You are on page 1of 19

Ionics (2022) 28:3321–3339

https://doi.org/10.1007/s11581-022-04574-8

ORIGINAL PAPER

Study of impacts of parameters identification methods


on model‑based state estimation for ­LiFePO4 battery
Shiyi Fu1 · Taolin Lv1 · Wen Liu1 · Lei Wu1 · Chengdong Luo1 · Jingying Xie1

Received: 12 December 2021 / Revised: 11 April 2022 / Accepted: 15 April 2022 / Published online: 12 May 2022
© The Author(s), under exclusive licence to Springer-Verlag GmbH Germany, part of Springer Nature 2022

Abstract
Model-based methods are widely used for online states estimation of electric vehicles (EVs) due to its accuracy and robust-
ness. Current research mainly focuses on improving estimation filters. However, there is less discussion on the parameters
identification methods. In this work, the parameters identification method is divided into two parts: formulation of the
regression model and application of identification algorithms. First, two methods of formulation of the regression model are
studied, respectively. Then, performances on parameters identification of forgetting factor recursive least squares (FFRLS),
optimal bounding ellipsoid (OBE), and linear Kalman filter (LKF) are discussed. Besides, cubature Kalman filter (CKF) is
selected for state of charge (SOC) estimation. In order to obtain the experimental data and verify the parameters identification
and states estimation accuracy, the Hybrid Pulse Power Characteristic (HPPC) test, urban dynamometer driving schedule
(UDDS) test, and new European driving cycle (NEDC) test are carried out. In addition, the maximum absolute error (MAE),
mean absolute error (MaE), and root mean square error (RMSE) are calculated for evaluating the SOC estimation accuracy.
When the parameters are identified by LKF, the best performance in MAE, MaE, and RMSE of SOC estimation is obtained.
Considering the estimated peak power fluctuation and the identified parameters fluctuation, the OBE algorithm is more
suitable for co-estimation of SOC and state of power (SOP).

Keywords  Lithium-ion battery · Regression model · Parameters identification · State of charge · State of power

Nomenclature EKF Extended Kalman filter


EVs Electric vehicles DEKF Dual extended Kalman filter
BMS Battery management system UKF Unscented Kalman filter
SOC State of charge CKF Cubature Kalman filter
SOP State of power RLS Recursive least squares
SOH State of health NC Noise compensating
SOE State of energy PF Particle filter
OCV Open circuit voltage UPF Unscented particle filter
ANN Artificial neural network HF H infinity filter
SVM Support vector machine NMC Lithium nickel manganese cobalt oxide
STSA Symbolic time series analysis LFP Lithium iron phosphate
EM Empirical model PNGV Partnership for new generation vehicles
ECM Electrochemical model HPPC Hybrid pulse power characteristic
EECM Electric equivalent circuit model FFRLS Forgetting factor recursive least squares
KF Kalman filter OBE Optimal bounding ellipsoid
LKF Linear Kalman filter RM1 The first regression model
RM2 The second regression model
UDDS Urban dynamometer driving schedule
* Jingying Xie
jyxie@hit.edu.cn NEDC New European driving cycle
MAE Maximum absolute error
1
Space Power Technology State Key Laboratory, Shanghai MaE Mean absolute error
Institute of Space Power-Sources, 2965#, Dongchuan Road, RMSE Root mean square error
Shanghai 200245, China

13
Vol.:(0123456789)
3322 Ionics (2022) 28:3321–3339

Introduction [12], symbolic time series analysis (STSA) [13], and their
derivatives have been developed to estimate SOC. However,
The world is currently facing an environmental crisis with such methods require a great quantity of data samples to
an increase in the energy demand and aggravation of green- train the estimation model. Once the data quality of the
house effect [1, 2]. To solve this problem, an enormous training set is poor and the amount of data is insufficient,
amount of research on renewable energy has been carried the estimation accuracy will be greatly reduced [14]. The
out in the past few decades. Today’s renewable energy not model-based estimation method is developed on the prem-
only affects national economy and policies, but also affects ise of accurately describing the electrical behavior of the
international trade relations [3, 4]. Lithium ion batteries have battery [15]. In order to use the model-based method to
the advantages of longer cycle life, lower life cycle cost, estimate SOC, the selection of battery model and the iden-
higher power, and energy density, making lithium ion bat- tification of model parameters are essential. Battery models
teries the focus of renewable energy research [5, 6]. So far, can be divided into empirical model (EM) [15], electro-
the lithium ion batteries have been widely used in electric chemical model (ECM) [16], and electric equivalent circuit
vehicles (EVs), energy storage, aerospace, and other fields. model (EECM) [17]. By introducing the constant phase ele-
In order to ensure that the performance of the battery ment (CPE), the fractional-order EECM is obtained and
during its service life meets the using requirements, a bat- may have a higher accuracy in battery dynamics simula-
tery manage system (BMS) is indispensable. BMS monitors tion [18]. EECM has an excellent trade-off between model
and predicts the state of the battery when it is working and complexity and accuracy [19], making it a better choice for
makes corresponding optimization decisions to maintain the model-based SOC estimation. Among numerous EECMs,
best performance of the battery [7, 8]. It should be noted that Thevenin model has been proved to be the best choice for
the battery state estimation is the most essential function of battery modeling [17]. Based on the model establishment,
a BMS, such as state of charge (SOC), state of power (SOP), algorithms like Kalman filter (KF) [20], extended Kalman
and state of health (SOH). As the states reflect the perfor- filter (EKF) [21], unscented Kalman filter (UKF) [22], par-
mance of the battery at each moment, the estimation of SOC ticle filter (PF) [23], unscented particle filter (UPF) [24],
and SOP are supposed to be fast and accurate. and H infinity filter (HF) [25] are applied to battery SOC
SOC estimation methods can be mainly divided into four estimation.
categories, including the coulomb counting method, the It is not hard to see that the main factors affecting the
look-up table-based method, the data-driven method, and accuracy of model-based SOC estimation method are as fol-
the model-based method. The coulomb counting method is lows: the performances of estimation algorithm like filters
the simplest and the most basic SOC estimation method. and the accuracy of model establishment. A lot of research
The SOC can be directly calculated by integrating the prod- has been carried out on estimation algorithms. Except for
uct of time and current, and its accuracy is affected by fac- studying the estimation performance of various filtering
tors such as sensor error, initial SOC error, and capacity algorithms, some auxiliary algorithms are also introduced
error. However, the coulomb counting method is an open- to improve these filters, such as adaptive algorithms [26, 27]
loop estimation method. Once these above-mentioned errors and multi-innovation algorithm [28, 29]. The establishment
occur, the estimated result will deviate from the reference of a battery model can be divided into two parts: fitting the
value and cannot be corrected. Hence, in order to enhance OCV model and identifying the lumped parameter elements
the performance of the coulomb counting method, other in EECMs. OCV tests have been comprehensively studied
methods are needed to calibrate the initial SOC and the [30]. Results show that OCV obtained from the incremen-
capacity [9]. Look-up table method estimates the SOC tal OCV test can improve the accuracy of SOC estimation.
based on measuring specific parameters of battery and sub- In another study, the impact of eighteen OCV model on
stitute it in mapping relationship between SOC and itself. SOC estimation were compared [31], and suggestions were
The monotone functional relationship between open circuit given for the selection of OCV models for Lithium Nickel
voltage (OCV) and SOC makes OCV an effective approach Manganese Cobalt Oxide (NMC) battery and Lithium Iron
to estimate SOC. However, the stable value of OCV requires Phosphate (LFP) battery. The LS-based algorithm is widely
a long time rest [10], which means that this kind of method used in battery parameters identification with its low com-
is not suitable for online SOC estimation. In the applica- puting cost. However, LS-based algorithm has the problem
tion of data-driven methods, the battery is regarded as a of uncontrollable divergence caused by the blowing-up
“black box.” The physical process of the battery itself is of covariance matrix [32] and the characteristic of noise-
not considered, and only the hidden relationship between sensitiveness [33]. Methods including Frisch scheme based
input and output is concerned. Algorithms such as artificial bias compensating RLS (FBCRLS) [33], adaptive forgetting
neural network (ANN) [11], support vector machine (SVM) recursive total least squares (AF-RTLS) [34], and combina-
tion of LS and variable projection algorithm (LSVPA) [35]

