Energy 293 (2024) 130555

Contents lists available at ScienceDirect

Energy
journal homepage: www.elsevier.com/locate/energy

Accurate capacity and remaining useful life prediction of lithium-ion

batteries based on improved particle swarm optimization and particle filter
Hui Pang a, *, Kaiqiang Chen a, Yuanfei Geng a, Longxing Wu b, Fengbin Wang a, Jiahao Liu a
a
School of Mechanical and Precision Instrument Engineering, Xi’an University of Technology, Xi’an, 710048, China
b
College of Mechanical Engineering, Anhui Science and Technology University, Chuzhou, 233100, China

A R T I C L E I N F O A B S T R A C T

Handling editor: A. Olabi Accurate prediction of capacity and remaining useful life (RUL) for lithium-ion batteries (LIBs) is crucial for
ensuring safe and reliable operation of electric vehicles. However, the battery capacity degradation and external
Keywords: environmental disturbances make it still challenging to achieve this goal. In this article, an accurate capacity and
Lithium-ion batteries RUL prediction method is proposed by combining improved particle swarm optimization (IPSO) with particle
Capacity prediction
filter (PF) algorithms. First, the parameters of particle swarm optimization (PSO) algorithm are adjusted by
Remaining useful life
adaptive weights to avoid the problem of local optimal solution. Subsequently, the optimal particle searched by
Particle filter
Improved particle swarm optimization IPSO is updated continuously by the PF algorithm to achieve a more accurate posterior estimation. Finally, the
proposed IPSO-PF method is verified by two independent and public datasets of NASA and CALCE batteries. The
results validate that the proposed method has high precision and generalizability in predicting the capacity and
RUL of LIBs even at various charging rates and battery types.

1. Introduction on the capacity and RUL prediction of batteries have received extensive
attention from researchers. Generally, these methods can be classified
Recently, lithium-ion batteries (LIBs) have become the dominant into two categories: data-driven methods and model-based methods
energy source for grid energy storage systems and electric vehicles due [8–10].
to their high energy density, high power density, cleanliness, and reli­ Data-driven methods do not consider the complex electrochemical
ability [1,2]. However, the battery performance inherently suffers from reactions inside LIBs but rely on learning models from data sources to
decrease over time due to occurrence of aging mechanisms such as forecast future changes in battery RUL. Conventional data-driven
active material loss and lithium inventory loss [3–5]. Moreover, inac­ methods mainly include support vector machine (SVM) [11,12], rele­
curate prediction of the current battery aging state can prevent the vance vector machine (RVM) [13], machine learning [14,15], neural
battery health management system from issuing risk warnings, and networks [16–18], and others. These methods can provide acceptable
persistent prediction errors may lead to premature battery failure, results of RUL estimation at the early stage of developing artificial in­
threatening the safety of the battery management system and potentially telligence technology. Subsequently, long-short-term memory (LSTM)
causing catastrophic accidents. models, which are deep learning models with powerful data processing
Remaining useful life (RUL) is one of crucial aging indicators for capabilities, have also been employed to predict battery life. For
various LIBs, which represents the number of cycles remaining from the example, Zhang et al. [19] proposed an LSTM recursive neural
current battery aging state to the failure threshold (capacity lower than network-based method for predicting the RUL of LIBs and verified its
70% of the rated capacity). The battery RUL may decrease mono­ effectiveness and superiority over the SVM method on different battery
tonically over time when the maximum number of cycles is known, datasets. Furthermore, Zhao et al. [20] developed a fused neural
primarily reflecting the degradation trend of the initial capacity of LIBs network model that utilized the broad learning system algorithm and
[6]. However, the accurate prediction of RUL is still challenging due to LSTM neural networks to predict the capacity and RUL of LIBs.
the complexity of the aging mechanism and the highly nonlinear rela­ Regrettably, data-driven methods cannot explain the battery aging
tionship of battery chemical behaviors [7]. Hence, a number of methods mechanism and may encounter prediction uncertainty issues. Moreover,

* Corresponding author.
E-mail address: huipang@163.com (H. Pang).

https://doi.org/10.1016/j.energy.2024.130555
Received 31 August 2023; Received in revised form 18 December 2023; Accepted 31 January 2024
Available online 9 February 2024
0360-5442/© 2024 Elsevier Ltd. All rights reserved.
H. Pang et al. Energy 293 (2024) 130555

the experimental data for battery aging is limited and the accuracy of 4. The effectiveness and generalizability of the proposed IPSO-PF
prediction results obtained through data-driven methods is heavily method are verified using two distinct battery datasets.
dependent on the training data, which may result in significant predic­
tion errors when predicting RUL for untested batteries under harsh The structure of this paper is outlined as follows. Section 2 describes
operating conditions. the theoretical foundation of relevant algorithms. The capacity and RUL
In contrast, model-based methods also referred to as filtering prediction framework based on the IPSO-PF method is described in
methods, are most straightforward methods to realize the estimation of Section 3. Section 4 introduces the two different public experimental
capacity and RUL by updating the relevant parameters. Specifically, this datasets and verifies the performance of the proposed IPSO-PF method
approach can be classified into Gaussian process-based filtering methods through a large number of capacity degradation and RUL prediction
and probability-based filtering methods. First, Gaussian process-based results. Section 5 presents the conclusions.
methods always exploited classical Kalman Filter (KF) models, such as
Extended Kalman Filter (EKF) [21–24] and Unscented Kalman Filter 2. Preliminaries
(UKF) [25–29], to investigate the battery state prediction and RUL
estimation. For instance, Yan et al. [23] proposed a 2.1. Particle filter
Lebesgue-sampling-based EKF approach for estimating the state of
health (SOH) and predicting battery RUL. To further improve the ac­ Particle filter (PF) is a widely utilized Bayesian probabilistic filtering
curacies of RUL prediction for LIBs, Xue et al. [27] put forward an in­ approach that is usually employed for estimating unknown parameters
tegrated algorithm that combined Adaptive UKF with Genetic in state space models characterized by non-linearity and non-Gaussian
Algorithm-optimized Support Vector Regression (GA-SVR) and verified characteristics. Instead of relying on assumptions about the distribu­
its efficacy on the NASA battery dataset. Moreover, Zheng et al. [28] tion type, the PF commonly represents the state space using a collection
suggested a new method integrating UKF and Relevance Vector of random samples called particles. During the updating process,
Regression to complete the prediction of battery RUL and short-term importance sampling and resampling techniques are applied to adjust
capacity. Unfortunately, these Gaussian process-based KF approaches the weights of the particles. This flexibility allows the PF to effectively
may fail to yield satisfactory outcomes when handling non-Gaussian handle nonlinearity and non-Gaussian characteristics, and enable ac­
distributed data due to the restrictions of Gaussian distributions [29]. curate state estimation without prior knowledge of the system model.
To address the challenges presented by Gaussian distributions, the Moreover, PF is adaptable to various model types and shows robustness
particle filter (PF) has emerged as a versatile probability-based filtering in the presence of uncertainty [33].
method in the academic community, especially in the field of battery The basic steps of the PF algorithm are described as follows:
state estimation [30–33]. A notable advantage of the PF method lies in
its ability to handle nonlinear systems characterized by non-Gaussian a Initialization
noise. For example, Ahwiadi et al. [34] carried out an enhanced parti­
cle method to reduce the impact of sample impoverishment in state For i = 1, …, N, sample x0 ∼ p(x0 ), and set k = 1.
(i)

