Automotive Innovation (2022) 5:359–375

https://doi.org/10.1007/s42154-022-00201-4

Adaptive Fitting Capacity Prediction Method for Lithium‑Ion Batteries

Xiao Chu1 · Fangyu Xue1 · Tao Liu1 · Junya Shao1 · Junfu Li1,2,3

Received: 27 February 2022 / Accepted: 19 September 2022 / Published online: 20 October 2022
© China Society of Automotive Engineers (China SAE) 2022

Abstract
Lithium-ion batteries have become the mainstream power source for electric vehicles because of their excellent performance.
However, lithium-ion batteries still experience aging and capacity attenuation during usage. It is therefore critical to accu-
rately predict battery remaining capacity for increasing battery safety and prolonging battery life. This paper first adopts
the metabolism grey algorithm and a simplified electrochemical model to predict battery capacity under different operating
conditions. To improve the prediction performance where the capacity changes nonlinearly, a decoupling analysis of battery
capacity loss is then conducted based on the simplified electrochemical model. Finally, an adaptive fitting method is devel-
oped for capacity prediction, aiming at improving the prediction accuracy at the inflection point of battery capacity diving.
The prediction results indicate that the developed adaptive fitting method can achieve high prediction accuracy under battery
capacity attenuation at different discharge stages with errors lower than 2.2%. And the battery capacity decay shows linear
variation, and the proposed method effectively forecast the inflection point of battery capacity diving.

Keywords Lithium-ion battery · Capacity prediction · Capacity diving · Adaptive fitting capacity prediction

Abbreviations 1 Introduction
BTS Battery test system
DPs Degradation patterns In recent years, the issue of air pollution caused by the burn-
EIS Electrochemical impedance spectroscopy ing of traditional energy sources has attracted widespread
GPR Gaussian process regression attention. As an alternative energy source, lithium-ion bat-
LSTM Long short-term memory teries are widely used in various industries [1, 2]. With the
MGA Metabolism grey algorithm popularity of lithium-ion batteries, problems such as the
RUL Remaining useful life spontaneous combustion of electric vehicles and the explo-
SEM Simplified electrochemical model sion of mobile phones emerged and caused massive losses to
TL Transfer learning their users. Battery capacity is a direct reflection of battery
life [3]. The premature end of battery life usually leads to
the collapse of the entire working system. To prevent such
accidents, it is necessary to predict battery capacity [4, 5].
There are many methods for capacity prediction of lith-
ium-ion batteries, and the advantages and limitations of
each method vary under different working conditions. At
present, prediction methods can be divided into two cat-
egories: data-driven prediction method and model-based
* Junfu Li prediction method [6]:
[email protected]
1
School of Automotive Engineering, Harbin Institute
(1) Data-driven prediction method.
of Technology, 2 West Wenhua Road, Huancui District,
Weihai 264209, China This method extracts the battery health features based on
2
Guangdong Guanghua Sci-Tech Co., Ltd., Shantou 515000, the collected battery external characteristics data to real-
China ize the battery capacity prediction. The data-driven method
3
School of Chemical Engineering and Chemistry, Harbin does not require the knowledge of battery mechanism, and
Institute of Technology, Harbin 150001, China

13
Vol.:(0123456789)
360 X. Chu et al.

the real-time prediction performance is better. Recently, and the error of late model forecast was reduced, realizing
some researchers used statistical mathematical methods to fast convergence [16]. Eddahech et al. used electrochemical
estimate battery SOH. Deng et al. proposed a data driven impedance spectroscopy (EIS) to analyze the aging mecha-
method based on random partial charging process and sparse nism of batteries and obtain low-frequency internal resist-
Gaussian process regression (GPR), which uses random ance. According to the relationship between internal resist-
capacity increment sequence as the battery health indicator. ance and cycle times, the remaining useful life (RUL) of the
Experiments showed that this method can achieve accurate battery can be predicted [17]. However, the EIS test has high
SOH estimation on three different types of batteries [7]. In requirements for equipment and is time-consuming, which is
addition to the battery SOH estimation method based on sta- not suitable for online implementation. Liu et al. studied the
tistical mathematics, the method based on machine learning modeling method of grey prediction algorithm (GM), used
algorithms is also the research direction of many research- computer modeling to simulate battery capacity attenuation,
ers. Deng et al. used a long short-term memory (LSTM) net- and carried out error accuracy analysis and evaluation, prov-
work to establish an SOH estimation model and compared ing that GM is only applicable to monotonically increasing
its performance with other machine learning algorithms [8]. or monotonically decreasing prediction objects [18].
It was found that the LSTM network had the best estimation To predict battery capacity using an electrochemical
accuracy and further improved the estimation accuracy by model, it is necessary to analyze the decline mechanism
degradation patterns (DPs) recognition and transfer learn- of capacity. The decline of battery capacity can be roughly
ing (TL) methods. However, data driven methods all have divided into three stages [19]: In the first stage, a large num-
some common shortcomings. The prediction result depends ber of lithium ions are consumed to generate SEI film, lead-
on the actual data obtained, and has certain uncertainty and ing to a rapid decline in capacity. In the second stage, the
randomness [9]. In the case of insufficient data, it is difficult battery capacity steadily decreases due to various side reac-
to guarantee prediction accuracy during the entire battery tions inside the battery. In the third stage, capacity declines
life cycle [10]. rapidly, probably because of lithium deposition and/or the
change of the internal crystal structure and the electrolyte
(2) Model-based prediction method. consumption at the later aging stage of lithium-ion batteries,
that is, the phenomenon of capacity diving. Capacity diving
This kind of method relies on the battery material prop- seriously affects the safe and reliable operation of the battery
erties, load conditions, and failure mechanisms to establish system. With the aging of the battery, when capacity div-
a model that reflects the battery degradation mechanism, ing occurs, the battery is unlikely to function properly. It is
thus achieving battery capacity prediction [11]. The model- particularly important to predict the occurrence of capacity
based prognostics have the advantages of high estimation diving. Rohr et al. studied the change of discharge curve to
accuracy and wide application range, but the process of reveal the decline law of battery capacity. The experiment
model building is complicated and requires huge compu- showed that if the slope inflection point of the discharge
tational resources. Model-based methods can be divided curve reached the cut-off voltage of the lower discharge
into two categories: electrochemical models [12–14] and limit, the battery capacity under the given discharge plan
external characteristic models. The external characteristic will decrease in the following aging process [20].
model can be divided into a neural network model and an The research on capacity diving is still at a preliminary
equivalent circuit model. Yang et al. used a back propagation stage and the research methods are not yet mature.
(BP) neural network and equivalent circuit model to predict This work firstly adopts the metabolism grey algorithm
the maximum available battery capacity, input equivalent (MGA) and a simplified electrochemical model (SEM) to pre-
circuit model parameters, and SOC to train BP, and output dict the battery capacity under different operating conditions
the predicted capacity [15]. The disadvantage of the equiva- with different known historical data, and the results achieved
lent circuit model is that the parameters lack physical and by the two prediction methods are compared. To solve the
chemical significance, and the prediction accuracy is low problem that the capacity prediction of the above method is
in the actual electric vehicle driving environment. On the not stable and lags in the late aging stage, an adaptive fitting
other hand, electrochemical models can accurately describe methods for capacity prediction is developed. The rest of this
the internal physicochemical reactions of cells. Based on work is organized as follows: Sect. 2 elaborates on the princi-
the single particle model of the battery, Cheng et al. pro- ples of capacity prediction based on MGA and SEM. Section 3
posed a battery electrochemical reduction model by using describes the experimental design process. Section 4 lists the
the Laplace transform, Pade approximation, and discretiza- results of capacity prediction based on MGA and SEM and
tion, which improved the calculation efficiency of the model. presents a comparative analysis. Section 5 proposes the adap-
Besides, substituting the Kalman filter into the model, the tive fitting method for capacity prediction, and analyses the
internal electrochemical status of the battery was analyzed, simulation results. Section 6 illustrates the main conclusions.