13
Ionics (2022) 28:3321–3339 3323

are studied to eliminate the noise effect on SOC estima- identification. Then, the cubature Kalman filter (CKF) is
tion. Three integrated model parameters identification and used for SOC estimation and the peak power is further esti-
SOC estimation methods have been compared in Ref. [36], mated. Furthermore, the capability of algorithms mentioned
including dual extended Kalman filter (DEKF), recursive above on obtaining reliable results of model parameters is
least squares EKF (RLS-EKF) and noise compensating discussed. And the impacts of parameters identification on
EKF (NC-EKF). The DEKF and NC-EKF are more robust state estimation is also analyzed.
to noise corruption than RLS-EKF. However, the complexity
of calculation process increase along with the innovation and
optimization of the algorithms. Hence, in practical appli- Battery model
cation, accuracy, robustness, and consumption of comput-
ing resource must be taken into account when selecting the Thevenin model is widely used due to its high accuracy and
method. low complexity. As shown in Fig. 1, the Thevenin model
SOP estimation methods can be divided into the Part- consists of OCV model and lumped parameter elements. The
nership for new generation vehicles (PNGV) Hybrid Pulse OCV model is represented by an ideal voltage source, which
Power Characteristic (HPPC) method and the model-based is actually a mapping function between OCV and SOC. The
method. The PNGV HPPC method only considers the con- series resistance R0 and an RC parallel network are used to
strain of voltage when estimating SOP. Besides, the battery predict the response of the battery to the load [42]. And by
dynamics is expressed in a simple form, and the battery state using Kirchhoff’s law, the dynamics of Thevenin model can
is not fully considered. However, the battery power capac- be written as Eq. (1).
ity is affected by many factors, such as the battery states {
Ut = UOCV + It R0 + UC
and the temperature. Thus, the PNGV-HPPC method may I U (1)
U̇ C = Ct − 𝜏C
generate overly optimistic results of the SOP estimation, 1

and further resulting in battery safety problems [37]. In the


model-based method, advanced battery model is introduced
to dynamically simulate the battery dynamics, by using the where Ut is the terminal voltage, UC is the polarization
EECM, the links of the model with battery state of energy voltage, UOCV is the OCV, and It is the load current.
(SOE), SOC, and voltage are strengthened [38, 39]. And by The battery SOC is defined by the ampere hour measure-
introducing the electrochemical-Thermal model, the impact ment method, which is expressed as follows:
of temperature on peak power estimation is also taken into 𝜂It,k Δt
account [40, 41]. SOCk = SOCk−1 + (2)
Qcap
In this work, we mainly focus on the parameters iden-
tification methods integrated with regression model. The where η is the Coulomb efficiency; △t is the sampling
formulating methods of two regression models are stud- time interval; Qcap is the available battery capacity; and k
ied. Based on this, forgetting factor recursive least squares is the time step.
(FFRLS), optimal bounding ellipsoid (OBE), and linear In order to estimate SOC by using model-based meth-
Kalman filter (LKF) algorithms are adopted for parameters ods, a state vector is defined as x = [SOC UC]T. Then by

Fig. 1  Thevenin model UC

C1
R0
It
R1
UOCV(SOC) Ut

13
3324 Ionics (2022) 28:3321–3339

combining Eq. (1)–(3), the discrete-time state space model Ut,k = a0 Ut,k−1 + Mk + a1 It,k + a2 It,k−1 (8)
is formulated as follows:
� � The second regression model (­ RM2) is formulated based
⎧ 1 0 � ⎡ 𝜂Δt
⎤ on discretization of differential elements of the battery
⎪ xk = � + ⎢ � Q cap � �� ⎥I
⎨ 0 exp − Δt
xk−1
⎢ R1 1 − exp − Δt ⎥ t,k dynamics in Eq. (1). We can get the following:
⎪ �
𝜏
� ⎣ 𝜏 ⎦
⎩ Ut,k = UOCV SOCk + It,k R0 + UC,k ⎧ It = UC + C1 dUC
(3) ⎪ R1 dt
⎨ UC = Ut − UOCV − It R0 (9)
⎪ dUC = dUt − dUOCV − R dIt
⎩ dt dt dt 0 dt

Parameter identification and state It has been proved that during a short time interval dtOCV
dU

estimation ≈ 0 [43]. Hence, the Eq. (9) can be further transformed to


the following:
In this section, the regression model and parameter identifica-
tion algorithm of battery parameters are discussed respectively. ( ) dU dI
Ut = UOCV + R0 + R1 It − 𝜏 t + 𝜏R0 t (10)
The derivation process of two regression models is given, and dt dt
three identification algorithms are given in detail. In addition,
By discretizing Eq. (11), the regression model can be
it should be noted that this work mainly studies the impact of
expressed as follows:
different parameters identification methods on state estimation,
( ) ( )
thus state estimation of all methods is based on CKF. Ut,k = UOCV,k + b0 It,k + b1 Ut,k − Ut,k−1 + b2 It,k − It,k−1
(11)
Regression model
where b0, b1 and b2 are related to R0, R1 and τ, and they
are expressed as follows:
The regression model of a system is written as follows:
⎧b = R +R
Y = AX + v (4) ⎪ 0 0 1
⎨ b1 = 𝜏 (12)
where Y denotes the output vector; X denotes the input ⎪ b2 = 𝜏R0
vector; A denotes the coefficient vector; and v denotes the ⎩
error term. The common point of these two regression models is
Obviously, battery is a nonlinear system. Hence, the that the terminal voltage of battery is taken as output, and
formulating of regression model is to find the linear rela- the voltage, current, and their linear combination are taken
tionship between measurement signals and battery model as input, φ1 = [Ut,k−1 1 It,k It,k−1]T for R ­ M1 and φ2 = [1 It,k
parameters. (Ut,k − Ut,k−1) ( It,k − It,k-1)]T for R
­ M2. And both of them have
The first regression model ­(RM1) is formulated based on the ability to extract the OCV from battery test. The param-
the transfer function of Eq. (1), which is written as follows: eter vector of the ­RM1 is defined as θ1 = [a0 Mk a1 a2]T, and
U(s) − UOCV (s) R0 + R1 + R0 𝜏s that of ­RM2 is defined as θ2 = [UOCV b0 b1 b2]T. Then the
= (5) regression models can be written as follows:
It (s) 1 + 𝜏s
{ T
By applying the bilinear transform s = 2(1 − z−1)/(1 + z−1) 𝜃1,k 𝜑1,k RM1
Ut,k = T
𝜑2,k RM2 (13)
/ △t, and change it into time series, Eq. (4) can be trans- 𝜃2,k
formed to:
where T the is matrix transpose operator.
Ut,k = a0 Ut,k−1 + UOCV,k − a0 UOCV,k + a1 It,k + a2 It,k−1 (6)
Algorithms for parameters identification
where a0, a1, and a2 are related to R0, R1, and τ, and they
are expressed as follows: The first algorithm is FFRLS, which is a variant of RLS. The
forgetting factor is employed to give less weight to the previ-
⎧ a0 = 2𝜏−Δt
⎪ 2𝜏+Δt
R1 Δt+R0 (2𝜏+Δt)
ous data, while more weight to current data. It can avoid the
⎨ a1 = 2𝜏+Δt (7) problem of classical RLS, which easily leads to saturation
⎪a = R1 Δt+R0 (−2𝜏+Δt)
[44]. The recursive process of FFRLS is given in Table 1.
⎩ 2 2𝜏+Δt
where λ is forgetting factor; K is gain vector; P is covari-
Define Mk = UOCV,k − a0UOCV,k−1, the regression model can ance matrix; and I is the unit matrix with the same order as
be expressed as follows: the parameter vector.

