Journal of Energy Storage 73 (2023) 108915

Contents lists available at ScienceDirect

Journal of Energy Storage

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

Research papers

Physics-informed ensemble deep learning framework for improving state of

charge estimation of lithium-ion batteries
Hanqing Yu a, Zhengjie Zhang a, Kaiyi Yang a, Lisheng Zhang a, Wentao Wang a, Shichun Yang a, *,
Junfu Li b, **, Xinhua Liu a, c, *
a
School of Transportation Science and Engineering, Beihang University, Beijing 102206, China
b
School of Automotive Engineering, Harbin Institute of Technology, Weihai 264209, China
c
Dyson School of Design Engineering, Imperial College London, London SW7 2AZ, UK

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

Keywords: With the advances in computer science, deep learning (DL) has been developed for battery management systems
Lithium-ion batteries (BMSs) with artificial intelligence. State of charge (SOC) estimation of lithium-ion batteries is the fundament and
State of charge estimation core of BMS, and improving the accuracy, robustness and generalization of model predictions is still challenging.
Physics-informed learning
Herein, this paper proposes a physics-informed ensemble deep learning (PIEDL) framework to enable the
Ensemble deep learning
Electrochemical model
physical information introduction and multi-model integration. Firstly, a battery simplified electrochemical
model (SEM) is used to quickly extract the physical information related to the battery SOC. Subsequently, the
open-circuit voltage and reaction polarization resistance from the SEM are integrated as key physical information
into the DL model and combined with the original input variables to construct the physics-informed deep
learning (PIDL) part of the framework. Then, DL models improved with different techniques are used as base
learners for the ensemble deep learning (EDL). At the second level of the EDL, a meta learner is used to integrate
multiple heterogeneous base models based on the blending strategy without any weight calculation. The results
show that PIEDL outperforms all base models and all models with fewer input variables, and improves the result
by more than 60 % relative to the original model with original inputs. Finally, the generalization of the trained
model is validated using different battery types. The PIEDL framework is not only important for improving the
performance and application scope of BMS, but also provides new ideas and methods for the field of DL.

1. Introduction ampere-hour integral, table look-up, model-based and data-driven

methods [6]. The first two categories are quite simple to implement
Lithium-ion batteries (LIBs) have currently been widely used in en­ and suitable for onboard implementation, but they can lead to lower
ergy storage devices such as electric vehicles, energy storage systems accuracy and stability of the results [7,8]. Model-based methods usually
and consumer electronics [1]. However, the sluggish development of incorporate a battery model with a non-linear state estimation algorithm
battery management system (BMS) technology is a key bottleneck hin­ [9]. Commonly used battery models for SOC estimation include equiv­
dering the development of these applications [2,3]. State of charge alent circuit models (ECMs) and electrochemical models [10]. As a
(SOC) estimation, as the core and foundation of BMS, figures promi­ closed-loop algorithm, this method offers better accuracy and stability.
nently in the optimal management of LIBs if this function can be well However, this approach can be limited by the accuracy of the battery
implemented [4]. However, the battery SOC is an internal state variable model and the performance of the non-linear state estimation algorithm.
that cannot be measured directly and can only be estimated from Therefore, many recent studies have made improvements in these two
measured data (e.g., voltage, current and temperature) [5]. Therefore, aspects for a better performance of SOC estimation [11,12]. Data-driven
designing accurate, stable and widely applicable algorithms for esti­ methods usually employ machine learning (ML) models to learn the non-
mating battery SOC still remains a challenge. linear relationship between the battery SOC and the input features of the
Currently, there are four primary methods for SOC estimation: model [13]. Such methods do not require physical modelling and have

* Corresponding authors at: School of Transportation Science and Engineering, Beihang University, Beijing 102206, China.
** Correspondence to: J. Li, School of Automotive Engineering, Harbin Institute of Technology, Weihai 264209, China.
E-mail addresses: [email protected] (S. Yang), [email protected] (J. Li), [email protected] (X. Liu).

https://doi.org/10.1016/j.est.2023.108915
Received 25 June 2023; Received in revised form 15 August 2023; Accepted 2 September 2023
Available online 27 September 2023
2352-152X/© 2023 Published by Elsevier Ltd.
H. Yu et al. Journal of Energy Storage 73 (2023) 108915

shown good accuracy, stability and applicability [14]. Deep learning of different types of models. Ensemble deep learning (EDL), which
(DL), an advancement of ML, has also been used to estimate battery SOC. combines the integration of EL and DL, has garnered widespread
Especially when considering the time-dependent characteristic of SOC, attention as a high-performing approach [32]. However, to the best of
recurrent neural networks (RNNs) have received much public attention, our knowledge, the investigations focusing on battery SOC estimation by
which excel in processing time series data [15]. Although the DL-based the EDL approach and even the combination with physical information
SOC estimation method has relatively good performance and is easy to are still in the research gap.
implement, its accuracy, stability and generalization still warrant To solve the above problems and to further explore the application of
improvement. In addition, there is an interesting finding: different PIDL and EDL methods in battery SOC estimation, this paper creatively
studies have come to different conclusions about the relationship be­ proposes a physics-informed ensemble deep learning (PIEDL) frame­
tween the good and bad performance of specific models. For instance, by work that combines PIDL and EDL methods. The performance is verified
comparing the accuracy of long short-term memory (LSTM), bidirec­ with the battery SOC estimation as the target. The main contributions of
tional LSTM (Bi-LSTM) and gated recurrent unit (GRU) for battery SOC the paper can be summarized in the following four aspects: (1) The
estimation, Ma et al. [16] found that the LSTM and Bi-LSTM performed pseudo-two-dimensions (P2D) model of LIBs is simplified to obtain a
almost identically, and LSTM was even slightly better than Bi-LSTM. simplified electrochemical model (SEM) to characterize the internal
However, Pang et al. [17] showed that Bi-LSTM was significantly bet­ behavior of the battery, and the key physical information related to the
ter than LSTM. battery SOC (i.e., open-circuit voltage OCV and reaction polarization
The studies presented above on ML-based SOC estimation have resistance Ract) is analyzed and inputted into the DL models, constituting
mapped current and voltage data directly to the battery SOC. A wealth of a PIDL method to achieve the SOC estimation; (2) An EDL framework is
informative data can facilitate efficient SOC estimation [18]. Particu­ proposed with the DL models improved by different techniques to meet
larly, physics-informed machine learning (PIML) has been widely re­ the requirement of base model heterogeneity, where the attention
ported recently [19]. This method focuses on combining a mechanistic mechanism (AM) can further enhance the interpretability of the DL
model with ML, thus having the advantages of both methods to further model, and a blending ensemble strategy is used to simply and effec­
improve the battery SOC estimation. Pure mechanism model-based tively integrate the different base models; (3) Through the comparison
methods rely too much on the model's accuracy, while pure data- of different algorithms and the analysis of multicollinearity, the kernel
driven methods often have unstable prediction results for a lack of ridge regression (KRR) with the best performance is selected as the meta
physical interpretability. Facing such challenges, the PIML naturally learner for the second layer of PIEDL; (4) Data from different battery
becomes a hot topic of research. Furthermore, physics-informed deep types are used as training and test datasets respectively for the first time,
learning (PIDL) can better learn the relation between data by using DL which validates the strong generalization ability of PIEDL and shows
models [17,20,21]. Tian et al. [20] proposed a PIDL, which used ECM to that the trained model can be applied to a wide range of application
provide auxiliary physical information for DL, thus improving the pre­ scenarios. The research findings of this paper can provide new insights
diction performance and generalization of the model. Variables with and approaches for the field of DL and are also important for improving
physical information may have a higher correlation with the target the performance and application scope of BMS.
variable. This physics-informed approach requires the information to be The remainder of this paper is organized as follows. Section 2 gives a
extracted and processed reasonably to meet the needs of ML. This detailed description of the methodology, including an overview of
approach has also been applied in the prediction of battery state of PIEDL, SEM, feature selection method, DL model, and ensemble method.
health (SOH) [22] and heat generation rate [17]. On the other hand, Section 3 presents the datasets used for experiments and describes the
Wang et al. [21] introduced fractional-order gradients into the RNN to acquisition of augmented data. Section 4 provides comprehensive re­
improve the back-propagation process, ultimately achieving satisfactory sults and in-depth discussion. Finally, conclusions are shown in Section
accuracy in SOC estimation. In this type of research, physical informa­ 5.
tion is used as a constraint to optimize and improve the ML model. The
design and optimization of the model are carried out by combining 2. Methodology
knowledge from physics and ML. This same thought has been used for
the prediction of battery voltage [23] and thermal runaway [24]. In 2.1. Overview
summary, the PIDL provides new ideas and methods for BMS solutions
by combining physical information with DL, thus effectively improving This subsection provides a general description of the PIEDL frame­
the performance of the model. work, as shown in Fig. 1. It can be seen that the PIEDL-based SOC
As mentioned earlier, different studies have drawn inconsistent estimation consists mainly of battery measurement data, SEM, feature
conclusions about the prediction performance of different ML models. selection, time series construction, and EDL framework. In addition to
This contradiction indicates that different ML models have their unique the directly measured data, the electrochemical model provides addi­
advantages in different aspects for different datasets [25]. More tional physical information for the construction of the proposed
importantly, the performance of different models cannot be evaluated in framework during the battery SOC estimation process. The proposed
advance for practical applications, which further increases the deploy­ PIEDL framework extremely improves the accuracy and generalization
ment requirements for model accuracy, robustness, and generalization. of the SOC estimation by combining the battery mechanism model, DL
To cope with these issues, multiple models can be integrated to achieve and EL. Specifically, the SEM provides the model with variables that
better SOC estimation, which is known as ensemble learning (EL). The vary over time and are relevant to the battery SOC, which allows the DL
average method is the simplest and most straightforward ensemble to incorporate physical information features. The EL framework in­
strategy, which averages the predictions of multiple basic models as the tegrates the results of multiple models, further improving the SOC
final result [26]. However, this approach has limited performance. estimation accuracy and generalization. It should be noted that in this
Further, the weights of each base model can be set and can be optimized framework, the base models of EDL can be determined by using a variety
by optimization algorithms [27,28]. Moreover, the ML algorithm can be of methods to improve an appropriate baseline model.
also used as the secondary learner to learn the weights of multiple base The BMS can sample battery signals including time, voltage, current
learners, such as random forest (RF) [29], back propagation neural and temperature during battery operation. During the testing of batte­
network [30] and LSTM [31]. It is important to note that homogeniza­ ries in the laboratory, accurate battery SOC data can be obtained based
tion of base models can lead to low or no improvement in the integrated on the ampere-hour integral method to train and test DL models, relying
results, which falls short of satisfactory results [25]. Therefore, diversity on accurate calibration and high sampling accuracy. These obtained
should be focused when selecting base models, to take more advantage measurements are further used for DL time series construction as well as

