Journal Pre-proof

Two-stage prediction method for capacity aging trajectories of lithium-ion batteries

based on Siamese-convolutional neural network

Jingcai Du, Caiping Zhang, Shuowei Li, Linjing Zhang, Weige Zhang

PII: S0360-5442(24)00719-9
DOI: https://doi.org/10.1016/j.energy.2024.130947
Reference: EGY 130947

To appear in: Energy

Received Date: 7 November 2023

Revised Date: 5 March 2024
Accepted Date: 6 March 2024

Please cite this article as: Du J, Zhang C, Li S, Zhang L, Zhang W, Two-stage prediction method for
capacity aging trajectories of lithium-ion batteries based on Siamese-convolutional neural network,
Energy (2024), doi: https://doi.org/10.1016/j.energy.2024.130947.

This is a PDF file of an article that has undergone enhancements after acceptance, such as the addition
of a cover page and metadata, and formatting for readability, but it is not yet the definitive version of
record. This version will undergo additional copyediting, typesetting and review before it is published
in its final form, but we are providing this version to give early visibility of the article. Please note that,
during the production process, errors may be discovered which could affect the content, and all legal
disclaimers that apply to the journal pertain.

© 2024 Published by Elsevier Ltd.

 Two-stage prediction method for capacity aging trajectories of lithium-ion batteries

based on Siamese-Convolutional neural network

Jingcai Du, Caiping Zhang*1, Shuowei Li, Linjing Zhang, Weige Zhang
National Active Distribution Network Technology Research Center (NANTEC), Beijing Jiaotong University, Beijing, 100044,
China
Abstract
Accurate battery lifespan prediction can guarantee battery safety in applications. Considering the complex nonlinear nature
of battery degradation, it is more challenging to predict battery aging trajectories utilizing less data. A two-stage prediction
method for battery capacity aging trajectories is proposed. The first stage is battery lifespan prediction based on the Siamese-
Convolutional neural network (Siamese-CNN). Taking the predicted lifespan as its prior information, the second stage is
predicting capacity aging trajectories with the Convolutional neural network (CNN). The method only requires partial
voltage-current data from 30 cycles, of which the similarity in different batteries is obtained by the Siamese-CNN. The
lifespan-unknown batteries in the testing set are fed into the Siamese-CNN to match the lifespan-known training set battery

f
oo
with the highest similarity. Then, CNN is employed to predict the capacity aging trajectories under different profiles. The
ratio of each cell’s predicted lifespan to the fixed length of the CNN output sequence is regarded as the interval for uniform
sampling to keep the series consistent. The prediction accuracy of the method is validated by an open dataset. It is

r
-p
demonstrated that the battery lifespan and capacity aging trajectory prediction achieve a mean absolute percentage error
(MAPE) of 2.44% and 1.28%, respectively.
re
Keywords: Lithium-ion battery; Battery lifespan prediction; Capacity aging trajectory prediction; Siamese-Convolutional
neural network.
lP

1 Introduction
Lithium-ion batteries are essential components of energy storage technology and are widely utilized in electric vehicles
and electronic devices[1,2,3]. However, in recent years, thermal runaway accidents of high-ratio lithium-ion battery systems
na

have been frequent[4]. To guarantee the safety and reliability of batteries, battery prediction and health management (PHM)
technology have been developed[5,6]. The existing research on battery health management includes two parts: battery
ur

lifespan prediction[7] and capacity aging trajectory prediction[8]. Both of them demand a massive amount of data and time
to train the model, which seriously affects the health management of batteries[9]. The predictions of battery lifespan and
Jo

capacity aging trajectory in the early-cycle stage under different profiles based on a small quantity of test data are the key to
ensuring the safe and efficient operation of battery systems[10]. It can also facilitate research on new materials and battery
design[11].
Extensive research has been carried out on battery lifespan prediction and capacity aging trajectory prediction. The
methods could be classified into two categories: model-based methods, and data-driven methods[12].
The model-based method consists of three types of models based on the information studied: electrochemical models[13]
containing the material properties of lithium-ion batteries, equivalent circuit models[14] reacting to the variation of lithium-
ion battery parameters, and empirical models[15] incorporating the profiles of lithium-ion batteries. Bhaskar[12] et al.
proposed that the capacity loss of the battery over a wide range of temperatures and discharge rates fits a physical model. The
model calculates the loss of lithium-ion based on two related mechanisms and achieves the estimation of battery capacity
loss. Zhang[16] et al. developed a Remaining Useful Life (RUL) prediction method based on the Box-Cox transformation
and Monte Carlo simulation. Chen[17] et al. proposed a gray neural network model combining a gray model and BP neural
network to predict the RUL of the battery. Cong[18] et al. improved the double exponential mathematical representation and
proposed an improved empirical capacity degradation model. The unscented particle filter (UPF) algorithm was employed to

*Corresponding author at: National Active Distribution Network Technology Research Center (NANTEC), Beijing Jiaotong University,
Beijing, 100044, China
E-mail address: [email protected] (Caiping Zhang)
1 / 16
predict the RUL. The model-based prediction method is complex, and the identification of parameters requires more
experiments, which makes it difficult to identify the parameters quickly and accurately.
The data-driven battery lifespan prediction method does not need complex physical models but focuses on mining
information related to the state of health (SOH) and battery lifespan from historical battery profile data. In practical
application, the data-driven method is gradually becoming a research hotspot as one of the most important methods for battery
lifespan prediction. Chen[19] et al. constructed a novel TOP model by combining CNN and Long Short-term Memory
(LSTM), where CNN was employed to extract features related to battery lifespan, and the features were used as the input of
LSTM to achieve early lifespan prediction using data from the first 100 cycles. Hong[20] et al. put forward an expansive
CNN and a one-dimensional CNN to achieve fast prediction of RUL, which is 25 times faster than the traditional conventional
methods. Yang[21] et al. proposed a hybrid network combining CNN and Bi-directional Long Short-Term Memory (Bi-
LSTM). The CNN extracted features from the charge and discharge currents. The Bi-LSTM calculated bidirectional time-
dependent information from the raw data and improved the accuracy of RUL capacity prediction. Zhang[10] et al. proposed
a hybrid parallel residual Convolutional neural network, which can effectively predict the battery RUL. The voltage, current,
and temperature data under charging profiles were fused through CNN, and the degradation features of different states of
aging were extracted by residual networks. Ding[22] et al. decomposed the charge and discharge current into sub-signals by

f
oo
wavelet packet decomposition to highlight the nonlinear information of battery degradation, and then the features were
extracted by CNN to evaluate the RUL of the battery. The above studies only realized the prediction of RUL, and most
training methods are end-to-end, that is, a set of specific features is associated with a set of target values. Although RUL

r
-p
predictions could be achieved with errors in a certain range, the full-life capacity aging trajectories of batteries are not
predicted. Therefore, the mentioned methods are more suitable for the widespread screening of aging batteries rather than the
re
accurate tracking of degradation.
To address this problem, Li[23] et al. developed a recurrent neural network-based sequence-to-sequence (RNN-S2S)
lP

model. The capacity and profile data of the first 100 cycles were utilized as input data to predict the capacity aging trajectory.
Saurabh Saxena[24] et al. designed a CNN model that applied the discharge voltage-capacity curves of the first 100 cycles
to predict the capacity aging trajectory. Zhang[25] et al. applied LSTM to RUL prediction and SOH prediction. The method
na

enabled multi-step prediction of capacity based on capacity sequences. Xu[26] et al. extracted the features from charge and
discharge profile data and selected the features with high correlation with capacity as final health features. The battery
ur

capacity trajectory was predicted based on an online sequential limit learning machine with beetle antenna search
optimization.
Jo