13
Ionics (2022) 28:3321–3339 3325

Table 1  Algorithm of FFRLS

The second algorithm is OBE. This algorithm uses ellip- CKF is an effective method for nonlinear system [47]. The
soid center and symmetric covariance matrix to characterize recursive process of CKF is given in Table 4.
the uncertainty of point estimation results and estimation where K is kalman gain vector; P is covariance matrix; Q
results, respectively. By continuously modifying the feasible is process noise covariance matrix; R is measurement noise
point estimation and the corresponding parameter bounded covariance matrix; ξi is the ith column of ξ; and n denotes
interval containing real parameters, the parameters can be the dimension of the state vector and n = 2.
effectively identified [45]. The recursive process of OBE is
shown in Table 2. SOP estimation
where λ is forgetting factor; K is gain vector; P is covari-
ance matrix; γ is the noise bound; 𝜎02 is the nonzero scalar; a Generally, the index of SOP estimation is actually battery
is a constant, and 0 < a < 1; α is defined as α = 1-λ. peak power. The peak power includes the peak charge power
The third algorithm is LKF. The Kalman filter was orig- and the peak discharge power, and they can be expressed as
inally designed for linear system. The principle of KF is follows:
minimizing the average of the square error, which provides { dis dis dis
a reliable recursive way to estimate the state of the system Pmin = Imin Ut
(14)
[46]. It can be seen from Eq. (8) and Eq. (11) that the regres- Pcha
max = I cha cha
max Ut
sion model has the ability to express the nonlinear battery
where Pdis max represent the peak discharge/charge
∕Pcha
system in a linear way. Hence, by defining θ as state vector, min
power; Imin ∕Imax represent the peak discharge/charge cur-
dis cha
the parameters can be identified by LKF. The recursive pro-
rent; Udis t represent the battery terminal voltage under
∕Ucha
cess of LKF is given in Table 3. t
corresponding condition. And it is not hard to see that the
where K is kalman gain vector; P is covariance matrix; Q
key to estimate peak power is the peak current.
is process noise covariance matrix; and R is measurement
Based on Eq. (1) and battery states estimated at time step
noise covariance matrix.
k, we have the expressions of battery voltage at time step
k + 1 shown as below:
CKF‑based SOC estimation
{ ( )
Ut,k+1 = UOCV SOC ( k+1 )+ It,k+1(R0 + UC,k+1
( ))
CKF is designed based on using the spherical radial volume
UC,k+1 = UC,k exp − Δt + R 1 − exp − Δt
It,k+1
rule to numerically compute multivariate moment integrals 𝜏 1 𝜏

encountered in the nonlinear Bayesian filter. The CKF is (15)


derivative-free. Compared with EKF, it omits the steps of The polarization voltage UC,k is a known quantity; thus,
solving Jacobian matrix and can be applied to higher dimen- once the terminal voltage Ut,k+1 is calculated by current
sional nonlinear systems. Compared with UKF, the square known state, the load current It,k+1 can be further estimated.
root covariance of CKF is always positive definite, which However, as expressed in Eq. (2), the functional relationship
avoids the numerical unsolvable problem. Therefore, the between ­SOCk+1 and Ik+1 cannot be ignored. Meanwhile,

13
3326 Ionics (2022) 28:3321–3339

Table 2  Algorithm of OBE

Table 3  Algorithm of LKF

13
Ionics (2022) 28:3321–3339 3327

Table 4  Algorithm of CKF

the UOCV(·) is a nonlinear function; thus, Eq. (15) cannot be According to [37], the UOCV(·) can be regarded as dif-
directly use to solve the load current. ferentiable at step k, with the first Taylor-series expansion,
Eq. (15) can be written as follows:

( ) ( ( ( )))
( ) Δt 𝜂Δt 𝜕UOCV (SOC) | Δt
Ut,k+1 = UOCV SOCk + UC,k exp − + It,k+1 |SOC=SOCk +R +R 1 − exp − (16)
𝜏 Qcap 𝜕SOC | 0 1
𝜏

13
3328 Ionics (2022) 28:3321–3339

Hence, the load current Ik+1 can be expressed as follows: where Ut,min and Ut,max represent the lower and upper cut-
( ) ( ) off voltage; Idis,vol
min,k+1
and Icha,vol
max,k+1
represent the estimated peak
Ut,k+1 − UOCV SOCk − UC,k exp − Δt 𝜏 current under voltage constrain. And it should be noted that
It,k+1 = ( ( )) the results are instantaneous.
𝜂Δt 𝜕UOCV (SOC) | Δt
| +R + R 1 − exp − For a state space model shown as Eq. (3), assume that in
Qcap 𝜕SOC |SOC=SOC k 0 1 𝜏
(17) continuous period L × △t, the system input keeps constant,
The battery must strictly operate within the cut-off volt- i.e.,
age range, which gives the following: uk+L = uk+L−1 = uk+L−2 = ⋯ = uk+1 = uk (19)
� �
⎧ Ut,min −UOCV (SOCk )−UC,k exp − Δt We have the estimated state results in step k + L as
⎪ I dis,vol = 𝜂Δt 𝜕UOCV (SOC) �

𝜏
� �� follows:
⎪ min,k+1 �SOC=SOCk +R0 +R1 1−exp − 𝜏

Δt


Qcap 𝜕SOC
� � (18) (∑L−1 )
Ut,max −UOCV (SOCk )−UC,k exp − Δt
⎪ I cha,vol = �
𝜏
� �� xk+L = ALk xk + AL−1−i
k
Bk uk (20)
𝜂Δt 𝜕UOCV (SOC) �
⎪ max,k+1 � +R +R 1−exp − Δt
i=0
⎩ Qcap 𝜕SOC � SOC=SOC k 0 1 𝜏

In this case, Eq. (15) can be written as follows:

( ) ( )
Δt L
Ut,k+1 = UOCV SOCk + UC,k exp −
( 𝜏 ) (21)
𝜂Δt 𝜕UOCV (SOC) | ( ( )) ∑L−1 ( )
Δt Δt L−1−i
+It,k+1 L × |SOC=SOCk +R0 +R1 1 − exp − exp −
Qcap 𝜕SOC | 𝜏 i=0 𝜏

Then the estimation of continuous peak current under peak discharge and charge current under SOC constrained,
voltage constrained can be expressed as follows: respectively.
� �L Considering the current limits, the instantaneous and con-
⎧ dis,vol Ut,min −UOCV (SOCk )−UC,k exp − Δt tinuous peak current can be further expressed as follows:
⎪ Imin,L =
𝜏
� � �� ∑ � �L−1−i
𝜕UOCV (SOC) �
⎪ L× Q𝜂Δt �SOC=SOCk +R0 +R1 1−exp − 𝜏
Δt L−1 Δt
i=0 exp − 𝜏

cap 𝜕SOC �
� � L
⎧ dis � �
Ut,max −UOCV (SOCk )−UC,k exp − Δt dis,vol dis,SOC
⎪ I cha,vol = 𝜏
⎪ min,k+1
I = max I I I
min min,k+1 min,k+1
⎪ max,L L× Q𝜂Δt
𝜕UOCV (SOC) �
� �
�SOC=SOCk +R0 +R1 1−exp − 𝜏
Δt
�� ∑L−1 �
Δt
�L−1−i
⎪ cha � �
⎩ 𝜕SOC � i=0 exp − 𝜏 cha,vol cha,SOC
⎪ Imax,k+1 = min Imax Imax,k+1 Imax,k+1
cap

(22) ⎨ dis � � (24)


dis,vol dis,SOC
Besides the voltage constrain, the battery should also ⎪ Imin,L = max Imin Imin,L Imin,L
⎪ cha � �
work within the allowed SOC range, which will also affect cha,vol cha,SOC
⎪ Imin,L = min Imax Imax,L Imax,L
the estimation of peak current. According to Eq. (2), when ⎩
­ OCk and the limits of SOC (i.e., the
we have the estimated S
where Imax represents the designed maximum charge cur-
­SOCmin and the ­SOCmax), the estimation of SOC constrained
rent, and Imin represents the designed minimum discharge
peak current can be expressed as follows:
current. Then, based on Eq. (14), Eq. (16), Eq. (21), and
⎧ dis,SOC SOCmin −SOCk Eq. (24), the peak power can be further estimated.
⎪ Imin,k+1 = 𝜂Δt∕Qcap
⎪ cha,SOC SOCmax −SOCk
⎪ Imax,k+1 = 𝜂Δt∕Qcap
⎨ dis,SOC SOCmin −SOCk (23)
⎪ Imin,L = L×𝜂Δt∕Qcap
⎪ I cha,SOC = SOCmax −SOCk
⎪ max,L L×𝜂Δt∕Qcap Table 5  Designed limit parameters

