Electrical Power and Energy Systems 156 (2024) 109764

Contents lists available at ScienceDirect

International Journal of Electrical Power and Energy Systems

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

A double broad learning approach based on variational modal

decomposition for Lithium-Ion battery prognostics
Xiaojia Wang a, b, *, Xinyue Guo a, b, Sheng Xu a, Xibin Zhao c
a
School of Management, Hefei University of Technology, Hefei 230009, China
b
Key Laboratory of Computer Network Research (Hefei University of Technology), Ministry of Education, Hefei 230009, China
c
School of Software, Tsinghua University, Beijing 100084, China

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

Keywords: Predicting the remaining life of lithium-ion battery equipment is becoming increasingly important as enterprises
Lithium-ion battery transition to smart manufacturing. Accurate prediction results can be used to effectively determine the battery’s
Remaining life prediction health status and improve operational safety. However, during the decline process, lithium-ion battery capacity
Capacity regeneration
regeneration occurs, resulting in significant fluctuations in the degradation data that can easily lead to insuffi­
Variational modal decomposition
Double broad learning
cient prediction accuracies. At the same time, a factor influencing the prediction results is the unification of
modal information and insufficient feature extraction of the battery capacity data in the prediction process.
Therefore, in this paper, a novel model based on variational modal decomposition and double broad learning
(VMD-DBL) is proposed. First, we use VMD to perform adaptive decomposition of the degraded data to form
intrinsic mode function (IMF) components and residual components to solve the data noise problem. Second,
these two modal data of the feature extraction and modal fusion are inputted into the trained DBL model. Finally,
the two modes are connected to the output layer to obtain the predicted result. The NASA dataset is used for
experimental validation in this paper, and the results show that our proposed method outperforms other methods
in terms of accuracy and feasibility.

increasingly essential in the field of battery development as a result of

1. Introduction the transition to smart manufacturing. One of the most popular and
frequently utilized modern energy sources is the lithium-ion battery.
Manufacturing is the core industry on which country depends, and it Lithium-ion batteries have a high energy density and a long cycle life
directly reflects the country’s productivity level. Operating machinery and cause little environmental contamination. However, the capacities
failure in the manufacturing process can result in increased of lithium-ion batteries will decrease and their performance will be
manufacturing costs, safety incidents, or even the complete breakdown diminished as the number of charge/discharge cycles rises. Moreover,
of the production process. Prediction and health management (PHM) failure to replace aging and failing batteries in time may result in sig­
can efficiently assess the condition of industrial machinery and improve nificant safety mishaps. To increase the reliability of battery operation,
the safety of machinery operations [1]. PHM is therefore essential for precise battery RUL prediction is important.
research on production processes. One of PHM’s key technologies and Model-based methods and data-driven methods are currently the two
fundamental issues is the accurate prediction of the remaining useful primary categories of prediction techniques [3]. To study the battery
life, which is a subject of intense study. Generally, the time period or performance degradation process, model-based prediction approaches
quantity of operations between the present instant and the moment of often rely on simulating the material properties, operating principles,
failure of a system or piece of equipment is known as the remaining electrochemical reactions, and other aspects of the battery [4]. Ac­
useful life (RUL), and it can be defined as [2]: cording to the differences in the modeling, two types of models can be
classified: those based on empirical degradation and those based on
T − t|T > t (1) degradation mechanisms. The battery empirical degradation model as­
where t represents the present instant and T is the device’s failure sumes that the battery’s capacity decays according to some innate
moment. The ability to predict the remaining life is becoming mathematical relationship. The degradation mechanism model analyses

* Corresponding author at: School of Management, Hefei University of Technology, Hefei 230009, China.
E-mail addresses: [email protected] (X. Wang), [email protected] (X. Guo), [email protected] (S. Xu), [email protected] (X. Zhao).

https://doi.org/10.1016/j.ijepes.2023.109764
Received 30 May 2023; Received in revised form 13 November 2023; Accepted 28 December 2023
Available online 12 January 2024
0142-0615/© 2023 The Author(s). Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-
nc-nd/4.0/).
X. Wang et al. International Journal of Electrical Power and Energy Systems 156 (2024) 109764