In the existing research, the methods of predicting the capacity aging trajectory iteratively are more widely utilized. The
biggest disadvantage of this type of method is that it demands real-time updating of battery degrading data and retraining the
models, which is not suitable for achieving fast and accurate lifespan prediction in the early-cycle stage of battery operation.
Some literature does not adopt the iterative method but requires more physical quantities (voltage, capacity, current,
temperature, internal resistance, etc.) and complete charge/discharge profile data. However, batteries are not always fully
charged and discharged in practical applications owing to usage behaviors or specific profiles, which make it difficult to
obtain complete charge and discharge data with capacity, temperature, and internal resistance[27]. Therefore, these methods
are limited in their implementation.
Considering the limitations of the mentioned methods, a novel two-stage prediction method of capacity aging trajectory
in the early-cycle stage based on Siamese-CNN is proposed. The highlights of this paper are as follows.
(1) Siamese-CNN is employed for the lifespan prediction of lithium-ion batteries, in which the similarity between
batteries is compared. Batteries with the highest similarity are considered to share the homologous lifespan.
(2) Taking the predicted lifespan as prior information, CNN is utilized for the prediction of battery capacity aging
trajectories under different profiles. The capacity aging trajectories can be obtained without updating any data after the first
30 cycles.
(3) The prediction effectiveness and accuracy of the two-stage method are validated using a public dataset. It is
demonstrated that the proposed method is capable of precisely predicting the capacity aging trajectory and achieving highly
2 / 16
accurate.
The paper is organized as follows. Section 2 introduces the experimental data briefly, and analyzes the selection basis and
construction process of input data in detail. Section 3 describes the theory and model structure of the battery lifespan
prediction and capacity aging trajectory prediction method. Section 4 presents the prediction results in detail and compares
the method with other methods. Section 5 summarizes the discussion and conclusion of this paper.
2 Data description

2.1 Dataset
In this paper, a data set contributed by K. A. Severson of MIT and P.M. Attia of Stanford University[28] is cited, which
consists of 124 batteries of a commercial lithium iron phosphate/graphite battery with a rated capacity of 1.1Ah and a rated
voltage of 3.3V. The charging profile for each battery is chosen from 72 different charging profiles. All batteries are
discharged under the same profile, which is constant current discharge to a cut-off voltage of 2 V at a 4C rate. Fig. 1(a)
illustrates the capacity aging trajectory for all batteries. The battery lifespan ranges roughly from 150 to 2300 cycles. At the
end of battery degradation, the degradation rate is significantly accelerated, showing the typically nonlinear degradation
process.

f
oo
(a) (b)
1.1 1.1
Discharge Capacity(Ah)

Discharge Capacity(Ah)
1.05 1.05

r
1 1

0.95
-p 0.95
Sample1
Sample2
Sample3
Sample4
re
0.9 0.9 Sample5
Sample6
0.85 0.85
0 500 1000 1500 2000 2500 0 500 1000 1500 2000
lP

Cycle Number Cycle Number

Fig. 1. Battery capacity aging trajectories. (a) Capacity aging trajectories of all batteries; (b) The lifespan of batteries under the same charging strategy.

Even under the same charging profile, there exists inconsistency in the aging paths of batteries. Taking six batteries under
na

the same charging profile as an example, as depicted in Fig. 1(b), the lifespans of these six batteries are quite different when
the batteries degrade to the same SOH. The lifespan of the No.1 battery is the shortest (800 cycles), and the lifespan of the
ur

No.6 battery is the longest (1895 cycles). Therefore, it is demonstrated that the lifespans of batteries under the same profile
could be different, which further makes it difficult to predict the battery capacity aging trajectory in the early-cycle stage.
Jo

2.2 Input data generation

The input voltage segments of the proposed model are determined by the aging mechanism of batteries. The performance
of the battery degrades irreversibly, resulting in a corresponding change in the open-circuit voltage (OCV) of the battery[29].
Meanwhile, battery capacity gradually decreases and the charge/discharge voltage curve changes significantly. As shown in
Fig. 2(a), the characteristic of the cycle-to-cycle evolution of the discharge voltage can be found as the color of the discharge
voltage curve changes gradually from blue to red with the cycle number increasing. In different cycles, the voltage curve
characteristics such as platform voltage value, duration, and curve end point are different.

Fig. 2. (a) Discharge voltage curve under different cycle numbers; (b) dQ/dV at C/10; (c) dQ/dV at 4C.

3 / 16
 Based on this principle, Dubarry and Liaw et al. proposed a battery aging mechanism diagnosis method based on
Incremental capacity analysis (ICA) and Differential voltage analysis (DVA)[30,31]. The change of peak and valley
characteristics of the IC curve can not only reflect the aging mechanism of the battery but also relate intimately to the battery
capacity. The extraction of IC curves is not limited to specific charging or discharging data. The condition is selected based
on the contribution of the data to the feature extraction from the IC curve. In this paper, we extract the IC curve for the
discharge data, this is because in the first 30 cycles of data used, the multi-step CC condition is used for charging, and the
ICs of multi-step CC are unstable under different conditions, while the stable CC discharge condition is more suitable for
extracting the IC features. Fig. 2(b) exhibits the battery IC curves of different cycles at 1/10 C rate. During the discharge
process, five-phase change reactions occur at the negative electrode of the battery, which are depicted as five capacity
increment peaks on the IC curve, the corresponding voltage range is 3.15~3.4V. Taking the IC curve of a fresh battery (cycle1)
as a reference, the peak height, peak position, and peak area of the IC curve change significantly as the battery ages.
However, the battery is hardly charged and discharged at the rate of 1/10C or even less in practice. A typical profile is a
low current charge/discharge test at certain cycle intervals. Nevertheless, low-current charging and discharging will lead to
capacity recovery, and frequent low-current tests may change the battery's degradation mode. Fig. 2(c) exhibits the IC curve
of 4C rate discharge. Due to the non-uniform embedding and detachment of lithium ions, only one main peak is characterized,

f
oo
and the other peaks cannot be characterized. The voltage range corresponding to the main peak is 2.8~3.3V. With the aging
of the battery, the peak height, peak area, and peak position have also changed, which indicates that the IC curve drawn by
the charge-discharge curve at a high rate can still reflect the aging of the battery. According to this analysis, the voltage

r
-p
interval of 2.8~3.3V is utilized as the input, corresponding to 1% to 95% state of charge (SOC) of the battery, which reflects
the performance degradation of the battery.
re
During the constant-current discharge, the BMS samples the voltage V(t) and the current I(t) and then calculates the
capacity Q by the ampere-hour integration method. The expression is as follows,
lP

t
Q(t)= ∫t v_end I(t)dt (1)
v_start

where tv_start is the moment of the selected voltage start point and tv_end is the moment of the selected voltage endpoint.
na