Items Upper Lower
where Idis,SOC
min,k+1
and Icha,SOC
max,k+1
represent the instantaneous
peak discharge and charge current under SOC constrained, Voltage 3.65 V (Umax) 2.5 V (Umin)
respectively; Idis,SOC and Icha,SOC represent the continuous Current 250 A (Imax)  − 250 A (Imin)
min,L max,L
SOC 90% ­(SOCmax) 10% ­(SOCmin)

13
Ionics (2022) 28:3321–3339 3329

Fig. 2  Parameters identification RM1 RM2 Reference


results of UDDS test: a–c R0; (a) 3 (b) 3 (c) 3
d–f R1; g–i τ  OBE FFRLS LKF

R0 (mΩ)

R0 (mΩ)

R0 (mΩ)
2 2 2

1 1 1

0 0 0
0 2 4 6 8 0 2 4 6 8 0 2 4 6 8
Time (h) Time (h) Time (h)
(d) 3 (e) 3 (f) 3
OBE FFRLS LKF

R1 (mΩ)

R1 (mΩ)

R1 (mΩ)
2 2 2

1 1 1

0 0 0
0 2 4 6 8 0 2 4 6 8 0 2 4 6 8
Time (h) Time (h) Time (h)
(g) 20 (h) 20 (i) 20
OBE FFRLS LKF
τ (s)

τ (s)

τ (s)
10 10 10

0 0 0
0 2 4 6 8 0 2 4 6 8 0 2 4 6 8
Time (h) Time (h) Time (h)

Fig. 3  Parameters identification RM1 RM2 Reference


results of NEDC test: a–c R0; (a) 3 (b) 3 (c) 3
d–f R1; g–i τ  OBE FFRLS LKF
R0 (mΩ)

R0 (mΩ)

R0 (mΩ)
2 2 2

1 1 1

0 0 0
0 5 10 0 5 10 0 5 10
Time (h) Time (h) Time (h)
(d) 3 (e) 3 (f) 3
OBE FFRLS LKF
R1 (mΩ)

R1 (mΩ)

R1 (mΩ)
2 2 2

1 1 1

0 0 0
0 5 10 0 5 10 0 5 10
Time (h) Time (h) Time (h)
(g) 20 (h) 20 (i) 20
OBE FFRLS LKF
τ (s)

τ (s)

τ (s)

10 10 10

0 0 0
0 5 10 0 5 10 0 5 10
Time (h) Time (h) Time (h)

Experimental study Parameters identification

In this work, a prismatic LFP battery with a nominal capac- The results of model parameters identification are shown
ity of 50 Ah was selected for testing. The HPPC test, urban in Figs. 2 and 3. It should be noted that the reference data
dynamometer driving schedule (UDDS) test, and new Euro- were offline extracted from HPPC test data. The legend
pean driving cycle (NEDC) test were carried out at room ­RM1 and ­RM2 refer to the first regression model and the
temperature (25 ℃). The experimental voltage and current second regression model, respectively. As can be seen
data were obtained from the Arbin SCTS battery test sys- from the figures, the identification results of R0 show good
tem. The designed limit parameters of the battery are listed convergence to the reference, while the results of R1 and τ
in Table 5. show divergence. When integration of regression models

13
3330 Ionics (2022) 28:3321–3339

Fig. 4  OCV identification RM1 RM2 Reference


results: a–c UDDS test; d–f (a) (b) (c)
NEDC test 3.5 3.5 3.5

OCV (V)

OCV (V)

OCV (V)
3 3 3

OBE FFRLS LKF


2.5 2.5 2.5
0 2 4 6 8 0 2 4 6 8 0 2 4 6 8
Time (h) Time (h) Time (h)

(d) (e) (f)


3.5 3.5 3.5
OCV (V)

OCV (V)

OCV (V)
3 3 3

OBE FFRLS LKF


2.5 2.5 2.5
0 5 10 0 5 10 0 5 10
Time (h) Time (h) Time (h)

and algorithms changes, the identification results changed. of model parameters, the results of online identified OCV
Among these three algorithms, identification results of the show deviation to the reference. Among all results, the OCV
OBE algorithm show the worst consistency to the reference, obtained by LKF from UDDS test show best consistency
the FFRLS follows, and the LKF is the best. In addition, when regression model changes, and the deviation with ref-
when integrated with ­RM2, the results are closer to the ref- erence is small.
erence compared with that integrated with R ­ M1. It should
be noted that the offline reference data are obtained by fit- Impact factors on SOC estimation
ting the dynamics of battery, the fitting errors are inevita-
ble. Hence, even the deviation between online identifica- OCV curve
tion results and the reference exists, it does not mean that
the online identification results are wrong. Meanwhile, the In order to verify the effectiveness of OCV data, the offline
model parameters may change when the battery working OCV data obtained from HPPC test and online OCV data
condition changes and battery aging occurs; thus, online obtained by LKF integrated with R ­ M2 from UDDS test are
parameters identification is indispensable. used for OCV-SOC curve fitting, respectively. The OCV
From Eqs. (9) and (12), we can know that the OCV can model used for curve fitting is shown as below:
also be extracted from the dynamic test, and the results are ( ) ( )
∑ ni ∑ nj x ∑nk x − ek
shown in Fig. 4. The reference OCV data are obtained from U(x) = ai + bj exp − + dk tanh
i=0 j=0 cj k=0 fk
HPPC test as well. It can be seen that when integrated with
­RM2, more stable OCV results are obtained compared to (25)
that integrated with R ­ M1. Besides, similar to the results

Fig. 5  OCV fitting results: Reference Fitted curve


a online OCV data based; b (a) 3.6 (b) 3.6
offline OCV data based
3.4 3.4
OCV (V)

OCV (V)

3.2 3.2

3 3

2.8 2.8

2.6 2.6

100 80 60 40 20 0 100 80 60 40 20 0
SOC (%) SOC (%)

13
Ionics (2022) 28:3321–3339 3331

Fig. 6  Results of voltage esti- Offline OCV-curve based Online OCV-curve based Reference
mation and corresponding error: (a) (b) 0.1 UDDS test
a–b UDDS test; c–d NEDC test 3.5 UDDS test

Voltage (V)
0

Error (V)
3 -0.1

-0.2
2.5
0 2 4 6 8 0 2 4 6 8
Time (h) Time (h)
(c) (d) 0.1
3.5 NEDC test NEDC test

Voltage (V)

Error (V)
0
3
-0.1

2.5 -0.2
0 5 10 0 5 10
Time (h) Time (h)

Fig. 7  Results of SOC estima- Offline OCV-curve based Online OCV-curve based Reference
tion and corresponding error: (a) 100 (b) 2
a–b UDDS test; c–d NEDC test 0

Error (%)
SOC (%)

-2
50
-4
-6
UDDS test UDDS test
0 -8
0 2 4 6 8 0 2 4 6 8
Time (h) Time (h)
(c) 100 (d)
0
Error (%)
SOC (%)