combined empirical knowledge elements with the battery storage tem­

Nomenclature perature, state of charge (SOC), and time mechanisms. It created a data-
driven model for the RUL prediction of batteries based on support vector
A radius of regression (SVR). To improve the accuracy and stability of the RUL of
B position of batteries, Wang et al. [16] used an artificial bee colony algorithm (ABC)
C further nomenclature continues down the page inside for SVR optimization. Ng et al. [17] proposed a naive Bayes (NB) model
the text box for the RUL prediction of batteries under various operating conditions
and compared it to the SVM. According to the experimental results, the
NB model has a more stable and accurate prediction performance. Such
methods excel at problems such as small sample data prediction, but
the change pattern of the battery performance throughout battery they significantly increase computational time and complexity when
operation using the essential mechanism of the lithium-ion battery’s dealing with large sample problems. Neural network methods are now
internal reaction. The battery empirical degradation model typically widely used in predicting the RUL of batteries. Based on an inverse
calls for the use of various functional forms to fit the battery’s capacity tangent function adaptive genetic algorithm combined with a back
decay trajectory, and the function with the best fit is chosen as the life propagation neural network (ATAGA-BP), Li et al. [18] proposed an
empirical model. Li et al. [5] use experimental test and the 2nd equi­ indirect RUL prediction method for online electric vehicle lithium-ion
valent circuit model to study the fading characteristic of the lithium-ion batteries. To a certain extent, BP neural networks can estimate how
battery, and provide a reasonable strategy to find the method that can long a device will last, but they are prone to hitting local optimums, and
evaluate and predict the battery cycle life. Bian et al. [6] proposed an their learning curves are slow to develop. To create a state of charge
open-circuit voltage model and Incremental Capacity Analysis (ICA) (SOC) estimate model for lithium-ion batteries, Lipu et al. [19] proposed
combination for the state of health (SOH) estimation of lithium batte­ an improved extreme learning machine (ELM) algorithm and enhanced
ries. The experimental results show that the method can perform accu­ ELM by searching for optimal value hidden layer neurons. The results
rate SOH estimation for different types of lithium batteries under partial from the experiments show how superior the suggested model is.
charging conditions. Based on the development mechanism of an Traditional machine learning, on the other hand, relies on specialized
interfacial film of a solid electrolyte, Xu et al. [7] developed a capacity knowledge and signal processing methods that, in the face of a deluge of
declination model that takes into account stress parameters such as the monitoring data, are challenging to analyze automatically.
average state of charge (SOC) and depth of discharge (DOD) to forecast Deep learning has powerful feature extraction capabilities.
the RUL. In addition, when the battery operating circumstances are Compared to machine learning, deep learning enables automatic feature
extreme, the fixed formula is difficult to adjust to forecast the RUL for extraction and unsupervised end-to-end learning. Li et al. [20] used a
the entire battery life cycle [8]. The empirical model ignores the bat­ one-dimensional convolutional neural network (CNN) to predict the
tery’s internal mechanism and predicts the RUL simply by developing a capacities of lithium-ion batteries. The convolutional kernel in CNN
mapping relationship between the historical data and battery health enables local weight sharing and the layer-by-layer mining of common
indicators (capacity and internal resistance) [9].Wang et al. [10] pro­ patterns among adjacent data, which is more practical. The results show
posed a novel feedback correction-adaptive Kalman filtering (FC-AKF) that the performance of the CNN algorithm exceeds that of other ma­
method is proposed for the online battery state co-prediction, which is chine learning algorithms. However, the CNN has limitations due to its
adaptive to the whole-life-cycle of the lithium-ion battery. inability to grasp time-series data. A recurrent neural network (RNN)
The degradation mechanism model evaluates the effect of the per­ permits neurons in the hidden layer to communicate with each other and
formance degradation on the battery’s internal and exterior state factors store the previous output result as information in the hidden layer,
before developing a battery degradation model. Wang et al. [11] resulting in coherent information. Catelani et al. [21] integrated
developed a state space model of the battery capacity and solved it using recurrent neural network (RNN)-based artificial intelligence estimation
a spherical cube particle filter. To establish the RUL of the battery, the methods with state-space estimation techniques to improve the accuracy
state space model can be extrapolated to a chosen failure threshold. To and precision of RUL estimation for lithium-ion batteries. RNN performs
anticipate the remaining life of a lithium battery, Xue et al. [12] pro­ well; however, it is susceptible to gradient disappearance or gradient
posed a data model-driven technique that combines a particle filter with explosion and cannot solve the problem of long-time dependence. Long
a genetic algorithm. Yang et al. [13] conducted SOC estimation exper­ Short Term Memory (LSTM) can solve the long-term dependency
iments on lithium batteries based on a simplified fractional impedance problem that is common in general recurrent neural networks. Infor­
model defined by Grunwald-Letnikov. The model-based technique uses mation in long time sequences can be efficiently delivered and expressed
expert knowledge for simulation modeling with a limited number of using LSTM. At the same time, LSTM can also solve the gradient ex­
parameter variables, which improves prediction capabilities. However, plosion problem in RNNs. KARA et al. [22] proposed a Convolutional
it is limited to the device’s dynamic data, is dependent on the model Neural Networks-Long Short Term Memory-Particle Swarm Optimiza­
selection and accuracy, has a low generalization ability and very com­ tion (CNN-LSTM-PSO) model for predicting the RUL of lithium-ion
plex model parameters and is computationally costly. batteries. The experimental results demonstrated that the method out­
Data-driven strategies for predicting the remaining life have received performed other relevant methods. Shi et al. [22] proposed a physical
much attention as artificial intelligence technology has advanced. This information long- and short-term memory (PI-LSTM) model to accu­
method is based on a very large amount of test data or sensor data rately simulate the lithium-ion battery degrading behavior and predic­
collected without requiring the device degradation principle analysis. It tion of RUL under various operating situations. While GRU simplifies the
establishes a direct link between the external parameters and interior LSTM, the computational process is more efficient. Xiao et al. [24]
states, resulting in accurate predictions. In the RUL prediction, the data- introduced a deep cyclic kernel deep learning technique that fully en­
driven approach overcomes the shortcomings of the model-based capsulates the GRU structure to predict the SOC of lithium-ion batteries.
approach. Data-driven techniques are currently classified into two The experimental results demonstrate that the method captures the or­
types: machine learning (ML) and deep learning (DL). ML-based dered matter and cyclic structure in sequential data and improves the
methods use collected data to continuously train the model to improve prediction accuracy.
its generalization capability and overcome the limitations of different The above methods are representative and are widely utilized in the
state data in degraded models. Patil et al. [14] proposed a classification field of remaining life prediction. However, there are still some issues
regression model based on a support vector machine (SVM) to achieve that need to be considered:
real-time RUL estimation for lithium-ion batteries. Liu et al. [15]