2
H. Yu et al. Journal of Energy Storage 73 (2023) 108915

Fig. 1. Schematic diagram of our proposed PIEDL framework for improving battery SOC estimation.

being fed into the mechanism model to obtain more characteristic previous research on the electrochemical modelling of LIBs [33], the
variables. external characteristics of the battery are the result of the combined
The DL-based data-driven approach typically learns information effect of several internal physicochemical reactions. The results of the
directly from the battery measurements to obtain a real-time mapping internal battery variables obtained by using the model calculations can
relationship with SOC. Although this approach is easy to implement, provide the DL with more physical information to better learn the
fusing physical information related to the battery mechanism with DL mapping between the input variables and the battery SOC. However,
can still further improve the accuracy of SOC estimation. Based on our high-quality features should be selected prior to input to further filter

3
H. Yu et al. Journal of Energy Storage 73 (2023) 108915

useful information. It is due to the fact that inputting less relevant var­ constructed into time series and then inputted into the trained EDL to
iables into the model not only increases the consumption of computa­ obtain the SOC estimation results of the battery, which is highly real-
tional resources, but also leads to a decrease in model performance. time and accurate.
Therefore, it is necessary to screen the model variables obtained from
the calculation. 2.2. Simplified electrochemical model
Combining the LIB measurements and the selected feature variables,
the time series needs to be constructed in order to be inputted into the During the operation of LIBs, many complex physicochemical pro­
algorithm. The number of feature variables selected from the mecha­ cesses are involved, hence researchers have proposed a number of
nism model is set to i. Considering that the three variables voltage, electrochemical models to describe the internal reaction processes. The
current and temperature are originally used as inputs, the number of most typical battery electrochemical model is the P2D model, which
input features is 3 + i. The output is the battery SOC, and its number is 1. provides a reliable explanation of the LIB properties from a physico­
Due to the time series characteristics, it is required to extract N one- chemical mechanism perspective [34]. However, the P2D model cannot
dimensional vectors sequentially in time order and then combine them be applied online due to the huge computational resources required.
into a two-dimensional matrix, i.e., each matrix has a sequence length of Therefore, in this paper, an SEM is used for battery electrochemical
N. Then, these matrix fragments are formed into a three-dimensional simulation to provide physical information related to the battery SOC
tensor with a batch size of M in time order to be inputted into the and to meet the requirements of online applications. SEM not only
RNNs that belong to DL. simulates the internal physicochemical processes of the LIB, but also has
Lastly, the EDL framework is used for battery SOC estimation. The a high computational efficiency comparable to that of ECM [35].
main concept of the framework proposed in this paper is that different The SEM still maintains the battery structure of the P2D model, i.e.,
technologies are adopted to improve the selected effective DL model, the sandwich structure, and since it is assumed that no electrochemical
and multiple different DL models can be obtained. This approach can reaction takes place in the current collector, the electrochemical
ensure heterogeneity between the base models concurrently. In this calculation domains are the negative electrode, separator and positive
paper, the best working LSTM model is used based on our previous electrode. The schematic diagram of the model is shown in the SEM
research results [16]. Considering the positive impact of model diversity section of Fig. 1, where ln, ls and lp indicate the thicknesses of the above
on integration performance, Bayesian optimization (BO), attention three components respectively. In the SEM constructed in this paper, the
mechanism (AM) and convolutional neural network (CNN) are chosen to internal mechanism of the battery is described as the basic working
be used in the first layer of EL to improve LSTM in terms of hyper­ process, solid-phase diffusion, reaction polarization, liquid-phase
parameter optimization, input feature importance and further feature diffusion and ohmic polarization. The detailed mathematical expres­
extraction. Moreover, dropout and root mean square propagation sions are shown in Table 1. The basic working process describes the
(RMSprop) methods are used to suppress the overfitting phenomenon change in the average lithium-ion concentration in the solid phase of the
and improve the convergence speed, respectively. Then, in the second electrode due to the transfer of lithium ions between the positive and
layer, the results of the three base models, i.e., BO-LSTM, AM-LSTM and negative electrodes during the charging and discharging process. Solid-
CNN-LSTM, are integrated using KRR as meta learner based on the phase diffusion is the concentration distribution resulting from the
blending strategy with no need for complex weight calculation. The diffusion of lithium ions in the radial direction inside the electrode
inputs to the meta learner are the prediction results of all base models particles. This model focuses on the lithium-ion concentration at the
and the output is the final prediction of the battery SOC. To quantita­ electrode particle surface, i.e., the difference between the particle sur­
tively assess the performance for SOC estimation, after the final battery face concentration and the average concentration calculated by the
SOC estimation results are obtained, the results are analyzed by using solid-phase diffusion equations. This is due to the fact that the lithium-
three error metrics, namely the average absolute error (AAE), root mean ion concentration at the electrode particle surface is related to the open-
square error (RMSE) and maximum absolute error (MAE), which are circuit potential (OCP). The OCP is the equilibrium potential of the
respectively defined as:

1 ∑ nr
Table 1
AAE = ̂i|
|SOCi − SOC (1) The mathematical expressions of each mechanism process for the SEM.
nr i=1
Physics SEM
√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
1 ∑ Basic working yavg (t) = y0 + I(t)t/Qp , xavg (t) = x0 − I(t)t/Qn
nr
RMSE = (SOCi − SOC ̂ i )2 (2) process and
[
1 12 τp
]
nr i=1 Δy' (tk+1 ) = Δy' (tk ) + I(tk ) − Δy' (tk ) (tk+1 − tk )
solid-phase τp 7 Qp
[ ]
diffusion 1 12 τn
Δx' (tk+1 ) = Δx' (tk ) + I(t ) − Δx' (tk ) (tk+1 − tk )
̂i|
MAE = max |SOCi − SOC (3) τn 7 Qn k
1≤i≤nr
2 τp 2 τn
Δy(t) = Δy′(t) + I(t), Δx(t) = Δx′(t) + I(t)
7 Qp 7 Qn
where nr is the number of samples for SOC estimation, and SOCi and ysurf (t) = yavg (t) + Δy(t), xsurf (t) = xavg (t) − Δx(t)
[ ]
SOC
̂ i are the reference and estimation values of SOC at the i-th time, [
Eocv (t) = Up ysurf (t) − Un xsurf (t)
]

respectively. AAE, RMSE and MAE can evaluate the error levels of Reaction 2RT(t)
{ [√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅ ] [√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
ln m2n (t) + 1 + mn (t) − ln m2p (t) + 1 +
average, standard deviation and maximum, respectively.
ηact (t) =
polarization F
}
In addition, the method in this paper can be divided into offline
]
mp (t)
training and online application in the process of practical realization. In
1 1
the offline training phase, the relationship between the battery SOC and mp (t) =
6Qp c0.5
( )0.5 ( )0.5 Pact I(t), mn (t) =
0 1 − ysurf ysurf
internal variables obtained from the simulation of LIBs using the SEM is
1 1
analyzed to screen high-quality variables with good correlation with ( )0.5 ( )0.5 Pact I(t)
6Qn c0.5 1 − xsurf xsurf
SOC. Time series are constructed by combining high-quality variables 0
[ ]
Liquid-phase 2RT(t) c0 + Δcp (t)
and the measured voltage, current and temperature data. They are then diffusion
ηcon (t) =
F
(1 − t+ )ln
c0 − Δcn (t)
imported into the EDL for training. In the online application phase, ( ) (
Δcp (t) = Pcon a 1 − e− t/τe I(t), Δcn (t) = Pcon b 1 − e− t/τe
)
I(t)
voltage, current and temperature data are measured in real time during Ohmic ηohm (t) = Rohm I(t)
the battery operation and the battery internal characteristics are simu­ polarization
lated in real time using the SEM. The data of the input variables are Terminal voltage Uapp (t) = Eocv (t) − ηact (t) − ηcon (t) − ηohm (t)