50
-5

NEDC test NEDC test


0 -10
0 5 10 0 5 10
Time (h) Time (h)

where U(x) denotes OCV; x denotes SOC; a, b, c, d, e part are offline identified from HPPC test data. Obviously,
and f are coefficients need to be fitted; ni is the number of large errors can be found in the results of offline OCV-curve-
constant terms, and nj and nk are the number of exponential based voltage and SOC estimation, the maximum absolute
function and hyperbolic tangent function terms, respectively voltage error exceeds 0.2 V, and the maximum absolute
[48–50]. SOC error is nearly 10%. In contrast, in the results of online
The curve fitting results are shown in Fig. 5. As can be OCV-curve-based voltage and SOC estimation, the maxi-
seen from the figure, the OCV model has high fitting accu- mum voltage error is limited within ± 0.05 V, and the maxi-
racy, and the trends of these two fitted curves are almost mum SOC error is limited within 2%. When using the offline
the same. By comparison, the online-based OCV curve has OCV-curve, both estimated voltage and SOC have large
higher voltage platform, and the numbers of data points deviation from the reference. These results indicate that the
using for curve fitting are more than the offline-based one. online OCV-curve has better adaptive for battery dynamic
In addition, the difference between the two voltage platforms characteristic, and the estimation accuracy of voltage and
may be caused by the difference between static and dynamic SOC is therefore much higher. Besides, online identification
characteristics of the battery. of OCV does not require a separate HPPC test. Considering
Figures 6 and 7 show voltage and SOC estimation results the battery aging phenomenon in whole life cycle, updating
with different OCV curves. The initial value of SOC in CKF the OCV curve regularly according to the OCV data identi-
is set as 80%. In addition, the model parameters using in this fied online will be helpful for accurate SOC estimation.

13
3332 Ionics (2022) 28:3321–3339

Fig. 8  Results of voltage (a) (b) 0.1


estimation and corresponding 3.5

Voltage (V)

Error (V)
error under UDDS test: a–b
OBE@RM 0
with OBE algorithm; c–d with 3
OBE@RM 2
1

FFRLS algorithm; e–f with Reference


LKF algorithm 2.5 -0.1
0 2 4 6 8 0 2 4 6 8
Time (h) Time (h)
(c) (d) 0.1
3.5

Voltage (V)

Error (V)
FFRLS@RM 1 0
3
FFRLS@RM 2
Reference
2.5 -0.1
0 2 4 6 8 0 2 4 6 8
Time (h) Time (h)
(e) (f) 0.1
3.5
Voltage (V)

Error (V)
LKF@RM 0
3 1
LKF@RM 2
Reference
2.5 -0.1
0 2 4 6 8 0 2 4 6 8
Time (h) Time (h)

Fig. 9  Results of voltage (a) (b) 0.1


estimation and corresponding 3.5
Voltage (V)

error under NEDC test: a–b


OBE@RM 0
with OBE algorithm; c–d with 3 1
OBE@RM 2
FFRLS algorithm; e–f with Reference
LKF algorithm 2.5 -0.1
0 2 4 6 8 10 12 0 2 4 6 8 10 12
Time (h)
(c) (d) 0.1
3.5
Voltage (V)

FFRLS@RM 1 0
3
FFRLS@RM 2
Reference
2.5 -0.1
0 2 4 6 8 10 12 0 2 4 6 8 10 12
Time (h)
(e) (f) 0.1
3.5
Voltage (V)

LKF@RM 1 0
3
LKF@RM 2
Reference
2.5 -0.1
0 2 4 6 8 10 12 0 2 4 6 8 10 12
Time (h)

Online parameters identification methods errors of OBE and FFRLS differ as the regression model
changes. Besides, in the initial and final stages of UDDS
Figures 8 and 9 show the voltage estimation and correspond- test and NEDC test, the numerical divergence of parameters
ing error under UDDS test and NEDC test, respectively. The identification results occurs, which further leads to the diver-
model parameters using in this part are obtained by different gence of voltage estimation. Besides, it can be found that the
identification methods. In addition, the online OCV-curve regression model has a great impact on voltage estimation
obtained in the “OCV curve” section is used. Combined accuracy. No matter what algorithm is used for parameters
with results in Figs. 2 and 3, we can see that the fluctuation identification, when integrated with R ­ M2, the voltage esti-
of parameters identification results may not always lead to mation error is smaller than that integrated with ­RM1. It
the increase of voltage estimation error, as its mean may infers that the ­RM2 is more effective in battery modeling.
be closer to the true value. As we can see, when OBE and The SOC estimation results and corresponding errors
FFRLS integrated with ­RM1, results of R1 and τ deviate from are shown in Fig. 10. The zoom figure of SOC estimation
the reference, and the results of LKF does not. Thus the errors is shown in Fig. 11. The initial value of SOC in CKF

13
Ionics (2022) 28:3321–3339 3333

Fig. 10  Results of SOC estima- OBE@RM 1 OBE@RM 2 FFRLS@RM1 FFRLS@RM2 LKF@RM1 LKF@RM2 Reference
tion and corresponding error: (a) 100 (b) 4
a–b UDDS test; c–d NEDC test 80
2

Error (%)
SOC (%)
60 0
40 -2
-4
20
UDDS test -6 UDDS test
0
0 2 4 6 8 0 2 4 6 8
Time (h) Time (h)

(c) 100 (d)


80 5

Error (%)
SOC (%)
60
0
40

20
NEDC test -5 NEDC test
0
0 2 4 6 8 10 12 0 2 4 6 8 10 12
Time (h) Time (h)

Fig. 11  Zoom figure of SOC OBE@RM


1
OBE@RM
2
FFRLS@RM
1
FFRLS@RM
2
LKF@RM
1
LKF@RM
2
estimation error: a UDDS test; (a) 20 (b) 20
b NEDC test UDDS test NEDC test

15
15

10
Error (%)

Error (%)
10

5
0

-5 0

0 0.01 0.02 0.03 0.04 0.05 0 0.01 0.02 0.03 0.04 0.05
Time (h) Time (h)

Table 6  SOC estimation errors in detail is set as 80%. It can be seen that the parameters identifica-
Test Parameters identifica- MAE MaE RMSE tion methods mainly affect the convergence speed and con-
tion method vergence accuracy of SOC estimation. And as the same as
voltage estimation, the divergence of parameters identifica-
UDDS OBE RM1 2.79% 1.37% 1.41%
tions will also lead to big error of SOC estimation, especially
RM2 3.42% 0.62% 1.00%
at the end of estimation. To comprehensively evaluate the
FFRLS RM1 4.92% 0.28% 0.38%
impacts on SOC estimation, the maximum absolute error
RM2 6.34% 0.41% 0.61%
(MAE), mean absolute error (MaE), and root mean square
LKF RM1 0.62% 0.19% 0.41%
error (RMSE) of SOC estimation are calculated, as listed in
RM2 0.76% 0.31% 0.89%
Table 6. It should be noted that, when calculating the MAE,
Offline / 1.64% 1.27% 1.36%
the data of the first 200 s are ignored, because the error in
NEDC OBE RM1 1.32% 0.62% 0.89%
this stage is mainly caused by the initial SOC error.
RM2 8.26% 0.74% 0.94%
According to the results of MaE and RMSE, it can be
FFRLS RM1 6.20% 0.56% 0.73%
found that no matter what online parameters identification
RM2 5.73% 0.66% 0.83%
method is used, higher precision results of SOC estimation
LKF RM1 1.05% 0.52% 0.67%
can be obtained compared to that integrated with offline
RM2 1.24% 0.59% 0.94%
parameters identification method. However, big MAEs occur
Offline / 1.56% 1.06% 1.16%
in the results of OBE and FFRLS-based SOC estimation,

13
3334 Ionics (2022) 28:3321–3339

Fig. 12  Results of SOC estima- OBE@RM 1 OBE@RM 2 FFRLS@RM1 FFRLS@RM2 LKF@RM1 LKF@RM2 Reference
tion and corresponding errors: (a) 100 (b)
0
a–b UDDS test; c–d NEDC test 80

SOC (%)
SOC (%)
60 -5
40
-10
20
UDDS test UDDS test
0 -15
0 1 2 3 4 5 6 7 0 1 2 3 4 5 6 7
Time (h) Time (h)