2
X. Wang et al. International Journal of Electrical Power and Energy Systems 156 (2024) 109764

• It is difficult for a single modal feature to capture the battery low accuracy and weak robustness of the single modal prediction are
degradation trend, and its robustness is low. For example, when addressed to enhance the prediction performance. In conclusion, the
using lithium batteries, the capacity regenerates, which means that following are the research contributions of this subject:
the capacity occasionally displays a modest rebound. Because of the
substantial variations in the degradation data caused by this phe­ • Because of the capacity regeneration problem of lithium-ion batte­
nomenon, it is difficult to solve the problem of the low prediction ries, degradation data can suddenly and dramatically increase,
accuracy with a single modal feature. creating anomalies. Additionally, the adaptive decomposition pro­
• The above methods are computationally demanding, costly, and cess of variational modal decomposition (VMD) has good robustness,
inflexible enough to address the problem. which can effectively circumvent data anomalies and modal con­
• Modalities are frequently associated with one another, and the above founding problems.
methods are more difficult to extract and fuse in the feature and data • BLS has the advantage of incremental learning. When new nodes are
layers. added to the network structure, it allows updating the network
weights with little computing overhead, saving computational time
The above content shows that multimodal data, as opposed to single- and cost.
modal data, can be used to obtain the strengths of different modalities to • DBL can process the information of both modes and effectively fuse
integrate information. It can also improve the data integrity and pre­ multimodal data. It can extract features better, thus improving the
diction accuracy. The fused information must retain its diversity and accuracy of the remaining life prediction of Li-ion batteries.
integrity while also maximizing the benefits of each modality and
minimizing information loss. The remainder of this work is structured as follows. In Section II, the
Decomposing and denoising the data information and fusing the fundamentals required for the model’s construction are introduced. In
information of each mode for output has become a new challenge in Section III, the model for predicting the remaining life based on the
predicting the remaining life. Chen et al. [25] propose a new procedure combination of variational modal decomposition and double-width
for SOH estimation of lithium-ion batteries by fusing current, voltage, learning is presented. In Section IV, the experimental data and metric
temperature, etc. parameters using vision transformer networks (VITs). design required for RUL predictions, as well as how the NASA lithium-
Liu et al.[26] proposed a hybrid model with deep belief network (DBN) ion battery dataset is used to validate model performance compared to
and the Kalman filter (KF) and use the battery electric current, terminal alternative methods, are discussed. Finally, in Section V, the experi­
voltage and temperature are used as the input of the proposed model. mental conclusions are provided, and potential future studies are
With the powerful nonlinear fitting capability of the DBN, the model can discussed.
extract relationship between the measurable parameters and battery
SOC. Yu et al. [27] proposed using empirical modal decomposition 2. Preliminaries
(EMD) to separate the main trend and fluctuation of the capacity decay
and then built RUL prediction models for the main trend and fluctuation 2.1. Variational modal decomposition (VMD)
separately. Chen et al. [28] investigated and developed an EMD, gray
relational analysis (GRA), and RNN hybrid data science model for The variational modal decomposition algorithm [34] is an adaptive,
lithium-ion battery RUL prediction. Zhang et al. [29] proposed a novel quasi-orthogonal, and completely non-recursive signal preprocessing
hybrid method for predicting future battery capacity and RUL by algorithm that converts the signal decomposition problem into an
combining improved variational modal decomposition (VMD), particle optimal solution problem for unconstrained problems [35]. This method
filtering (PF), and Gaussian process regression (GPR). has the following benefits: 1) it can determine the number of decom­
Because of the large number of parameters to be optimized, deep posed modes adaptively based on the actual environment; 2) it de­
learning typically requires a significant amount of time and machine modulates the IMF to the corresponding fundamental frequency band so
resources for optimization. In contrast, a breadth learning (BLS) method that the finite bandwidth and optimal center frequency of each mode are
proposed by Philip Chen [30] that acquires data features through feature adaptively matched in the search element seeking process; and 3) it can
layers can effectively and efficiently update the system for newly added effectively divide the frequency domain characteristics of the signal,
data. It can achieve comparable performance while avoiding deep adaptively separate the intrinsic mode fraction (IMF) and then obtain
learning requirements. In comparison to deep learning, BLS has a the optimal center frequency. The power failure effect and modal mixing
simpler structure, faster training speed, and higher accuracy, as well as issues can be resolved using VMD in comparison to the empirical mode
the benefit of incremental learning. BLS is now used in a wide range of decomposition (EMD) method. It can also reduce the non-smoothness of
fields. Moreover, Chen et al. [31] demonstrated the universal approxi­ the time series with a high complexity and strong nonlinearity and
mation capability of BLS, with this method producing better results in a enhance the generalization capabilities of the overall prediction model.
variety of applications. Meanwhile, some progress has been made in Hence, the nonstationary battery capacity series data in this paper are
multimodal data processing by combining the improved BLS method decomposed using VMD in the manner described below.
with other algorithms. Wang et al. [32] fused two sets of isomorphic
multimodal material data for feature fusion, then used width learning of (1) A constrained variational model is constructed. The original
cascaded feature nodes for feature learning, and finally obtained mate­ signal C is decomposed into K components, and its corresponding
rial classification results. This method shortens the training time while constrained variational function is:
improving the classification performance. Jia et al. [33] proposed a BLS- { }
based fusion feature learning framework and validated its stability and ∑k ⃦ ⃦
min ⃦∂t [(δ(t) + j/πt ) ∗ uk (t) ]e− j ωt ⃦2 (2)
speed on the Cornell University crawl dataset and the University of {uk },{ωk }
k=1
k 2

Washington RGB-D dataset.