estimation. Sun et al. [35] proposed a hybrid method based on Un­

scented Particle Filter (UPF) and optimized multi-kernel Relevance b Importance sampling
Vector Machine (RVM). This method could employ RVM to provide
multiple prediction abilities and obtain the kernel parameters and ⃒
⃒ (i)
For i = 1, …, N, sample ̂
x k ∼ q(xk ⃒x0:k-1 ,y1:k ), and set ̂
(i) (i) (i) (i)
x 0:k ≜(̂
x 0:k ,̂
x k ).
weights of RVM through grid searching, then correct the initial esti­
mation of UPF by residual prediction to enhance the prediction perfor­ For i = 1, …, N, evaluate the importance weights as:
mance. Additionally, Yang et al. [36] adopted a comprehensive method ( ⃒ ) ( ⃒
⃒ (i) (i) ⃒ (i)
)
for RUL prediction of LIBs based on the optimal combination strategy p yk ⃒xk p xk ⃒xk− 1
(1)
(i) (i)
ω
̃ k = ωk− 1 ( ⃒ )
(OCS) of particle resampling and UPF and then validated the superiority (i) ⃒ (i)
p xk ⃒xk− 1 , y1:k
and robustness of the proposed method through comparison with
mainstream methods. To suppress the measurement noise of current and
voltage, Wei et al. [37] investigated a battery RUL prediction and SOH c For i = 1, …, N, normalize the importance weight as:
estimation method with constant current and voltage as input. The PF /
( ) ( ) ∑ ( )
was subsequently employed to estimate the impedance degradation
N
ωk x(i) (2)
(i)
= ω k x ωk x(i)
parameter and the effectiveness of the proposed method was verified
0:k 0:k 0:k
i=1
through experimental analysis. Nonetheless, the degradation of particles
in the PF algorithm may lead to increased evaluation errors due to the
d Resampling
tendency of most weights to approach zero, with only a small number of
significant weights remaining [38].
Resample particles based on the importance weights, where particles
Based on the above analysis, this paper proposes an IPSO-PF method
with higher weights are more likely to be selected. Set a valid sample
for capacity and RUL prediction of LIBs. Different from the PF method
number Nt as the threshold, the formula for estimating particle degra­
that uses prior distribution as the importance distribution, the proposed
dation can be defined as:
IPSO-PF method improves the particle degradation by using the calcu­
/
lation results of IPSO as the importance distribution and the estimation ∑N
( i )2
accuracy of battery capacity and RUL is further improved. The main Neff = 1 ωk (3)
contributions of this work can be summarized as: i=1

If Neff < Nt, then regenerate new particles with weights of 1/N.
1. A double exponential empirical model is proposed to describe the
degradation trend of the capacity of LIBs. e Estimation
2. An improved particle swarm optimization with an adaptive weight is
presented to solve the problem of local optimal solution. Obtain the estimation results as:
3. An IPSO-PF fusion method is proposed to accurately predict the ca­
pacity and RUL of LIBs.

2
H. Pang et al. Energy 293 (2024) 130555

∑
N the algorithm. Improving algorithm performance through parameter
xk =
̂ ωik ⋅xik (4) adjustment and enhancing population diversity are two main directions
for its improvement. For the PSO, the most significant adjustable
i=1

parameter is the inertial weight. In this paper, an improved particle

f Repeat swarm optimization (IPSO) with an adaptive weight is proposed. The
adaptive weight refers to the use of larger ω in the early stages of the
Repeat step b to step e for each time step until the end of the time algorithm to improve global search capability and optimization speed,
horizon. and the use of smaller ω in the later stages to improve local search
The flowchart of the PF algorithm is summarized in Fig. 1(a). capability. The improved formula is expressed by:

ω = ωmin + (ωmax − ωmin ) ∗ e− t/tmax

+ σ ∗ betarnd(p, q) (6)
2.2. Improved particle swarm optimization
where ωmax represents the initial inertial weight; ωmin represents the
Most real-world optimization problems involve various complexities, inertial weight of the maximum iteration times; t is the current iteration
including nonlinearity, discontinuity, and mixed variable types. There­ number; tmax is the maximum iteration number; betarnd is a random
fore, the metaheuristic optimization technique inspired by natural pro­ number generator produced by MATLAB software, which can enhance
cesses has been developed to solve complex and large-scale optimization global search capability and accuracy; σ is the inertia adjustment factor.
problems. The usual metaheuristic optimization algorithms in current Herein, σ = 0.1, ωmax = 0.9, ωmin = 0.4, p = 1 and q = 3.
research are particle swarm optimization algorithm, genetic algorithm, To achieve particle diversity in the initial search phase and converge
ant colony algorithm and simulated annealing algorithm. Among these, to the global optimum quickly in the later stage, a dynamic adjustment
the Particle Swarm Optimization (PSO) algorithm has been extensively of the learning factors, c1 and c2 using the tangent function to balance
employed due to its robust global search capability and extensive global and local search more effectively is proposed. The tangent func­
applicability [39]. Drawing inspiration from the social behavior of bird tion is expressed as follows:
flocking, PSO replicates the natural interactions amongst individuals in ⎧ ( t )0.6 ))
⎪ c1 = (c1 start − c1 end ) × tan (0.875 × (1 −
a swarm to explore optimal solutions. Through iterative updates, the ⎪
⎨ T
+ c1 end
algorithm adjusts the positions and velocities of particles to obtain the ( )0.4 )) (7)
⎪
optimal solution within the entire swarm and the individual optimal ⎩ c2 = (c2 start − c2 end ) × arctan (2.8 × (1 − t
⎪
+ c2 end
T
solution for each particle. Furthermore, randomization is incorporated
into PSO to enhance its exploration capability, while an inertia weight is
where t is the current iteration number; T is the maximum iteration
employed to strike a balance between local and global searches. As
number; c1 start and c2 start are initial acceleration constant; c1 end and
depicted in Fig. 1(b), the diagram outlines the general procedure for
c2 end are final acceleration constants. In this paper, c1 start = 2.5,
searching for the optimal solution. The updated formula for the particles
c2 start = 1, c1 end = 0.5 and c2 end = 3.
during each iteration can be described as:
The fitness function curves for the two PSO algorithms as a function
{ ( ) ( )
vik+1 = ω⋅vik + c1 ⋅r1 ⋅ Pbest − xki + c2 ⋅r2 ⋅ Gbest − xik of iteration number are shown in Fig. 1(c).
i i i
(5)
xk+1 = xk + vk+1
3. The proposed capacity and RUL prediction method
where ω represents the inertial weight parameter; k is the current iter­
3.1. Battery degradation model
ation number; i represents the i-th particle; νik and xik represent the ve­
locity and position, respectively; c1 and c2 represent the learning factors
Repeated use of the battery can result in battery capacity degrada­
that are non-negative constants; r1 and r2 are random numbers of [0,1];
tion and power reduction, which changes the internal parameters of
Pbest and Gbest represent the individual best and global best, respectively.
battery. Therefore, it is necessary to establish an appropriate battery
However, the PSO has some limitations, such as premature conver­
capacity degradation model to analyze the performance decay law of the
gence and the tendency to fall into a local optimum in the later stages of
battery. Among various potential models, the dual-exponential empir­
ical degradation model for battery capacity has emerged as a front­
runner, demonstrating superior performance in global regression [40].
Therefore, the dual-exponential model is selected as the battery capacity
degradation model for LIBs in this paper, which is expressed by:
yk = ak ⋅ exp(bk ⋅ k) + ck ⋅exp(dk ⋅ k) (8)

where k is the cycle number; yk denotes the battery capacity at cycle k; bk

and dk represent the decay factor of fast degradation process and slow
degradation process, respectively; ak and ck represent the correlation
factor of fast degradation process and slow degradation process,
respectively; The decay factor and correlation factor indicate the rate
and influence degree of degradation process respectively.
It should be clarified that this study assumes the model parameters of
ak, bk, ck and dk are time-varying considering the complexity and time­
liness of battery degradation. Herein, xk represents the model param­
eter, which is defined as:

xk = [ak bk ck dk ]T (9)
Considering the impact of external noise and uncertainty, a state-
Fig. 1. (a) the general schematic of the PF; (b) the schematic of IPSO. The
speed direction of finial speed is composed of various weights; (c) the fitness
space model is constructed as follows:
function curves for two algorithms.

3
H. Pang et al. Energy 293 (2024) 130555

xk+1 = xk + ωk (10) where ̂

y is the experimental value.

yk = ak ⋅ exp(bk ⋅ k) + ck ⋅ exp(dk ⋅ k) + vk (11) Step 4. IPSO optimization

The IPSO algorithm updates the velocity and position of each particle
where ωk and vk represent process noise and observation noise respec­ to approach the true state, the formula is as follows:
tively and they are assumed to be uncorrelated zero-mean white noise. { ( ) ( )
vik+1 = ω⋅vik + c1 ⋅r1 ⋅ Pbest − xik + c2 ⋅r2 ⋅ Gbest − xki
i
(15)
xk+1 = xik + vik+1
3.2. Proposed IPSO-PF method

The challenge of particle degeneracy is commonly found in the

standard PF algorithm, where the variance of particle weights progres­ Step 5. Weight normalization
sively increases during iterative time steps. A promising avenue for /
∑
N
enhancing the effectiveness of the PF method is offered by integrating i i
ωk = ωk ωjk (16)
the improved particle swarm optimization (IPSO) and the particle filter j=1
(PF). First, the IPSO iteratively adjusts the velocity and position of
particles to locate the optimal value within the search space, and the PF
approximates the accurate posterior probability distribution of the sys­ Step 6. Resample
tem by updating positions and weights of the particles. Next, the particle /
with the highest fitness value represents the optimal value within the ∑
N
( i )2
search space of the IPSO algorithm and the one with the maximum Neff = 1 ωk (17)
i
weight denotes the most probable state of the system. Furthermore, the
IPSO and PF algorithms employ distinct motion mechanisms. The IPSO
updates positions and velocities by pursuing individual and global op­
tima. In contrast, in the PF, particles first update their positions using a Step 7. State estimation to achieve capacity prediction:
motion model and then refine their weights using a measurement model. ∑
N
Thus, leveraging these shared characteristics, the IPSO algorithm holds xk =
̂ ωik xik (18)
the potential to enhance the performance of the standard PF algorithm. i=1

The IPSO-PF method is utilized to predict the capacity and RUL of

batteries by establishing a discrete state space model represented in Eq.
(12). In this model, xk = [ak bk ck dk ]T serves as the state variable, rep­ Step 8. RUL prediction
resenting the internal battery characteristics, while the battery capacity
yk acts as the observation variable, reflecting the measurable attributes. 1) The capacity degradation model for k + j cycles is defined as:
{
xk+1 = xk + ωk yik+j = ak ⋅ exp[bk ⋅ (k + j)] + ck ⋅exp[dk ⋅ (k + j)] (19)
(12)
yk = ak ⋅exp(bk ⋅k) + ck ⋅exp(dk ⋅k) + vk