(c) 100 (d) 0


80

SOC (%)
SOC (%)
60 -5

40
-10
20
NEDC test NEDC test
0 -15
0 2 4 6 8 10 0 2 4 6 8 10
Time (h) Time (h)

which are mainly caused by the divergence of parameters. Impact on peak power estimation
Besides, when integrated with ­RM1, the MAE, MaE, and
RMSE are decreasing compared to that integrated with ­RM2. According to the “SOP estimation” section, the batteries
It further shows that more accurate battery model can be are supposed to be operated within the voltage, current, and
established based on ­RM2. And by comparing the results SOC limits. Hence, in the case of peak power estimation,
obtained from different algorithms, it can be found that when we shall only consider the data corresponding to the SOC
LKF is selected for parameters identification, the MAE, range in 10 ~ 90%.
MaE, and RMSE are smaller than that of OBE and LKF. It In this part, the initial value of SOC in CKF is set as
means that the LKF-CKF-based SOC estimation method has 100%, and the initial value of reference SOC is 90%. The
the best performance in this comparative study. results of SOC estimation and corresponding errors are

LKF OBE FFRLS


(a) 0 (b) 500 (c) 0 (d) 500
Peak discharge current (A)

Peak discharge current (A)


Peak charge current (A)

Peak charge current (A)

400 400
-200 -200

300 300
-400 -400
200 200

-600 -600
100 100
UDDS test UDDS test NEDC test NEDC test
-800 0 -800 0
0 2 4 6 0 2 4 6 0 2 4 6 8 10 0 2 4 6 8 10
Time (h) Time (h) Time (h) Time (h)

104 104 104 104


(e) 0 (f) 15 (g) 0 (h) 15
Peak discharge current (A)

Peak discharge current (A)


Peak charge current (A)

Peak charge current (A)

-5 10 -5 10

-10 5 -10 5

-15 0 -15 0
UDDS test UDDS test NEDC test NEDC test

0 2 4 6 0 2 4 6 0 2 4 6 8 10 0 2 4 6 8 10
Time (h) Time (h) Time (h) Time (h)

Fig. 13  Results of instantaneous peak current estimation: a–d under voltage constrained; e–h under SOC constrained

13
Ionics (2022) 28:3321–3339 3335

shown in Fig. 12. It can be seen that when integrated with Based on analysis in the “SOP estimation” section, we
the ­RM1-based parameters identification method, the big have the results of instantaneous peak current and peak
initial SOC estimation error cannot be corrected, and big power estimation, as shown in Fig. 13. We can see that the
error exists in whole process of estimation. It means that estimation results are almost the same when parameters
the SOC estimation method no longer has any effect, and identification algorithm changes. Meanwhile, the estimated
other battery state estimation will be affected. In contrast, peak discharge current is larger than the estimated peak
the ­RM2-based method show high accuracy of SOC estima- charge current. It is caused by the difference between bat-
tion, and the 10% initial error can be converged quickly. The tery voltage and the upper/lower cutoff voltage. In the SOC
initial SOC error can be corrected in short time and the SOC range of 10 ~ 90%, the battery voltage is remaining around
estimation error can be limited within low range (± 2% in 3.2 V, while the upper cut-off voltage is 3.65 V and the lower
UDDS test and ± 2.5% in NEDC test). These results indicate cutoff voltage is 2.5 V. Which means that in the terms of
that the R
­ M2 has better adaptability to battery working con- voltage constraint, the discharge current that the battery can
dition and is more helpful for parameters identification and support is greater than the charging current.
SOC estimation. In this case, in the following analysis of For EVs, the actions such as brake, acceleration, and
peak power estimation, we mainly focus on the ­RM2-based uphill are not instantaneous, but need to last for a period of
method. time. Hence, the capability of continuous charge/discharge
of the battery must be taken into account. In this paper, we
calculated the continuous peak current in 30 s, and the cor-
responding continuous peak power is also calculated.
0.3
10
-3
According to Eq. (21), the voltage constrained peak cur-
0.25 6 rent estimation is mainly affected by two factors: the product
dOCV/dSOC (V)

0.2 4 of continuous duration, value of OCV-SOC derivative and


0.15 2 reciprocal of capacity, and the value of parameters. As
0.1 0
80 60 40 20
shown in Fig. 14, in the range of SOC in 10 ~ 90%, the deriv-
0.05 ative value of OCV-SOC function is very small. And when
0
100 90 80 70 60 50 40 30 20 10 0
multiplied by the reciprocal of capacity, the order of magni-
SOC (%) tude of this value will be only ­10−9, then the impact caused
by continuous duration can be ignored. According to results
Fig. 14  Derivative of OCV-SOC function in the “Parameters identification” section, the order of

LKF OBE FFRLS

(a) 0 (b) 500 (c) 0 (d) 500


Peak discharge current (A)

Peak discharge current (A)


Peak charge current (A)

Peak charge current (A)

400 400
-200 -200

300 300
-400 -400
200 200

-600 -600
100 100
UDDS test UDDS test NEDC test NEDC test
-800 0 -800 0
0 2 4 6 0 2 4 6 0 5 10 0 5 10
Time (h) Time (h) Time (h) Time (h)

(e) 0 (f) 6000 (g) 0 (h)6000


Peak discharge current (A)

Peak discharge current (A)

5000 5000
Peak charge current (A)

Peak charge current (A)

-1000 -1000
4000 4000
-2000 -2000
3000 3000
-3000 -3000
2000 2000
-4000 -4000
1000 1000
-5000 0 -5000 0
UDDS test UDDS test NEDC test NEDC test
-6000 -1000 -6000 -1000
0 2 4 6 0 2 4 6 0 5 10 0 5 10
Time (h) Time (h) Time (h) Time (h)

Fig. 15  Results of continuous peak current estimation: a–d under voltage constrained; e–h under SOC constrained

13
3336 Ionics (2022) 28:3321–3339

LKF OBE FFRLS

0 300 0 300
(a) (b) (c) (d)
Peak discharge current (A)

Peak discharge current (A)


-50 250 -50 250

Peak charge current (A)

Peak charge current (A)


-100 200 -100 200

-150 150 -150 150

-200 100 -200 100

-250 50 -250 50
UDDS test UDDS test NEDC test NEDC test
-300 0 -300 0
0 2 4 6 0 2 4 6 0 2 4 6 8 10 0 2 4 6 8 10
Time (h) Time (h) Time (h) Time (h)

(e) (f) (g) (h)

SOC SOC SOC SOC

Current Current Current Current

Voltage Voltage Voltage Voltage

UDDS test UDDS test NEDC test NEDC test

0 2 4 6 0 2 4 6 0 2 4 6 8 10 0 2 4 6 8 10
Time (h) Time (h) Time (h) Time (h)

Fig. 16  Results of continuous peak current estimation and corresponding constraint signal under multi-constraints: a–d peak power estimation
results; e–h constraint signals

LKF OBE FFRLS


(a) 0 (b)1000 (c) 0 (d)1000
Peak discharge power (w)

Peak charge power (w)

Peak charge power (w)


-200 800 -200 800
Peak discharge power (w)

-400 600 -400 600

-600 400 -600 400

-800 200 -800 200


UDDS test UDDS test NEDC test NEDC test
-1000 0 -1000 0
0 2 4 6 0 2 4 6 0 2 4 6 8 10 0 2 4 6 8 10
Time (h)
Time (h) Time (h) Time (h)

(e) (f) (g) (h)

SOC SOC SOC SOC

Current Current Current Current

Voltage Voltage Voltage Voltage


UDDS test UDDS test NEDC test NEDC test

0 2 4 6 0 2 4 6 0 2 4 6 8 10 0 2 4 6 8 10
Time (h) Time (h) Time (h) Time (h)

Fig. 17  Results of continuous peak power estimation and corresponding constraint signal under multi-constraints: a–d peak power estimation
results; e–h constraint signals