4
H. Yu et al. Journal of Energy Storage 73 (2023) 108915

electrode material and is an intrinsic property of the material that can be two variables, the value of PCC approaches 1 or − 1, and when there is no
derived from priori measurement and calculation [36]. And, open- linear relationship between the two variables, the value of PCC ap­
circuit voltage (OCV) is the difference between the OCPs of the posi­ proaches 0.
tive and negative electrodes. Reaction polarization reflects the charge KCC is used to measure the degree of ordinal correlation between
exchange caused by electrochemical reactions at the junction of the solid two variables and is used to assess the rank correlation of non-linear
and liquid phases and is usually described by the Butler-Volmer equa­ variables. When two samples (xi, yi) and (xj, yj) are in the same order,
tion. Liquid-phase diffusion is caused by the time for the transport of i.e., xi > xj and yi > yj, or xi < xj and yi < yj, they are considered to be
lithium ions moving through the electrolyte being much greater than the consistent. Conversely, when the order is different, i.e., xi > xj and yi <
time for the transport of electrons in the external circuit. The liquid- yj, or xi > xj and yi < yj, the two samples are considered inconsistent. n1
phase lithium ions are consumed in one electrode region, and the con­ and n2 are set to denote the number of concordant pairs and discordant
centration decreases; while in the other electrode region, lithium ions pairs, respectively. The KCC is defined as:
accumulate, and the concentration increases. This concentration n1 − n2
gradient eventually leads to the creation of a concentration polarization KCC = (5)
nr (nr − 1)/2
overpotential. Ohmic polarization describes the polarization of positive
and negative charges (i.e., lithium ions and electrons) during transport, SCC is used to measure the degree of monotonic correlation between
which follows Ohm's law. Ultimately, the battery terminal voltage can two variables and is equal to the Pearson correlation between the rank
be broken down into the OCV, reaction polarization overpotential, values of the two variables. The variables X and Y are sorted according to
concentration polarization overpotential and ohmic polarization over­ the same rules (i.e., in the same ascending or descending order) to obtain
potential. Our previous studies have given a detailed and clear expla­ the corresponding ranks for xi and yi as r(xi) and r(yi) respectively. A
nation of SEM [33,37,38], which will not be repeated herein. difference set D can be obtained by subtracting the corresponding ele­
Furthermore, all the parameters used in the SEM can be found in Section ments, where di = r(xi) − r(yi). Then, SCC can be calculated by the
3. In summary, SEM is an effective electrochemical tool for describing following equation:
the internal physicochemical reaction processes of LIBs. By under­ ∑nr
standing how a LIB works, simulations can be carried out for the internal 6 di2
reaction processes of the battery, which are further used to guide battery SCC = 1 − ( i=1 ) (6)
nr n2r − 1
design and management.
MI is essentially a measure of the degree of association between two
systems and can represent the amount of information shared between
2.3. Selection of feature variables two variables. It can be used in the feature selection process to assess the
correlation between two variables and performs well when dealing with
Although using as many feature variables as possible as inputs would high-dimensional data. The MI approach measures the strength of cor­
seem to maximize the model performance, this is not the case in fact relation by calculating the amount of information between two vari­
[39]. As the number of features increases, some irrelevant features will ables. A higher value of mutual information indicates a stronger
introduce more noise, thus reducing the performance to a certain extent. correlation between them. The MI coefficient (MIC) of two variables can
In addition, this can lead to a significant increase in computational re­ be defined as follows [41]:
sources [40]. Therefore, in order to improve the performance of the
model, high-quality feature variables need to be selected as inputs. To ∑∑ p(x, y)
MIC = p(x, y)log (7)
maximally combine the advantages of different feature selection x y
p(x)p(y)
methods, two methods are adopted in this paper, including the corre­
lation coefficient (CC) method based on statistical principle and the where p (x) and p (y) are marginal probability distributions, and p (x, y)
mutual information (MI) method based on information theory. is joint probability distributions.
The CC method is used to measure the degree of linear correlation By calculating the above-mentioned different coefficients separately,
between two variables. In the process of selecting features, the CC can be the feature variable with the better correlation is finally selected to be
used to assess the linear correlation between a feature variable and the added to the input after a comprehensive consideration. When the result
target variable. The final result obtained is between − 1 and 1, and the of the CC is higher than 0.3 [39] or the MIC is relatively large, the
sign of the CC describes the positive or negative correlation between the variable can be recorded and selected. Then, a stepwise regression with
two variables. The magnitude of its absolute value indicates the strength forward selection is used to determine the validity of the variables. A
of the correlation. The characteristic with the strongest correlation regression is first performed with the original inputs, and then new
should be selected. Three commonly used CCs are the Pearson correla­ feature variables are gradually added to observe changes in overall
tion coefficient (PCC), Kendall correlation coefficient (KCC) and model performance. Finally, the selected feature variables in this way
Spearman correlation coefficient (SCC). Given existing an n-dimensional are both significant and free of multicollinearity. Following this part, the
sample (X, Y), where X represents a vector of features and Y represents a new feature variables selected from the SEM can then be combined with
vector of target variables (i.e., SOC in this paper). xi and yi represent data the original variables (i.e., voltage, current and temperature) to
from one sampling point, i.e., the feature value and SOC value of the i-th construct the time series together for input to the DL models.
sample, respectively.
PCC is used to measure the degree of linear correlation between two
2.4. Deep learning models
variables and can be defined as the ratio of the covariance and standard
deviation of the two vectors. It can be calculated by the following
DL models can be used to effectively construct mappings between
equation:
input variables and battery SOC. Time series prediction, especially SOC
∑nr estimation, can be well approached by using RNN models [42]. Our
(xi − x)(yi − y)
Cov(X, Y) previous studies have shown that LSTM, as a typical RNN, works better
i=1
PCC = √̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅ = √̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅ (4)
Var(X)Var(Y) ∑nr ∑nr than other models [16]. Therefore, this paper constructs DL models
(xi − x)2 (yi − y)2
i=1 i=1 based on LSTM, which possesses universality. In addition, for different
types of problems, other DL models can surely be used as a basis, such as
where Cov(⋅) and Var(⋅) represent covariance and variance respectively. multi-layer perceptron (MLP) [43] and CNN [44]. Considering the EL
x and y are the average. When there is a linear relationship between the framework proposed in this paper, different techniques are used to

5
H. Yu et al. Journal of Energy Storage 73 (2023) 108915

further improve the LSTM. Here, BO, AM and CNN are employed to optimization interval, and d* is the optimal combination. J(⋅) is the loss
improve the original model in terms of hyperparameter optimization, function, which is identified here as mean square error (MSE). In sum­
input feature importance and further feature extraction, respectively. mary, the hyperparameter selection process of BO-LSTM reduces
training costs and ensures that the final model has better performance.
2.4.1. LSTM
The LSTM network is a special type of RNN designed to solve the 2.4.3. AM-LSTM
gradient disappearance problem that exists in long sequence data. For DL models, different inputs tend to carry different amounts of
Compared to traditional RNNs, the LSTM uses a gate mechanism to information, i.e., different importance. Placing more attention on vari­
effectively control the information flow and the forgetting mechanism, ables that possess high importance allows the information contained in
thus better capturing the dependencies between time series. In an LSTM the inputs to be more fully exploited [48]. AM can be employed to
unit, three gates control the update of the unit state and the hidden state, extract important information from the input data to help LSTM better
namely the input gate, forget gate and output gate. At time t, each unit capture temporal relationships and ultimately achieve accurate pre­
receives input xt and the hidden state ht− 1 from the previous moment as dictions [49]. In essence, AM is an adaptive gating mechanism that
input. The input gate uses a sigmoid activation function to select which learns and integrates the attention levels corresponding to different
information should be added to the unit state. The forget gate selects variables during the model training process. Specifically, the input
which information should be removed from the unit state by a sigmoid variables are weighted at each sampling point so that the impact of
activation function. The output gate determines which part of the unit different variables on the model output can be accurately distinguished
state should be output and can be seen as implemented by a sigmoid and the model can automatically enhance useful features and reduce
function plus a tanh function. Within an LSTM unit, the individual states useless ones. By doing this, the impact of different inputs on the pre­
are calculated as follows: diction results of the model can be quantified simultaneously, thus
( ) increasing the interpretability of the DL model.
ft = σ Wf xt + Uf ht− 1 + bf (8)
The specific process of the AM-LSTM model is that after receiving the
input, each variable is multiplied by a number before being inputted into
it = σ(Wi xt + Ui ht− 1 + bi ) (9)
the LSTM model. Here, the number multiplied by each variable is a
parameter in the model that needs to be trained for optimization. At
ot = σ(Wo xt + Uo ht− 1 + bo ) (10)
each time step, an attention function is used to calculate the weights of
(
gt = tanh Wg xt + Ug ht− 1 + bg
)
(11) each feature and a weighted sum is applied to obtain the final repre­
sentation of the variables inputted into the LSTM. Specifically, the
ct = ft ⊗ ct− 1 + it ⊗ gt (12) attention function is calculated in three steps: first, the attention score is
calculated, using the dot product; then, the score is normalized to obtain
ht = ot ⊗ tanh(ct ) (13) the attention weight of each variable commonly by using the softmax
activation function; finally, the input vectors are weighted and summed
where W and U are the weights obtained by training, and b is the bias of according to the attention weights to obtain the fused variable repre­
every gate. The subscripts of W, U and b represent different layers. ⊗ sentation. It is to be noted that in an AM-LSTM model, the AM and the
represents the element-wise multiplication. LSTM model are independent of each other and can adopt different
combination relationships to adapt to different tasks [49,50]. In sum­
2.4.2. BO-LSTM mary, AM can automatically enhance the information in the input data
The different combinations of hyperparameters can directly affect to make the constructed model more effective.
the performance of a network model, including prediction accuracy and
computational cost [17]. The BO algorithm is an optimization method 2.4.4. CNN-LSTM
based on the Bayes theorem that can be used to optimize the hyper­ Under normal conditions, features for DL cannot be easily extracted,
parameters of a neural network to improve the model performance. This which leads to poor model performances. Therefore, the LSTM can be
avoids the inaccuracy and complexity of manual tuning as well as the further improved from the perspective of feature extraction. Here, the
parametric problems that exist in metaheuristic optimization algorithms CNN is chosen as the front layer of the LSTM due to its ability to excel in
[45,46]. According to existing research, the BO algorithm can converge extracting potential features from sample data through convolutional
within a minimum number of iterations [47], so it is chosen here for the kernels, i.e., constituting the CNN-LSTM model, where the CNN is used
optimization of model hyperparameters. to extract features from the data and the LSTM is used to learn time-
The core thought of BO is to optimize based on a priori and posterior dependent relationships. The introduction of CNN can significantly
distribution. In hyperparameter optimization of neural networks, each reduce the number of model parameters, speed up training, improve
sample corresponds to an objective value, i.e., the loss function. BO uses prediction accuracy and improve generalization performance [51]. As
the results of the loss function to gradually adjust the hyperparameters battery SOC estimation is a one-dimensional time series problem, one-
so that the network performs better on the training samples. Since the dimensional convolution is used in the convolution layer. In summary,
probabilistic agent model of the objective function is assumed to be the CNN-LSTM model can efficiently process sequential data and extract
consistent with the Gaussian process, the BO algorithm observes the time series information, resulting in accurate and fast predictions with
prior probabilities of the hyperparameters through the Gaussian process strong generalization capability.
model and updates the posterior distribution. The Gaussian process
model is then continuously updated based on new results for each 2.4.5. Training of the DL models
objective function. Then, a set of the most desirable hyperparameters is After the time series are constructed, the DL models can be trained to
selected for training and model performance is recorded until the determine the optimized set of model parameters by minimizing the loss
specified requirements are met. The relationship between the optimal function. In this paper, the root mean square propagation (RMSprop)
hyperparameter combination and the objective function is presented optimizer is adopted to train the model, which is a technique with good
below: convergence speed and convergence properties [52]. The loss function is
set as the mean squared error (MSE) of the predicted result with respect
d* = argminJ(d), d ∈ (0, D) (14) to the true value. The simple yet effective dropout algorithm [53] is also
used to avoid model overfitting during the training process. Specifically,
where d represents a hyperparameter combination, D represents the
a neuron in an LSTM layer has a probability of p to be randomly disabled