The detailed steps for battery RUL prediction using the IPSO-PF
method are outlined below. 2) The obtained posterior probability distribution function is estimated
Step 1. Obtain the observed values and define the fitness function. as:
Conventional particle filters use suboptimal importance functions, ( )
( ⃒ ) ∑ N
leading to suboptimal particle sampling. By incorporating the latest p yk+j ⃒y1:k ≈ ωik δ yk+j − yik+j (20)
measurement value into the sampling process, the particle filter’s sam­ i=1

pling capability can be enhanced. Thus, the fitness function is defined as

follows: where δ( ⋅) is the Dirac function.
[ ]
1 3) Battery capacity for k + j cycles
fitness = exp − (yNew − yPred )2 (13)
2Rk
∑
N
y k+j =
̂ ωik yik+j (21)
where Rk is the measurement noise variance, yNew is the latest observed i=1
value, and yPred is the predicted value.
Step 2. Initialization 4) The RUL of LIBs is determined by the number of cycles required for
The particle count, process noise variance, measurement noise the current available capacity to decline to 70% of the rated capacity.
variance, failure threshold, prediction start point, and initial values of The value of j, obtained from Eq. (16), represents the RUL.
i
model parameters a0, b0, c0 and d0 are configured. The initial particle ̂
x0,
yik+j = ak ⋅ exp[bk ⋅ (k + j)] + ck ⋅ exp[dk ⋅ (k + j)] = 0.7 (22)
i = 1, 2, …N is generated based on the prior distribution p(x0), and the
particle weights are assigned as 1/N. The failure threshold is set at 70% Incorporating the proposed dual-exponential empirical model, Fig. 2
of the battery’s rated capacity. However, in the case of the B7 battery illustrates the specific procedure of the IPSO-PF method for predicting
from the NASA dataset [41], which has not degraded to this threshold, it the capacity and RUL of LIBs.
is set to 80% of the rated capacity.
4. Experiment procedure and results analysis
Step 3. Update weight
The weight of the current particle is updated according to the most
4.1. Experimental dataset description
recent measurement as follows:
{ }
1 1 In this section, two separate battery datasets are utilized to verify the
ωik = √̅̅̅̅̅̅̅̅ exp − (̂y k − yk )2 (14)
2πR 2R effectiveness and versatility of the proposed IPSO-PF method in

4
H. Pang et al. Energy 293 (2024) 130555

Fig. 2. The research flowchart of the proposed capacity and RUL prediction method.

predicting the capacity and RUL of LIBs. In detail, dataset I is collected near the electrode during the battery’s internal reactions, which
from the NASA Prognostics Center of Excellence (PCoE) containing weakens its overall response [42]. To better evaluate the effect of bat­
experimental data for four 18650 cylindrical batteries: B5, B6, B7, and tery RUL estimation, the end-of-life (EOL) of the battery is herein
B18 at a temperature of 24 ◦ C [41]. These 18650 batteries have a defined as the threshold when its capacity deteriorates to 70% of the
cathode material of LiMnCo and an anode material of graphite with a rated capacity, precisely 1.4 Ah. However, for the B6 battery, the
rated capacity of 2Ah. The current and voltage distributions during the degradation only reached 80% of its initial capacity, exactly 1.51 Ah,
single charge-discharge process are illustrated in Fig. 3(a). Besides, without getting the defined threshold.
Fig. 3(b) depicts the capacity degradation curves, showcasing the bat­ Dataset II is obtained from the Center for Advanced Life Cycle En­
tery capacity changes with the number of charge-discharge cycles. It gineering (CALCE) [43], which utilizes a CS2 prismatic battery as the
needs to be emphasized that each charge-discharge process consists of test sample. The CS2 battery has a cathode material of LiCoO2 and an
three operational phases: charging, resting, and discharging. A summary anode material of graphite. To discuss the effectiveness of the proposed
of the experimental configuration is provided in Table 1. Further details IPSO-PF method, three types of batteries CS2-34, CS2-36, and CS2-37
regarding the testing procedures are provided as follows: with a rated capacity of 1.1 Ah are selected. These batteries under­
went charging and discharging cycles at 24 ◦ C. The experimental setup
(1) Charging process: The LIB undergoes a constant current charging conditions can be found in Table 2. The specific experimental procedure
(CCC) mode with a current of 1.5 An until reaching the upper cut- is as follows:
off voltage of 4.2 V. Subsequently, the battery is transferred to The current and voltage distributions and capacity degradation
constant voltage charging (CVC) mode until the charging current curves for each battery are depicted in Fig. 4(a) and (b). C-rate repre­
is reduced to 0.02 A. After the charging process, the battery is left sents the current rate in battery test, which is the measurement of the
to rest for 20 min. charge and discharge current with respect to its nominal capacity.
(2) Discharging process: The lithium-ion battery carries out constant Initially, a constant current of 0.5C (0.55A) is applied to the LIBs until
current discharge (CCD) at a current of 2A until the voltage the voltage reaches the upper voltage limit of 4.2 V. Subsequently, the
reaches a predetermined lower cut-off voltage (see Fig. 3(a)). charging process changes to constant voltage mode until the charging
current is reduced to 0.05 A, indicating the completion of charging.
The capacity demonstrated in Fig. 3(b) does not continuously Following this, the CS2-34 battery performs a constant current discharge
decrease but exhibits a self-recovery phenomenon. This intriguing at 0.5C (0.55A) until the voltage is dropped to 2.7 V, while the CS2-36
occurrence occurs during brief idle periods following battery charging or and CS2-37 batteries are discharged at a constant current of 1 C (1.1A)
discharging, leading to a temporary increase in capacity. The funda­ until the voltage reaches 2.7 V.
mental reason behind this phenomenon is the accumulation of reactants In Fig. 4(b)–a similar phenomenon can be observed, where the

Fig. 3. NASA batteries [41] (a) Current and voltage conditions of single charge-discharge; (b) The battery capacity degradation curve.

5
H. Pang et al. Energy 293 (2024) 130555

Table 1
NASA battery data [41].
Battery number Charging current Charging cut-off voltage Discharging current End-off voltage Rated capacity Temperature

B5 1.5 A 4.2 V 2A 2.7 V 2 Ah 24 ◦

C
B6 1.5 A 4.2 V 2A 2.5 V 2 Ah 24 ◦
C
B7 1.5 A 4.2 V 2A 2.2 V 2 Ah 24 ◦
C
B18 1.5 A 4.2 V 2A 2.5 V 2 Ah 24 ◦
C

Table 2
CALCE battery data [43].
Battery number Charging current Charging cut-off voltage Discharging current End-off voltage Rated capacity Temperature

CS2-34 0.5C 4.2 V 0.5C 2.7 V 1.1 Ah 24 ◦ C

CS2-36 0.5C 4.2 V 1C 2.7 V 1.1 Ah 24 ◦ C
CS2-37 0.5C 4.2 V 1C 2.7 V 1.1 Ah 24 ◦ C

Fig. 4. CALCE batteries [43] (a) Current and voltage conditions of single charge-discharge; (b) The battery capacity degradation curve.

