUKF based battery SoE estimator has good robust can greatly improve the inner inhibited state estimation
performance against erroneous initial SoEs. What's more, in performance.
considering that battery open circuit voltage is sensitive to its SoE
estimation, the three-dimensional response surface model method 4.2.2. Verifying the SoP estimator
obtains desired prediction precision for battery OCV under uncer-
tainty aging levels between 0 cycle and 1500 cycles. (1) Case 04: SoP estimation under temperature 10  C.

(2) Case 02: SoE estimation under temperature 25  C Through the on-line estimations of battery parameters and SoE,
we can calculate battery available power capability with the
Estimation results of battery terminal voltage and SoE for cell detailed operating process in Fig. 1. The available peak current
operated at the temperature of 25  C are plotted in Fig. 4, the initial capability and power capability prediction results for cell operated
SoE are incorrectly set to 1.0 and 0.8 respectively. The initial at the temperature of 10  C are plotted in Fig. 6 and its initial SoE is
parameter matrices for the AUKF algorithm can be found in Table 6. set to 1.0. In considering that the proposed method can achieve
As expect, we nd that the maximum estimation errors for accurate battery SoE regardless how erroneous the initial SoE does,
battery terminal voltage and SoE are less than 2%. It suggests that the SoP estimation under initial SoE of 0.8 has not been considered
the proposed method can ensure the high estimation accuracy for here. Fig. 6(a) plots battery peak charge current capability, Fig. 6(b)
cell operating under uncertainty temperature and aging level. plots battery peak discharge current capability, Fig. 6(c) plots bat-
tery peak charge power capability, Fig. 6(d) plots battery peak
(3) Case 03: SoE estimation under temperature 40  C discharge power capability.
Fig. 6(a) shows that battery charge current capability is less than
The estimation results of battery terminal voltage and SoE for 120A (2C rate), Fig. 6(b) shows that battery discharge current
cell operated at the temperature of 40  C are plotted in Fig. 5, the capability is less than 240A (4C rate). What's more, in most case its
initial SoE are incorrectly set to 1.0 and 0.8 respectively. The initial charge current capability is close to 80A and its discharge current
parameter matrices for the AUKF algorithm can be found in Table 6. capability is close to 190A at the temperature of 10  C. Fig. 6(a) and
As expect, from Fig. 5, we can also nd that the maximum (b) also show that battery charge currents capability are less than
estimation errors for battery terminal voltage and SoE are less than its discharge current capability. This is because that in most cases
2%. It suggests that the proposed method can ensure the high battery terminal voltage is over 3.2 V, as a result, there is 0.4 V
estimation accuracy for cell operating under uncertainty temper- deviations during the charging of a battery while 0.6 V deviation for
ature and aging level. discharging process of a battery. Additionally, through Eq. (37), we
Based on the three cases, we can conclude that the AUKF algo- can nd that battery polarization voltage also affects the prediction
rithm based SoE estimator work well against uncertain operating result of battery current. When battery SoE closes to 0.95, battery's
temperatures and aging levels of battery cells. The uncertainty of peak charge current capability is less than 60A at 10  C. However, as
battery model and its parameters can be projected through the RLS battery discharging operation goes on, especially for SoE less than
method. The accurate OCV can be predicted by the three-demission 0.9, the absolute value of battery peak charge current capability
W. Zhang et al. / Journal of Power Sources 289 (2015) 50e62 59

Fig. 4. SoE estimation result with initial SoE of 0.8 and 1.0 at 25  C (Case02): (a) Voltage estimation; (b) Voltage estimation error; (c)SoE estimation; (d)SoE estimation error.

increases obviously. This is because that LiFePO4 lithium-ion bat- the model-based method and the constraint from SoE-based power
tery has a at OCV platform characteristic. Fig. 2 shows that battery prediction has not been activated owing that battery SoE has not
OCV reduces greatly when the SoE less than 0.9, while if battery SoE reached 0.1 or 1.0.
is between 0.2 and 0.9, its OCV changes slowly.
Fig. 6(c) and (d) show that battery charge power capability is (2) Case 05: SoP estimation under temperature 25  C.
close to 250 W and battery discharge power capability is close to
650 W in most cases. The above prediction results are obtained by The available peak current capability and SoP prediction results

