Journal of Energy Chemistry 94 (2024) 716–739

Contents lists available at ScienceDirect

Journal of Energy Chemistry

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

Review

Insights and reviews on battery lifetime prediction from research to

practice
Xudong Qu a,1, Dapai Shi a,⇑,1, Jingyuan Zhao b,⇑,1, Manh-Kien Tran c, Zhenghong Wang a, Michael Fowler c,
Yubo Lian d, Andrew F. Burke b
a
Hubei Longzhong Laboratory, Hubei University of Arts and Science, Xiangyang 441000, Hubei, China
b
Institute of Transportation Studies, University of California Davis, Davis, CA 95616, USA
c
Department of Chemical Engineering, University of Waterloo, Waterloo, Ontario N2L 3G1, Canada
d
BYD Automotive Engineering Research Institute, Shenzhen 518118, Guangdong, China

a r t i c l e i n f o a b s t r a c t

Article history: The rising demand for energy storage solutions, especially in the electric vehicle and renewable energy
Received 7 February 2024 sectors, highlights the importance of accurately predicting battery health to enhance their longevity
Revised 13 March 2024 and reliability. This article comprehensively examines various methods used to forecast battery health,
Accepted 13 March 2024
including physics-based models, empirical models, and equivalent circuit models, among others. It delves
Available online 20 March 2024
into the promise of data-driven prognostics, utilizing both conventional machine learning and cutting-
edge deep neural network techniques. The advantages and limitations of hybrid models are thoroughly
Keywords:
analyzed, with a focus on the benefits of integrating diverse data sources to improve prognostic precision.
Batteries
Lifetime
Through practical case studies, the article showcases the effectiveness and flexibility of these approaches.
Health It also critically addresses the challenges encountered in applying battery health prognostics in real-
Machine learning world scenarios, such as issues of scalability, complexity, and data anomalies. Despite these challenges,
Deep learning the article underscores the emerging opportunities brought about by recent technological, academic, and
Field research advancements. These include the development of digital twin models for batteries, the use of
Real world data-centric AI and standardized benchmarking, the potential integration of blockchain technology for
enhanced data security and transparency, and the synergy between edge and cloud computing to boost
data analysis and processing. The primary goal of this article is to enrich the understanding of current
battery health prognostic techniques and to inspire further research aimed at overcoming existing hur-
dles and tapping into new opportunities. It concludes with a visionary perspective on future research
directions and potential developments in this evolving field, encouraging both researchers and practi-
tioners to explore innovative solutions.
Ó 2024 Science Press and Dalian Institute of Chemical Physics, Chinese Academy of Sciences. Published
by ELSEVIER B.V. and Science Press. All rights reserved.

1. Introduction ous operating conditions are crucial [2]. Precise lifetime prediction
has numerous benefits throughout the battery’s life cycle, such as
1.1. Background and significance of battery lifetime prognostics expediting product development, optimizing manufacturing pro-
cesses, reducing warranty and insurance costs, enabling timely
Lithium-ion batteries are utilized across a wide range of indus- maintenance, minimizing upfront capital expenses, and improving
tries, including consumer electronics, electric vehicles (EVs), rail, charging and discharging control for extended battery life. Addi-
marine, and grid storage systems [1]. To enhance the performance tionally, accurate lifetime prediction fosters the potential for
and cost-effectiveness of batteries, accurate estimation of their second-life applications, facilitating the development of a circular
state of health (SOH) and reliable lifetime predictions under vari- economy in battery manufacturing, reuse, and recycling [3]. How-
ever, predicting battery end-of-life is challenging due to intrinsic
⇑ Corresponding authors. factors such as manufacturing variations and pack design, as well
E-mail addresses: [email protected] (D. Shi), [email protected] (J. Zhao). as extrinsic factors such as temperature and usage intensity. While
1
These authors contributed equally to this work. existing research covers methods for estimating SOH and lifetime

https://doi.org/10.1016/j.jechem.2024.03.013
2095-4956/Ó 2024 Science Press and Dalian Institute of Chemical Physics, Chinese Academy of Sciences. Published by ELSEVIER B.V. and Science Press. All rights reserved.
X. Qu, D. Shi, J. Zhao et al. Journal of Energy Chemistry 94 (2024) 716–739