Assuming the interval of discrete sampling is ΔV, there are N sampling points in the sampling voltage interval, N= (Vstart
- Vend)/ΔV. The sampling curve is encoded into a 2×N matrix M, as illustrated follows, which is adopted as the input of the
ur

model.
V1 V2 … VN
Jo

𝑀 = [Q Q … Q ] (2)
1 2 N

where VN and QN are the voltage and capacity at the N-th sampling point, respectively.
Regardless of the voltage sampling from any position, the capacity corresponding to the sampling start point is set to 0,
that is, the input capacity sequence is the discharge capacity corresponding to the voltage sampling interval, which does not
depend on the complete discharge curve, making the input data more flexible and easier to be extracted in the actual BMS.
Each battery is constructed as a 2×N matrix according to the above criteria. The scale variance of different inputs can
negatively affect the optimization of Convolutional networks, resulting in the optimal direction being the same as the negative
gradient direction. To eliminate this impact, the input matrix is normalized to the maximum and minimum.
3 Proposed Methodology
A novel two-stage capacity aging trajectory prediction method for lithium-ion batteries is proposed. The first stage is the
battery lifespan prediction based on the Siamese-CNN, and taking the predicted lifespan as its prior information, the second
stage is to predict the capacity aging trajectory with the CNN, which are presented in Sections 3.1 and 3.2. Fig. 3 demonstrates
the overview of the proposed method.
The framework of the battery lifespan prediction based on Siamese-CNN is illustrated in the green box of Fig. 3. Firstly,
the discharge voltage and current of each cell in the first 30 cycles are preprocessed based on the rules in Section 2.2. Then
the preprocessed data of two batteries are fed into the Siamese-CNN which consists of three Convolutional layers, three

4 / 16
Pooling layers, and two Dense layers to extract features. By calculating the Euclidean distance (ED) between the features,
the two batteries with smaller ED are considered to share a homologous lifespan. It is noted that in the model training process,
two lifespan-known batteries in the training set are inputted into the model, whereas in the model testing process or practical
applications, the inputs to the model are the data of one lifespan-unknown battery in the testing set and one lifespan-known
battery in the training set, respectively.
The framework of the battery capacity aging trajectory prediction with CNN is illustrated in the yellow box of Fig. 3. The
ratio of the predicted lifespan of each cell to the fixed length of the CNN output sequence is regarded as the interval for
uniform sampling. The battery cycle data are sampled according to the interval and fed into the CNN to realize the battery
capacity aging trajectory prediction.

Data Battery cycle life prediction based on Siamese-CNN

pre-processing
Cell 1

...
...
Battery single-cycle
data 2D 2D 2D 2D 2D 2D Diverse
Conv Pooling Conv Pooling Conv Pooling Euclidean Lifespan
distance
Share weights Homologous

f
Lifespan

oo
...
Cell 2

...
...
2D 2D 2D 2D 2D 2D
Conv Pooling Conv Pooling Conv Pooling
.. .
Cn

r
C2

V1 V2
Q1 Q2
.. .. .. VQ
N
N
C1
-p
Capacity trajectory prediction based on CNN

...
Cycle
re
Vend

Q(V ) =  I (t )dt

...
Vstart

...
2D 2D 2D 2D 2D 2D Capacity
Conv Pooling Conv Pooling Conv Pooling
lP

Fig. 3. The overview of the proposed method for battery lifespan prediction and capacity aging trajectory prediction.
na

3.1 Battery lifespan prediction based on Siamese-CNN

Siamese neural network [32] was first employed for the verification of signatures written on a pen-input tablet and has
ur

been widely adopted for feature matching recently. It is a similarity measurement method, which is utilized to compare and
match new unknown samples. More specifically, the Siamese-CNN[33] is trained to learn features, making the dissimilarity
Jo

between samples in the same class small and the dissimilarity between samples in different classes large. In this paper, the
Siamese-CNN is employed to realize the battery lifespan prediction under different profiles. The structure of the proposed
Siamese-CNN is indicated in Fig. 4.
CNN mapping function
...
...

Diverse
Convolution Pooling Euclidean Lifespan
distance
share weights
Homologous
Lifespan
CNN mapping function
...
...
...

Convolution Pooling

Fig. 4. The structure of the Siamese-CNN.

The Siamese neural network is usually composed of two neural network branches sharing weights, which extracts the
features from the two input values. The neural network can be LSTM, CNN, or any other network. Since CNN can
comprehensively and automatically capture features hidden in curves[34], this paper employs CNN as the neural network of
the Siamese neural network as depicted in the blue box in Fig. 4. CNN is a classical deep-learning algorithm that has achieved

5 / 16
great success in various scenarios, which is generally composed of Convolution layers, Pooling layers, and Dense layers. The
main function of the Convolution layer is to slide convolution kernels over the input data to obtain a set of features through
convolution operations, and each convolution kernel implies the extraction of different features. This process is called Feature
Mapping. The formula is as follows.
𝐾
(l+1)
𝑍 (l+1) (i , j) = f (∑ 𝑍 l (i , j) ⨂𝜔k + 𝑏 (l+1) ) (3)
k=1

l l+1 (l+1)
where Z (i , j) and Z (i , j) are the input and output of the (l+1) layer, respectively. 𝜔k denotes the weight of the neurons
(l+1)
in the (𝑙 + 1)th layer. K represents the number of convolution kernels and 𝑏 indicates the bias. ⨂ depicts the
convolution operation, and f (·) signifies the nonlinear activation function.
To solve the over-fitting problem, a Pooling layer is usually added behind a Convolution layer to reduce the
dimensionality of features from the Convolutional layer and the number of parameters. The commonly used pooling methods
are average pooling and maximum pooling. The formula is as follows.
1
p

𝑦 (i , j)k = [ ∑ (o(m,n)k )p ]

f
(4)

oo
(m,n)∈Ri,j

where 𝑦 (i , j)k denotes the output of the pooling operator at the position (i , j). o(m,n)k is the eigenvalue at position (𝑚,
𝑛). When 𝑝 is equal to 1, it is average pooling and when 𝑝 tends to infinity, it is maximum pooling.

r
-p
The Convolution layer and Pooling layer guarantee the translation invariance of CNN, that is, the system produces the
same response no matter how the input is translated. Convolution assures that the features are still detected while pooling
re
maintains as consistent a representation as possible. Next, two Dense layers are connected to the Pooling layer to stitch the
extracted features into a one-dimensional vector for similarity calculation. Finally, the layers listed above are summarized to
lP

constitute the CNN mapping function Gw.