13
Adaptive Fitting Capacity Prediction Method for Lithium‑Ion Batteries 361

2 Method for Capacity Prediction â = [a, b]T = (BT B)−1 BT Y (6)

2.1 Battery Capacity Prediction Based Substituting the estimated values of the parameter â into
on Metabolism Grey Algorithm the model, the solution of the GM (1, 1) model is.
obtained as follows:
MGA does not have high requirements on the amount of data
b −at b
( )
in modeling. It can be used to effectively track the changes X̂ (1) (t) = X (0) (1) − e + (7)
a a
within a small amount of existing data. In this work, MGA
is used to capture the variation rule of model parameters of Since X (1) is the accumulated data, to restore it, the fol-
SEM and achieve an accurate prediction of battery capacity. lowing equation is obtained:
The specific establishment process of MGA (1, 1) is as fol-
b −at
( )
lows [18]: X̂ (0) (k + 1) = (1 − ea ) X (0) (1) − e , k = 1, 2, ⋯ , N
a
Grey differential equations in grey theory: (8)
dX Thus, the grey prediction sequence is described as
+ aX = b (1)
dt (0)
( (0) (0) (0)
)
X̂ = X̂ (1), X̂ (2), ⋯ , X̂ (N) (9)
where a represents the development coefficient, and b rep-
resents the grey action amount.(
During the prediction of MGA (1, 1), GM (1, 1) estab-
For the sequence X(0) = X(0) (1), X(0) (2), ⋅ ⋅ ⋅X(0) (N) ,
)
lished by the known initial data sequence predicts a value,
where, X(0) is the original data sequence and the length of X
(0) and then the predicted value is added to the known sequence.
is n. X(1) is supposed to be
[ generated by the accumulation The earliest data are removed and the equal dimension of
of X (0). And if Z(1) (k) = 12 X(0) (k) + X(1) (k − 1) , it can be
]
the sequence is maintained so that the metabolism can be
expressed as
predicted one by one and supplemented successively until
X 0 (k) + aZ (1) (k) = b (2) the prediction goal is completed.

The above is the basic form of the grey GM (1, 1) model.

The solution process is as follows. The vec- 2.2 Battery Capacity Prediction Based on Simplified
tor composed of original data is presented as Electrochemical Model
X(0) = X(0) (1), X(0) (2), ⋅ ⋅ ⋅X(0) (n) , and X (1) is the sequence
( )

generated by X (0) through one accumulation, as shown below This method substitutes the initial stoichiometric number of
negative and positive electrodes x0, y0, the maximum vari-
X(1) = X(1) (1), X(1) (2), ⋅ ⋅ ⋅X(1) (n) (3)
( )
ation range of a stoichiometric number of the negative and
positive electrodes Dx, Dy, capacities of effective active
To solve the model, the grey numbers a and b need to be
material in positive and negative electrodes Qp, Qn, total
solved first, which can be represented by â = [a, b]T.
battery capacity Q and other parameters into SEM. Select-
Assume that.
(1) (0) ing different lengths of raw data, each parameter based on
⎡ −Z (2) 1 ⎤ ⎡ X (2) ⎤
(1)
⎢ −Z (3) 1 ⎥ ⎢ X(0) (3) ⎥ the grey prediction algorithm is estimated. Then the battery
Let: B = ⎢ ⎥,Y = ⎢ ⎥. terminal voltage is simulated using the identified parameters,
⋮ ⋮⎥ ⎢ (0)⋮ ⎥
⎢ (1)
⎣ −Z (N) 1 ⎦ ⎣ X (N) ⎦ and the battery capacity is calculated by the ampere-hour
Taking the above B and Y into the basic form of GM (1, 1), integral method.
the error sequence ε can be obtained, and ε can be expressed SEM is based on the single-particle model, taking the
by B and Y as: effects of the liquid-phase diffusion process and the liquid-
phase potential distribution effect into account and replac-
𝜀 = Y − B̂a (4) ing the original partial differential equation describing the
Solving the square s of the error sequence ε, it can be solid- and liquid-phase processes with algebraic equation
obtained that approximation. The specific equations of SEM and the cor-
responding physical processes, the definitions of param-
s = 𝜀T 𝜀 = (Y − B̂a)T (Y − B̂a) (5) eters used in the model, and the definitions of variables in
the model are shown in the Table 1, Table 2 and Table 3.
According to the principle of least squares, the estimated An excitation response analysis method is used to identify
value of the parameter â can be expressed as follows: model parameters and the detailed processes can be found
in the previous work [21, 22].