Fig. 5. SoE estimation result with initial SoE of 0.8 and 1.0 at 40  C (Case03): (a) Voltage estimation; (b) Voltage estimation error; (c)SoE estimation; (d)SoE estimation error.
60 W. Zhang et al. / Journal of Power Sources 289 (2015) 50e62

Fig. 6. SoP prediction result with initial SoE of 1.0 (10  C): (a) Peak charge current; (b) Peak discharge current; (c) Peak charge power; (d) Peak discharge power.

for cell operated at the temperature of 25  C are plotted in Fig. 7, the approximates 800 W in most cases. It is noted that when battery
initial SoE is set to 1.0. Fig. 7(a) plots battery peak charge current SoE is bigger than 0.9, the peak discharge power capability of
capability, Fig. 7(b) plots battery peak discharge current capability, battery will be bigger than 1000 W. However, due to that the
Fig. 7(c) plots battery peak charge power capability, Fig. 7(d) plots 1000 W of power design limits, the maximum discharge power is
battery peak discharge power capability. controlled to 1000 W. On the other hand, at the end of the pre-
Fig. 7(a) shows that battery charge current capability is less diction process, the peak discharge power capability will reach zero
than 150A (2.5C rate), Fig. 7(b) shows that battery discharge cur- owing to the limits from the SoE-based power predictions. Namely,
rent capability is less than 300A (5C rate). What's more, in most when battery SoE less than lower design limits of 0.1, battery's
case its charge current capability is close to 105A and its discharge discharge power is reduced to zero, which is helpful for avoiding
current capability is close to 250A. Compared with the prediction discharge behavior of battery.
results presented in Fig. 6, we can nd that the available energy
and power capacity at temperature of 25  C are bigger than those (3) Case 06: SoP estimation under temperature 40  C.
at 10  C.
Fig. 7(c) and (d) show that battery charge power capability ap- The available peak current capability and SoP prediction results
proximates 350 W and battery discharge power capability for cell operated at the temperature of 40  C are plotted in Fig. 8, the

Fig. 7. SoP prediction result with initial SoE of 1.0 (25  C): (a)Peak charge current; (b) Peak discharge current; (c) Peak charge power; (d) Peak discharge power.
W. Zhang et al. / Journal of Power Sources 289 (2015) 50e62 61

Fig. 8. SoP estimation result with initial SoE of 1.0 (40  C): (a)Peak charge current; (b) Peak discharge current; (c) Peak charge power; (d) Peak discharge power.