6
H. Yu et al. Journal of Energy Storage 73 (2023) 108915

during training by setting its input to zero and leaving it unchanged Table 2
during the testing phase. This algorithm reduces the complex co- Specifications of the tested batteries.
adaptive relationships between neurons, thus further improving the Battery type LG 18650HG2 Panasonic 18650PF
generalization of the network model.
Cathode material NMC NCA
In addition, the hyperparameters can be determined comprehen­ Anode material Graphite + SiO Graphite
sively based on the results of multiple BO optimization. Herein, the Nominal capacity (Ah) 3.0 2.9
number of random searches is 10 and the number of BO iterations is 20. Nominal voltage (V) 3.6 3.6
In the time series construction, the sequence length N and batch size M
are set to 15 and 128, respectively. The number of epochs is set to 200,
Madison [57]. The test conditions for both types of batteries are similar
which means that all training samples are repeated at most 200 times.
in that they refer to several standard drive cycle profiles such as US06,
Also, the early stopping technology is used to ensure that training is
HWFET, UDDS and LA92. All profiles are obtained by calculating from
stopped when the monitoring metric stops improving. This technique
real electric vehicles and scaling according to the battery pack structure.
can also be used to combat overfitting when using iterative optimization
It should be noted that a current under zero in the dataset indicates
methods. During the complete training process, the data in the test set is
discharge and above zero indicates charging. Furthermore, as a
not involved in the optimization of the model parameters. In the LSTM
groundbreaking piece of research, this paper does not verify the validity
model, there are three main types of hyperparameters that need to be
of the proposed framework for different ambient temperatures, so the
determined, namely the learning rate, the number of nodes and the
data selected here are all under test conditions at an ambient tempera­
dropout rate. Based on the optimization results, the learning rate is set to
ture of 25 ◦ C. In fact, it is simple to consider the temperature. The SEM
0.0003; the number of nodes in the first layer LSTM is set to 32; the
developed in this paper maintains the extensibility that the
number of nodes in the second layer LSTM is set to 16; all dropout rates
electrochemical-thermal coupling model can be constructed by coupling
are set to 0.3.
SEM and thermal model. And, similar treatment has been carried out in
the study that has been done [17]. In addition, to ensure consistency and
2.5. Ensemble estimation method
stability in the training of the DL model, the data was preprocessed so
that all data were kept at a sampling rate of 1 Hz, and inputted into the
To address the instability of single model prediction performance,
DL model after a maximum-minimum normalization.
this paper uses an EL approach to further integrate the results of multiple
Specifically, for DL, a training dataset and a test dataset should be
learners to obtain more accurate and robust battery SOC estimation
constructed from the measurement dataset. First, to verify the general­
results. EL is a meta-learning technique, which improves generalization
ization ability to the different conditions for the same battery material
and adaptability through the properties and performance of the algo­
system, both the training and test datasets were constructed to be
rithm itself. Specifically, EL improves the accuracy and robustness of a
generated from the NMC battery. The UDDS condition as shown in Fig. 2
model by incorporating the results of multiple base learners, i.e., meta
(a) was used for training. The LA92 as shown in Fig. 2(b) was used for
features, to train the meta learner. To obtain better prediction perfor­
testing. Each row in the figure shows in turn, from left to right, the
mance, a second-level meta learner is introduced to automatically
voltage, current, temperature and SOC obtained from testing in one
determine the weights of each base model in order to achieve the inte­
condition. Here the reference values for SOC are calculated using the
gration of the results. In addition, to avoid overly complex integration
ampere-hour integral method. The high accuracy of the test equipment
algorithms, the simple and effective blending method is chosen, which
and the experimentally calibrated initial SOC values ensure that the
also offers good interpretability [54].
reference values of the calculated SOC are valid. Further, in order to
In the first stage of offline training, the learning for the first layer of
verify the extremely strong generalization ability of the proposed
base learners is performed based on the input features obtained, i.e.,
framework, further experiments were carried out for different types of
voltage, current, temperature and the features selected from SEM. Then,
batteries. The training dataset was still UDDS data obtained from the
in the second stage, the predictions from the training set are stacked in
NMC battery. The test dataset used US06 data obtained from the NCA
parallel as the meta feature set as the input to the meta learner in the
battery, as shown in Fig. 2(c). Herein, the training and testing datasets
second layer. Finally, considering the multicollinearity property among
are different in terms of conditions and battery material systems.
the inputs, KRR is selected as the meta learner in the second layer after
several comparisons [55], which is further demonstrated and discussed
in Subsection 4.3. It is important to note that the trained base learners 3.2. Acquisition of enhanced data
and meta learners are saved for quick use online. During the online
application phase, the measurement data are inputted into the SEM to In our proposed PIEDL framework, some of the variables from the
obtain battery mechanism-related variables containing physical infor­ battery SEM output containing mechanism-dependent physical infor­
mation. Then, the measured data and the feature variables selected from mation are inputted into DL models as enhanced input data. The SEM
the SEM are inputted into three base models, i.e., BO-LSTM, AM-LSTM has been verified in our previous studies [33,37]. Herein, the collected
and CNN-LSTM, to obtain three SOC predictions. Finally, these three experimental data can be used to slightly evaluate the voltage response
predictions are inputted into the meta learner and the final results of the of the SEM. It is clear from Fig. S1 that the voltage variations from the
battery SOC estimation can be obtained online. SEM are in high agreement with the experimental data generated by the
dynamic test conditions. Despite some deviations, the battery SEM has
3. Datasets and data acquisition captured the main trends in voltage variation and these deviations are
within acceptable ranges. Therefore, the SEM can be used to effectively
3.1. The selection and construction of datasets simulate the operating behavior of the battery and the simulation results
can be further used for subsequent studies. It is important to note that
To validate the performance of the proposed framework, batteries of the model errors are not within the scope of this paper. In addition,
two different material systems were used in this paper, i.e., their cathode reference values for the parameters required for the model simulation
materials are NMC and NCA respectively. The detailed battery basic can be found in the references [33, 36]. In this paper, the fitted SEM
information is presented in Table 2. The datasets used are all open parameter set is obtained by further fine-tuning some parameters.
source. The LG 18650HG2 battery with NMC as the cathode material The SEM simulation results of the battery internal characteristics are
was tested at McMaster University [56]; the Panasonic 18650PF battery shown in Fig. 3, taking the UDDS condition of the training dataset as an
with NCA as the cathode material was tested at University of Wisconsin- example. Fig. 3(a) shows the OCV curve obtained from the simulation,

7
H. Yu et al. Journal of Energy Storage 73 (2023) 108915

Fig. 2. Battery datasets for training and testing: (a) UDDS for NMC battery; (b) LA92 for NMC battery; (c) US06 for NCA battery. Each row in the figure shows the
voltage, current, temperature and SOC in turn, from left to right, obtained from testing in one condition.

Fig. 3. SEM simulation results of the battery internal characteristics: (a) OCVs; (b) OCVc; (c) Δx; (d) Δy; (e) ηact; (f) ηcon; (g) ηohm; (h) Rx; (i) Ry; (j) Ract; (k) Rcon;
(l) Rohm.

namely OCVs. In fact, it should not be used as an input to DL models, the disadvantages of the ampere-hour integral method. Therefore, this
because the OCV in the simulation is originally obtained by its rela­ paper calculates the corresponding OCV value, namely OCVc, through
tionship with the SOC-related variables [36]. At this time, using it again the formula of terminal voltage, and the result is shown in Fig. 3(b).
as an input to predict SOC through DL models does not reasonably avoid Since the terminal voltage is OCV minus all overpotentials, the

8
H. Yu et al. Journal of Energy Storage 73 (2023) 108915

calculated OCV is the measured terminal voltage plus the overpotentials representativeness can be selected, thus improving the accuracy of DL
from the simulation. This avoids a direct link between OCV and SOC and models. The CC method based on statistical principles gives three results
improves the effectiveness and robustness of the models. The deviations depending on the calculation method: PCC, KCC and SCC. Based on the
of the particle surface concentration from the average concentration for MI method of information theory principles, the corresponding coeffi­
the negative and positive electrodes can then be obtained as Δx and Δy, cient MIC can be obtained after calculation. These four results for the 12
respectively, as shown in Fig. 3(c) and (d). The simulation results for the variables extracted from the SEM are shown in Fig. 4. The results for
reaction polarization overpotential ηact, concentration difference polar­ both the OCVs obtained from the simulation and the OCVc calculated are
ization overpotential ηcon and ohmic polarization overpotential ηohm are very good and both have high correlations with the battery SOC. The
shown in Fig. 3(e), (f) and (g), respectively. These process components results for the CCs are close to 1 and the MICs are around 3. These
within the battery describe the response of the battery to SOC variations indicate that both OCVs and OCVc have a very high relationship with the
and current excitation at different time scales, which are likely to SOC. This is due to the inherent function relationship that exists between
contain physical information of relatively high relevance to SOC. the SOC and OCV of LIBs. However, the relationship between SOC and
Furthermore, the response within the battery is influenced by current, OCV is not fixed and it can be influenced by a number of factors [58],
and there is no direct relationship between the battery SOC and which makes it unreliable to use this relationship to look up tables
instantaneous current values. Therefore, by dividing the variables of directly to obtain SOC. Therefore, it is not enough to adopt only one
internal battery processes such as solid phase diffusion by the corre­ function. However, it is not practical to calibrate the battery SOC-OCV
sponding current values, the corresponding resistance values can be functions in all cases. This would require a large number of experi­
obtained. In particular, for solid phase diffusion processes, the variables ments and could not cover all cases. Therefore, the method proposed in
Δx and Δy, which react to their extent, are both dimensionless numbers. this paper can be used to learn from big data using DL in order to
However, the results obtained from dividing them by the current can minimize the effect of noise and make the battery SOC estimation more
still be considered to have the corresponding physical meaning of accurate, given that the SOC-OCV curve is not very accurate.
resistance. In terms of electromagnetism, the unit of current is Ampere As can be seen from the results, the correlation of OCVs is better than
(A), which is used to describe the amount of charge passing through the that of OCVc. This is because the formula of OCVs simulated by the model
cross section of a wire per unit of time. Hence, the reciprocal of Ampere, is directly related to SOC, while OCVc obtained from the calculation
i.e., Ampere− 1 (A− 1), can be understood as a unit describing the resis­ takes on some noise from other formulas in the model. However, for the
tance of a wire, indicating the ability of a substance to resist the passage reasons mentioned in Subsection 3.2, the OCV obtained by model
of current, and is used to describe the electrical properties of a simulation is obtained by directly looking up the table of SOC-related
conductor. It should be noted that when the current is 0, dividing by the variables. There is therefore a direct relationship between this variable
current is meaningless. Therefore, the resistance value in this case is set and the SOC, which is calculated using the ampere-hour integral
to 0. This can ensure the real time of the model while minimizing invalid method. In this way, the disadvantages of traditional methods cannot be
information. The simulation results for the resistance values for each avoided. Therefore, the input to DL models should not be OCVs, but
process are shown in Fig. 3(h)–(l). rather OCVc, which will improve the effectiveness and robustness of the
models. It should be noted that even if the correlation between OCV and
4. Results and discussion battery SOC is high, it cannot be used directly to estimate battery SOC.
This is because the relationship is not fixed and can be influenced by
In this section, feature variable selection results are first shown, many factors [58]. This makes using this relationship to look up tables
which identify high-quality variables in the SEM. Then, considering directly to obtain SOCs unreliable, so a DL approach to big data is
whether and how many variables containing physical information are needed to minimize the impact of noise. Furthermore, for the variable,
added as one factor, and whether the LSTM is improved and further electrochemical polarization resistance Ract, the CCs show better results
integrated as another factor, the battery SOC estimation results are overall and the MIC is also significantly larger than variables except for
presented comprehensively. Next, the effects of different meta learners OCV, which indicates that there is also a correlation between Ract and
are investigated and the extremely strong generalization of the trained SOC. The polarization process of an electrochemical reaction is
model based on the PIEDL framework to different types of batteries is controlled by the Butler-Volmer equation, where the electrode compo­
verified. Finally, discussions and outlooks are presented. sition, charge state and current influence its overpotential behavior
[59]. The overpotential is divided by the current to obtain the corre­
4.1. Feature variable selection results sponding resistance to remove the influence of the current. The char­
acteristics of the electrodes are fixed for a given battery. Therefore, there
In this subsection, by analyzing the correlation between the target is some correlation between Ract and SOC, which can be a potentially
variable and different feature variables, the feature variables with strong competitor for the input variable. However, it is less correlated

Fig. 4. The coefficient calculation results of the variables in SEM for selecting feature variables.