13
362 X. Chu et al.

Table 1  Simplified electrochemical model

Related process description Model equations

Active ion diffusion in solid phase ysurf (t) = yavg + Δy(t)

xsurf (t) = xavg + Δx(t)
2 𝜏p
Δy(t) = ΔyI (t) + 7 Qp
∙ I(t)
2 𝜏n
Δx(t) = ΔxI (t) + ∙ I(t)
7 Qn
( ) ( ) 1 [ 12 𝜏p ( ) ( )] ( )
ΔyI tk+1 = ΔyI tk + 𝜏 ∙ 7 Q ∙ I tk − ΔyI tk ∙ tk+1 − tk
p p
[ ( )] (
𝜏n
ΔxI tk+1 = ΔxI tk + 𝜏1 ∙ 12
( ) ( ) ( ) )
7 Qn
∙ I tk − ΔxI tk ∙ tk+1 − tk
p

The difference of lithium ion concentration in the liquid phase between posi- Δc tk+1 = Δc tk + 1 ∙ Pcon ∙ I tk − Δc tk ∙ tk+1 − tk
( ) ( ) [ ( ) ( )] ( )
tive and negative fluid 𝜏e

Concentration polarization overpotential 𝜂 = 2RT 1 + t

( )( c0 +Δc0 (t) )
ln
con F + c0 −Δc(t)

Reaction polarization
[ (√ ) (√ )]
2RT
𝜂act = F
∙ ln m2n + 1 + mn + ln m2p + 1 + mp
1
mp (t) = 0.5 0.5 0.5 ∙ Pact,p ∙ I(t)
6∙Qp ∙(1−ysurf ) (ysurf ) c0
1
mn (t) = 0.5 0.5 0.5 ∙ Pact,n ∙ I(t)
6∙Qn ∙(1−xsurf ) (xsurf ) c0

Ohmic polarization 𝜂ohm = Rohm ∙ I(t)

Terminal voltage
( ) ( )
U0 (t) = U ref,p ysurf − U ref ,n xsurf − 𝜂con (t) − 𝜂act (t) − 𝜂ohm (t)

Table 2  Parameter definition in SEM The flow chart of SEM and MGA methods are shown in
Fig. 1.
Parameter (U.) Definitions

y0, x0( – ) Initial stoichiometric number of positive/

negative electrodes 3 Accelerated Aging Test of Batteries
Qp, Qn (A s) Capacities of effective active material in
positive/negative electrodes 26,650 lithium-ion batteries manufactured by Delipu Bat-
yofs ( – ) Positive and negative polarity offset tery Technology Co., Ltd. are used in the accelerated aging
τp, τn (s) Positive/negative solid phase diffusion test. The nominal capacity of the battery is 5A·h, of which
time constant
the negative electrode material is graphite and the positive
τe (s) Liquid phase diffusion time constant
electrode material is ­LiNi1/3Co1/3Mn1/3O2.
Pcon (mol ­m−3 ­A−1) Liquid phase diffusion proportional
coefficient
A battery test system (BTS) is used to conduct acceler-
Pact ­(m−1.5 ­mol0.5 s) Reaction polarization coefficient
ated aging tests on lithium-ion batteries at room temperature
Rohm (Ω) Internal resistance
and the cycle data such as voltage, current, and capacity of
Uref (V) Reference voltage
lithium-ion batteries are recorded at a sampling frequency
cmax,p,cmax,n (mol ­m−3) Maximum lithium concentration in posi-
of 1 s. The experiment is mainly divided into five parts:
tive/negative electrodes constant current 0.02 C-rate discharge, parameter identifica-
tion, EIS, reference performance test, and accelerated aging
Table 3  Variable definition in SEM

Variate (U.) Definitions

ysurf/ yavg, xsurf/ xavg (–) Surface/average lithium concentration fraction of positive/negative
electrode
Qall (A s) Discharge capacity
Dy, Dx (–) Maximum variation range of stoichiometric number of the positive/
negative electrodes
ηcon (V) Concentration polarization potential
ηact (V) Reaction polarization potential
ηohm (V) Ohmic polarization potential
U0/UOCV (V) Terminal voltage/Open circuit voltage
Uset (V) Stated discharge cut-off voltage

13
Adaptive Fitting Capacity Prediction Method for Lithium‑Ion Batteries 363

Fig. 1  The flow chart of SEM

and MGA methods

test. During the aging process of the battery, the parameters the battery is charged to the charge cut-off voltage of 4.2 V
of the battery model at different aging cycles are identified. at 1C constant current rate, and then charged at constant
Figure 2 shows the process of the battery accelerated aging voltage until the current dropped below 0.05C. After the
test. charging process, the battery is stood for 5 min, and then dis-
An accelerated aging test is conducted on twelve batter- charged to the discharge cut-off voltage of 2.5 V at constant
ies under different aging conditions. In each aging cycle, discharge currents of 1C, 2C, and 3C, respectively. After

Fig. 2  Schedule of battery accelerated aging test

13
364 X. Chu et al.