The issue of lithium-ion battery’s remaining life forecasting will be ∑
k

addressed, and an extensive investigation will be performed in light of s.t. uk = C (3)

the aforementioned analysis. This study offers a residual life prediction k=1

approach based on the combination of variational modal decomposition where uk is the kth IMF component obtained from the decomposi­
and double-width learning, taking into account the drawbacks of tion; {ωk } = {ω1 , ω2 , ..., ωk } is the central spectrum of each mode; ∂t is
existing methods with regard to the robustness and forecast accuracy, as the derivative function of t; δ(t) is the shock function; and * is the sign of
well as the problem of capacity regeneration during battery decline. The the convolution calculation.

3
X. Wang et al. International Journal of Electrical Power and Energy Systems 156 (2024) 109764

Fig. 1. The structure of the BLS.

(2) The Lagrange multiplier operator λ and penalty factor are intro­ 2.2. Broad learning system (BLS)
duced to change the variational problem with constraints into a
variational problem with unconstrained conditions. The The broad learning system (BLS) is an incremental learning network
augmented Lagrange expression is as follows. improved on the basis of Prof. Jun-Long Chen’s [30] RVFL planar
network, and its network model structure is shown in Fig. 1 BLS has four
∑
k ⃦ ⃦2
L({uk }, {ωk }, λ ) = α ⃦∂t [(δ(t) + j/πt ) ∗ uk (t) ]e− jωtk ⃦ basic network structures: 1) input, 2) feature layer, 3) reinforcement
layer, and 4) output. In contrast to RVLF, BLS first transforms the input
2
k=1
⃦ ⃦2 〈 〉 (4) data into a feature layer via linear mapping and then uses a nonlinear
⃦ ∑k ⃦ ∑k
⃦ ⃦ mapping activation function to form the reinforcement layer. Finally,
+⃦C(t) − uk (t) ⃦ + λ(t), C(t) − uk (t)
⃦ ⃦
k=1 2 k=1 the feature and enhancement layers are connected as the input and
output layers, respectively, and the network weights of the output layer
{ }{ }{ } are obtained directly by the pseudoinverse method of ridge regression.
(3) Repeated iterations are performed to initialize u1k , ω1k , λ1
It is supposed that the input dataset X has N samples, each sample has
and n = 1. The individual intrinsic mode functions (IMFs) and
K dimensions, and Y is the output matrix belonging to RN×C . For n
their central frequencies are continuously updated using the
feature mappings, each mapping generates b nodes, and the output ex­
alternating direction method of multipliers (ADMM) to eventu­
pressions of the mapping feature layer are of the form (5) and (6) shown
ally solve the saddle point search and obtain the optimal solution.
as:
(4) The model is updated until the condition is satisfied.

Fig. 2. The structure of DBL.

4
X. Wang et al. International Journal of Electrical Power and Energy Systems 156 (2024) 109764

Fig. 3. RUL prediction process based on VDM-DBL.

Zi = φ(XWei + βei ), i = 1, 2, …, n (5) 2.3. Double broad learning (DBL)

Z n =Δ [Z1 , Z2 , ..., Zn ] (6) To address the fusion problem of two modalities, double broad
learning (DBL) combines two BLS units, the structure of which is shown
where Wei and βei are random matrices with corresponding di­
in Fig. 2 DBL network training is divided into three stages. First, Mo­
mensions and biases and ϕ(⋅) is the activation function. Wei ,βei can be
dality 1 and Modality 2 are inputted into their respective BLS units, and
obtained through a sparse autoencoder (SAE) and are more efficient
high-dimensional features are extracted to generate feature nodes and
than random initialization methods, with the aim of obtaining better
enhancement nodes. The high-dimensional features of the two modal­
features Zn .
ities are then linked in the fusion node layer by nonlinear mapping.
Similarly, the enhancement output matrix is calculated by (7) and (8)
Finally, these three components are linked to the DBL network output
to obtain:
layer by learning the output weight matrix and calculating with the
( )
Hj = ξj Z n Whj + βhj , j = 1, 2, ..., m (7) generalized inverse of ridge regression to achieve an effective output to
the target.
H m =Δ [H1 , H2 , …Hm ] (8) We suppose the number of samples in the network input is N. The
number of feature nodes and the number of enhanced nodes of Modal
where Whj and βhj are the randomly generated weights and biases, Sample 1 are N1 and N2 , and the number of feature nodes and the
respectively; ξj (⋅) is the activation function, and an S-type nonlinear number of enhanced nodes of Modal Sample 2 are N3 and N4 , respec­
activation function is generally chosen; and m is the number of rein­ tively.
forcement layer nodes. Therefore, the total characteristics of the two modalities can be
The feature layer and the enhancement layer are connected to form expressed as:
the input layer, and the output layer is connected to them. As a result, [ ]
when new data are entered into the system, the BLS model does not need (9)
N×(N +N N×(N +N
F N×(N1 +N2 +N3 +N4 ) = Ah 1 2 ) , Af 3 4 )
to retrain the entire model, greatly improving computational efficiency
and enabling incremental learning. where Ah and Af are the total features of Modality 1 and Modality 2.
The output features of the fused node layer can be calculated from F
by nonlinear mapping, and finally, the output Y of the DBL network is
obtained from the network weights W:

5
X. Wang et al. International Journal of Electrical Power and Energy Systems 156 (2024) 109764

Fig. 4. . Lithium-ion battery structural diagrams.