The input of the Siamese-CNN[35] is a pair of samples (X1, X2). Firstly, X1 and X2 are alternately input into the above
CNN mapping function Gw(X) to extract features Gw(X1) and Gw(X2). Then, the similarity of the extracted features is calculated,
na

and D(Gw(X1), Gw(X2)) is obtained, where D(·) is a similarity measure function. The commonly adopted similarity measure
functions mainly include ED, cosine distance, and cross-correlation functions. In this paper, ED is employed to assess the
ur

similarity. The expression is as follows.

h
Jo

Ew (X1 , X2 ) = D(Gw (X1 ), Gw (X2 ))+b=√∑ (Gw (X1 )i -Gw (X2 )i )2 +b (5)
i=1

where b denotes the bias factor, h represents the number of features. Finally, ED is fed into a Dense layer to realize the
judgment of whether X1 and X2 share the homologous lifespan. The output of the last Dense layer is the label Y, where Y=1
indicates that X1 and X2 share the homologous lifespan, and Y=0 signifies that they have widely varying lifespans.
Before training the model, the hyperparameters of the network are initialized, including the depth of the network (the
number of hidden layers), the number of Convolution kernels, and the size of Convolution kernels. One of the commonly
adopted pooling methods, known as the average pooling, is selected. To avoid the problem of gradient disappearance and
improve the training rate of the network, Rectified Linear Units (ReLU) is applied as the activation function. To avoid
overfitting, a portion of neurons is randomly ignored during training using Dropout. The multivariate cross-entropy is utilized
as the loss function to evaluate the accuracy of the model. Adam optimizer is employed to train the parameters in
backpropagation to update the weights. During the training process, the hyperparameters are adjusted continuously to
optimize the performance of the model. Table 1 signifies a set of hyperparameters when the model performance is optimal.
Table 1 Optimal hyperparameters
Hyp erp arameters
Siamese Nu mb er of layers 3
2 DCNN
n etwo rk Nu mb er of Convo lu tion 32

6 / 16
 k ern els 64
128
Kernel Size 5*5
Nu mb er of layers 1
Dense
Size 64
Nu mb er of layers 2
Dense
Size 64

The data set is divided into a training set, a validation set and a testing set according to the ratio of 6:2:2. Suppose there
are m batteries in the training set, n batteries in the validation set, and t batteries in the testing set. Select battery numbered k
(traink, k ≤ m) from the training set as Sample1, and battery numbered s (trains, s ≤ m, s ≠ k) as Sample2. The lifespans of
traink and trains are Lk and Ls , respectively. The input of the Siamese-CNN is the partial voltage-current data of traink and
trains in the first 30 cycles, which is a voltage matrix constructed according to the principles in Section 2.2, with a matrix
size of 2 × 420 × 30. The output is the label of traink and trains. The AE of lifespans of the two batteries is ΔL=∣Lk - Ls∣.
ΔL is chosen as the decision to determine whether traink and trains are similar. When ΔL<y, traink, and trains are considered
to be similar, the corresponding output label is (1,0); when ΔL ≥ y, traink, and trains are considered not similar, the
corresponding output label is (0,1). Sequentially, we need to judge whether a battery in the training set is similar to other

f
oo
batteries in turn and determine the output labels. Finally, n × (n-1)/2 sets of labels are obtained. The training set and the
validation set are used to train the model.

r
The lifespan of the battery x (testx) in the testing set needs to be predicted. The input matrix is constructed based on the
-p
principles in Section 2.2 and fed into the trained-well model. The output is m+n probability values, which represent the
similarity probability between testx and each battery in the training set and verification set. The larger the probability value,
re
the more similar testx is to the battery in the training set and verification set. The battery with the highest output probability
value (traini, i is the battery number in the training set or verification set) is identified as the battery that is most similar to
lP

testx, and the lifespan of testx is the same as the lifespan of traini. Since the lifespan of traini is known, the lifespan of testx
can be obtained. Finally, the lifespan prediction of the lifespan-unknown test battery is achieved.
na

3.2 Capacity aging trajectory prediction based on CNN

The AE of the two batteries' lifespans is selected as the evaluation index to judge whether they are similar in the training
ur

process. In supervised learning, the aforementioned decision-making belongs to weak supervision, that is, it is not a completely
correct decision. The capacity aging trajectory may be quite different despite the matched battery being homologous to the
Jo

target battery in terms of lifespan. Therefore, it cannot be assumed that the capacity aging trajectory of the target battery is
identical to the capacity aging trajectory of the matched battery. In these circumstances, capacity aging trajectory prediction is
indispensable.
The battery partial discharge voltage-current data are preprocessed based on the rules in section 2.2 and fed into the CNN.
As depicted in the yellow box in Fig. 3, CNN consists of three Convolutional layers, three Pooling layers, and two Dense layers.
Similar to the first-stage Siamese-CNN, the second-stage CNN extracts features from the input data automatically which are
relevant to the capacity aging trajectory. The two Dense layers with different structures output the cycle number sequence and
capacity sequence, respectively.
It is noted that the Dense layer in the CNN determines that the output length of the model must be consistent, as illustrated
in the yellow box in Fig. 3. Assume that the range of battery life in the training set is Lmin~Lmax. In the proposed method, the
length of the output sequence corresponds to the battery with the shortest lifespan (Lmin) in the training set. However, various
lifespans of the batteries result in different lengths of cycle-capacity sequences. Uniform sampling is applied on all batteries to
compress their lengths of cycle-capacity sequences into Lmin to meet the requirement of the same output length. The ratio of
the predicted lifespan of each cell to the fixed length of the CNN output sequence is regarded as the interval for uniform
sampling.
In this model, the pooling method is average pooling. To avoid the problem of gradient disappearance and improve the
training rate of the network, ReLU is adopted as the activation function. To avoid overfitting, a portion of neurons is randomly
7 / 16
ignored during training using Dropout. MAE is utilized as the loss function to evaluate the error between the predicted value
and the real value, which is commonly employed in regression tasks. The weights are updated by training the parameters in
backpropagation using the Adam optimizer. Equally, the model also requires determining the optimal hyperparameters, and the
optimal results are illustrated in Table 2. The training set, validation set, and testing set for the capacity aging trajectory
prediction model are the same as those of the first-stage battery lifespan prediction.
Table 2 Optimal hyperparameters
Nu mb er of layers 3
32
Nu mb er of Convo lu tion
2 DCNN 64
k ern els
128
Kernel Size 5*5
Nu mb er of layers 2
Dense -cycle
Size 419
Dense - Nu mb er of layers 2
capacity Size 419

4 Results and Discussion

f
It is extremely valuable for new batteries to realize battery lifespan prediction and capacity aging trajectory prediction in

oo
the early-cycle stage using a small amount of data. There is no definitive quantitative equation to accurately describe the
relationship between lithium-ion battery lifespan and the discharge voltage-current curves. The features that can accurately

r
describe the battery cycle and capacity aging trajectory have not been extracted completely. In the proposed method, the
-p
features are extracted automatically from the voltage-current data directly by CNN instead of manually extracting features.
re
Although these features mentioned by the reviewer (such as peak height, peak area, and peak position) vary with battery aging,
it is still questionable whether these features can fully characterize the aging state and process of the battery. Meanwhile, there
lP

is no clear method to determine which manually extracted features are optimal. In this case, the extracted features are more
comprehensive, avoiding the problem of incompleteness or redundancy in manual feature extraction. In this section, the
performance and results of the proposed prediction method are analyzed.
na

4.1 Battery lifespan prediction results

Five metrics are utilized to assess the performance of the battery lifespan predictions in the early-cycle stage.
ur

The Absolute error (AE) in cycles describes the absolute deviation of the predicted value from the true value.
AE = yi - ŷi
Jo

(6)
The Relative Error (RE) indicates the percentage of absolute error in the true value.
y - ŷi
RE = i × 100% (7)
yi
The root mean square error (RMSE) in cycles is a stability index.
N
1
RMSE = √ ∑ (yI - ŷi )2 (8)
n
i=1

The mean absolute percentage error (MAPE) is the precision assessment of RE.
N
1 y - ŷi
MAPE = ∑| i | ×100% (9)
N yi
i=1

The mean absolute error (MAE) is the precision assessment of absolute error.
N
1
MAE = ∑| yI - ŷI | (10)
N
i=1

In Eqs. (6) to (10), N represents the number of samples, yi denotes the true value, and ŷi indicates the predicted value.