Table 4  Accelerated aging test under different aging conditions

Battery number Discharge C-rate Charge cut-off voltage (V) DOD (%) Temperature (℃)

3-1, 3-2 1C 4.2 100 25

3-3, 3-4 1C 4.2 50 25
3-5, 3-6 2C 4.2 100 25
3-7, 3-8 2C 4.2 50 25
4-5, 4-6 3C 4.2 100 25
4-7, 4-8 3C 4.22 50 25

being stood for another 5 min, the battery would enter the As for the five batteries shown in Figs. 3, 4, 5, 6, the
next cycle. The above process is repeated 20 times for each prediction accuracy of battery capacity with a stronger lin-
aging process. The specific aging conditions of each battery ear variation tendency turns out to be higher under different
in tests are shown in Table 4. aging conditions. In particular, for the capacity prediction of
battery 3–5 in Fig. 5, the error of capacity prediction based
on MGA is about 1% under different known data lengths.
4 Results of Capacity Prediction Based When the known data length was 7, the error is only 0.449%,
on MGA and SEM indicating a relatively high prediction accuracy in the whole
life cycle. The error of capacity prediction based on SEM is
Results of battery capacity prediction based on MGA and about 2%. As shown in Fig. 4(a), (d), the linear tendency of
SEM under different aging conditions are shown in Figs. 3, capacity change is obvious, and simultaneously, the errors
4, 5, 6 and average errors of two methods for capacity pre- of the two methods are relatively low, both below 3%. When
diction are compared in Table 5. The symbol n is the length the capacity changes nonlinearly, the prediction error rise
of known data. By adjusting n, the capacity prediction accu- and would become unstable in different cycles of the bat-
racy of both methods in various aging conditions and dif- tery. Besides, the gap in prediction accuracy under differ-
ferent cycles is verified. It can be seen from Table 5 that the ent lengths of known data would increase. For example, the
prediction accuracy of the two methods varies under differ- capacity prediction errors of batteries 3–7, 3–8, 4–7, and
ent aging conditions and known data lengths. 4–8 are above 3%, the maximum capacity prediction error

Fig. 3  Capacity Prediction of 3-1,3-2 (1C/100%DOD)

13
Adaptive Fitting Capacity Prediction Method for Lithium‑Ion Batteries 365

Fig. 4  Capacity Prediction of 3-3,3-4 (1C/50%DOD)

Fig. 5  Capacity Prediction of 3-5,3-6 (2C/100%DOD)

is 16.457% at the later aging stage of battery degradation, of battery 3–2 shown in Fig. 3, the early prediction error
and prediction results exhibited hysteresis. reaches over 7% due to the capacity rising trend of the first
Inevitably, there are also cases where the prediction error and second data points. The prediction errors of batteries
at the early or middle aging stage of degradation is higher 3–7 and 3–8 are also greater than that at the later aging
than that at the later aging stage influenced by outliers or stage because the data at the middle aging stage fluctuate
noise data. As capacity prediction results based on the MGA

13
366 X. Chu et al.

Fig. 6  Capacity Prediction of 3-8 (2C/50%DOD)

significantly and is located at the inflection point of capac- is 1.707% with an initial data length of 8. Compared with
ity change. similar operating conditions in this work, the average predic-
In Ref. [14], MGA is also adopted to predict battery tion errors for batteries 3–2, 3–3, and 3–4 using MGA are
capacity based on NASA's recorded data sets. The aver- 1.455%, 1.952%, and 2.038%, respectively, which indicates
age prediction error under 4A discharge current at 24℃

Table 5  Average error of the capacity prediction based on MGA/SEM

3–1 n 5 8 10
MGA avg error 4.523% 3.276% 3.531%
SEM avg error 6.938% 1.446% 2.369%
3–2 n 5 7 9
MGA avg error 7.429% 1.455% 2.193%
SEM avg error 4.550% 4.353% 6.558%
3–3 n 8 12 17
MGA avg error 1.952% 1.888% 4.490%
SEM avg error 1.278% 3.168% 7.025%
3–4 n 8 10 12
MGA avg error 2.038% 1.704% 2.462%
SEM avg error 1.927% 5.177% 6.388%
3–5 n 5 7 9
MGA avg error 1.078% 0.449% 1.028%
SEM avg error 2.279% 1.701% 1.560%
3–6 n 6 8 11
MGA avg error 2.539% 4.436% 4.552%
SEM avg error 1.947% 2.893% 3.267%
3–7 n 10 25 32
MGA avg error 4.523% 8.022% 4.000%
SEM avg error 10.045% 9.015% 3.863%
3–8 n 15 20 30
MGA avg error 8.573% 11.600% 3.375%
SEM avg error 7.237% 9.894% 7.975%
4–7 n 15 20 30
MGA avg error 5.790% 9.886% 16.457%
SEM avg error 4.610% 7.713% 13.178%
4–8 n 15 22 30
MGA avg error 3.360% 4.012% 14.493%
SEM avg error 4.674% 4.285% 6.522%