[ ( ) ]
Y = F N×(N1 +N2 +N3 +N4 ) , ξ FN×(N1 +N2 +N3 +N4 ) ⋅WtN×(N1 +N2 +N3 +N4 ) + βNt ⋅N5 {r(t), i1 (t), ..., im (t) } = VMD(C) (13)
[ N×(N +N +N +N ) N×N ]
= F 1 2 3 4
,T 5 ⌢n+1 ∑
where u k (ω) is the current residual and C⌢(ω) − =k u i (
⌢
(10) i∕ ω) is the
result of Wiener filtering.
3. Prediction model based on VMD-DBL
(4) Then, the DBL network is trained. Combining the benefits of BLS
In this topic, the residual life prediction of industrial equipment incremental learning, double-width learning is chosen to extract
using multimodal data will be addressed and investigated. Based on the and fuse the features of the two decomposed components (1*168-
introduction of VMD and DBL algorithms in the previous paper, the BLS dimensional data), and the number of iterations is set as the
is developed and improved to some extent in this paper, and a residual iteration termination condition to obtain the prediction results.
life prediction method based on variational modal decomposition with To obtain the prediction results, the training set data are labeled,
double broad learning (VMD-DBL) is proposed to compensate for the and the total feature extraction and fusion expressions are known
shortcomings of the single-modal prediction’s low accuracy and poor from (14) as:
robustness and improve the prediction performance. The method first R(t + 1) = BLS1 (r(t) ) (14)
decomposes the raw capacity data into an intrinsic modal function (IMF)
that reflects the random fluctuation of battery capacities at different ( )
∑
M ∑
t
scales and a residual component that reflects the overall degradation IM (t + 1) = BLS2 gi (t) (15)
trend. This step makes it easier to solve the battery capacity regeneration M=1 t=1

problem, which improves the prediction accuracy. The two components Therefore, the total characteristics of the two modes can be com­
are then separately inputted into the DBL subnetwork for feature bined as follows:
extraction to obtain the fusion node layer’s output features. Finally, the ⎡ { }N×(N3 +N4 ) ⎤
network weights W are used to calculate the prediction results. Fig. 3 ( ∑4 ) ∑
M

depicts the method’s flow chart. F N×

i=1
Ni
= ⎣{R(t + 1) }N×(N 1 +N2 )
, IM (t + 1) ⎦ (16)
M=1
As shown in Fig. 3, the proposed method in this paper can be divided
into three main stages: original data VMD, DBL network training, and We let the number of generated fusion nodes be N5 and the output
optimized VMD-DBL model prediction. The steps are shown below. characteristics of the fusion node layer be:
( ( ∑4 ) ( ∑4 ) )
(1) The lithium-ion battery discharge capacity degradation data is T N×N5 = ξ FN× N
i=1 i ⋅W
t
N5 × N
i=1 i
+ bNt ⋅N5 (17)
obtained.
(2) Using t-period as the starting point for prediction, periods are where ξ is the activation function and the number of nodes of the
used as the training set, and the data after are used as the test set, fusion layer is N5 .
with t being the current moment.
(3) VMD is used to decompose the capacity degradation data, and the (5) Remaining life prediction is performed based on VMD-DBL. With
alternating direction multiplier method is used to continuously the optimized model, the output is obtained based on the network
update each IMF and its central frequency to determine the weight assignment, and the expressions are
optimal saddle point. Finally, two modal characteristics of the [ ( ∑4 ) ]
residual component and IMF component are obtained, namely: Y = FN× i=1
Ni
, T N×N5 W (18)
∑ ⌢ ⌢
⌢n+1 C⌢(ω) − =k u i (ω) + λ (ω)/2
(11)
i∕
uk ( ω) =
1 + 2α(ω − ωk ) (6) The estimated RUL values are evaluated for the calculation of the
∫∞ ⌢ index to effectively and quantitatively evaluate the performance
⌢n+1 ω|u (ω) |2 dω index of the entire model, and the mean absolute error (MAE),
ωk (ω) = ∫0 ∞ ⌢ (12)
0
|u (ω) |2 dω mean absolute percentage error (MAPE), root mean square error

6
X. Wang et al. International Journal of Electrical Power and Energy Systems 156 (2024) 109764

Fig. 5. NASA Battery Capacity Trends.

(RMSE), and remaining life error (RULe) are introduced in this

paper to evaluate the results of the remaining life prediction
model for lithium-ion batteries. The evaluation equations are
illustrated below.
∑m ⃒
⃒
⃒
⌢⃒
MAE = (1/m) ⃒yi − y i ⃒
⃒ ⃒ (19)
i=1

⃒ ⃒
⃒ ⌢⃒
⃒ y
m ⃒ i − y ⃒
i⃒
∑
MAPE = (1/m) ⋅100% (20)
i=1
yi

√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
)2̅
∑ m (
⌢
RMSE = (1/m) yi − y i (21)
i=1

⃒ ⃒
RULe = ⃒RULreal − RULpre ⃒ (22)

where m is the total number of sample data, yi denotes the true RUL
of lithium batteries at the ith moment, y i denotes the predicted RUL of
⌢

the model at the ith moment.

The algorithm framework based on VMD-DBL is shown in Algorithm
1.
Algorithm 1. Algorithm for VMD-DBL.