8 / 16
 Fig. 5 illustrates the lifespan prediction results of batteries in the testing set based on Siamese-CNN. In Fig. 5(a), the
horizontal axis is the index of the battery in the testing set, and the vertical axis is the lifespan of the battery. The blue rectangular
signifies the real lifespans of the batteries, and the red rectangular signifies the predicted lifespans. Fig. 5(b) and Fig. 5(c)
illustrate the AEs. Fig. 5(d) and Fig. 5(e) demonstrate the REs. It can be observed that the AEs of 9 batteries are less than 10
cycles, 6 batteries within 11~20 cycles, 4 batteries within 21~30 cycles, and 5 batteries within 31~40 cycles. The predicted
lifespans of some batteries in the testing set are close to the real values, and the others are almost identical. In general, all AEs
of lifespan prediction are within 40 cycles, and all REs are less than 5.62%. The MAPE, RMSE, and MAE are 2.5%, 21 cycles,
and 17 cycles. The low MAPE and MAE reflect the high accuracy of the lifespan prediction model, and the small RMSE
symbolizes the strong stability and robustness of the model. These results fully demonstrate that the proposed battery lifespan
prediction method in the early-cycle stage can automatically extract the features that dominate the battery lifespan from the
partial voltage and current data. The model has high accuracy and strong stability for the lifespan prediction of various batteries
under different profiles. It is worth noting that the capacity aging trajectory prediction method further reduces the error of
battery lifespan prediction. This will be analyzed in Section 4.2.
(a)
2500
Real lifespan

f
Predicted lifespan
2000

oo
Cycle Number

1500

1000

r
500

0
0 5 10 15 20
-p 25
re
Battery Index
lP
na
ur
Jo

Fig. 5. The results and errors of lifespan predictions based on the first-stage Siamese-CNN. (a) The battery lifespan prediction results; (b) AEs;
(c) Boxplot of AEs; (d) REs; (e) Boxplot of REs.

4.2 Capacity aging trajectory prediction results

The proposed method of capacity aging trajectory prediction is non-iterative. The capacity aging trajectory can be obtained
through one-step prediction without updating the data. The performance of the proposed method is verified using 24 batteries
with various lifespans under different profiles. Due to the limitation of paper length, the predicted and real capacity aging
trajectory of 3 batteries in the testing set are exhibited in Fig. 6. These 3 batteries are charged under 3 different charge profiles
providing different lifespans. As depicted in Fig. 6(a) to Fig. 6(c), every two figures represent the predicted results of the
capacity aging trajectory of one battery. Fig. 6(a-1), Fig. 6(b-1), and Fig. 6(c-1) signify the comparison between the real and
predicted capacity aging trajectory of these 3 batteries. The predicted capacity aging trajectories match the real capacity aging
trajectories well. Fig. 6(a-2), Fig. 6(b-2), and Fig. 6(c-2) illustrate the REs between true and predicted values of the capacity
aging trajectories of these 3 batteries. Fig. 6(d) denotes the REs of the predicted results of all test batteries’ capacity aging
trajectories. For the battery with the best result of capacity aging trajectory prediction, the RE is less than 0.5%, and for the
worst result, the RE is less than 4%. Fig. 6(e) provides the MAPEs of the predicted results of the capacity aging trajectory. The
MAPEs of 23 batteries are all less than 1%, and the MAPE of one battery is more than 1%, which is only 1.28%. In general,
the results indicate that the proposed method only with the partial voltage and current data in the first 30 cycles can accurately

9 / 16
predict the capacity aging trajectories and the predicted trajectories match well with the actual trajectories. The low MAPEs
reflect the high accuracy of the model.
It is noted that the accuracy of battery lifespan prediction based on the CNN is further improved compared with the
prediction results based on the Siamese-CNN. The AEs of lifespan predictions are within 37 cycles, and the REs of lifespan
predictions are less than 5%. The MAPE, RMSE, and MAE are 2.5%, 21 cycles, and 17 cycles, respectively.
(a-1) (b-1) (c-1)
1.1 Start of Prediction 1.1 Start of Prediction 1.1 Start of Prediction

1.05 1.05 1.05

Capacity(Ah)

Capacity(Ah)
Capacity(Ah)
Target data Target data Target data
1 Prediction 1 Prediction 1 Prediction

0.95 0.95 0.95

0.9 0.9 0.9

0 100 200 300 400 500 0 200 400 600 800 0 500 1000 1500 2000
Cycle Number Cycle Number Cycle Number
(a-2) (b-2) (c-2)
2.5
0.5
0.8 2

f
oo
0.6 1.5

RE(%)
RE(%)

RE(%)

0.4 0 1
0.2 0.5

r
-0.2
0

0 100 200 300 400 500

-0.5

0 200
-p 400 600 800
0
-0.5
0 500 1000 1500 2000
re
Cycle Number Cycle Number Cycle Number

(d)
4
lP

3
na

1
RE(%)

ur

0
Jo

-1

-2

-3

-4
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
Battery Index
(e) (f)
1.5
MAX=1.29
1.2

1 1
MAPE(%)

MAPE(%)

0.8
3=0.69
0.5 0.6

0.4 2=0.44
1=0.29
0 0.2 MIN=0.17
0 5 10 15 20 25
Battery Index Battery Index

10 / 16
Fig. 6. Capacity aging trajectory prediction results. (a-1), (b-1), and (c-1) depict the comparison between the true and predicted values of the capacity aging
trajectories; (a-2), (b-2), and (c-2) illustrate the REs between true and predicted values of the capacity aging trajectories; (d) the REs of the predicted results
of the capacity aging trajectories for all test batteries; (e) and (f) the MAPEs and their boxplot of the predicted results of the capacity aging trajectories; (g-
1), (g-2) and (g-3) the result of the final lifespan predictions.

4.3 Feature Visualization

CNN can effectively serve as an information distillation pipeline, iteratively transforming the input data to filter out

f
irrelevant information and amplify and refine the information related to the battery lifespan. Through the visualization

oo
technique, it is possible to analyze how the features are trained at each layer during the network training process. Feature
visualization refers to exhibiting the output feature maps of a given input image from each intermediate layer (including

r
Convolutional and Pooling layers, etc.). The feature visualization results can help us diagnose potential problems and improve
-p
the current network. In the visualization result, there may be some pure black feature maps, which are called dead feature
maps. These dead feature maps cannot provide effective information, and the locations of them in the fixed Convolutional
re
layer remain the same for different input data. Therefore, the corresponding Convolution kernels can be eliminated from the
network, which serves the purpose of model compression, more distinctive feature extraction results, fewer invalid features,
lP

clearer feature maps, and less blurring.