13
Adaptive Fitting Capacity Prediction Method for Lithium‑Ion Batteries 367

the effectiveness of the developed methods for capacity comparing the changes of the battery terminal voltage curve
prediction. before and after being affected by a certain process (diffu-
From the capacity prediction curve, it can be demon- sion process, polarization process) and according to the time
strated that when the battery capacity changes linearly, the variation of terminal voltage from charge cut-off voltage to
two methods could effectively track the variation tendency discharge cut-off voltage, the capacity loss is calculated by
of battery capacity and obtain high-precision results. How- the ampere-hour integral method.
ever, not all the battery capacities at different cycles show As a complex chemical system, there are also some
linear changes, especially at the later aging stage of bat- side reactions during battery use in addition to the main
tery aging shown in Figs. 5(f) and 6(c). The capacity would chemical reactions. The mechanisms leading to capacity
dive due to the coupling effect of various aging processes loss of batteries are diverse at different aging stages. The
inside the battery. Both methods fail to effectively identify aging mechanisms leading to a linear capacity change are
the inflection point of battery capacity change, resulting in SEI growth and structure transition of active materials.
hysteresis at the later aging stage, which makes both meth- The main side effects of the nonlinear capacity change
ods less reliable and applicable. are lithium plating and the shedding of active materials.
The degradation modes can be generalized as the loss of
lithium ions and active materials. In addition, the polariza-
5 Adaptive Fitting Method for Capacity tion overpotential will increase due to battery aging, which
Prediction leads to an advanced cut-off voltage during discharging
process and decreases battery capacity. Initial stoichio-
From the simulation results of capacity prediction based metric number of negative and positive electrodes (x0 and
on MGA and SEM in the previous section, it can be con- y0) indicates the amount of reactive lithium ions inside the
cluded that the two methods had drawbacks such as unsta- battery. Parameters of Qp and Qn indicate the amount of
ble capacity prediction accuracy in different cycles of the effective active materials in each electrode. The variations
battery aging, and hysteresis of capacity prediction at the of the above parameters can reflect the aging mechanisms
later aging stage. According to the analysis, the main rea- to some extent. By tracking the changes in SEM param-
son for these problems is that the capacity of the battery eters, accurate prediction of capacity change throughout
would decline rapidly under the influence of various aging the whole life cycle can be achieved.
reactions at the later aging stage, that is, the phenomenon Combined with the above analysis, the total capacity
of diving. On the one hand, the above two methods fail loss is divided into the following parts: the capacity loss Qt
to effectively identify the inflection point of the battery caused by the internal polarization reaction of the battery
capacity diving. On the other hand, the capacity fluctua- (reaction polarization, ohmic polarization, and concentration
tion at the inflection point also has a certain impact on the polarization), the capacity loss QLLI caused by the reduction
subsequent battery capacity prediction. of active lithium ions, and the capacity loss QLAM caused by
the damage of active materials.

5.1 Decoupling Analysis of Battery Capacity Loss

Qloss = Qt + QLLI + QLAM (10)
where Qloss is the total capacity loss of the battery, and its
To improve the accuracy of battery capacity prediction and value is the difference between the battery capacity QN and
address the problem of capacity prediction hysteresis after the initial battery capacity Q0 under different cycles.
capacity diving, a new capacity loss decomposition method
is proposed in this section. The total capacity loss of the Qloss = Q0 − QN (11)
battery is decoupled according to the change of battery ter-
The detailed calculation process of capacity loss of each
minal voltage under different conditions. By analyzing the
part is as follows.
proportion of capacity loss in each part to the total capacity
loss, the dominant aging mechanism in various stages of the
5.1.1 Calculation Process of Qt
battery is determined, providing a strong guarantee for the
precise evaluation of battery SOH assessment.
The following formula can be used to calculate the capacity
In this section, the battery capacity loss is decoupled
loss Qt caused by overpotential:
based on SEM. When simulating the battery terminal volt-
age, the diffusion process of active lithium ions in the bat- Qt = Δt ∗ I (12)
tery and the influence of polarization are considered in the
SEM. In the process of capacity decomposition, the voltage
response method and single variable method are adopted. By

13
368 X. Chu et al.

where ∆t is the time that battery terminal voltage takes to the fixed length of M. After a new measurement, the first
discharge to the cut-off voltage under different conditions, data of the above queue is removed, and the remaining M-1
and I is the discharging current. data are moved forward in turn, and the newly sampled data
The expression of battery terminal voltage U1N consider- are inserted as the tail of the new queue. Then arithmetic
ing ohmic polarization is: operations are performed on this queue to obtain the result of
this measurement. In signal processing, the moving average
U1N = EOCV (N) − 𝜂con (13)
N N N
− 𝜂act − 𝜂ohm filter is often used for denoising. Assuming that the input is
x and the output is L, the calculation formula of the moving
The expression of battery terminal voltage U2N without
average filter is as follows:
considering ohmic polarization is:
x(m) + x(m − 1) + ... + x(m − M + 1)
U2N = EOCV (N) (14) L(m) = (18)
M
Eocv (N) is the ideal electromotive force of the battery in The window width M of the sliding filter is set to be 5.
the nth charge–discharge cycle. 𝜂con N
, 𝜂act
N
and 𝜂ohm
N
are the The methods for capacity prediction based on MGA and
polarization overpotential of the Nth cycle. tN1 and tN2 are SEM could not effectively identify the inflection point of
the discharge time of U1N and U2N , which discharged from the battery capacity and the capacity prediction hysteresis at
charge cut-off voltage to the discharge cut-off voltage. Thus, the later aging stage of the battery. Therefore, to address this
the capacity loss is: problem, an adaptive fitting method for capacity prediction
is proposed in this section. By capturing the characteristics
Qt = ||tN1 − tN2 || ∗ I (15)
of battery capacity data changes, the functional relationship
between battery cycle times N and battery capacity is estab-
5.1.2 Calculation Process of QLLI lished to achieve an accurate prediction of battery capacity.
The difference method is an approximate numerical solu-
Under the discharge condition of 0.02C, the polarization tion of differential equations. In this section, the method
effect has a weak influence so it is not considered. x0, y0, is used to replace the differential with the finite difference
Qn and Qp are used to simulate the terminal voltage of the and the derivative with the finite difference quotient. The
battery. The terminal voltage does not consider the diffusion multi-order difference joint method is adopted to select the
process of lithium-ion in the solid phase and the influence of appropriate capacity curve fitting. The error method uses
overpotential. QLLI is calculated as: the error between the fitting curve and the actual capacity
curve as the judgment basis to set different fitting methods,
QLLI = ||tN − t0 || ∗ I (16) and selects the one with the smallest fitting error as the final
fitting result.
where tN and t0 are the discharge time at the Nth cycle and After the simulation analysis, it was found that the above
the initial cycle of the battery at the current of 0.02C. two methods have certain defects when applied alone. The
capacity prediction based on the difference method has prob-
lems of hysteresis and low prediction accuracy at the mid-
5.1.3 Calculation Process of QLAM
dle and later aging stage of the battery. When the length of
known data is selected at the early aging stage of the battery,
QLAM can be calculated by subtracting the other partial
the fitting formula obtained by the error method is often non-
capacity losses from the total capacity loss.
linear fitting, which caused the value of the capacity predic-
QLAM = Qloss − QLLI − Qt (17) tion to drop too fast at the later aging stage or even increase
in some cases. It is inconsistent with the actual situation.
Therefore, the above two methods are combined in this sec-
tion, and the adaptive fitting method for capacity prediction
5.2 Adaptive Fitting Method for Capacity Prediction
is proposed based on the difference-error fusion method, to
Based on Difference‑Error
obtain a more accurate prediction of battery lifespan. The
method takes the length of the known data as a prerequisite
For the drawbacks of the two capacity predicting methods in
for using the difference method or the error method.
Sect. 4, to reduce the impact of data fluctuations and better
The flow chart of adaptive fitting method is shown in
extract the change law of data, the moving average filtering
Fig. 7.
method is used to filter the experimental data in this section.
Moving average filtering is based on statistical laws, and
the continuous sampling data are considered as a queue with