initial SoE is set to 1.0. Fig. 8(a) plots battery peak charge current (ii) Resulting from the uncertain operation conditions and aging
capability, Fig. 8(b) plots battery peak discharge current capability, levels of battery cell, it is hard to accurately quantify the
Fig. 8(c) plots battery peak charge power capability, Fig. 8(d) plots model parameters with the archived measured data. There-
battery peak discharge power capability. fore, an online quantication method for model structure
Fig. 8(a) shows that battery charge current capability is less than and parameter is proposed based on the recursive least
150A (2.5C rate), Fig. 8(b) shows that battery discharge current square method and autoregressive exogenous model.
capability is less than 300A. Fig. 8(c) and (d) show that battery (iii) An adaptive unscented Kalman lter is employed to develop
charge power capability is close to 350 W and battery discharge a joint estimator for battery SoE and SoP. This method
power capability is close to 800 W in most cases. Compared with adaptively adjusts the values of the process and measure-
the cell with the same aging level operated at 25  C, battery ment noise covariance. The robustness of the proposed joint
operated 40  C performs a little bigger available energy. Namely, estimator has been veried against different operating
with the same current proles, the end of battery SoE operated temperatures and aging levels. The result indicates that the
40  C is bigger than 0.1. Thus, its discharge power capability is estimation errors of voltage and SoE are less than 2% even if
bigger than zero. give large erroneous initial states of joint estimator, which
It is noted that the above battery SoP predictions are carried out makes the SoP predict more accurate and reliable for the
for battery cell with 1000 aging cycles, thus the predicted current electric vehicles application.
and power capability are much less than its design limits. However,
if we predict current and power for fresh cell, its available current Acknowledgments
and power capability would be bigger than the predictions pre-
sented in Figs. 6e8. What's more, the above predictions are carried This work was supported by the National High Technology
out for 1s, when the last time is more than 1s, the peak power will Research and Development Program of China (Grant Numbers
reduced accordingly. 2013BAG24B02).
Based on the three cases, we can conclude that the reliable
battery SoE estimator is helpful for calculating battery SoP, the
References
erroneous initial SoE can be accurately corrected. Through the real-
time battery model and accurate battery SoE, we can obtain reliable [1] C.C. Chan, The state of the art of electric and hybrid vehicles, Proc. IEEE 90 (2)
battery SoP for ensuring the safety and reliable operation of bat- (2007) 247e275.
teries used in electric vehicles. [2] Z. Ma, J. Jiang, W. Shi, W. Zhang, C. Mi, Investigation of path dependence in
commercial lithium-ion cells for pure electric bus applications: aging mech-
anism identication, J. Power Sources 274 (2015) 29e40.
5. Conclusions [3] F. Sun, R. Xiong, H. He, Estimation of state-of-charge and state-of-power
capability of lithium-ion battery considering varying health conditions,
J. Power Sources 259 (2014) 166e176.
Battery state of energy and available power are critical functions [4] Y. Wang, C. Zhang, Z. Chen, A method for joint estimation of state-of-charge
of battery management system for electric vehicles. For achieving and available energy of LiFePO4 batteries, Appl. Energy 135 (2014) 81e87.
accurate joint estimations, the main work can be summarized. [5] R. Xiong, F. Sun, H. He, T. Nguyen, A data-driven adaptive state of charge and
power capability joint estimator of lithium-ion polymer battery used in
electric vehicles, Energy 63 (2013) 295e308.
(i) In considering that battery energy and available power are [6] G. Plett, Extended Kalman ltering for battery management systems of LiPB-
inuenced by its open circuit voltage (OCV), an improved based HEV battery packs, Part 1. Background, J. Power Sources 134 (2) (2004)
response surface model based OCV model has been deter- 252e261.
[7] G. Plett, Extended Kalman ltering for battery management systems of LiPB-
mined through the systematic analysis for the relationships based HEV battery packs, Part 2. Modeling and identication, J. Power Sources
among battery OCV and other factors. 134 (2) (2004) 262e276.
62 W. Zhang et al. / Journal of Power Sources 289 (2015) 50e62