Fig. 7 denotes the visualization results of the features extracted from two input batteries with similar lifespans based on
na

the Siamese-CNN. Fig. 7(a)-(c) represent the feature visualization results of battery1. Fig. 7(d)-(f) and Fig. 7(g)-(i) indicate
the feature visualization results of battery2, which are similar and different to battery1, respectively. These heat maps indicate
ur

that the convolutional kernels in the Convolutional layer can extract features of different scales and focus from unordered
data, and these abstract features describe the hidden attributes of input data from different perspectives. Batteries with similar
Jo

lifespans share a high degree of similarity in their features, so the features in Fig. 7 exhibit little variation. Conversely, the
features extracted from two batteries with dissimilar lifespans tend to differ significantly, which is the basis for the battery
lifespan prediction. Finally, these multi-channel feature maps are weighted through a Dense layer, where each feature map
contributes to the final result. Even if the eigenvalues of features are small or similar, it aids in further improving the accuracy
of lifespan prediction.
(a) (b) (c)
0.35
0.05
0.3 0.2
0.25 0.04
0.15
0.2
0.03
0.15 0.1
0.1 0.02
0.05
0.05
0 0.01

11 / 16
 (d) (e) (f)
0.35
0.05
0.3 0.2
0.25 0.04
0.15
0.2
0.03
0.15 0.1

0.1
0.05 0.02
0.05
0 0.01

(g) (h) (i)

0.05
0.15
0.3
0.04

0.1 0.03
0.2

0.02
0.1 0.05
0.01

f
oo
0 0 0

Fig. 7. Feature maps of the Siamese-CNN. (a)-(c) Feature maps of the three Convolutional layers of battery 1;(d)-(f) feature maps of the three Convolutional
layers of battery 2(similar with battery 1); (g)-(i) feature maps of the three Convolutional layers of battery 2(different form battery 1).

r
-p
In the task of capacity aging trajectory prediction, the CNN iteratively transforms the input data to filter out irrelevant
information and amplify and refine the information related to capacity and number of cycles. Fig. 8 demonstrates the feature
re
visualization results of three Convolutional layers of a battery in the testing set. There are a large amount of dark blue low
values in Fig. 8(a)-(c). In the blue box of Fig. 3, the first column of the model input data matrix is the number of cycles, and
lP

the rest of the columns are filled with cumulative capacity data. The number of cycle values is much larger than the capacity
data of the batteries used in this paper, which makes the cycle number features with large values (yellow points in Fig. 8) and
many capacity features with small values (dark blue points in Fig. 8) displayed in one CNN feature map. Finally, the accurate
na

capacity aging trajectories are obtained based on these minor feature differences by convolution and pooling operations,
which further demonstrates the effectiveness and robustness of CNN in extracting features.
ur

(a) (b) (c)

8 15
Jo

40
6
10 30
4
20
5
2
10

0 0

Fig. 8 Feature maps of three Convolutional layers of the CNN.

4.4 Comparisons
In recent studies, many methods analyzed the dataset in this paper to achieve battery lifespan prediction. In this section,
four papers are selected for comparative analysis, and the results are depicted in Table 3.
Table 3 Prediction results in comparison with other papers

Input data Lifespan Trajectory

MAE RMSE AE RE
Models Input-cycles Input-signal MAPE(%) MAPE(%)
(cycles) (cycles) (cycles) (%)

20 V-I 19 24 3.1 -42~46 -5.91~7.22 1.4

Siamese-CNN 30 V-I 17 21 2.5 -37~37 -5~5 1.28
40 V-I 15 19 2.29 -37~35 -4.99~4.8 0.87

12 / 16
 CNN-LSTM 30 V-I 113 139 17.97 -270~282 -48.42~48.14 17.97
70 18 25 3.02 -27~84 -2.65~9.85
TOP-NET[17] V-I-T ━
80 19 25 3.43 -37-53 1.96~10.88
Full[28] 100 V-I N/A 118 14.1 -540~490 N/A ━
RF-ABC-RNN[36] 100 V-I- -R N/A 80 9.8 -200~100 N/A ━
DNN[37] 1 V-I- -T N/A 57 6.46 N/A N/A ━
MOGP[38] 100 N/A 134 7.87 N/A N/A 6.53
CNN[39] 100 ━ 2.65

It is a challenging issue to effectively predict battery life using limited battery test data before battery deployment. If too
much cycle data is required, the practicality of methods is questionable even if they can improve prediction accuracy. This is
because the battery with extensive cycles before deployment will cause aging, and its performance will decline compared to
a new battery, which is unacceptable to manufacturers and consumers. Therefore, it is necessary to find a balance between
prediction accuracy and practicality to achieve high-accuracy battery life prediction without performance decline due to over
testing. Based on the analysis, we conducted a comparison to study the accuracy of battery life prediction under different

f
oo
numbers of cycles as inputs and determined the most suitable cycle number.
The first 20, 30, and 40 cycles of discharge data are fed into the established model to predict the battery lifespan and
capacity degradation trajectory, respectively. The prediction results are shown in Table 3. When using the first 20 cycles of

r
-p
data as input, the maximum AE value increased from 5% in the first 30 cycles to 7.22%, even though the maximum AE value
increased by only 8 cycles. The result indicates that the model’s prediction accuracy for the same batteries decreases in the
re
first 20 cycles. When using the first 40 cycles of data as input, the model’s prediction accuracy is the highest in three cases.
However, compared with the first 30 cycles, the improvement in prediction accuracy is not significant. Ten more cycles are
lP

required, which may cause a certain degradation of the battery performance. The slight improvement in accuracy brought by
the increased cycles can be ignored compared to the large economic and time costs it brings. Considering comprehensively,
na

we choose the first 30 cycles of data as input, which is reasonable.

To further highlight the superiority of the proposed network structure in lifespan prediction and capacity aging
trajectory prediction, we input the same data into the network of CNN+LSTM, which is commonly used in lifetime
ur

prediction, and analyze the prediction results comparatively analyze the prediction results in comparison with the
proposed Siamese-CNN. It is evident from Table 3 that the CNN-LSTM's metrics are worse than the Siamese-
Jo