battery capacity periodically increases in the process of continuous

reduction. Moreover, compared with the NASA batteries, the battery 1∑ n
MAE = |yi − ̂y i | (23)
capacity demonstrates higher fluctuation levels. Similarly, the EOL of n i=1
the CALCE battery is defined as the threshold when its capacity degrades √̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
to 70% of the rated capacity, precisely 0.81 Ah. 1∑ n
RMSE = (yi − ̂y i )2 (24)
n i=1
4.2. Prediction of battery capacity degradation trend
where n represents the number of samples, yi represents the estimated
To verify the accuracy of the proposed IPSO-PF method for capacity data from different models and ̂ y i represents the experimental test data.
prediction, two battery datasets are used to perform the simulation. The The MAE and RMSE fitting errors for the four batteries are provided
first simulation experiment is conducted using the battery capacity in Table 3. For battery B5, the PF fitting method yields an MAE error of
degradation dataset from NASA [41]. To illustrate the advantages of the 0.0060 and an RMSE of 0.0083, while the IPSO-PF fitting method ach­
proposed IPSO-PF method over the traditional PF, the capacity estima­ ieves an MAE error of 0.0027 Ah and an RMSE of 0.0036. Meanwhile,
tion results of battery B5, B6, B7 and B18 are shown in Fig. 5, and the fitting error calculations are also performed on battery B6, B7 and B18,
comparison of fitting errors is present on the right side of Fig. 5. where the proposed IPSO-PF method consistently demonstrates higher
In Fig. 5, the actual capacity is represented by a purple solid line, the accuracy with smaller errors than the PF method, which further in­
estimated capacity by the PF method is shown as an orange solid line, dicates that the proposed IPSO-PF method offers superior precision in
and the estimated capacity by the IPSO-PF method is depicted as a blue battery estimation and prediction, making it a more effective alternative
solid line. Notably, the capacity degradation process appears highly to the PF method.
irregular, with indications of capacity growth. It can be seen that, To further illustrate the generalization ability of the proposed IPSO-
although the estimated capacities from both methods align closely, the PF method, another battery capacity degradation dataset collected from
capacity estimated from the proposed IPSO-PF method can better fit the the CALCE [43] is utilized to further validate the accuracy of the
actual capacity. This conclusion can be further verified in the error curve IPSO-PF method. The capacity tracking results for battery CS2-34,
on the right side of Fig. 5. To be specific, the fluctuation range of errors CS2-36, and CS2-37 obtained using the PF and IPSO-PF methods are
for the PF method is greater than that of the IPSO-PF method. Overall, shown in Fig. 6, along with the comparison of capacity errors on the
based on the observations from Fig. 5, the IPSO-PF method demonstrates right side of Fig. 6.
superior performance and smaller errors in estimating battery capacity In Fig. 6, the purple solid line, orange solid line, and green solid line
compared with the PF method, particularly for battery B5 and B6. have the same meaning as the corresponding curves in Fig. 5. Yet, the
To quantitatively assess the accuracy of the IPSO-PF method, two green solid line here represents the capacity estimated by the IPSO-PF
performance indicators, namely the Mean Absolute Error (MAE) and the method. From Figs. 6(a-1), Fig. 6(b-1), and Figs. 6(c-1), both the pro­
Root Mean Square Error (RMSE) are utilized to evaluate the estimation posed IPSO-PF and the conventional PF methods demonstrate effective
and prediction outcomes, which are expressed by battery capacity estimation capabilities even though the capacity
changes appear rugged with two distinctive stages. However, upon
closer inspection of the zoomed-in image in Fig. 6, it is evident that the

6
H. Pang et al. Energy 293 (2024) 130555

Fig. 5. The comparison results of the proposed IPSO-PF and PF method in tracking the capacity of the NASA cells.

Table 3
Fitting errors of four cells.
Battery cell B5 B6 B7 B18

Method PF IPSO-PF PF IPSO-PF PF IPSO-PF PF IPSO-PF

MAE 0.0060 0.0027 0.0058 0.0035 0.0069 0.0037 0.0049 0.0026
RMSE 0.0083 0.0036 0.0084 0.0058 0.0091 0.0053 0.0073 0.0043

estimated capacity values obtained by the IPSO-PF algorithm align more battery capacity, providing more precise battery capacity estimation
closely with the true values. Observing the right side of Fig. 6, the ca­ with lower MAE and RMSE errors. This reaffirms the IPSO-PF method as
pacity estimation errors fluctuate within the range of − 0.02V to 0.04 Ah. a reliable and effective method for predicting the capacity of LIBs.
In comparison to the capacity error fluctuation range depicted in Fig. 5,
both methods show decreased accuracy in predicting the CS2 series 4.3. RUL prediction and validation at various starting points
batteries, which may be related to the discrepancies in the databases.
However, it is worth noting that the green solid line exhibits a smaller 4.3.1. RUL prediction for NASA lithium-ion battery
fluctuation range and it is closer to zero. This implies that the IPSO-PF To evaluate the effectiveness of the proposed IPSO-PF method, the
method yields smaller capacity estimation errors compared to the PF NASA battery capacity degradation dataset [41] is utilized to conduct
method and it is more suitable for battery RUL prediction applications. RUL prediction experiments and illustrate the predictive performance of
Table 4 compares the MAE and RMSE values for battery capacity the proposed IPSO-PF method against the PF method.
estimation using the CALCE battery dataset. Similar to Table 3, the PF Before commencing the experiments, three important regression
method yields an MAE of 0.0049 and an RMSE of 0.0066 for the CS2-34 metrics as RMSE provided in (24), Absolute Error (AE), and Relative
battery, while the IPSO-PF fitting method results in an MAE of 0.0032 Error (RE) are introduced to assess the predictive performance of the
and an RMSE of 0.0045 for the same battery. Similar results are obtained overall degradation process.
for battery CS2-36 and CS2-37. Additionally, the proposed IPSO-PF Herein, both AE and RE are respectively given by
method has higher accuracy and broader applicability in estimating

7
H. Pang et al. Energy 293 (2024) 130555

Fig. 6. The comparison results of the proposed IPSO-PF and PF method in tracking the capacity of the CALCE cells.