4. Experimental and analysis

4.1. Experimental dataset and preprocessing

As shown in Fig. 4, the composition of a lithium-ion battery includes

the positive electrode, negative electrode, diaphragm, organic electro­
lyte, and cell case. The active materials in the positive and negative
electrodes are typically active materials, while the separator is a high-
molecular film that has been specially shaped. The organic electrolyte
acts as a carrier for ionic transportation within the battery. Over time,
factors such as reduced active material content, increased impurities,
and changes in crystal structure can lead to a decline in performance.
The most immediate manifestation of this is a significant decrease in
actual capacity. Therefore, predicting the remaining life of a lithium-ion
battery based on capacity deterioration data is crucial. Fig. 6. NASA battery VMD decomposition effect.
We use the NASA [dataset] [1] Battery DataSet dataset, Model
18650, with a rated capacity of 2 Ah, which contains three aged lithium
cobalt-acid batteries, B5, B6, and B7, obtained through cyclic charge and
discharge experiments, in this paper. Its aging process is split into two

7
X. Wang et al. International Journal of Electrical Power and Energy Systems 156 (2024) 109764

Fig. 7. RUL prediction based on VMD-DBL for four battery groups. Fig. 8. Comparison of the prediction results of different models.

8
X. Wang et al. International Journal of Electrical Power and Energy Systems 156 (2024) 109764

Fig. 9. Comparison of the RUL estimation performance of different models with other papers.

Table 1 Table 2
RUL estimated performance evaluation of NASA batteries. Performance evaluation of RUL estimation for different models of NASA
Battery number MAE MAPE RMSE RULe
batteries.
Battery number Model MAE MAPE RMSE RULe
B5 0.0064 0.0047 0.0088 0
B6 0.0070 0.0053 0.0111 1 B5 VMD-DBL 0.0064 0.0047 0.0088 0
B7 0.0052 0.0035 0.0074 1 VMD-BLS 0.0293 0.0213 0.0304 7
BLS 0.0582 0.0431 0.0663 4
B6 VMD-DBL 0.0070 0.0053 0.0111 1
stages: charging and discharging. Charging is performed with a constant VMD-BLS 0.0237 0.0181 0.0254 5
current of 1.5 A until the battery reaches 4.2 V and then with a fixed BLS 0.0283 0.0219 0.0337 4
B7 VMD-DBL 0.0052 0.0035 0.0074 1
voltage until the battery current drops to 20 mA; discharging is done
VMD-BLS 0.0304 0.0205 0.0313 12
with a constant current until B5, B6, and B7 reach the voltage setting BLS 0.0633 0.0437 0.0772 –
values of 2.7 V, 2.5 V, and 2.2 V, respectively. Fig. 5 depicts the aging
trend of each battery model.
A 30 % capacity decay is recognized as the battery’s aging threshold In Fig. 7, the dark blue line represents the lithium-ion battery’s ca­
during repeated charging. The lithium-ion battery undergoes a physical pacity degradation curve during the charge/discharge cycle, the pink
and chemical relaxation reaction during this process, resulting in a line represents the RUL estimation curve based on VMD-DBL, and the
sudden increase in capacity. This special case also presents a research short black line represents the failure threshold. Fig. 7 shows that the
challenge for Li-ion battery RUL prediction. VMD-DBL-based remaining life prediction curve fits the capacity
First, the dataset is decomposed using VMD individually. The degradation curve well. The evaluation results in Table I are obtained
training and test sets are then separated and trained using the BLS al­ using the performance evaluation formula proposed above.
gorithm to create a VMD-BLS-based model for estimating lithium-ion As shown in Table 1, the remaining life error values of Batteries B5,
battery health. The network’s parameters, such as internal weights B6 and B7 are 0, 1 and 1, respectively, which are basically consistent
and bias, are then adjusted through training, and estimation results are with the actual values. Taking B5 as an example, the MAE of the test set
obtained from the trained model using the test set data. Finally, the is 0.0064, MAPE is 0.0047 Ah, and RMSE is 0.0088 Ah. Therefore, it is
estimation results are used to assess the performance of the VMD-BLS- proven that the VMD-DBL network model proposed in this paper has a
based health state estimation method. The number of VMD sequences good prediction effect on different models of lithium-ion batteries.
is set to four, and the bandwidth limit is set to 300. Fig. 6 depicts the
results of the VMD decomposition of B5, B6, and B7, which are
commonly used in NASA lithium batteries. As shown in the figure, the 4.3. Comparison and analysis of results
residual component essentially reflects the general trend of battery
degradation as the number of cycles increases, IMF1 is the capacity The effectiveness of the VMD-DBL model proposed in this paper is
regeneration component, which roughly coincides with the position of validated by comparing it to the BLS and VMD-BLS network models. The
battery capacity regeneration, and IMF2 and IMF3 are the random graph below compares the RUL prediction results of the three network
fluctuation components caused by noise, measurement error, and other models for B5, B6, and B7.
disturbing factors. The figure shows how VMD can effectively decouple Table 2 shows the detailed prediction data of the above models for
the original battery capacity data into multiple data components, the three lithium-ion batteries, which proves that the VMD-DBL model
reducing the complexity and non-smoothness of the original data and has the best prediction performance in terms of the MAE, MAPE, RMSE,
improving the accuracy of subsequent model training. and RULe.
The evaluation index shows the single prediction model’s substantial
prediction error. The data in the table demonstrate that VMD signal
4.2. Model prediction based on VMD-DBL decomposition can effectively lessen the impact of the non-smoothness
of the capacity degradation data on the prediction by showing that the
Different failure thresholds of 70 %, 75 % and 70 % are set for these prediction accuracy of the VMD-BLS method, which is predicted by BLS
three cells. Fig. 7 shows the fitting of the remaining life prediction re­ after VMD multiscale decomposition, is significantly improved. The
sults for B5, B6 and B7 of VMD-DBL to the real test results.Fig. 8.Fig. 9.. VMD-DBL technique shown in this study achieves the lowest prediction