13
Adaptive Fitting Capacity Prediction Method for Lithium‑Ion Batteries 369

shown in Fig. 8. From the simulation images, QLLI always

occupies the main part of the total battery capacity loss and
the proportion continued to increase, from the initial value
of 49.84% to 90.65%. Qt also shows an overall upward trend,
but it decreases after the 180th–200th cycle. This phenom-
enon is related to the internal factors that cause the battery
capacity to recover. The value of QLAM remains positive,
and after the 140th cycle, it gradually increases with the
increasing cycles, which is also in line with the actual char-
acteristics of battery aging.
In summary, it is feasible to decompose the total capacity
loss into capacity loss caused by polarization overpotential,
active lithium ions, and active materials.
The adaptive fitting prediction method is used to pre-
dict QLLI, because this part has a high consistency with the
change law of battery capacity loss. Qt and QLAM do not
show regularity. Thus, the adaptive fitting prediction method
could not be used for capacity prediction, and the MGA
Fig.7  Flow chart of adaptive fitting method method is used for capacity prediction instead. In the fol-
lowing simulation diagrams, Fig. 9 shows the results of QLLI
5.3 Results of Capacity Prediction of battery 3–1. In other aging conditions, only the remain-
ing capacity prediction simulation images of the battery are
The change curve of the capacity loss of each part in the shown.
whole life of the battery after the capacity decomposition is

Fig. 8  Capacity loss decomposition caused by: (a) polarization overpotential, (b) active lithium ions, and (c) active material

Fig. 9  Simulation image of capacity loss prediction due to reduction of active lithium ions

13
370 X. Chu et al.

It can be seen from Fig. 9 that QLLI shows a high pre- can effectively identify the battery diving mechanism and
diction accuracy in the whole battery life. However, Qt address the problem of capacity prediction hysteresis.
and QLAM are not regular, so the prediction accuracy is Nevertheless, for batteries 4–7 and 4–8 in Fig. 14(b) and
unsatisfactory. (e), when the length of known data is at the middle aging
Results of battery capacity prediction with different stage, the prediction error is relatively large. The prediction
known data lengths under various aging conditions are error of Fig. 14(b) reaches the maximum of more than 8%.
shown in Figs. 10, 11, 12, 13, 14. The average error of But the error of the adaptive fitting capacity method is still
capacity prediction is shown in Table 6, where n is the length slightly reduced compared with the errors of both MGA and
of known data. SEM methods. Although the prediction errors of batteries
Primarily, it can be seen from Fig. 10 that the first data 3–7 and 3–8 in Fig. 13(b) and (e) are more than 8% due to
point of the capacity of battery 3–2 has a great impact on the influence of the subsequent battery capacity increase and
the battery capacity prediction. From the 20th cycle to the decrease, the capacity predictions are in accordance with the
40th cycle, the battery capacity shows an upward trend. The battery capacity change law, and the hysteresis of capac-
analysis shows that in this cycle period, a stable SEI film ity prediction is effectively avoided, making the prediction
is gradually formed, which improves the battery capacity. accuracy higher than the capacity prediction results based
Under different aging conditions, the prediction accu- on MGA and SEM.
racy of the adaptive fitting method is also improved with The battery capacities of batteries 3–7, 3–8, 4–7, and
the increase in linear trend. For the six batteries shown in 4–8 in the figures all show non-linear changes, and the pre-
Figs. 10, 11, 12, most of the prediction errors under dif- diction accuracy of batteries 3–7 and 3–8 in the mid-aging
ferent known data lengths are within 2.2%. Compared period is higher than that of batteries 4–7 and 4–8. This
to the two methods in Sect. 4, the prediction accuracy is is because the data fluctuations of batteries 3–7 and 3–8
manifested to have been further improved. When the bat- at the inflection point of the curve are relatively slighter,
tery capacity shows non-linear changes, the battery error which makes the capacity variation trend better tracked.
is mostly controlled within 5%. Especially in the capacity On the other hand, the capacity changes of battery 4–7
prediction at the later aging stage of batteries 4–7 and 4–8, from the 460th cycle to the 540th cycle are gentle and
the errors are 0.817% and 4.900%, respectively. In contrast located at the inflection point of capacity drop. But its
to the above two methods in Sect. 4, the prediction accu- capacity fluctuated wildly before the 460th cycle, which
racy is also enhanced significantly. It means that the method interferes with subsequent capacity predictions. The
capacity of battery 4–8 fluctuates greatly after the 300th

Fig. 10  Capacity Prediction of 3-1,3-2 (1C/100%DOD)

13
Adaptive Fitting Capacity Prediction Method for Lithium‑Ion Batteries 371

Fig. 11  Capacity Prediction of 3-3,3-4 (1C/50%DOD)

Fig. 12  Capacity Prediction of 3-5,3-6 (2C/100%DOD)

13
372 X. Chu et al.