9
H. Yu et al. Journal of Energy Storage 73 (2023) 108915

than OCV due to the fact that the relationship between Ract and battery effects on it. This is also the case in other relevant studies [20]. The
SOC is too complex for the model to learn the information in it. Except PIEDL framework can be divided into two aspects: the fusion of physical
for OCV and Ract, all of the variables extracted from the SEM have CCs information and the EL for integrating DL models. Specifically, the
less than 0.3. The MICs are also significantly lower than the selected PIEDL first performs an initial estimation of the battery SOC by adding
variables, with values of almost 0. Thus, they all barely correlate with features containing physical information and then integrates the results
the SOC. Diffusion processes, both in the solid and liquid phases, tend to with multiple DL models to obtain more accurate battery state estima­
be dependent on the lithium-ion concentration and require a certain tion. As can be seen in Fig. 5, in the horizontal direction, we compare the
response time, with a large time lag. The ohmic process comes mainly results of DL models using the original inputs and the inputs added OCV,
from the contribution of the contact resistance and intrinsic resistance of Ract, and OCV and Ract, respectively. It can be concluded that the pro­
the porous electrodes and is almost constant with SOC [60]. In conclu­ posed estimation based on physical information can track the changes in
sion, after screening, Ract and the calculated OCVc can serve as feature SOC well, and the result is better after adding all relevant variables. For
variables for the DL models. all models, i.e., LSTM, BO-LSTM, AM-LSTM, CNN-LSTM and EDL, the
After correlation analysis, alternative variables are further added to prediction results are improved with OCV or Ract added to the input.
the model by the forward stepwise regression, to determine the effec­ And, the addition of both variables results in a more accurate battery
tiveness of the selected variables. Based on the original LSTM model, the SOC estimation than the addition of single features. In the longitudinal
results of inputting only voltage, current and temperature are used as a direction, for all inputs, i.e., original input, added OCV, added Ract, and
baseline, and new variables are gradually added to observe the change added OCV and Ract, BO, AM and CNN can all improve the prediction of
in the performance of the model. Both the training and test datasets used the LSTM, and the best prediction result can be achieved after EL.
here are from the NMC battery. It should be noted that the OCV in the Moreover, the EL is equivalent to combining multiple results, which can
test refers to the results from the calculation. For the reason explained effectively reduce errors and achieve further performance improve­
above, the simulated OCV should not be used as the input. Hence, there ments over each single model. These findings can be also derived from
is no further experiment for it. The results of stepwise addition regres­ Table 4. The PIEDL framework proposed in this paper has the best
sion for battery SOC estimation are shown in Fig. S2. The specific error prediction results, i.e., EDL #4 in Fig. 5. The fluctuations in MAE are
statistics are summarized in Table 3. It can be seen that the selected mainly caused by randomness. Specifically, in comparison with the
variables OCV and Ract both significantly improve prediction perfor­ LSTM model using the original input, the PIEDL enables the following
mance when they are added to the model input, with the addition of improvements in battery SOC estimation results: AAE is from 1.57 % to
OCV being particularly effective. Thus, both variables contain physical 0.58 %; RMSE is from 1.89 % to 0.76 %; MAE is from 5.85 % to 2.31 %.
information relevant to battery SOC and can lead to more accurate SOC The error relative reductions are 63 %, 60 % and 61 %, respectively,
estimation results. For AAE and RMSE, the results are better than adding validating the excellent effectiveness of the framework in improving the
none new input variables only when these two variables are added. In all accuracy of battery SOC estimation.
other cases, the results become worse. However, it is not only the The real-time of the PIEDL can also be discussed. For the SEM
addition of OCV and Ract that can make the MAE smaller. The SOC simulation, owing to a series of simplifications, the time required to
estimation MAEs are also slightly reduced when Δx, Δy, ηact and ηohm are simulate a time step is negligibly short [61]. Moreover, for existing
added, probably due to the fact that the MAE are very sensitive to out­ models, after adding OCV and Ract to the model input, there is no in­
liers. This slight reduction in error does not have sufficient reference crease in computational resources due to the low dimensionality of the
value. original inputs and the low complexity of the model. Therefore, the
calculation for physical information only adds a small amount of
computational cost [62]. The time consumed for DL prediction is also
4.2. SOC estimation results
little [63]. Thus, our method can obtain accurate battery state estima­
tion results in a short time based on the trained model, thus meeting the
Considering the applicability of the PIEDL framework proposed in
real-time requirement of electric vehicles. To sum up, the PIEDL
this paper for battery SOC estimation, the dataset of NMC batteries was
framework can significantly improve the accuracy of battery SOC esti­
used to verify the accuracy. The results of the battery SOC estimation
mation in terms of both physical information and integration of multiple
under various conditions are shown in Fig. 5, where #1 indicates that
DLs.
the model uses the original inputs, i.e., voltage, current and tempera­
The weights that AM assigns to each input can be used to analyze the
ture; #2 indicates that the feature variable OCV is added to the model
contribution of single input features to the model, which enhances the
input; #3 indicates that the feature variable Ract is added; #4 indicates
interpretability of the DL model. Specifically, the AM-LSTM model can
that OCV and Ract are added together. Statistical errors for all cases are
adaptively weigh the input features to improve the model performance
shown in Table 4, and their visual results are shown in Fig. S3. It should
and efficiency. All 11 variables extracted from the SEM (the simulated
be noted that there are some results missing at the beginning and end of
OCV is not used as input) that can be used as features are added to the
all the predicted results. This is because the RNN-type models often use
inputs and the AM-LSTM is used to obtain the weights of each variable
sliding windows to generate input and output sequences in the process of
while predicting the SOC results. The SOC estimation results for all the
time series prediction. In this process, for parts of the data at the
newly obtained variables added to the AM-LSTM model are shown in
beginning and end of the dataset, predictions cannot be made because
Fig. 6(a). It is clear from the results that adding all the variables to the
there is not enough prior or later data to generate a complete sequence of
input obtains a worse prediction result than only adding OCV and Ract.
inputs and outputs. Therefore, when constructing time series, in order to
Therefore, inputting irrelevant variables into DL models not only in­
ensure the integrity of input and output, these parts of data are usually
creases the computational cost but also leads to poorer model perfor­
directly eliminated. In the actual application process, because the data is
mance. The weight contributions of each variable after normalization
continuously generated and the time scale is large, it will not cause bad

Table 3
Battery SOC estimation error of stepwise variable addition.
Added variable None OCV Δx Δy ηact ηcon ηohm Rx Ry Ract Rcon Rohm

AAE (%) 1.57 0.99 1.67 1.76 1.82 1.84 1.99 2.03 2.11 1.40 2.17 1.99
RMSE (%) 1.89 1.25 2.06 2.14 2.24 2.33 2.31 2.54 2.51 1.70 2.75 2.38
MAE (%) 5.85 4.41 5.50 5.63 5.78 7.77 5.43 7.41 6.44 4.67 7.75 6.00

10
H. Yu et al. Journal of Energy Storage 73 (2023) 108915

Fig. 5. SOC estimation results and absolute errors of all the experiments. #1, #2, #3 and #4 indicate that the model inputs are original, with OCV added, with Ract
added, and with OCV and Ract added, respectively.

11
H. Yu et al. Journal of Energy Storage 73 (2023) 108915

Table 4 conclusions to some extent.

Statistical results of SOC estimation error for all cases. #1, #2, #3 and #4
indicate that the model inputs are original, with OCV added, with Ract added,
4.3. Comparison of different meta learner models
and with OCV and Ract added, respectively.
#1 #2 #3 #4 It is important to select a suitable second-level meta learner to
Metrics: AAE (%) integrate the base model prediction results, which can also have a sig­
LSTM 1.57 0.99 1.40 0.87 nificant impact on the EL result. A number of commonly used regression
BO-LSTM 1.44 0.92 1.24 0.81
models are chosen as options for the meta learner to investigate the
AM-LSTM 1.28 0.88 1.25 0.76
CNN-LSTM 1.27 0.84 1.11 0.75
specific impact of the learner on integration performance, including
EDL 1.09 0.69 1.01 0.58 adaptive boosting (AdaBoost), eXtreme gradient boosting (XGBoost),
Metrics: RMSE (%) RF, k-nearest neighbors regression (KNR), support vector regression
LSTM 1.89 1.25 1.70 1.10 (SVR), kernel ridge regression (KRR), MLP, GRU and LSTM. The deter­
BO-LSTM 1.75 1.16 1.57 1.02
mination of hyperparameters is manually adjusted according to the
AM-LSTM 1.67 1.18 1.56 1.01
CNN-LSTM 1.59 1.05 1.40 0.97 prediction results of randomly selected hyperparameters. The perfor­
EDL 1.36 0.88 1.23 0.75 mances of different meta learners are tested in the PIEDL framework (i.
Metrics: MAE (%) e., EDL #4). The SOC estimation errors on the NMC battery dataset are
LSTM 5.85 4.41 4.67 3.71 given in Fig. 7. After comparison, it can be seen that KRR has the best
BO-LSTM 4.29 3.65 6.43 2.92
AM-LSTM 5.28 4.39 5.18 3.62
results for all three metrics, i.e., AEE, RMSE and MAE. Interestingly,
CNN-LSTM 5.64 2.83 3.95 2.50 GRU and LSTM, which belong to RNN, do not perform well. The direct
EDL 3.79 2.46 3.43 2.36 reason is due to the mismatch between data complexity and model ca­
pacity [25]. The prediction results of base learners are in general close to
the actual battery SOC, and the meta learning in EL has less training
obtained from AM are shown in Fig. 6(b). It can be found that the
data. Hence, the low complexity and the small data volume of the
contribution of the selected features is relatively high, with OCV being
dataset is not suitable for learning with an overly complex model that
the highest. It indicates that the selected features have an important
would otherwise produce large overfitting. And, the LSTM with more
impact on the model performance, and confirms the previous
complexity than GRU has worse performance, indicating that overly

Fig. 6. (a) SOC estimation results of AM-LSTM with all the variables obtained from SEM added to the input. (b) The weight contribution of each variable after
normalization obtained from AM.