13
Ionics (2022) 28:3321–3339 3337

magnitude of R0 and R1 is ­10−3, which is much larger than (0.19% in UDDS test and 0.52% in NEDC test), and
­1 0 − 9 . � F u ��
� r t h e r m o r e�, w�h e n m u l t i p l i e d w i t h RMSE (0.41% in UDDS test and 0.67% in NEDC test).
∑L−1 L−1−i
It implies that the LKF-based parameters identification
1-exp − 𝜏 Δt
i=0
exp − Δt
 , the impact of R1 will
𝜏
method is more suitable for SOC estimation.
be amplified. Hence, the parameters identification will have
3. For co-estimation of SOC and peak power, results show
a great impact on the estimation of continuous peak
that ­RM1-based methods are ineffective. Violent fluctua-
current.
tion of peak power estimation results is obtained based
The results of continuous peak current estimation under
on LKF algorithm (528.4 ~ 1000 W in UDDS test and
voltage constraints and SOC constraints are shown in
451.4 ~ 1000 W in NEDC test). By comparison, the OBE
Fig. 15. As the analysis above, the estimated voltage con-
algorithm is more suitable for co-estimation.
strained peak current varies as the parameters identification
algorithm changes. Especially when LKF algorithm is used,
Our future work will continue on improving the perfor-
the estimated results fluctuate violently. While the results
mance of battery models and states estimation methods, and
of estimated SOC constrained peak current are almost the
the impacts on states estimation of battery aging and tem-
same.
perature will be taken into account.
The results of continuous peak current, continuous peak
power under multi-constraints, and the constraint signals
are shown in Figs. 16 and 17. For peak discharge current Funding  This work was supported by the National Key R&D Program
and peak discharge power estimation, current dominates of China (2019YFC1907801-4) and the Shanghai Rising-Star Program
the multi-constraints. And for peak charge current and peak (N0. 21QB1401400, NO. 22QA1406400).
charge power estimation, voltage dominates the multi-con-
straints, which confirms to the voltage characteristic of LFP Declarations 
battery. The frequent jumping of constraint signals occurs in
LKF-based method, which is mainly caused by the fluctua- Conflict of interest  The authors declare no competing interests.
tion of estimated results. The peak power varies from 528.4
to 1000 W in UDDS test, varies from 451.4 to 1000 W in
NEDC test, which is unreasonable. In contrast, the results References
obtained from OBE and FFRLS-based method are more sta-
1. Kaygusuz K (2012) Energy for sustainable development: a case
ble. And based on Figs. 2 and 3, the results of τ obtained of developing countries. Renewable Sustainable Energy Rev
by FFRLS algorithm also show fluctuation; thus, the OBE 16(2):1116–1126. https://​doi.​org/​10.​1016/j.​rser.​2011.​11.​013
algorithm is more suitable for peak power estimation. 2. Li L, Lin J, Wu N, Xie S, Meng C, Zheng Y, Wang X, and Zhao Y
(2020) Review and outlook on the international renewable energy
development. Energy Built Environ. Review and outlook on the
international renewable energy development
Conclusion 3. Zheng H, Song M, Shen Z (2021) The evolution of renewable
energy and its impact on carbon reduction in China. Energy
In this paper, parameters identification methods for LFP bat- 237:121639. https://​doi.​org/​10.​1016/j.​energy.​2021.​121639
4. Sattich T, Freeman D, Scholten D, Yan S (2021) Renewable
tery based on integration of regression models and three energy in EU-China relations: policy interdependence and its
well-known algorithms are discussed. Combined with the geopolitical implications. Energy Policy 156:112456. https://​doi.​
CKF algorithm, the battery SOC and peak power are further org/​10.​1016/j.​enpol.​2021.​112456
estimated based on experimental data obtained from UDDS 5. Tarascon JM, Armand M (2001) Issues and challenges facing
rechargeable lithium batteries. Nature. https://​doi.​org/​10.​1142/​
test and NEDC test. And the following conclusions can be 97898​14317​665_​0024
drawn: 6. Diouf B, Pode R (2015) Potential of lithium-ion batteries in
renewable energy. Renewable Energy 76:375–380. https://​doi.​
1. For battery parameters identification, the online identifi- org/​10.​1016/j.​renene.​2014.​11.​058
7. Lu L, Han X, Li J, Hua J, Ouyang M (2013) A review on the key
cation method is more accurate than the offline identifi- issues for lithium-ion battery management in electric vehicles. J
cation method. And for the two regression models men- Power Sources 226:272–288. https://​doi.​org/​10.​1016/j.​jpows​our.​
tioned in this paper, the R
­ M2 is more suitable parameters 2012.​10.​060
identification. 8. Dai H, Jiang B, Hu X, Lin X, Wei X, and Pecht M (2020)
Advanced battery management strategies for a sustainable energy
2. For SOC estimation, the impacts of parameters identifi- future: multilayer design concepts and research trends. Renewable
cation methods are mainly reflected in the difference of Sustainable Energy Rev :110480. https://​doi.​org/​10.​1016/j.​rser.​
convergence speed and convergence accuracy. Among 2020.​110480
these algorithms, LKF has the best performance in MAE 9. Zhang S, Guo X, Dou X, Zhang X (2020) A data-driven coulomb
counting method for state of charge calibration and estimation
(0.62% in UDDS test and 1.05% in NEDC test), MaE