Fig. 13  Capacity Prediction of 3-7,3-8 (2C/50%DOD)

Fig. 14  Capacity Prediction of 4-7,4-8 (3C/50%DOD)

cycle, and the changing trend became vague. Besides, The quality of the capacity prediction results should be
there are distinct peaks and troughs of capacity changes judged from the two aspects of accuracy and applicability
locally, which also have a strong impact on the subsequent under different working conditions. Overall, compared with
capacity prediction accuracy. the first two methods for capacity prediction in Sect. 4, the
adaptive fitting method for capacity prediction has enhanced

13
Adaptive Fitting Capacity Prediction Method for Lithium‑Ion Batteries 373

Table 6  Average error of difference-error method

3–1 n 5 9 11
Avg error 1.348% 4.197% 2.061%
3–2 n 5 7 9
Avg error 3.932% 4.812% 7.705%
3–3 n 6 10 14
Avg error 1.598% 2.161% 2.093%
3–4 n 6 8 10
Avg error 2.181% 1.754% 1.975%
3–5 n 4 6 9
Avg error 0.450% 1.123% 0.893%
3–6 n 5 9 12
Avg error 1.948% 3.517% 1.468%
3–7 n 13 26 32
Avg error 4.099% 8.470% 6.070%
3–8 n 15 26 33
Avg error 6.396% 8.857% 7.308%
4–7 n 5 16 35
Avg error 3.432% 8.208% 0.817%
4–8 n 10 20 36
Avg error 3.586% 3.620% 4.900%

the capacity prediction accuracy at different aging stages, above two methods could achieve prediction with errors less
which indicates the reliability and applicability of the algo- than 2% where the curves of capacity change show a linear
rithm have been improved. The phenomenon that the pre- trend. However, when the battery capacity changes nonlin-
diction accuracy decreases with the increase of the known early, the prediction errors increase significantly, and the
data length is ameliorated. Although the prediction errors maximum capacity prediction error even reaches more than
of batteries 3–7 and 3–8 are not significantly reduced com- 15% at the later aging stage. The battery capacity prediction
pared to the first two methods for capacity prediction, in error is also high under the same aging condition with dif-
terms of prediction effect, the adaptive fitting method for ferent known historical data.
capacity prediction effectively identifies the diving point To solve the above problems, this paper conducts a
of battery capacity and avoids the hysteresis, solving the decoupling analysis of battery capacity loss. The dominant
hidden safety problems in the practical application of the mechanism of battery capacity loss is clarified in different
MGA and SEM methods. In addition, the gap of prediction cycles. And an adaptive fitting method of capacity predic-
errors in the same aging condition with different known data tion is developed. When the capacity changes nonlinearly,
lengths is reduced, which also indicates that the adaptive fit- the inflection point of the battery capacity diving could be
ting method for capacity prediction could effectively weaken effectively identified, and the error at the later aging stage
the influence of noise data and improve the robustness of of the battery declines to lower than 7.5%. Besides, the
capacity prediction. prediction errors of different known data lengths are also
reduced, which proves that the method effectively improves
the robustness of capacity prediction.
6 Conclusions The drastic fluctuation at the inflection point of battery
capacity will affect the subsequent prediction, and the accu-
As a chemical power source, the performance of lithium-ion racy of the developed adaptive fitting method under different
batteries inevitably degrades over time. Accurate prediction working conditions remains to be improved, which leaves
of residual battery capacity is related to the safety and ech- room for future research.
elon recycling of lithium-ion batteries. Lithium-ion battery
Acknowledgements This work was financially supported by China
capacity under different aging conditions is predicted in this Postdoctoral Science Foundation (2021M690740) and the Weihai Sci-
work. First, the metabolism grey prediction algorithm-based entific Research and Innovation Funds (2019KYCXJJYB09).
method and simplified electrochemical model-based method
are used to predict the battery capacity, respectively. The

13
374 X. Chu et al.