Fig. 7. SOC estimation error results of different meta learners with PIEDL framework.

12
H. Yu et al. Journal of Energy Storage 73 (2023) 108915

complex models are not suitable for use as the meta learner. In addition, This method sacrifices unbiasedness to reduce the variance of the esti­
given the random nature of ML in the base model, the prediction process mation and enhance the stability of the estimation. Moreover, regula­
of the meta learner would not be suitable to be considered as a time rization methods have been shown to enhance stacked generalization
series problem, otherwise the model would have difficulty capturing the [65] and can achieve similar results to the EDL proposed in this paper.
patterns in the data. For other regression algorithms, it can also be seen Further, KRR better copes with multicollinearity by introducing a kernel
by the experimental results that the performance is not as good as KRR. function that maps the samples from the original space to the higher
Moreover, we have also tested different meta learners for the EL model dimensional space [66].
under the original inputs (i.e., EDL #1), and the error results are shown
in Table S1. The same conclusions can be reached to further validate the 4.4. Generalization of trained model for different battery types
reliability and generalizability of the above conclusions. In summary,
KRR has the best SOC estimation result as meta learner, and is selected The above SOC estimation results are obtained based on experi­
as the meta learner in the PIEDL framework. mental data from the same NMC battery. However, different types of
As can also be found from the results in Fig. 7, although the results batteries can have different properties, such as different materials, ca­
are better on the whole after integration, the results are different when pacities and shapes. And, it is obviously expensive to test all types of
using different models as meta learner. Moreover, sometimes the results batteries. Moreover, for an unknown new battery, sometimes there is not
become worse after integration. These phenomena are mainly caused by necessarily the opportunity and cost to conduct detailed testing.
multicollinearity, which refers to the existence of a high degree of cor­ Therefore, the generalization capability of the model prediction needs to
relation between multiple independent variables. In ML, multi­ be improved so that it can be trained to be applicable to a wider range of
collinearity can reduce the reliability and stability of the model, thus scenarios. In particular, for batteries of different material systems, the
affecting its estimation performance. The problem of multicollinearity is differences in electrochemical properties make the application of the
particularly prominent in EL. This is because the predictions of the base model especially difficult. Here, data from two publicly available data­
learners are all close to the true values of the target variable. Hence, sets are used as training and test sets, respectively, from batteries pro­
there is bound to be multicollinearity between the outputs of multiple duced by two different manufacturers. The batteries differ in cathode
base learners. Verification of multicollinearity is carried out by means of material, anode material and nominal capacity, and their specifications
the variance inflation factor (VIF) and correlation analysis. The VIF of can be accessed in Table 2. The NMC battery is used for training and the
the input variable Xi (i.e., the predicted outcome of BO-LSTM, AM-LSTM NCA battery for testing. In this way, we have verified for the first time
or CNN-LSTM) for a given meta learner can be calculated by the the ability of the trained PIEDL to generalize across different types of
following equation [64]: batteries.
The experimental error results are shown in Fig. 8. The prediction
1
VIF = (15) result of the LSTM model with only voltage, current and temperature as
1 − R2i
inputs (i.e., LSTM #1) is poor, which demonstrates that predicting the
where Ri is the coefficient of determination of a linear regression model SOC of different types of batteries is indeed very challenging. The
with Xi as the dependent variable and the input variables of meta learner addition of physical information (i.e., OCV and Ract) and the use of the
other than Xi as the independent variables. It is generally accepted that EDL framework for the models can each significantly reduce SOC error.
there is no multicollinearity when the VIF is less than 10, strong mul­ Moreover, the physical information improves the model performance
ticollinearity when the VIF is between 10 and 100, and severe multi­ more. Ultimately, the PIEDL proposed in this paper (i.e., EDL #4) ach­
collinearity when the VIF is greater than 100. In the framework of EL, ieves better SOC estimation results. AAE and RMSE are both within 2 %
the VIF calculation results for each meta feature of the meta learner for and MAE is less than 3.5 %, which meets the basic requirements for
four different input cases are shown in Table 5. It can be seen that all the battery SOC estimation accuracy in the industry. This result validates
results are greater than 400, indicating that there is severe multi­ the ability of PIEDL to generalize to battery types and indicates that the
collinearity between all the meta features of the meta learner. Mean­ model has better reliability in practical applications.
while, it can be observed that the multicollinearity is the most severe In existing battery state estimation methods, the generalizability
when both SOC-related variables with physical information are added, validation of different battery types is obtained by retraining the model
and far exceeds the results of the other three input cases. This also in­ [28,67], which does not really contribute much to the practical appli­
dicates that the predictions of the base model are closer to the true cations for different battery types. Recently, Ruan et al. [44] have used
values of each other at this time, i.e., the DL performance can be transfer learning to achieve more satisfactory estimation results with
improved by adding physical information. For the correlation analysis only a small amount of data. However, the model is still trained for a
method, herein, the PCC is adopted for the analysis, and the results are small amount of time during the use of different types of batteries. The
shown in Table S2. All results are greater than 0.998, which also shows PIEDL proposed in this paper can achieve similar results without further
that there is a strong correlation between the variables, i.e., very strong training. This is mainly due to the fact that the added physical infor­
multicollinearity. mation and the use of the EDL framework can make better use of the
KRR stands out exactly because of its superior ability to resolve the physicochemical properties of LIBs and the historical data, and incor­
multicollinearity problem. Ridge regression (RR) reduces the effect of porate the predictions of multiple DL models. Ultimately, good accuracy,
covariance by adding a regularization term to reduce the over­ robustness and generalization are achieved, allowing this model to be
dependence on the independent variables, while retaining all features. widely used in different types of batteries and providing a new solution
for battery management.

Table 5 4.5. Discussions and outlook

VIF calculation results for each meta feature of the meta learner for four different
input cases. BO-LSTM, AM-LSTM and CNN-LSTM respectively represent the
DL methods have been extensively used, yet the accuracy, robust­
meta feature of the meta learner, corresponding to the prediction result of the
ness, stability and generalization of the target task can be further
base model as the input to the meta learner.
improved due to the lack of physical information in the data and the
Added variable None OCV Ract OCV&Ract
limitations of a single model. The PIEDL framework proposed in this
BO-LSTM 445 914 410 1250 paper can improve performance from both perspectives by introducing
AM-LSTM 1854 746 894 4220 physical information and the structure of integrating multiple single
CNN-LSTM 1254 963 504 3148
models. Herein, the effectiveness of the PIEDL framework is validated

13
H. Yu et al. Journal of Energy Storage 73 (2023) 108915

Fig. 8. SOC estimation error results of the training set and the test set using the data of different types of batteries. #1 indicates that the model uses the original
inputs, and #4 indicates that OCV and Ract are added to the model input together.

using the SOC estimation of batteries as the target task. However, the deeper prospect.
framework is not limited to this, but applies to a variety of ML tasks with
high applicability and flexibility. The physical model, the base learners 5. Conclusions
and the meta learner in EL can all be tailored to the specific application
scenario to suit the target task. Nevertheless, a minor drawback of the In this paper, a physics-informed ensemble deep learning (PIEDL)
PIEDL framework is its complexity due to the integration of multiple DL framework is proposed to incorporate physical information and inte­
models and mechanistic models. In the future, the demand for compu­ grate multiple models, with a focus on battery state of charge (SOC)
tational resources can be further reduced through model simplification estimation. The accuracy, robustness and generalization of the proposed
and structural optimization for online use. Computing power can also be framework are rigorously evaluated and validated. The introduced
developed through the development of technologies such as cloud physical information is obtained by a simplified electrochemical model
computing. (SEM) simulation, which enhances the quality of the input features for
Due to the extremely strong generalization capability of the proven deep learning (DL) models. Specifically, the open-circuit voltage (OCV)
PIEDL framework, it has become possible to develop a foundation mode derived from the measured terminal voltage, along with the simulated
with high versatility for battery SOC estimation based on the framework. overpotentials, and the reaction polarization resistance Ract are identi­
A foundation model can be trained and optimized based on large fied as high-quality features and integrated into the DL model's input.
amounts of data and ML algorithms that can be used to predict and Both individual and combined incorporation of these features can lead
optimize the performance and behavior of various systems [68]. For to improved SOC estimation, with the combined approach yielding the
battery SOC estimation, the foundation model can be developed through best results. To further enhance the predictive capability of the long
the physical features of the battery and a large amount of historical data short-term memory (LSTM) model, the ensemble deep learning
to suit all application scenarios (including battery type, temperature, approach is employed. Bayesian optimization (BO), attention mecha­
working condition and ageing). When applying the developed founda­ nism (AM), and convolutional neural network (CNN) techniques are
tion model, it can be used directly or fine-tuned to meet all application utilized to refine the LSTM predictions. And, the integration of these
scenarios. However, considering online application scenarios for PIEDL techniques yields improved results over the separate improvement
or even foundation models, simplifying models or increasing computing models. Overall, the proposed PIEDL framework demonstrates en­
power is also a key technical challenge. In particular, for the deployment hancements in both physical information utilization and model inte­
on the mobile side such as the vehicle side, end-cloud collaboration gration, and the combination of these aspects yields the best
technology can be used to achieve the high real-time application of performance. Compared to the prediction result of LSTM with the
complex models [28]. However, limited by some current technical original input, the PIEDL framework achieves a relative reduction of 63
bottlenecks, a whole-process toolchain should be provided for the entire %, 60 %, and 61 % in terms of average absolute error (AAE), root mean
development framework to enable large-scale and rapid deployment square error (RMSE) and maximum absolute error (MAE), respectively,
with limited hardware conditions. highlighting the significantly improved accuracy of battery SOC
In summary, the PIEDL framework encompasses two very promising estimation.
categories of technology: the fusion of artificial intelligence (AI) and Furthermore, the suitability of the kernel ridge regression (KRR)
science, and the paradigm construction of foundation models. Firstly, as model as the meta-learner is established by comparing various models
AI techniques continue to develop, they are also widely used in scientific and investigating the multicollinearity between the base models' out­
domains. At the same time, scientific computing can enhance the puts. The extremely strong generalization capability of the proposed
interpretability of AI due to its rigor. The PIEDL framework is a com­ framework is further verified using diverse data from different types of
bination of scientific knowledge and DL techniques to build more batteries for training and testing. Remarkably, the experiment results
effective data-driven models and provide new insights and methods for meet the fundamental accuracy requirements for battery SOC estima­
many areas including battery SOC estimation. Secondly, research tion. The PIEDL framework holds the potential for online application,
related to large-scale DL models is gradually growing in popularity as leveraging technologies such as cloud computing. Especially for the
the data size continues to increase and the computational power con­ mobile side, the computing framework of end-cloud collaboration
tinues to improve. This type of model has good performance and an emerges as an effective approach for implementation.
extremely wide range of applications, and the PIEDL framework can In summary, the PIEDL framework represents a significant contri­
serve as a basis for such research due to its improved model performance bution to the intersection of artificial intelligence and scientific
and strong generalization capability. With the development of technol­ research. It tackles the challenges of incorporating physical information
ogy and the expansion of application scenarios, the research and and consolidating multiple models. With continuous technological
application based on the PIEDL framework will have an even wider and innovation and expansion of application scenarios, the prospects for