9
X. Wang et al. International Journal of Electrical Power and Energy Systems 156 (2024) 109764

Table 3 References
Comparison of the results of this paper with other literature.
[1] Ochella S, Shafiee M, Dinmohammadi F. Artificial intelligence in prognostics and
Battery number Model MAE RMSE RULe
health management of engineering systems. Eng Appl Artif Intell 2022;108:
B5 CNN-LSTM-DNN[36] 0.0083 0.0145 – 104552.
VMD-PSO-GRU[37] 0.0118 0.0148 2 [2] Zio E, Maio FD. A data-driven fuzzy approach for predicting the remaining useful
SADE-MESN[38] 0.0129 0.0172 3 life in dynamic failure scenarios of a nuclear system. Reliab Eng Syst Safe 2010;95
(1):49–57.
VMD-DBL 0.0064 0.0088 1
[3] Chen Z, Wu M, Zhao R, Guretno F, Yan R, Li X. Machine Remaining Useful Life
Prediction via an Attention-Based Deep Learning Approach. IEEE Trans Ind
Electron 2020;68(3):2521–31.
error and the maximum accuracy, demonstrating the method’s capacity [4] Hu X, Xu L, Lin X, Pecht M. Battery lifetime prognostics. Joule 2020;4(2):310–46.
to successfully separate and fuse residual components and IMF compo­ [5] Li X, Wang Q, Yang Y, Kang J. Correlation between capacity loss and measurable
parameters of lithium-ion batteries. Int J Electr Power Energy Syst 2019;110:
nents after VMD decomposition, enhancing the prediction accuracy and
819–26.
stability. [6] Bian X, Wei Z, Li W, Pou J, Sauer DU, Liu L. State-of-Health Estimation of Lithium-
To further demonstrate the advantages of the VMD-DBL prediction Ion Batteries by Fusing an Open Circuit Voltage Model and Incremental Capacity
Analysis. IEEE Trans Power Electron 2022;37(2):2226–36.
method proposed in this paper, the prediction results of the method in
[7] Xu B, Oudalov A, Ulbig A, Andersson G, Kirschen DS. Modeling of lithium-ion
this paper are compared with those of the models proposed in recent battery degradation for cell life assessment.IEEE. Trans Smart Grid 2018;9(2):
related literature using the same training set of B5 batteries in the NASA 1131–40.
dataset as an example, and the comparison results are shown in Table 3. [8] Li H, Fu L, Zhang Y. A novel hybrid data-driven method based on uncertainty
quantification to predict the remaining useful life of lithium battery. J Energy
To verify the generalization performance of the method in this paper, Storage 2022;52:104984.
the performance metrics of B6 and B7 are compared, and the results are [9] Cheng L, Wan Y, Zhou Y, Gao W. Operational reliability modeling and assessment
shown in the following figure: of battery energy storage based on lithium-ion battery lifetime degradation. J Mod
Power Syst Clean Energy 2022;10(6):1738–49.
[10] Wang S, Takyi-Aninakwa P, Fan Y, Yu C, Jin S, Fernandez C, et al. A novel feedback
5. Conclusion correction-adaptive Kalman filtering method for the whole-life-cycle state of
charge and closed-circuit voltage prediction of lithium-ion batteries based on the
second-order electrical equivalent circuit model. Int J Electr Power Energy Syst
In this paper, we suggest a new VMD-DBL model approach for pre­ 2020;139:108020.
dicting the residual life of lithium-ion battery devices to expand RUL- [11] Wang D, Yang F, Tsui K, Tsui K-L, Zhou Q, Bae SJ. Remaining Useful Life Prediction
related research. In this work, two decomposed modal components are of Lithium-ion Batteries Based on Spherical Cubature Particle Filter. IEEE Trans
Instrum Meas 2016;65(6):1282–91.
chosen as prediction feature inputs based on the signal decomposition of
[12] Xue Z, Zhang Y, Cheng C, Ma G. Remaining useful life prediction of lithium-ion
capacity degradation data by VMD, and DBL is used for feature fusion batteries with adaptive unscented kalman filter and optimized support vector
and prediction to obtain the prediction results. To validate the effec­ regression. Neurocomputing 2020;376:95–102.
[13] Yang Q, Xu J, Li X, Xu D, Cao B. State-of-health estimation of lithium-ion battery
tiveness of the chosen method, tests are performed on three distinct
based on fractional impedance model and interval capacity. Int J Electr Power
battery models, and the performance evaluation is also compared to Energy Syst 2020;119:105883.
other current and related literature. The findings show that the VMD- [14] Patil MA, Tagade P, Hariharan KS, Kolake SM, Song T, Yeo T, et al. A novel
DBL model suggested in this paper effectively overcomes the issue of multistage Support Vector Machine based approach for Li ion battery remaining
useful life estimation. Appl Energy 2015;159:285–97.
battery capacity data noise, improves the prediction accuracy signifi­ [15] Liu K, Peng Q, Teodorescu R, Foley AM. Knowledge-Guided Data-Driven Model
cantly, and BLS’s advantage of incremental learning has saved our With Transfer Concept for Battery Calendar Ageing Trajectory Prediction. IEEE/
computation time and cost. This method provides better RUL prediction CAA J Autom Sin 2023;10(1):272–324.
[16] Wang Y, Ni Y, Lu S, Wang J, Zhang X. Remaining Useful Life Prediction of Lithium-
for lithium batteries. Ion Batteries Using Support Vector Regression Optimized by Artificial Bee Colony.
IEEE Trans Veh Technol 2019;68(10):9543–53.
CRediT authorship contribution statement [17] Ng SSY, Xing Y, Tsui KL. A naive Bayes model for robust remaining useful life
prediction of lithium-ion battery. Appl Energy 2014;118:114–23.
[18] Li N, He F, Ma W, Wang R, Jiang L, Zhang X. An Indirect State-of-Health Estimation
Xiaojia Wang: Conceptualization, Supervision, Funding acquisition, Method Based on Improved Genetic and Back Propagation for Online Lithium-Ion
Writing – review & editing. Xinyue Guo: Conceptualization, Method­ Battery Used in Electric Vehicles. IEEE Trans Veh Technol 2022;71(12):12682–90.
[19] Lipu MSH, Hannan MA, Hussain A, Saad MH, Ayob A, Uddin MN. Extreme
ology, Software, Writing – original draft. Sheng Xu: Investigation. Xibin
Learning Machine Model for State-of-Charge Estimation of Lithium-Ion Battery
Zhao: Supervision, Formal analysis. Using Gravitational Search Algorithm. IEEE Trans Ind Appl 2019;55(4):4225–34.
[20] Li C, Xiao F, Fan Y, Yang G, Tang X. An Approach to Lithium-Ion Battery SOH
Estimation Based on Convolutional Neural Network. J Electrochem Soc 2020;35
Declaration of competing interest
(19):4106–19.
[21] Catelani M, Lorenzo C, Fantacci R, Patrizi G, Picano B. Remaining Useful Life
The authors declare that they have no known competing financial Estimation for Prognostics of Lithium-Ion Batteries Based on Recurrent Neural
interests or personal relationships that could have appeared to influence Network. IEEE Trans Instrum Meas 2021;70:3524611.
[22] Kara A. A data-driven approach based on deep neural networks for lithium-ion
the work reported in this paper. battery prognostics. Neural Comput Appl 2021;33:13525–38.
[24] Xiao F, Li C, Fan Y, Yang G, Tang X. State of charge estimation for lithium-ion
Data availability battery based on Gaussian process regression with deep recurrent kernel. Int J
Electr Power Energy Syst 2021;124:106369.
[25] Chen L, Xie S, M.lopes A, Bao X. A vision transformer-based deep neural network
Data will be made available on request. for state of health estimation of lithium-ion batteries. Int J Electr Power Energy
Syst 2023;152:109233.
[26] Liu D, Li L, Song Y, Wu L, Peng Y. Hybrid state of charge estimation for lithium-ion
Acknowledgement battery under dynamic operating conditions. Int J Electr Power Energy Syst 2019;
110:48–61.
This work was supported by the National Natural Science Foundation [27] Yu J. State of Health Prediction of Lithium-Ion Batteries: Multiscale Logic
Regression and Gaussian Process Regression Ensemble. Reliab Eng Syst Safe 2018;
of China-Guangdong Joint Fund (No. U2001201), the National Natural
174:82–95.
Science Foundation of China (62376082, 61876055). The authors [28] Chen JC, Chen T-L, Liu W-J, Cheng CC, Li M-G. Combining empirical mode
gratefully acknowledge their supports. decomposition and deep recurrent neural networks for predictive maintenance of
lithium-ion battery. Adv Eng Inform 2021;50:101405.
[29] Zhang C, Zhao S, He Y. An Integrated Method of the Future Capacity and RUL
Prediction for Lithium-Ion Battery Pack. IEEE Trans Veh Technol 2021;71(3):
2601–13.