Declarations 17. Eddahech, A., Briat, O., Woirgard, E., Vinassa, J.M.: Remain-
ing useful life prediction of lithium-batteries in calendar ageing
for automotive applications. Microelectron. Reliab. 52(9–10),
Conflict of interest On behalf of all the authors, the corresponding au-
2438–2442 (2012)
thor states that there is no conflict of interest.
18. Chen, L., Lin, W., Li, J., Tian, B., Pan, H.: Prediction of lithium-
ion battery capacity with metabolic grey model. Energy 106,
662–672 (2016)
19. Han, X., Languang, L., Zheng, Y., Feng, X., Li, Z., Li, J., Ouyang,
M.: A review on the key issues of the lithium ion battery degrada-
References tion among the whole life cycle. eTransportation (2019). https://​
doi.​org/​10.​1016/j.​etran.​2019.​100005
1. Xiong, R., Chen, H., Wang, C., et al.: Towards a smarter hybrid 20. Rohr, S., Müller, S., Baumann, M., et al.: Quantifying uncertain-
energy storage system based on battery and ultra capacitor: a criti- ties in reusing lithium-ion batteries from electric vehicles. Proced.
cal review on topology and energy management. J. Clean. Prod. Manuf. 8, 603–610 (2017)
202, 1228–1240 (2018) 21. Li, J., Wang, D., Deng, L., Cui, Z., Lyu, C., Wang, L., Pecht, M.:
2. Xiong, R., Yu, Q., Shen, W., et al.: A sensor fault diagnosis Aging modes analysis and physical parameter identification based
method for a lithium-ion battery pack in electric vehicles. IEEE on a simplified electrochemical model for lithium-ion batteries. J.
Trans. Power Electron. 34(10), 9709–9718 (2019) Energy Storage 31, 101538 (2020). https://​doi.​org/​10.​1016/j.​est.​
3. Li, X., Hu, L.H., Lv, C.X., Hu, J.: Prediction modeling and simu- 2020.​101538
lation of lithium-ion battery capacity. J. Syst. Simul. 26(08), 22. Li, J.F., Wang, L.X., Lyu, C., Liu, E.H., Xing, Y.J., Pecht, M.:
1733–1740 (2014) A parameter estimation method for a simplified electrochemical
4. Bubna, P., Brunner, D., Advani, S.G., et al.: Prediction-based model for Li-ion batteries. Electrochim. Acta 275, 50–58 (2018)
optimal power management in a fuel cell/battery plug-in hybrid
vehicle. J. Power Sour. 195(19), 6699–6708 (2010) Springer Nature or its licensor holds exclusive rights to this article under
5. Wang, S.L., Jin, S.Y., Bai, D.K., et al.: A critical review of a publishing agreement with the author(s) or other rightsholder(s);
improved deep learning methods for the remaining useful life author self-archiving of the accepted manuscript version of this article
prediction of lithium-ion batteries. Energy Rep. 7, 5562–5574 is solely governed by the terms of such publishing agreement and
(2021) applicable law.
6. Xiong, R., Gong, X.Z., Mi, C., et al.: A robust state-of-charge
estimator for multiple types of lithium-ion batteries using adaptive
extended Kalman filter. J. Power Sour. 243(6), 805–816 (2013) Xiao Chu received his M.E. degree in Vehicle
7. Deng, Z.W., Lin, X.K., Cai, J.W., Hu, X.S.: Battery health estima- Engineering from Harbin Institute of Tech-
tion with degradation pattern recognition and transfer learning. nology Weihai, China, in 2022. He currently
J. Power Sour. (2022). https://​doi.​org/​10.​1016/j.​jpows​our.​2022.​ works for BYD Ferdi Battery Company. His
231027 research interests include lithium-ion battery
8. Deng, Z.W., Hu, X.S., Li, P., Lin, X.K.: Data-driven battery state aging modeling, capacity attenuation estima-
of health estimation based on random partial charging data. IEEE tion, and life prediction.
Trans. Power Electron. 37(5), 5021–5031 (2022)
9. Shih, C., Hsu, Y., Yeh, J., Lee, P.: Grey number prediction using
the grey modification model with progression technique. Appl.
Math. Model. 35(3), 1314–1321 (2011)
10. Lin, C., An, J.J., Wang, H.M., Zhang, M.: Remaining useful life
prediction for lithium-ion battery by combining an improved par- Fangyu Xue is pursuing his B.E. degree in
ticle filter with sliding-window grey model. J. Electrotech. 35(6), the School of Automotive Engineering, Har-
1181–1188 (2020) bin Institute of Technology, Weihai, China,
11. Chen, Z., Shi, N., Yufan Ji, M., Niu, Y.W.: Lithium-ion batteries majoring in Vehicle Engineering. His
remaining useful life prediction based on BLS-RVM. Energy 234, research interests include the technology of
121269 (2021). https://​doi.​org/​10.​1016/j.​energy.​2021.​121269 health feature extraction and life prediction
12. Rahman, M.A., Anwar, S., Izadian, A.: Electrochemical model for lithium-ion batteries.
parameter identification of a lithium-Ion battery using particle
swarm optimization method. J. Power Sources 307, 86–97 (2016)
13. Tian, J.P., Xiong, R., Shen, W.X., Sun, F.C.: Fractional order bat-
tery modelling methodologies for electric vehicle applications:
recent advances and perspectives. Sci. China-Technol. Sci. 63(11),
2211–2230 (2020) Tao Liu received his Ph.D. degree in Vehicle
14. Liu, L., Zhu, J., Zheng, L.: An effective method for estimating Engineeringn from Harbin Institute of Tech-
sstate of charge of lithium-ion batteries based on an electrochemi- nology. He is currently a professor in the
cal model and Nernst equation. IEEE Access 8, 211738–211749 School of Automotive Engineering at Harbin
(2020) Institute of Technology Weihai. His main
15. Yang, D., Wang, Y.J., Pan, R., et al.: Comparison of prognostic research fields are automotive lightweight,
algorithms for estimating remaining useful life of batteries[C], 8th special vehicles, and automotive passive
International Conference on Applied Energy (ICAE) 183, 18–30 safety. He serves as a member of the China
(2015) Society of Automotive Engineers.
16. Cheng, Q.Y., Zhang, X., Gao, Y.H., Guo, B.J.: Estimation of
lithium-ion battery state based on reduced order electrochemical
model. Battery 51(02), 110–113 (2021)

13
Adaptive Fitting Capacity Prediction Method for Lithium‑Ion Batteries 375

Junya Shao is currently pursuing her B.E. Junfu Li received his Ph.D. degree in Electri-
degree in Vehicle Engineering in the School cal Engineering from Harbin Institute of
of Automotive Engineering at Harbin Insti- Technology. He is currently working as a
tute of Technology Weihai. Her research lecturer in the School of Automotive Engi-
interests include aging mechanism analysis, neering at Harbin Institute of Technology
remaining useful life prediction, and cascade Weihai. He mainly engages in research on
utilization of lithium-ion batteries. lithium-ion battery electrochemical mode-
ling, state estimation, and cascade
utilization.

13