14
H. Yu et al. Journal of Energy Storage 73 (2023) 108915

research and application of the PIEDL framework are expected to grow [12] X. Xu, S. Tang, H. Ren, X. Han, Y. Wu, L. Lu, X. Feng, C. Yu, J. Xie, M. Ouyang,
W. Liu, Y. Yan, Joint state estimation of lithium-ion batteries combining improved
even more profound and extensive.
equivalent circuit model with electrochemical mechanism and diffusion process,
Journal of Energy Storage 56 (2022), 106135, https://doi.org/10.1016/j.
CRediT authorship contribution statement est.2022.106135.
[13] R. Li, B. Xiong, S. Zhang, X. Zhang, Y. Li, H. Iu, T. Fernando, A novel one
dimensional convolutional neural network based data-driven vanadium redox flow
Hanqing Yu: Conceptualization, Methodology, Investigation, Vali­ battery modelling algorithm, Journal of Energy Storage 61 (2023), 106767,
dation, Visualization, Writing – original draft. Zhengjie Zhang: Meth­ https://doi.org/10.1016/j.est.2023.106767.
[14] X. Li, Z. Huang, J. Tian, Y. Tian, State-of-charge estimation tolerant of battery
odology, Investigation, Writing – review & editing. Kaiyi Yang: aging based on a physics-based model and an adaptive cubature Kalman filter,
Investigation, Validation. Lisheng Zhang: Conceptualization, Method­ Energy 220 (2021), 119767, https://doi.org/10.1016/j.energy.2021.119767.
ology. Wentao Wang: Validation. Shichun Yang: Supervision. Junfu [15] J. Chen, Y. Zhang, J. Wu, W. Cheng, Q. Zhu, SOC estimation for lithium-ion battery
using the LSTM-RNN with extended input and constrained output, Energy 262
Li: Investigation, Writing – review & editing. Xinhua Liu: Funding
(2023), 125375, https://doi.org/10.1016/j.energy.2022.125375.
acquisition. [16] B. Ma, S. Yang, L. Zhang, W. Wang, S. Chen, X. Yang, H. Xie, H. Yu, H. Wang,
X. Liu, Remaining useful life and state of health prediction for lithium batteries
based on differential thermal voltammetry and a deep-learning model, J. Power
Declaration of competing interest Sources 548 (2022), 232030, https://doi.org/10.1016/j.jpowsour.2022.232030.
[17] H. Pang, L. Wu, J. Liu, X. Liu, K. Liu, Physics-informed neural network approach for
heat generation rate estimation of lithium-ion battery under various driving
The authors declare that they have no known competing financial conditions, Journal of Energy Chemistry 78 (2023) 1–12, https://doi.org/10.1016/
interests or personal relationships that could have appeared to influence j.jechem.2022.11.036.
[18] J. Tian, C. Chen, W. Shen, F. Sun, R. Xiong, Deep learning framework for lithium-
the work reported in this paper. ion battery state of charge estimation: recent advances and future perspectives,
Energy Storage Materials 61 (2023), 102883, https://doi.org/10.1016/j.
Data availability ensm.2023.102883.
[19] B. Huang, J. Wang, Applications of physics-informed neural networks in power
systems - a review, IEEE Trans. Power Syst. 38 (2023) 572–588, https://doi.org/
Data will be made available on request. 10.1109/TPWRS.2022.3162473.
[20] J. Tian, R. Xiong, J. Lu, C. Chen, W. Shen, Battery state-of-charge estimation amid
dynamic usage with physics-informed deep learning, Energy Storage Materials 50
Acknowledgments (2022) 718–729, https://doi.org/10.1016/j.ensm.2022.06.007.
[21] Y. Wang, X. Han, D. Guo, L. Lu, Y. Chen, M. Ouyang, Physics-informed recurrent
This work was financially supported by the National Natural Science neural network with fractional-order gradients for state-of-charge estimation of
lithium-ion battery, IEEE J. Radio Freq. Identif. 6 (2022) 968–971, https://doi.org/
Foundation of China (No. 52102470). 10.1109/JRFID.2022.3211841.
[22] S. Kohtz, Y. Xu, Z. Zheng, P. Wang, Physics-informed machine learning model for
battery state of health prognostics using partial charging segments, Mech. Syst.
Appendix A. Supplementary data
Signal Process. 172 (2022), 109002, https://doi.org/10.1016/j.
ymssp.2022.109002.
Supplementary data to this article can be found online at https://doi. [23] R.G. Nascimento, M. Corbetta, C.S. Kulkarni, F.A.C. Viana, Hybrid physics-
org/10.1016/j.est.2023.108915. informed neural networks for lithium-ion battery modeling and prognosis, J. Power
Sources 513 (2021), 230526, https://doi.org/10.1016/j.jpowsour.2021.230526.
[24] S.W. Kim, E. Kwak, J.-H. Kim, K.-Y. Oh, S. Lee, Modeling and prediction of lithium-
References ion battery thermal runaway via multiphysics-informed neural network, Journal of
Energy Storage. 60 (2023), 106654, https://doi.org/10.1016/j.est.2023.106654.
[25] C. Lin, J. Xu, J. Hou, Y. Liang, X. Mei, Ensemble method with heterogeneous
[1] J.T. Frith, M.J. Lacey, U. Ulissi, A non-academic perspective on the future of
models for battery state-of-health estimation, IEEE Trans. Industr. Inform. (2023)
lithium-based batteries, Nat. Commun. 14 (2023) 420, https://doi.org/10.1038/
1–9, https://doi.org/10.1109/TII.2023.3240920.
s41467-023-35933-2.
[26] B. Gou, Y. Xu, X. Feng, An ensemble learning-based data-driven method for online
[2] X. Hu, Z. Deng, X. Lin, Y. Xie, R. Teodorescu, Research directions for next-
state-of-health estimation of lithium-ion batteries, IEEE Trans. Transp. Electrific. 7
generation battery management solutions in automotive applications, Renew.
(2021) 422–436, https://doi.org/10.1109/TTE.2020.3029295.
Sustain. Energy Rev. 152 (2021), 111695, https://doi.org/10.1016/j.
[27] J. Meng, L. Cai, D.-I. Stroe, J. Ma, G. Luo, R. Teodorescu, An optimized ensemble
rser.2021.111695.
learning framework for lithium-ion battery state of health estimation in energy
[3] H. Yu, L. Yang, L. Zhang, J. Li, X. Liu, Parallel battery pack charging strategy under
storage system, Energy 206 (2020), 118140, https://doi.org/10.1016/j.
various ambient temperatures based on minimum lithium plating overpotential
energy.2020.118140.
control, IScience 25 (2022), 104243, https://doi.org/10.1016/j.isci.2022.104243.
[28] B. Ma, L. Zhang, H. Yu, B. Zou, W. Wang, C. Zhang, S. Yang, X. Liu, End-cloud
[4] Z. Cui, L. Wang, Q. Li, K. Wang, A comprehensive review on the state of charge
collaboration method enables accurate state of health and remaining useful life
estimation for lithium-ion battery based on neural network, Intl J of Energy
online estimation in lithium-ion batteries, Journal of Energy Chemistry 82 (2023)
Research 46 (2022) 5423–5440, https://doi.org/10.1002/er.7545.
1–17, https://doi.org/10.1016/j.jechem.2023.02.052.
[5] Y. Tian, R. Lai, X. Li, L. Xiang, J. Tian, A combined method for state-of-charge
[29] A. Banerjee, A. Sarkar, S. Roy, P.K. Singh, R. Sarkar, COVID-19 chest X-ray
estimation for lithium-ion batteries using a long short-term memory network and
detection through blending ensemble of CNN snapshots, Biomedical Signal
an adaptive cubature Kalman filter, Appl. Energy 265 (2020), 114789, https://doi.
Processing and Control 78 (2022), 104000, https://doi.org/10.1016/j.
org/10.1016/j.apenergy.2020.114789.
bspc.2022.104000.
[6] R. Xiong, J. Cao, Q. Yu, H. He, F. Sun, Critical review on the battery state of charge
[30] X. Lin, J. Wu, Y. Wei, An ensemble learning velocity prediction-based energy
estimation methods for electric vehicles, IEEE Access 6 (2018) 1832–1843, https://
management strategy for a plug-in hybrid electric vehicle considering driving
doi.org/10.1109/ACCESS.2017.2780258.
pattern adaptive reference SOC, Energy 234 (2021), 121308, https://doi.org/
[7] J.G. Qu, Z.Y. Jiang, J.F. Zhang, Investigation on lithium-ion battery degradation
10.1016/j.energy.2021.121308.
induced by combined effect of current rate and operating temperature during fast
[31] G. Liu, X. Zhang, Z. Liu, State of health estimation of power batteries based on
charging, Journal of Energy Storage. 52 (2022), 104811, https://doi.org/10.1016/
multi-feature fusion models using stacking algorithm, Energy 259 (2022), 124851,
j.est.2022.104811.
https://doi.org/10.1016/j.energy.2022.124851.
[8] Z. Ni, Y. Yang, A combined data-model method for state-of-charge estimation of
[32] Y. Cao, T.A. Geddes, J.Y.H. Yang, P. Yang, Ensemble deep learning in
lithium-ion batteries, IEEE Trans. Instrum. Meas. 71 (2022) 1–11, https://doi.org/
bioinformatics, Nat Mach Intell. 2 (2020) 500–508, https://doi.org/10.1038/
10.1109/TIM.2021.3137550.
s42256-020-0217-y.
[9] Y. Zheng, M. Ouyang, X. Han, L. Lu, J. Li, Investigating the error sources of the
[33] J. Li, D. Wang, M. Pecht, An electrochemical model for high C-rate conditions in
online state of charge estimation methods for lithium-ion batteries in electric
lithium-ion batteries, J. Power Sources 436 (2019), 226885, https://doi.org/
vehicles, J. Power Sources 377 (2018) 161–188, https://doi.org/10.1016/j.
10.1016/j.jpowsour.2019.226885.
jpowsour.2017.11.094.
[34] M. Doyle, T.F. Fuller, J. Newman, Modeling of galvanostatic charge and discharge
[10] J. Meng, D.-I. Stroe, M. Ricco, G. Luo, R. Teodorescu, A simplified model-based
of the lithium/polymer/insertion cell, J. Electrochem. Soc. 140 (1993) 1526–1533,
state-of-charge estimation approach for lithium-ion battery with dynamic linear
https://doi.org/10.1149/1.2221597.
model, IEEE Trans. Ind. Electron. 66 (2019) 7717–7727, https://doi.org/10.1109/
[35] M. Mastali Majdabadi, S. Farhad, M. Farkhondeh, R.A. Fraser, M. Fowler,
TIE.2018.2880668.
Simplified electrochemical multi-particle model for LiFePO4 cathodes in lithium-
[11] P. Shrivastava, P.A. Naidu, S. Sharma, B.K. Panigrahi, A. Garg, Review on
ion batteries, J. Power Sources 275 (2015) 633–643, https://doi.org/10.1016/j.
technological advancement of lithium-ion battery states estimation methods for
jpowsour.2014.11.066.
electric vehicle applications, Journal of Energy Storage 64 (2023), 107159,
https://doi.org/10.1016/j.est.2023.107159.