overall degradation prediction results, validating the effectiveness of the

Table 4
proposed method in forecasting the capacity at each cycle. Importantly,
Fitting errors of three cells.
the proposed IPSO-PF method generates minimal AE and RE, signifying
Battery cell CS2-34 CS2-36 CS2-37 its capability to achieve highly accurate RUL predictions with negligible
Method PF IPSO-PF PF IPSO-PF PF IPSO-PF errors. In conclusion, the IPSO-PF method, as an enhanced PF method,
MAE 0.0049 0.0032 0.0049 0.0027 0.0048 0.0025 exhibits remarkable adaptability in RUL prediction for LIBs, which can
RMSE 0.0066 0.0045 0.0063 0.0037 0.0065 0.0039
effectively capture the intricate nonlinear degradation curve of
batteries.
⃒ ⃒
AE = ⃒Tcycle − Pcycle ⃒ (25)
4.3.2. RUL prediction for CALCE lithium-ion battery
⃒ ⃒
⃒Tcycle − Pcycle ⃒ To further assess the accuracy of the proposed IPSO-PF method, this
RE = (26) study also conducts the experimental validation of the CALCE lithium-
Tcycle
ion battery capacity degradation dataset [43]. Both the PF and
IPSO-PF methods are utilized to predict the RUL for battery CS2-34,
where Tcycle denotes the actual number of cycles experienced by the
CS2-36, and CS2-37. To better illustrate the estimation robustness, the
battery throughout its true lifetime, and Pcycle represents the predicted
predictive results initiated at K = 80 and K = 400, are illustrated in
number of cycles until failure.
Fig. 9. The blue, purple, and orange solid lines correspond to the curves
Figs. 7 and 8 depict the experimental outcomes showcasing the ca­
shown in Fig. 7. The solid blue line in Fig. 9 exhibits notable variations in
pacity trajectories of battery B5, B6, B7, and B18. It should be noted that
the capacity of the CS2 battery, reflecting distinct capacity growth.
the prediction starting points designated as K = 58 and K = 78 are
Moreover, both the PF and IPSO-PF methods can successfully capture
adopted for the RUL verification. Moreover, the blue solid line repre­
the trend of capacity degradation. Notably, the orange solid line can
sents the actual measured capacity of LIBs, the purple solid line depicts
approximate the blue solid line at the failure threshold closely, out­
the capacity degradation curve predicted by the PF method and the
performing the purple solid line. This observation confirms that the
orange solid line corresponds to the capacity degradation curve pre­
proposed IPSO-PF method can be capable of predicting the actual
dicted by the proposed method. More importantly, the green dashed line
behavior of the three batteries with higher precision.
denotes the prediction start point, and the black solid line signifies the
Table 6 provides the comparisons of the RMSE, AE, and RE for three
battery capacity failure threshold. From the zoomed-in sections of
battery ensembles. It is observed that the three indicators by the pro­
Figs. 7 and 8, it is evident that the orange solid line closely aligns with
posed IPSO-PF are all definitely less than those by the PF method, which
the blue solid line near the failure threshold. This result indicates that
means that the proposed method can accurately forecast the battery RUL
the proposed method achieves a predictive trend that is almost in
for each cycle. Specifically, for battery CS34 at K = 80, the IPSO-PF
accordance with the real situation of the four batteries.
method outperforms the PF method by predicting the RUL with an
Moreover, some detailed quantitative indicators comparing the
improvement of 34 cycles, which highlights the superior predictive ac­
predictive performance of four battery sets are presented in Table 5. The
curacy of the IPSO-PF method.
related results reveal that the proposed method consistently outperforms
Furthermore, the three batteries are tested again by shifting the
the PF method, exhibiting lower RMSE. The RMSE effectively depicts the
prediction starting point to further verify the accuracy of the proposed

8
H. Pang et al. Energy 293 (2024) 130555

Fig. 7. RUL prediction results at Cycle 58 for NASA batteries.

Fig. 8. RUL prediction results at Cycle 78 for NASA batteries.

IPSO-PF method. The result reveals that the proposed method exhibits 5. Conclusion
the smallest RMSE, AE, and RE, indicating highly accurate RUL pre­
diction with minimal error. In summary, the IPSO-PF method proves its In this paper, an accurate fusion method is proposed to enable ca­
adaptability to the nonlinear degradation characteristics of lithium-ion pacity and RUL prediction with high precision and good robustness for
battery capacity, excelling in prediction accuracy and holding signifi­ LIBs. First, a dual-exponential empirical model is presented to describe
cant potential for widespread application in RUL prediction. the degradation trend of initial capacity in LIBs. Next, an IPSO algorithm
with an adaptive weight is presented to solve the problem of local

9
H. Pang et al. Energy 293 (2024) 130555

Table 5
Prediction effect comparison of NASA batteries.
Indicators Methods B5 B6 B7 B18

K = 58 K = 78 K = 58 K = 78 K = 58 K = 78 K = 58 K = 78

RMSE PF 0.047 0.024 0.044 0.040 0.049 0.056 0.042 0.029

IPSO-PF 0.035 0.016 0.036 0.030 0.033 0.038 0.035 0.018
AE PF 16 6 8 7 24 15 9 7
IPSO-PF 10 2 6 3 9 11 5 4
RE PF 0.129 0.048 0.074 0.064 0.193 0.120 0.092 0.072
IPSO-PF 0.081 0.016 0.055 0.027 0.0725 0.088 0.051 0.041

Fig. 9. RUL prediction results at different cycles for CALCE batteries.

optimal solution. Subsequently, the optimal particle searched by IPSO is

Table 6
updated continuously by the PF algorithm to achieve a more accurate
Prediction effect comparison of CALCE batteries.
posterior estimation. Finally, the proposed IPSO-PF method is verified
Indicators Methods CS34 CS36 CS37 by database of NASA and CALCE batteries. The experimental results
K= K= K= K= K= K= demonstrate the proposed method has better precision and generaliz­
80 400 80 400 80 400 ability, achieving a maximum MAE for capacity estimation below 4%
RMSE PF 0.031 0.036 0.051 0.021 0.022 0.028 and a maximum RMSE below 6%, while also maintaining an RMSE in
IPSO-PF 0.025 0.028 0.029 0.012 0.011 0.022 RUL prediction within 5%. Furthermore, even if the prediction threshold
AE PF 63 46 56 40 25 27 range is expanded, the capacity and RUL prediction accuracy of the
IPSO-PF 29 26 22 7 4 5
RE PF 0.106 0.077 0.096 0.067 0.037 0.040
IPSO-PF method remain higher than the standard PF method, which
IPSO-PF 0.049 0.043 0.036 0.011 0.006 0.007 provides a theoretical foundation for ensuring the safe and reliable
operation of electric vehicles.
Future work may investigate a more effective and streamlined