CNN's in both the battery lifespan prediction results and the capacity aging trajectory prediction results, which
further proves the superiority of the proposed Siamese-CNN in lifespan and capacity aging trajectory prediction.
Chen[17] achieved RUL prediction using voltage-current-temperature data in the first 50 cycles with an AE range of -
32~61 cycles. The input physical quantities include not only voltage and current but also temperature. Severson, Kristen
A[28] employed the voltage-current data in the first 100 cycles to realize the lifespan prediction. Although the AE is the
smallest among the four methods (-18~22 cycles), the method entails too much data and manual feature extraction. Most
importantly, capacity aging trajectory prediction is not achieved. Zhang[36] utilized voltage-current-capacity-internal
resistance data in the first 100 cycles to achieve RUL prediction with an AE range of -200~100 cycles. The AE of prediction
is the largest. Too many physical quantities are required, and the internal resistance needs to be obtained through additional
test experiments. Besides, it is also necessary to extract features manually. Hsu[37] developed a novel-architecture DNN with
a special Convolutional training strategy to predict the residual life of a battery to a MAPE of 6.46%, using voltage-current-
capacity-temperature data in only one cycle. Although its input data is the least among the listed papers, its MAPE reaches
6.46% and RMSE is 57 cycles, which is twice the errors of our proposed method. For the battery lifespans in a range of a
few hundred cycles, the prediction accuracy is too poor to meet the requirement of accurate prediction.
Chen[38] established an aging end-point early predictor of Li-ion batteries based on the multi-output Gaussian process
(MOGP). The model’s input is the first 100 cycles of data and the MAPE reaches 7.87%. Meanwhile, a degradation trajectory
predictor was constructed based on the ‘prompt-learning’ paradigm neural network, and the predicted life of the MOGP was
13 / 16
used as the prompt input the MAPE is 6.53%. Lin[39] proposed a transferred CNN-based strategy to predict the battery aging
trajectories using the capacity of the first 100 cycles and the MAPE is within 2.65%. The two studies achieve capacity aging
trajectory prediction, but the prediction accuracy is lower than the method proposed in this paper and requires more input
data.
Compared with the above methods, the method proposed in this paper has the following advantages:
(1) Early battery lifespan prediction can be achieved by partial voltage and current data in the first dozens of cycles of
each battery without any additional data such as capacity, temperature, internal resistance, etc. Hence, data collection and
storage can be reduced in practice.
(2) Both capacity aging trajectory prediction and battery lifespan prediction are achieved, while other papers only achieve
battery lifespan prediction. It can be adopted not only for the widespread screening of batteries but also for accurate tracking
of degradation.
(3) Unlike the existing method of manually selected features, the proposed method extracts features automatically and
directly from the voltage-current base on CNN. The extracted features are more comprehensive, avoiding the problem of
incompleteness or redundancy in manual feature extraction.

f
4.5 Transfer learning for different battery

oo
In practice, it is time-consuming and expensive to develop a new life prediction model using different datasets. Acquiring
a large amount of battery aging data requires a significant amount of time and financial resources. It is meaningful if the

r
developed prediction model can be migrated to different datasets with a small amount of test data through transfer learning
(TL)[40]. As an optimization method that can be accomplished through parameter fine-tuning, TL applies the knowledge
-p
learned from the source domain to the target domain, significantly reducing the data requirements. The purpose of fine-tuning
re
is to adjust the parameters of the pre-trained network using a small amount of target samples. A Lithium Cobalt Oxide (LCO)
battery dataset with a rated capacity of 5Ah is chosen as the target sample for TL. Due to the small size of the LCO dataset,
lP

TL is implemented by freezing the convolutional layers and training the fully connected layers to fine-tune the model
parameters. Fig. 9 depicts the results of LCO battery lifespan prediction based on TL. It can be observed that the AEs are
na

within 25 cycles, and all REs are less than 5%. Fig. 10 depicts the results of capacity aging trajectory prediction based on TL.
The predicted trajectories match well with the actual trajectories. And all REs are less than 2%. The results demonstrate that
ur

the developed model exhibits good transferability.

(a) (b) (c)
Jo

700 5
Real lifespan
Predicted lifespan 20
600
AE(Cycle)
Cycle Number

500 10
RE(%)

400 0
0
300
200
-10

100 -20 -5
0 0 2 4 6
1 2 3 4 5 0 2 4 6
Battery Index Battery Index Battery Index

Fig. 9. The results and errors of lifespan predictions based on TL. (a) The battery lifespan prediction results; (b) AEs; (c) REs.

(a)
5 Start of Prediction

4.8
Capacity(Ah)

4.6
Target data
4.4 Prediction

4.2

4
0 200 400 600
Cycle Number

14 / 16
 Fig. 10. Capacity aging trajectory prediction results based on TL. (a) depict the comparison between the true and predicted values of the capacity aging
trajectories; (b) illustrate the RE; (c) the boxplot of REs of all batteries.

5 Conclusions
In this paper, a novel two-stage prediction method for capacity aging trajectories based on Siamese-CNN is proposed.

f
Siamese-CNN is employed for the lifespan prediction of lithium-ion batteries. Then, taking the predicted lifespan as its prior

oo
information, CNN is employed to predict the capacity aging trajectories of batteries under different profiles. Only partial
voltage and current data of each cycle are needed, not depending on the complete charge and discharge curve and any data

r
after the first 30 cycles. The accuracy of the proposed method is validated by a well-known dataset with widely varying
-p
lifespans under 72 profiles. According to the results, the battery lifespan prediction based on Siamese-CNN achieves a MAPE
of 2.44%, and the capacity aging trajectory prediction based on CNN achieves a MAPE of 1.28%.
re
The proposed method demands no additional physical quantities as inputs besides partial voltage and current data, without
consuming excessive time and resources on extensive cycle testing, which is of greater value in the rapid assessment of
lP

battery capacity aging trajectory under different profiles before practical application. Compared with the existing methods
on the same dataset, the developed model not only diminishes the cost of battery tests but also improves the accuracy of the
na

capacity aging trajectory.

Data availability
ur

The datasets used in this study are available at https://data.matr.io/1.