10
X. Wang et al. International Journal of Electrical Power and Energy Systems 156 (2024) 109764

[30] Chen CLP, Liu Z. Broad Learning System: An Effective and Efficient Incremental [34] Dou C, Wu D, Yue D, Jin B, Xu S. A Hybrid Method for False Data Injection Attack
Learning System Without the Need for Deep Architecture. IEEE Trans Neural Detection in Smart Grid Based on Variational Mode Decomposition and OS-ELM.
Networks Learn Syst 2018;29(1):10–24. CSEE J Power Energy Syst 2020;8(6):1697–707.
[31] Chen CLP, Liu Z, Feng S. Universal Approximation Capability of Broad Learning [35] Ahajjam MA, Licea DB, Ghogho M, Kobbane A. Experimental investigation of
System and Its Structural Variations. IEEE Trans Neural Networks Learn Syst 2019; variational mode decomposition and deep learning for short-term multi-horizon
30(4):1191–204. residential electric load forecasting. Appl Energy 2022;326:119963.
[32] Wang Z, Xu X, Liu H, Sun F. Multi-modal broad learning for material recognition. [36] Saha B, Goebel K. Battery data set. NASA Ames Prognostics Data Repository; 2007.
CAAI Transactions on Intelligent Systems 2020;15(4):787–94. [37] Zraibi B, Okar C, Chaoui H, Mansouri M. Remaining Useful Life Assessment for
[33] Jia C, Liu H, Xu X, Sun F. Multi-modal information fusion based on broad learning Lithium-Ion Batteries Using CNN-LSTM-DNN Hybrid Method. IEEE Trans Veh
method. CAAI Transactions on Intelligent Systems 2019;14(1):150–217. Technol 2021;70(5):4252–61.
[38] Liu Q, Xing Y. Remaining life prediction of lithium-ion battery based on VMD-PSO-
GRU model. Environ Sci Technol 2023;12(1):236–46.

11