10
H. Pang et al. Energy 293 (2024) 130555

method for predicting the RUL of batteries, which combines filtering and prediction of electric vehicle lithium batteries. Renew Sustain Energy Rev 2022;
156:111843. https://doi.org/10.1016/j.rser.2021.111843.
data-driven methods. Additionally, a promising research direction
[9] Li X, Yuan C, Wang Z, Xie J. A data-fusion framework for lithium battery health
would be to establish an RUL prediction model for battery packs, condition Estimation Based on differential thermal voltammetry. Energy 2022;239:
starting from the RUL prediction model for individual cells. 122206. https://doi.org/10.1016/j.energy.2021.122206.
[10] Guo F, Wu X, Liu L, Ye J, Wang T, Fu L, Wu Y. Prediction of remaining useful life
and state of health of lithium batteries based on time series feature and Savitzky-
Ethical approval Golay filter combined with gated recurrent unit neural network. Energy 2023;270:
126880. https://doi.org/10.1016/j.energy.2023.126880.
[11] Patil MA, Tagade P, Hariharan KS, Kolake SM, Song T, Yeo T, Doo S. A novel
Not applicable.
multistage Support Vector Machine based approach for Li ion battery remaining
useful life estimation. Appl Energy 2015;159:285–97. https://doi.org/10.1016/j.
Funding apenergy.2015.08.119.
[12] Wang J, Zhang S, Li C, Wu L, Wang Y. A data-driven method with mode
decomposition mechanism for remaining useful life prediction of lithium-ion
This work is supported by the Artificial intelligence technology batteries. IEEE Trans Power Electron 2022;37(11):13684–95. https://doi.org/
project of Xi’an Science and Technology Bureau (No. 21RGZN0014). 10.1109/TPEL.2022.3183886.
[13] Guo W, He M. An integrated method for bearing state change identification and
prognostics based on improved relevance vector machine and degradation model.
Availability of data and materials IEEE Trans Instrum Meas 2022;71:1–14. https://doi.org/10.1109/
TIM.2022.3161705.
[14] Li X, Yu D, Vilsen SB, Store DI. The development of machine learning-based
All data generated and analyzed during this study are included in this remaining useful life prediction for lithium-ion batteries. J Energy Chem 2023.
published article. The data that support the findings of this study are https://doi.org/10.1016/j.jechem.2023.03.026.
available on request from the corresponding author. [15] Tong Z, Miao J, Tong S, Lu Y. Early prediction of remaining useful life for Lithium-
ion batteries based on a hybrid machine learning method. J Clean Prod 2021;317:
128265. https://doi.org/10.1016/j.jclepro.2021.128265.
CRediT authorship contribution statement [16] Yang X, Zheng Y, Zhang Y, Wong DS-H, Yang W. Bearing remaining useful life
prediction based on regression shapalet and graph neural network. IEEE Trans
Instrum Meas 2022;71:1–12. https://doi.org/10.1109/TIM.2022.3151169.
Hui Pang: Conceptualization, Data curation, Funding acquisition, [17] Ma G, Zhang Y, Cheng C, Zhou B, Hu P, Yuan Y. Remaining useful life prediction of
Investigation, Methodology, Writing – original draft, Writing – review & lithium-ion batteries based on false nearest neighbors and a hybrid neural network.
editing. Kaiqiang Chen: Data curation, Methodology, Validation, Appl Energy 2019;253:113626. https://doi.org/10.1016/j.apenergy.2019.113626.
[18] Cheng Y, Wang C, Wu J, Zhu H, Lee CK. Multi-dimensional recurrent neural
Writing – original draft. Yuanfei Geng: Conceptualization, Data cura­ network for remaining useful life prediction under variable operating conditions
tion, Formal analysis, Methodology, Validation, Writing – original draft. and multiple fault modes. Appl Soft Comput 2022;118:108507. https://doi.org/
Longxing Wu: Formal analysis, Investigation, Writing – review & 10.1016/j.asoc.2022.108507.
[19] Zhang Y, Xiong R, He H, Pecht MG. Long short-term memory recurrent neural
editing. Fengbin Wang: Methodology, Writing – review & editing.
network for remaining useful life prediction of lithium-ion batteries. IEEE Trans
Jiahao Liu: Data curation, Writing – review & editing. Veh Technol 2018;67(7):5695–705. https://doi.org/10.1109/TVT.2018.2805189.
[20] Zhao S, Zhang C, Wang Y. Lithium-ion battery capacity and remaining useful life
prediction using board learning system and long short-term memory neural
Declaration of competing interest network. J Energy Storage 2022;52:104901. https://doi.org/10.1016/j.
est.2022.104901.
[21] Xiong R, Sun F, Chen Z, He H. A data-driven multi-scale extended Kalman filtering
The authors declare that they have no known competing financial based parameter and state estimation approach of lithium-ion polymer battery in
interests or personal relationships that could have appeared to influence electric vehicles. Appl Energy 2014;113:463–76. https://doi.org/10.1016/j.
the work reported in this paper. apenergy.2013.07.061.
[22] Pang H, Guo L, Wu L, Jin J, Zhang F, Liu K. A novel extended Kalman filter-based
battery internal and surface temperature estimation based on an improved electro-
Data availability thermal model. J Energy Storage 2021;41:102854. https://doi.org/10.1016/j.
est.2021.102854.
[23] Yan W, Zhang B, Zhao G, Tang S, Niu G, Wang X. A battery management system
Data will be made available on request.
with a Lebesgue-sampling-based extended Kalman filter. IEEE Trans Ind Electron
2018;66(4):3227–36. https://doi.org/10.1109/TIE.2018.2842782.
References [24] He H, Xiong R, Zhang X, Sun F, Fan J. State-of-charge estimation of the lithium-ion
battery using an adaptive extended Kalman filter based on an improved Thevenin
model. IEEE Trans Veh Technol 2011;60(4):1461–9. https://doi.org/10.1109/
[1] Pang H, Wu L, Liu J, Liu X, Liu K. Physics-informed neural network approach for
TVT.2011.2132812.
heat generation rate estimation of lithium-ion battery under various driving
[25] Chang Y, Fang H, Zhang Y. A new hybrid method for the prediction of the
conditions. J Energy Chem 2023;78:1–12. https://doi.org/10.1016/j.
remaining useful life of a lithium-ion battery. Appl Energy 2017;206:1564–78.
jechem.2022.11.036.
https://doi.org/10.1016/j.apenergy.2017.09.106.
[2] Zhang Y, Feng F, Wang S, Meng J, Xie J, Ling R, Yin H, Zhang K, Chai Y. Joint
[26] Zhang Y, Tu L, Xue Z, Li S, Tian L, Zheng X. Weight optimized unscented Kalman
nonlinear-drift-driven Wiener process-Markov chain degradation switching model
filter for degradation trend prediction of lithium-ion battery with error
for adaptive online predicting lithium-ion battery remaining useful life. Appl
compensation strategy. Energy 2022;251:123890. https://doi.org/10.1016/j.
Energy 2023;341:121043. https://doi.org/10.1016/j.apenergy.2023.121043.
energy.2022.123890.
[3] Wu L, Liu K, Liu J, Pang H. Evaluating the heat generation characteristics of
[27] Xue Z, Zhang Y, Cheng C, Ma G. Remaining useful life prediction of lithium-ion
cylindrical lithium-ion battery considering the discharge rates and N/P ratio.
batteries with adaptive unscented kalman filter and optimized support vector
J Energy Storage 2023;64:107182. https://doi.org/10.1016/j.est.2023.107182.
regression. Neurocomputing 2020;376:95–102. https://doi.org/10.1016/j.
[4] Feng F, Song B, Xu J, Na W, Zhang K, Chai Y. Multiple time scale state-of-charge
neucom.2019.09.074.
and capacity-based equalisation strategy for lithium-ion battery pack with passive
[28] Zheng X, Fang H. An integrated unscented kalman filter and relevance vector
equaliser. J Energy Storage 2022;53:105196. https://doi.org/10.1016/j.
regression approach for lithium-ion battery remaining useful life and short-term
est.2022.105196.
capacity prediction. Reliab Eng Syst Saf 2015;144:74–82. https://doi.org/
[5] Wu L, Pang H, Geng Y, Liu X, Liu J, Liu K. Low-complexity state of charge and
10.1016/j.ress.2015.07.013.
anode potential prediction for lithium-ion batteries using a simplified
[29] Zhang J, Xia C. State-of-charge estimation of valve regulated lead acid battery
electrochemical model-based observer under variable load condition. Int J Energy
based on multi-state Unscented Kalman Filter. Int J Electr Power Energy Syst 2011;
Res 2022;46(9):11834–48. https://doi.org/10.1002/er.7949.
33(3):472–6. https://doi.org/10.1016/j.ijepes.2010.10.010.
[6] Lin CP, Ling MH, Cabrera J, Yang F, Yu DYW, Tsui KL. Prognostics for lithium-ion
[30] Hong S, Qin C, Lai X, Meng Z, Dai H. State-of-health estimation and remaining
batteries using a two-phase gamma degradation process model, vol. 214.
useful life prediction for lithium-ion batteries based on an improved particle filter
Reliability engineering & system safety; 2021, 107797. https://doi.org/10.1016/j.
algorithm. J Energy Storage 2023;64:107179. https://doi.org/10.1016/j.
ress.2021.107797.
est.2023.107179.
[7] Sadabadi KK, Jin X, Rizzoni G. Prediction of remaining useful life for a composite
[31] Walker E, Rayman S, White RE. Comparison of a particle filter and other state
electrode lithium ion battery cell using an electrochemical model to estimate the
estimation methods for prognostics of lithium-ion batteries. J Power Sources 2015;
state of health. J Power Sources 2021;481:228861. https://doi.org/10.1016/j.
287:1–12. https://doi.org/10.1016/j.jpowsour.2015.04.020.
jpowsour.2020.228861.
[32] Restaino R, Zamboni W. Comparing particle filter and extended kalman filter for
[8] Li P, Zhang Z, Grosu R, Deng Z, Hou J, Rong Y, Wu R. An end-to-end neural
battery State-Of-Charge estimation. In: IECON 2012-38th Annual Conference on
network framework for state-of-health estimation and remaining useful life