15
H. Yu et al. Journal of Energy Storage 73 (2023) 108915

[36] J. Li, M. Zhao, C. Dai, Z. Wang, M. Pecht, A mathematical method for open-circuit using CNN-LSTM and UKF, Batteries 9 (2023) 114, https://doi.org/10.3390/
potential curve acquisition for lithium-ion batteries, J. Electroanal. Chem. 895 batteries9020114.
(2021), 115488, https://doi.org/10.1016/j.jelechem.2021.115488. [52] Z. Tong, J. Miao, S. Tong, Y. Lu, Early prediction of remaining useful life for
[37] H. Yu, J. Li, Y. Ji, M. Pecht, Life-cycle parameter identification method of an Lithium-ion batteries based on a hybrid machine learning method, J. Clean. Prod.
electrochemical model for lithium-ion battery pack, Journal of Energy Storage 47 317 (2021), 128265, https://doi.org/10.1016/j.jclepro.2021.128265.
(2022), 103591, https://doi.org/10.1016/j.est.2021.103591. [53] P. Baldi, P. Sadowski, The dropout learning algorithm, Artif. Intell. 210 (2014)
[38] H. Yu, K. Yang, L. Zhang, W. Wang, Z. Zhang, S. Li, S. Chen, S. Yang, J. Li, X. Liu, 78–122, https://doi.org/10.1016/j.artint.2014.02.004.
Simplified numerical modeling and analysis of electrolyte behavior in multiple [54] C.-H. Chen, K. Tanaka, M. Kotera, K. Funatsu, Comparison and improvement of the
physical fields for lithium-ion batteries, Journal of Energy Storage 72 (2023), predictability and interpretability with ensemble learning models in QSPR
108536, https://doi.org/10.1016/j.est.2023.108536. applications, J. Chem. 12 (2020) 19, https://doi.org/10.1186/s13321-020-0417-9.
[39] Q. Zhao, Y. Shan, C. Xiang, J. Wang, Y. Zou, G. Zhang, W. Liu, Predicting power [55] H. Abdulla, M. Maalouf, I. Barsoum, H. An, Truncated Newton kernel ridge
conversion efficiency of binary organic solar cells based on Y6 acceptor by machine regression for prediction of porosity in additive manufactured SS316L, Appl. Sci.
learning, Journal of Energy Chemistry 82 (2023) 139–147, https://doi.org/ 12 (2022) 4252, https://doi.org/10.3390/app12094252.
10.1016/j.jechem.2023.03.030. [56] P. Kollmeyer, C. Vidal, M. Naguib, M. Skells, LG 18650HG2 Li-ion battery data and
[40] X. Fang, S. Bagui, S. Bagui, Improving virtual screening predictive accuracy of example deep neural network xEV SOC estimator script, Mendeley Data 3 (2020),
Human kallikrein 5 inhibitors using machine learning models, Comput. Biol. Chem. https://doi.org/10.17632/cp3473x7xv.3.
69 (2017) 110–119, https://doi.org/10.1016/j.compbiolchem.2017.05.007. [57] P. Kollmeyer, Panasonic 18650PF Li-ion battery data, Mendeley Data 1 (2018),
[41] L. Ni, D. Wang, J. Wu, Y. Wang, Y. Tao, J. Zhang, J. Liu, F. Xie, Vine copula https://doi.org/10.17632/wykht8y7tg.1.
selection using mutual information for hydrological dependence modeling, [58] Q. Shi, Z. Jiang, Z. Wang, X. Shao, L. He, State of charge estimation by joint
Environ. Res. 186 (2020), 109604, https://doi.org/10.1016/j. approach with model-based and data-driven algorithm for lithium-ion battery,
envres.2020.109604. IEEE Trans. Instrum. Meas. 71 (2022) 1–10, https://doi.org/10.1109/
[42] K. Yang, Y. Tang, S. Zhang, Z. Zhang, A deep learning approach to state of charge TIM.2022.3199253.
estimation of lithium-ion batteries based on dual-stage attention mechanism, [59] V.J. Ovejas, A. Cuadras, State of charge dependency of the overvoltage generated
Energy 244 (2022), 123233, https://doi.org/10.1016/j.energy.2022.123233. in commercial Li-ion cells, J. Power Sources 418 (2019) 176–185, https://doi.org/
[43] B. Ma, L. Zhang, W. Wang, H. Yu, X. Yang, S. Chen, H. Wang, X. Liu, Application of 10.1016/j.jpowsour.2019.02.046.
deep learning for informatics aided design of electrode materials in metal-ion [60] Y.-H. Kao, M. Tang, N. Meethong, J. Bai, W.C. Carter, Y.-M. Chiang, Overpotential-
batteries, Green Energy & Environment. (2022), https://doi.org/10.1016/j. dependent phase transformation pathways in lithium iron phosphate battery
gee.2022.10.002. electrodes, Chem. Mater. 22 (2010) 5845–5855, https://doi.org/10.1021/
[44] H. Ruan, Z. Wei, W. Shang, X. Wang, H. He, Artificial Intelligence-based health cm101698b.
diagnostic of Lithium-ion battery leveraging transient stage of constant current and [61] H. Yu, L. Zhang, W. Wang, S. Li, S. Chen, S. Yang, J. Li, X. Liu, State of charge
constant voltage charging, Appl. Energy 336 (2023), 120751, https://doi.org/ estimation method by using a simplified electrochemical model in deep learning
10.1016/j.apenergy.2023.120751. framework for lithium-ion batteries, Energy 278 (2023), 127846, https://doi.org/
[45] J. Ma, Y. Ding, J.C.P. Cheng, F. Jiang, V.J.L. Gan, Z. Xu, A Lag-FLSTM deep 10.1016/j.energy.2023.127846.
learning network based on Bayesian Optimization for multi-sequential-variant [62] Y. LeCun, Y. Bengio, G. Hinton, Deep learning, Nature 521 (2015) 436–444,
PM2.5 prediction, Sustain. Cities Soc. 60 (2020), 102237, https://doi.org/ https://doi.org/10.1038/nature14539.
10.1016/j.scs.2020.102237. [63] D. Roman, S. Saxena, V. Robu, M. Pecht, D. Flynn, Machine learning pipeline for
[46] L. Du, R. Gao, P.N. Suganthan, D.Z.W. Wang, Bayesian optimization based dynamic battery state-of-health estimation, Nat Mach Intell. 3 (2021) 447–456, https://doi.
ensemble for time series forecasting, Inform. Sci. 591 (2022) 155–175, https://doi. org/10.1038/s42256-021-00312-3.
org/10.1016/j.ins.2022.01.010. [64] S. Fei, M.A. Hassan, Y. Xiao, X. Su, Z. Chen, Q. Cheng, F. Duan, R. Chen, Y. Ma,
[47] F. He, J. Zhou, Z. Feng, G. Liu, Y. Yang, A hybrid short-term load forecasting model UAV-based multi-sensor data fusion and machine learning algorithm for yield
based on variational mode decomposition and long short-term memory networks prediction in wheat, Precision Agric. 24 (2023) 187–212, https://doi.org/
considering relevant factors with Bayesian optimization algorithm, Appl. Energy 10.1007/s11119-022-09938-8.
237 (2019) 103–116, https://doi.org/10.1016/j.apenergy.2019.01.055. [65] J. Beemer, K. Spoon, L. He, J. Fan, R.A. Levine, Ensemble learning for estimating
[48] Z. Niu, Z. Yu, W. Tang, Q. Wu, M. Reformat, Wind power forecasting using individualized treatment effects in student success studies, Int. J. Artif. Intell. Educ.
attention-based gated recurrent unit network, Energy 196 (2020), 117081, https:// 28 (2018) 315–335, https://doi.org/10.1007/s40593-017-0148-x.
doi.org/10.1016/j.energy.2020.117081. [66] X. Qiu, P.N. Suganthan, G.A.J. Amaratunga, Short-term electricity price forecasting
[49] Z. Zeng, G. Jin, C. Xu, S. Chen, Z. Zeng, L. Zhang, Satellite telemetry data anomaly with empirical mode decomposition based ensemble kernel machines, Procedia
detection using causal network and feature-attention-based LSTM, IEEE Trans. Computer Science. 108 (2017) 1308–1317, https://doi.org/10.1016/j.
Instrum. Meas. 71 (2022) 1–21, https://doi.org/10.1109/TIM.2022.3151930. procs.2017.05.055.
[50] X. Liu, Q. Shi, Z. Liu, J. Yuan, Using LSTM neural network based on improved PSO [67] S. Guo, L. Ma, A comparative study of different deep learning algorithms for
and attention mechanism for predicting the effluent COD in a wastewater lithium-ion batteries on state-of-charge estimation, Energy 263 (2023), 125872,
treatment plant, IEEE Access 9 (2021) 146082–146096, https://doi.org/10.1109/ https://doi.org/10.1016/j.energy.2022.125872.
ACCESS.2021.3123225. [68] K. Chen, T. Han, J. Gong, L. Bai, F. Ling, J.-J. Luo, X. Chen, L. Ma, T. Zhang, R. Su,
[51] W. Wang, B. Ma, X. Hua, B. Zou, L. Zhang, H. Yu, K. Yang, S. Yang, X. Liu, End- Y. Ci, B. Li, X. Yang, W. Ouyang, FengWu: pushing the skillful global medium-range
cloud collaboration approach for state-of-charge estimation in lithium batteries weather forecast beyond 10 days lead. http://arxiv.org/abs/2304.02948, 2023.

16