Jo

Acknowledgements
This work was supported by the Fundamental Research Funds for the Central Universities(Grant No. 2022JBZY040),
and the National Outstanding Youth Science Fund Project of the National Natural Science Foundation of China (Grant No.
52222708).
References
[1] Xiong R, Zhang Y, He H, Zhou X, Pecht MG. A double-scale, particle-filtering, energy state prediction algorithm for lithium-ion batteries. IEEE
Trans Ind Electron 2017;65:1526–38.
[2] Yang S, Zhang C, Jiang J, Zhang W, Chen H, Jiang Y, et al. Fast screening of lithium-ion batteries for second use with pack-level testing and machine
learning. eTransportation 2023;17:100255.
[3] Qian Ch, Xu B, Chang L. Convolutional neural network based capacity estimation using random segments of the charging curves for lithium-ion
batteries. Energy 2021; 227:120333.
[4] Li S, Zhang C, Du J, Cong X, Zhang L, Jiang Y, et al. Fault diagnosis for lithium-ion batteries in electric vehicles based on signal decomposition and
two-dimensional feature clustering. Green Energy and Intelligent Transportation 2022;1(1):100009.
[5] Kong J, Yang F, Zhang X, Pan E, Peng Z, Wang D. P Voltage-temperature health feature extraction to improve prognostics and health management
of lithium-ion batteries. Energy 2021;223:120114.
[6] Zhang Y, Li Y. Prognostics and health management of Lithium-ion battery using deep learning methods: A review. Renewable and Sustainable
Energy Reviews 2022;161:112282
[7] Saha B, Goebel K, Poll S, Christophersen J. Prognostics methods for battery health monitoring using a bayesian framework. IEEE Transactions on
Instrumentation and Measurement 2009;58:291–96.
[8] Richardson RR, Osborne MA, Howey DA. Battery health prediction under generalized conditions using a Gaussian process transition model. Journal
of Energy Storage 2019;23:320–28.
[9] Liu X, Zhang X, Chen X, Zhu G, Yan Ch, Huang J, et al. A generalizable, data-driven online approach to forecast capacity degradation trajectory of
lithium batteries. Journal of Energy Chemistry 2022;68:548-55.
[10] Zhang Q, Yang L, Guo W, Qiang J, Peng CH, Li Q, et al. A deep learning method for lithium-ion battery remaining useful life prediction based on
sparse segment data via cloud computing system. Energy 2022;241:122716.
15 / 16
[11] Li, Y, Liu K, Foley AM, Zülke A, Berecibar M, Nanini-Maury E, et al. Data-driven health estimation and lifetime prediction of lithium-ion batteries:
A review. Renewable and Sustainable Energy Reviews 2019;113:109254.
[12] Bhaskar S, Kai G, Jon Chn. Comparison of Prognostic Algorithms for Estimating Remaining Useful Life of Batteries. Transactions of the Institute of
Measurement and Control 2009; 31(3-4):293-308.
[13] Li w, Cao D, Jost D, Ringbeck F, Kuipers M, Frie F,et al. Parameter sensitivity analysis of electrochemical model-based battery management systems
for lithium-ion batteries. Applied Energy 2020;269:115104.
[14] Hu X, Li S, Peng H. A comparative study of equivalent circuit models for Li-ion batteries. J Power Sources 2012;198:359–67.
[15] Yang, Y, Hu X, Qing D, Chen F. Arrhenius Equation-Based Cell-HEalth Assessment: Application to Thermal Energy Management Design of a HEV
NiMH Battery Pack. Energies 2013;196(6): 2709-25.
[16] Zhang Y, Xiong R, He H. Lithium-ion battery remaining useful life prediction with Box-Cox transformation and Monte Carlo simulation. IEEE
Transactions on Industrial Electronics 2019;66:1585-97.
[17] Chen L, Ding Y, Liu B, Wu Sh, Wang Y, Pan H. Remaining useful life prediction of lithium-ion battery using a novel particle filter framework with
grey neural network. Energy 2022;244:122584.
[18] Cong X, Zhang C, Jiang J, Zhang W, Jiang Y. A Hybrid Method for the Prediction of the Remaining Useful Life of Lithium-Ion Batteries With
Accelerated Capacity Degradation. IEEE transactions on vehicular technology 2020;69: 12775-85.
[19] Chen D, Zhang W, Zhang C, Sun B, Cong X, Wei Sh,et al. A novel deep learning-based life prediction method for lithium-ion batteries with strong
generalization capability under multiple cycle profiles. Applied Energy 2022;327:120114.
[20] Hong J, Lee D, Jeong E, Yi Y. Towards the swift prediction of the remaining useful life of lithium-ion batteries with end-to-end deep learning.
Applied Energy 2020;278:115646.
[21] Yang H, Wang P, An Y, Shi C, Sun X, Wang K,et al. Remaining useful life prediction based on denoising technique and deep neural network for
lithium-ion capacitors. eTransportation 2020;5:100078.
[22] Ding P, Liu X, Li H, Huang Z, Zhang K, Shao L, et al. Useful life prediction based on wavelet packet decomposition and two-dimensional
convolutional neural network for lithium-ion batteries. Renewable and Sustainable Energy Reviews 2021;148:111287.
[23] Li W, Sengupta N, Dechent P, Howey D, Annaswamy A, Sauer DU. One-shot battery degradation trajectory prediction with deep learning. J Power
Sources 2021;506: 230024.

f
[24] Saxena S, Ward L, Kubal J, Lu W, Babinec S. A convolutional neural network model for battery capacity fade curve prediction using early life data.

oo
J Power Sources 2022;542:231736.
[25] Zhang W, Li X, Li X. Deep learning-based prognostic approach for lithiumion batteries with adaptive time-series prediction and on-line validation.
Measurement 2020;164:108052.
[26] Xu, T, Peng Zh, Wu L. A novel data-driven method for predicting the circulating capacity of lithium-ion battery under random variable current.

r
Energy 2021;218:119530.

1521-34.
-p
[27] Tian J, Xiong R, Shen W, Lu J, Yang X. Deep neural network battery charging curve prediction using 30 points collected in 10 min. Joule 2021;5:

[28] Severson KA, Attia PM, Jin N, Perkins N, Jiang B, Yang Z, et al. Data-driven prediction of battery cycle life before capacity degradation. Nature
re
Energy 2019;4:383-91.
[29] Gao Y, Yang S, Jiang J, Zhang C, Zhang W, Zhou X, et al. The Mechanism and Characterization of Accelerated Capacity Deterioration for Lithium-
Ion Battery with Li(NiMnCo) O2 Cathode. Journal of the Electrochemical Society 2019;166: 1623-35.
lP

[30] Dubarry M, Svoboda V, Hwu R, Yann LB. Incremental Capacity Analysis and Close-to-Equilibrium OCV Measurements to Quantify Capacity Fade
in Commercial Rechargeable Lithium Batteries. Electrochemical and Solid-State Letters 2006; 9(10):454.
[31] Dubarry M, Liaw BY. Identify capacity fading mechanism in a commercial LiFePO4 cell. J Power Sources 2009;194(1):541-49.
[32] Liu B, Yu X, Zhang P, Yu A, Fu Q, Wei X. Supervised Deep Feature Extraction for Hyperspectral Image Classification. IEEE Transactions on
na

Geoscience and Remote Sensing 2018;56:1909-21.

[33] Huang L, Chen Y. Dual-Path Siamese CNN for Hyperspectral Image Classification With Limited Training Samples. IEEE Geoscience and Remote
Sensing Letters 2021;18:518-22.
[34] Xie Q, Liu R, Huang J,S Jianhui. Residual life prediction of lithium-ion batteries based on data preprocessing and a priori knowledge-assisted CNN-
ur

LSTM. Energy 2023;281:128232.

[35] Sun X, Han G, Guo L, Yang H, Wu X, Li Q. Two-stage aware attentional Siamese network for visual tracking. Pattern Recognition 2022;124:108502.
[36] Zhang Y, Peng Zh, Guan Y, Wu L. Prognostics of battery cycle life in the early-cycle stage based on hybrid model. Energy 2021;221:119901.
[37] Hsu C, Xiong R, Chen N, Li J, Tsou N. Deep neural network battery life and voltage prediction by using data of one cycle only. Applied Energy
Jo

2022;306:118134.
[38] Chen Z, Chen L, Ma Z, Xu K, Zhou Y, Shen W. Joint modeling for early predictions of Li-ion battery cycle life and degradation trajectory. Energy
2023;277:127633.
[39] Lin M, You Y, Meng J, Wang W, Wu J, Stroe D.I. Lithium-ion battery degradation trajectory early prediction with synthetic dataset and deep learning.
Journal of Energy Chemistry 2023;85:534-546.
[40] Ma G, Xu S, Jiang B, Cheng C, Yang X, Shen Y, et al. Real-time personalized health status prediction of lithium-ion batteries using deep transfer
learning. Energy & Environmental Science 2022;15:4083-4094.

16 / 16
HIGHLIGHTS
1. A Siamese-CNN is employed for battery lifespan prediction and capacity aging
trajectory.
2. Only partial voltage-current data from the first 30 cycles are required.
3. Achieved 2.44% test error for battery lifespan prediction and 1.28% for capacity
aging trajectory prediction.

f
r oo
-p
re
lP
na
ur
Jo
Declaration of interests

☒ The authors declare that they have no known competing financial interests or personal relationships
that could have appeared to influence the work reported in this paper.

☐ The authors declare the following financial interests/personal relationships which may be considered
as potential competing interests:

of
ro
-p
re
lP
na
ur
Jo