11
H. Pang et al. Energy 293 (2024) 130555

IEEE Industrial Electronics Society; 2012. p. 4018–23. https://doi.org/10.1109/ regression. IEEE Trans Ind Electron 2017;65(7):5634–43. https://doi.org/
IECON.2012.6389247. 10.1109/TIE.2017.2782224.
[33] Qiu X, Wu W, Wang S. Remaining useful life prediction of lithium-ion battery [38] Wang D, Yang F, Tsui K-L, Zhou Q, Bae SJ. Remaining useful life prediction of
based on improved cuckoo search particle filter and a novel state of charge lithium-ion batteries based on spherical cubature particle filter. IEEE Trans Instrum
estimation method. J Power Sources 2020;450:227700. https://doi.org/10.1016/j. Meas 2016;65(6):1282–91. https://doi.org/10.1109/TIM.2016.2534258.
jpowsour.2020.227700. [39] Hao X, Wang S, Fan Y, Xie Y, Fernandez C. An improved forgetting factor recursive
[34] Ahwiadi M, Wang W. An enhanced particle filter technology for battery system least square and unscented particle filtering algorithm for accurate lithium-ion
state estimation and RUL prediction. Measurement 2022;191:110817. https://doi. battery state of charge estimation. J Energy Storage 2023;59:106478. https://doi.
org/10.1016/j.measurement.2022.110817. org/10.1016/j.est.2022.106478.
[35] Sun X, Zhong K, Han M. A hybrid prognostic strategy with unscented particle filter [40] Ma Q, Zheng Y, Yang W, Zhang Y, Zhang H. Remaining useful life prediction of
and optimized multiple kernel relevance vector machine for lithium-ion battery. lithium battery based on capacity regeneration point detection. Energy 2021;234:
Measurement 2021;170:108679. https://doi.org/10.1016/j. 121233. https://doi.org/10.1016/j.energy.2021.121233.
measurement.2020.108679. [41] Saha B, Goebel K. Battery data set. NASA Ames prognostics data Repository. In:
[36] Yang J, Fang W, Chen J, Yao B. A lithium-ion battery remaining useful life NASA Ames research Center. CA: Moffett Field; 2007. Available: https://ti.arc.
prediction method based on unscented particle filter and optimal combination nasa.gov/tech/dash/pcoe/prognostic-data-repository/#battery.
strategy. J Energy Storage 2022;55:105648. https://doi.org/10.1016/j. [42] Wu L, Lyu Z, Huang Z, Zhang C, Wei C. Physics-based battery SOC estimation
est.2022.105648. methods: recent advances and future perspectives. J Energy Chem 2024;89:27–40.
[37] Wei J, Dong G, Chen Z. Remaining useful life prediction and state of health https://doi.org/10.1016/j.jechem.2023.09.045. 2024/02/01/.
diagnosis for lithium-ion batteries using particle filter and support vector [43] Pecht M. Battery data set. CALCE. Maryland, MD: CALCE Battery Research Group;
2017. Available: https://web.calce.umd.edu/batteries/index.html.

12