13
3338 Ionics (2022) 28:3321–3339

of lithium-ion battery. Sustainable Energy Technol Assess 27. Duan W, Song C, Chen Y, Xiao F, Peng S, Shao Y, and Song S
40:100752. https://​doi.​org/​10.​1016/j.​seta.​2020.​100752 (2020) Online parameter identification and state of charge estima-
10. Zheng F, Xing Y, Jiang J, Sun B, Kim J, Pecht M (2016) Influence tion of battery based on multitimescale adaptive double Kalman
of different open circuit voltage tests on state of charge online filter algorithm. Math Probl Eng 2020 https://​doi.​org/​10.​1155/​
estimation for lithium-ion batteries. Appl Energy 183:513–525. 2020/​95026​05
https://​doi.​org/​10.​1016/j.​apene​rgy.​2016.​09.​010 28. Wang Y, Tian J, Sun Z, Wang L, Xu R, Li M, Chen Z (2020) A
11. Hannan MA, Lipu MSH, Hussain A, Saad MH, Ayob A (2018) comprehensive review of battery modeling and state estimation
Neural network approach for estimating state of charge of lithium- approaches for advanced battery management systems. Renewable
ion battery using backtracking search algorithm. IEEE Access Sustainable Energy Rev 131:110015. https://​doi.​org/​10.​1016/j.​
6:10069–10079. https://​doi.​org/​10.​1109/​ACCESS.​2018.​27979​76 rser.​2020.​110015
12. Anton JCA, Nieto PJG, Viejo CB, Vilán JAV (2013) Support vec- 29. Sassi HB, Errahimi F, Najia ESS (2020) State of charge estimation
tor machines used to estimate the battery state of charge. IEEE by multi-innovation unscented Kalman filter for vehicular applica-
Trans Power Electron 28(12):5919–5926. https://d​ oi.o​ rg/1​ 0.1​ 109/​ tions. J Energy Storage 32:101978. https://​doi.​org/​10.​1016/j.​est.​
TPEL.​2013.​22439​18 2020.​101978
13. Li Y, Chattopadhyay P, Xiong S, Ray A, Rahn CD (2016) 30. Lin C, Yu Q, Xiong R (2017) A study on the impact of open
Dynamic data-driven and model-based recursive analysis for circuit voltage tests on state of charge estimation for lithium-ion
estimation of battery state-of-charge. Appl Energy 184:266–275. batteries. Appl Energy 205:892–902. https://​doi.​org/​10.​1016/j.​
https://​doi.​org/​10.​1016/j.​apene​rgy.​2016.​10.​025 apene​rgy.​2017.​08.​124
14. Liu Q, Liu J, Le W, Guo Z, He Z (2019) Data-driven intelligent 31. Yu Q-Q, Xiong R, Wang L-Y, Lin C (2018) A comparative study
location of public charging stations for electric vehicles. J Cleaner on open circuit voltage models for lithium-ion batteries. Chin J
Prod 232:531–541. https://​doi.​org/​10.​1016/j.​jclep​ro.​2019.​05.​388 Mech Eng 31(1):1–8. https://d​ oi.o​ rg/​10.1​ 186/s​ 10033-​018-0​ 268-8
15. Shrivastava P, Soon TK, Idris MYIB, Mekhilef S (2019) Overview 32. Fortescue TR, Kershenbaum LS, Ydstie BE (1981) Implemen-
of model-based online state-of-charge estimation using Kalman tation of self-tuning regulators with variable forgetting factors.
filter family for lithium-ion batteries. Renewable Sustainable Automatica 17(6):831–835. https:// ​ d oi. ​ o rg/ ​ 1 0. ​ 1 016/ ​ 0 005-​
Energy Rev 113:109233. https://​doi.o​ rg/​10.​1016/j.​rser.​2019.​06.​ 1098(83)​90117-6
040 33. Wei Z, Meng S, Xiong B, Ji D, Tseng KJ (2016) Enhanced
16. Newman J, Tiedemann W (1975) Porous-electrode theory with online model identification and state of charge estimation for
battery applications. AIChE J 21(1):25–41. https://​doi.​org/​10.​ lithium-ion battery with a FBCRLS based observer. Appl Energy
1002/​aic.​69021​0103 181:332–341
17. Hu X, Li S, Peng H (2012) A comparative study of equivalent 34. Wei Z, Zhao J, Xiong R, Dong G, Pou J, Tseng KJ (2018) Online
circuit models for Li-ion batteries. J Power Sources 198:359–367. estimation of power capacity with noise effect attenuation for
https://​doi.​org/​10.​1016/j.​jpows​our.​2011.​10.​013 lithium-ion battery. IEEE Trans Ind Electron 66(7):5724–5735
18. Wang Y, Gao G, Li X, and Chen Z (2020) A fractional-order 35. Wei Z, Zhao D, He H, Cao W, Dong G (2020) A noise-tolerant
model-based state estimation approach for lithium-ion battery model parameterization method for lithium-ion battery manage-
and ultra-capacitor hybrid power source system considering load ment system. Appl Energy 268:114932
trajectory. J Power Sour 449:227543 36. Wei Z, Zhao J, Zou C, Lim TM, Tseng KJ (2018) Comparative
19. Nikdel M (2014) Various battery models for various simulation study of methods for integrated model identification and state
studies and applications. Renewable Sustainable Energy Rev of charge estimation of lithium-ion battery. J Power Sources
32:477–485. https://​doi.​org/​10.​1016/j.​rser.​2014.​01.​048 402:189–197. https://​doi.​org/​10.​1016/j.​jpows​our.​2018.​09.​034
20. Xu L, Wang J, Chen Q (2012) Kalman filtering state of charge 37. Plett GL (2004) High-performance battery-pack power estimation
estimation for battery management system based on a stochas- using a dynamic cell model. IEEE Trans Veh Technol 53(5):1586–
tic fuzzy neural network battery model. Energy Convers Manage 1593. https://​doi.​org/​10.​1109/​TVT.​2004.​832408
53(1):33–39. https://​doi.​org/​10.​1016/j.​encon​man.​2011.​06.​003 38. Xiong R, He H, Sun F, Liu X, Liu Z (2013) Model-based state
21. Plett GL (2004) Extended Kalman filtering for battery manage- of charge and peak power capability joint estimation of lithium-
ment systems of LiPB-based HEV battery packs: part 1. Back- ion battery in plug-in hybrid electric vehicles. J Power Sources
ground J Power Sources 134(2):252–261. https://​doi.​org/​10.​ 229:159–169. https://​doi.​org/​10.​1016/j.​jpows​our.​2012.​12.​003
1016/j.​jpows​our.​2004.​02.​031 39. Wang Y, Pan R, Liu C, Chen Z, Ling Q (2018) Power capability
22. Peng S, Chen C, Shi H, Yao Z (2017) State of charge estimation evaluation for lithium iron phosphate batteries based on multi-
of battery energy storage systems based on adaptive unscented parameter constraints estimation. J Power Sources 374:12–23
Kalman filter with a noise statistics estimator. IEEE Access 40. Wei Z, Quan Z, Wu J, Li Y, Pou J, Zhong H (2021) Deep deter-
5:13202–13212. https://​doi.​org/​10.​1109/​ACCESS.​2017.​27253​01 ministic policy gradient-drl enabled multiphysics-constrained
23. Ye M, Guo H, Cao B (2017) A model-based adaptive state of fast charging of lithium-ion battery. IEEE Trans Ind Electron
charge estimator for a lithium-ion battery using an improved adap- 69(3):2588–2598
tive particle filter. Appl Energy 190:740–748. https://​doi.​org/​10.​ 41. Zou C, Hu X, Wei Z, Wik T, Egardt B (2017) Electrochemical
1016/j.​apene​rgy.​2016.​12.​133 estimation and control for lithium-ion battery health-aware fast
24. Wang Y, Chen Z (2020) A framework for state-of-charge and charging. IEEE Trans Ind Electron 65(8):6635–6645
remaining discharge time prediction using unscented particle fil- 42. Zhang X, Zhang W, Lei G (2016) A review of li-ion battery equiv-
ter. Appl Energy 260:114324 alent circuit models. Trans Electr Electron Mater 17(6):311–316.
25. Xiong R, Yu Q, Lin C (2017) A novel method to obtain the open https://​doi.​org/​10.​4313/​TEEM.​2016.​17.6.​311
circuit voltage for the state of charge of lithium ion batteries in 43. He H, Xiong R, Guo H (2012) Online estimation of model param-
electric vehicles by using H infinity filter. Appl Energy 207:346– eters and state-of-charge of LiFePO4 batteries in electric vehicles.
353. https://​doi.​org/​10.​1016/j.​apene​rgy.​2017.​05.​136 Appl Energy 89(1):413–420. https://​doi.​org/​10.​1016/j.​apene​rgy.​
26. Ouyang Q, Ma R, Wu Z, Xu G, Wang Z (2020) Adaptive square- 2011.​08.​005
root unscented Kalman filter-based state-of-charge estimation for 44. Xiong R, Sun F, Gong X, Gao C (2014) A data-driven based adap-
lithium-ion batteries with model parameter online identification. tive state of charge estimator of lithium-ion polymer battery used
Energies 13(18):4968. https://​doi.​org/​10.​3390/​en131​84968

13
Ionics (2022) 28:3321–3339 3339

in electric vehicles. Appl Energy 113:1421–1433. https://​doi.​org/​ 48. Albertus P, Couts J, Srinivasan V, and Newman J (2008) II. A
10.​1016/j.​apene​rgy.​2013.​09.​006 combined model for determining capacity usage and battery size
45. Xiong R, Li L, Yu Q, Jin Q, Yang R (2020) A set membership for hybrid and plug-in hybrid electric vehicles. J Power Sources
theory based parameter and state of charge co-estimation method 183(2):771–782. https://​doi.​org/​10.​1016/j.​jpows​our.​2008.​05.​012
for all-climate batteries. J Cleaner Prod 249:119380. https://​doi.​ 49. Christensen J, Newman J (2006) Stress generation and fracture in
org/​10.​1016/j.​jclep​ro.​2019.​119380 lithium insertion materials. J Solid State Electrochem 10(5):293–
4 6. He H, Xiong R, Zhang X, Sun F, Fan J (2011) State-of-charge 319. https://​doi.​org/​10.​1007/​s10008-​006-​0095-1
estimation of the lithium-ion battery using an adaptive extended 50. Christensen J, Srinivasan V, Newman J (2006) Optimization of
Kalman filter based on an improved Thevenin model. IEEE Trans lithium titanate electrodes for high-power cells. J Electrochem Soc
Veh Technol 60(4):1461–1469. https://​doi.​org/​10.​1109/​T VT.​ 153(3):A560. https://​doi.​org/​10.​1149/1.​21725​35
2011.​21328​12
47. Arasaratnam I, Haykin S (2009) Cubature kalman filters. IEEE Publisher's note Springer Nature remains neutral with regard to
Trans Autom Control 54(6):1254–1269. https://​doi.​org/​10.​1109/​ jurisdictional claims in published maps and institutional affiliations.
TAC.​2009.​20198​00

13

You might also like