[8] G. Plett, Extended Kalman ltering for battery management systems of LiPB- Li-ion batteries in electric vehicles by a hardware-in-loop approach, Energies
based HEV battery packs, Part 3. State and parameter estimation, J. Power 5 (5) (2012) 1455e1469.
Sources 134 (2) (2004) 277e292. [21] W. Waag, C. Fleischer, D.U. Sauer, Adaptive on-line prediction of the available
[9] S. Sepasi, R. Ghorbani, B.Y. Liaw, Improved extended Kalman lter for state of power of lithium-ion batteries, J. Power Sources 242 (2013) 548e559.
charge estimation of battery pack, J. Power Sources 255 (2014) 368e376. [22] R. Isermann, U. Baur, W. Bamberger, P. Kneppo, H. Siebert, Comparison of six
[10] M.H. Chang, H.P. Huang, S.W. Chang, A new state of charge estimation method on-line identication and parameter estimation methods, Automatica 10 (1)
for LiFePO4 battery packs used in robots, Energies 6 (4) (2013) 2007e2030. (1974) 81e103.
[11] J. Wang, J. Guo, L. Ding, An adaptive Kalman ltering based State of Charge [23] S. Yuan, H. Wu, C. Yin, State of charge estimation using the extended Kalman
combined estimator for electric vehicle battery pack, Energ. Convers. Manag. lter for battery management systems based on the ARX battery model, En-
50 (2009) 3182e3186. ergies 6 (1) (2013) 444e470.
[12] Saeed Sepasi, Reza Ghorbani, Bor Yann Liaw, A novel on-board state-of-charge [24] C. Yi-Hsien, S. Wu-Yang, K. Jia-Cheng, Online estimation of internal resistance
estimation method for aged Li-ion batteries based on model adaptive and open-circuit voltage of lithium-ion batteries in electric vehicles, J. Power
extended Kalman lter, J. Power Sources 245 (2014) 337e344. Sources 196 (8) (2011) 3921e3932.
[13] R. Xiong, F. Sun, X. Gong, C. Gao, A data-driven based adaptive state of charge [25] M.T. Sabet, P. Sarhadi, M. Zarini, Extended, Unscented, Kalman lters for
estimator of lithium-ion polymer battery used in electric vehicles, Appl. En- parameter estimation of an autonomous underwater vehicle, Ocean Eng. 91
ergy 113 (2014) 1421e1433. (2014) 329e339.
[14] S.J. Moura, N.A. Chaturvedi, K. Miroslav, Adaptive Partial Differential Equation [26] Y. Tian, B. Xia, W. Sun, Z. Xu, W. Zheng, A modied model based state of
Observer for Battery State-of-Charge/State-of-Health Estimation Via an Elec- charge estimation of power lithium-ion batteries using unscented Kalman
trochemical Model, J. Dyn. Syst.-T ASME 136 (1) (2014) 11011e11015. lter, J. Power Sources 270 (2014) 619e626.
[15] F. Sun, R. Xiong, A novel dual-scale cell state-of-charge estimation approach [27] H.F. Dai, X.Z. Wei, Z.C. Sun, Design and implementation of a UKF-based SOC
for series- connected battery pack used in electric vehicles, J. Power Sources estimator for LiMnO2 batteries used on electric vehicles, PRZEGLAD ELEK-
274 (2015) 582e594. TROTECHNICZNY 88 (1B) (2012) 57e63.
[16] X. Liu, J. Wu, C. Zhang, Z. Chen, A method for state of energy estimation of [28] J. Zhang, C. Xia, State-of-charge estimation of valve regulated lead acid battery
lithium-ion batteries at dynamic currents and temperatures, J. Power Sources based on multi-state Unscented Kalman Filter, Int. J. Elec. Power 33 (3) (2011)
270 (2014) 151e157. 472e476.
[17] K. Mamadou, A. Delaille, E. Lemaire-Potteau, Y. Bultel, The state-of-energy: a [29] R. Xiong, F. Sun, Z. Chen, H. He, A data-driven multi-scale extended Kalman
new criterion for the energetic performances evaluation of electrochemical ltering based parameter and state estimation approach of lithium-ion
storage devices, ECS Trans. 25 (35) (2010) 105e112. polymer battery in electric vehicles, Appl. Energy 113 (2014) 463e476.
[18] K. Mamadou, E. Lemaire, A. Delaille, D. Riu, S.E. Hing, Y. Bultel, Denition of a [30] Saurabh S. Bisht, Mahendra P. Singh, An adaptive unscented Kalman lter for
State-of-Energy Indicator (SoE) for electrochemical storage devices: applica- tracking sudden stiffness changes, Mech. Syst. Signal PR 49 (1e2) (2014)
tion for energetic availability forecasting, J. Electrochem. Soc. 159 (8) (2012) 181e195.
A1298eA1307. [31] Chingiz Hajiyev, Halil Ersin Soken, Robust adaptive unscented Kalman lter
[19] L.P. Gregory, High-performance battery-pack power estimation using a dy- for attitude estimation of pico satellites, Int. J. Adapt. Control 28 (2) (2014)
namic cell model, IEEE Trans. Veh. Technol. 53 (2004) 1586e1593. 107e120.
[20] R. Xiong, H. He, F. Sun, K. Zhao, Online estimation of peak power capability of

Вам также может понравиться