using laboratory data, there is a need to address the unique chal- cinctly outlines the practical challenges of battery prognostics for
lenges associated with field data and second-life applications. Field real-life applications, spotlighting potential avenues and
data exhibits irregular cycling patterns, varying operating condi- opportunities.
tions, and degradation mechanisms influenced by specific paths,
making accurate predictions difficult [4,5]. Prognostics based on
2. Methodologies for battery lifetime prognostics
real-world data remain an open research challenge, particularly
for industrial sectors such as automotive manufacturers that face
2.1. Battery modelling for prognostics
tight profit margins and strict certification requirements.
Although gathering fleet data may entail additional costs and
A reliable model captures the complex electrochemical behav-
efforts, having a fundamental set of laboratory degradation mea-
ior and degradation mechanisms of batteries, allowing for accurate
surements can offer valuable insights into the impact of operating
performance and degradation prediction under various operating
parameters on battery degradation. Achieving accuracy in capacity
conditions [6]. Battery aging, a multifaceted phenomenon, is sub-
and resistance estimation, with a target of at least 95% accuracy
ject to the influences of both external and internal factors. External
and ideally 98%, is essential for accurate lifetime prediction and
aging primarily hinges on parameters such as the battery’s charge–
the development of battery management systems focused on
discharge rates, ambient temperature, frequency of charging and
health monitoring. Defining the necessary levels of accuracy and
discharging, and the intervals between discharges [7]. Conversely,
confidence for such systems is an urgent research objective that
internal aging predominantly stems from the depletion of active
requires attention.
lithium ions, loss of active material, and escalation in internal
Overall, addressing the challenges associated with predicting
resistance [2]. The charging and discharging process essentially
battery lifetime in real-world applications is crucial for advancing
entails the insertion and extraction of lithium ions within the
battery technology, enabling sustainable energy systems, and
anode and cathode materials. However, improper usage and the
meeting the demands of various industries.
passage of time can culminate in the loss of active materials
between the battery’s electrodes, with the battery’s capacity
1.2. Contribution and structure of the paper
directly contingent upon the quantity of active material housed
within. The diminution of lithium inventory is intricately linked
This paper strives to explore comprehensively various
to factors such as the formation of the solid electrolyte interface
approaches for battery lifetime prediction, with a keen focus on
(SEI) layer, electrolyte decomposition, and lithium plating. Notably,
laboratory and field data. The primary objective is to evaluate the
battery aging is intimately entwined with the choice of battery
advantages, limitations, and associated challenges of each method
materials; for example, while lithium titanate utilized as an anode
while highlighting the potential of a hybrid approach that amalga-
circumvents the formation of an SEI layer, graphite anodes invari-
mates physics-based and data-driven models. It endeavors to fur-
ably form an SEI layer [8]. Understanding the intricate mechanisms
nish detailed insights into the computational complexity, data
governing battery aging not only facilitates more accurate predic-
requirements, and accuracy of different life prediction methods
tions of battery health but also presents challenges in model con-
in a laboratory environment. Furthermore, this paper deliberates
struction and computational efficiency. Regarding modeling
on the challenges and opportunities of implementing these meth-
methodologies, battery models can be broadly categorized into
ods on real-world field data, accentuating the potential benefits of
physics-based models, empirical models, and equivalent circuit
hybrid models integrating physics-based and data-driven
models (ECMs), as shown in Fig. 1. Each approach offers unique
approaches. Notably, it highlights the intricacies of battery lifetime
advantages and challenges, and their selection depends on the
estimation for second-life applications. In its essence, this paper
specific requirements of the prognostic application (Table 1). In
offers a deep understanding of diverse life prediction approaches,
this section, we discuss the importance of these battery modeling
their practical implementations, and the associated challenges
for prognostics and provide an overview of different modeling
and opportunities. Through the exploration of various methods’
approaches.
strengths and weaknesses and by advocating a hybrid approach,
it aims to contribute significantly to battery prognostics and bol-
ster the development of accurate and reliable lifetime prediction 2.1.1. Physics-based models
models. Playing a vital role in battery health prognostics, physics-based
The paper begins by presenting various life prediction models encapsulate the underlying electrochemical processes and
approaches grounded in laboratory data, including empirical aging degradation mechanisms within batteries [12,13]. Grounded on
maps, purely data-driven models, and physics-based models. The fundamental principles, these models enable predictions of bat-
pros and cons of each technique are meticulously explored, consid- tery health and RUL, using operating conditions and historical
ering aspects such as computational complexity, data require- data [6].
ments, and accuracy. The focus then shifts to life predictions A deeper physical understanding allows for the definition of
based on field data, analyzing the hurdles and prospects of apply- more precise degradation metrics beyond just capacity or resis-
ing these methods to real-world scenarios. Key challenges inherent tance. These may encompass ’degradation modes’ such as loss of
in both physics-based and data-driven methods are underscored. active material and loss of lithium inventory, both tied to funda-
Hybrid strategies, which blend physics-based and data-driven mental mechanisms such as SEI growth or lithium plating
approaches, might be a promising solution to these challenges. [14,15]. In a study focused on predicting battery health, a
These hybrid methods potentially offer a balance between preci- physics-based model was employed, accounting for three primary
sion, resilience in the face of scarce or subpar data, and broad degradation mechanisms (loss of active materials, slip of the cath-
applicability. The discussion then turns to the intricacies of esti- ode, and loss of lithium inventory) [9]. Utilizing nonlinear least
mating lifespan for second-life applications, including evaluating squares method, constraints were applied to the fitting coefficients
repurposed batteries in the absence of historical data, gauging to achieve the most accurate prediction results. Identified through
SOH beyond the critical turning point where degradation might differential voltage analysis, these modes can subsequently be
accelerate towards the end of life, and the economic analysis con- employed for predictions. Another significant category within
trasting testing costs for repurposed batteries with the potential physics-based models comprises those based on porous-electrode
revenue from precise SOH estimation. In conclusion, the paper suc- theory [16]. These models are designed to encompass degradation
717
X. Qu, D. Shi, J. Zhao et al. Journal of Energy Chemistry 94 (2024) 716–739

Fig. 1. Modelling for battery health estimation and lifetime prediction. (a) Collection battery data (voltage, current, temperature) and preprocess any anomalous data
(abnormal, noise, missing values) to select suitable features for the model; (b) the battery model encompasses empirical models, electrochemical models, and equivalent
circuit models; (c) apply filtering algorithms to process test results; (d) output the model’s predictive outcomes.

Table 1
Characteristics of physics-based models, empirical models, and equivalent circuit models.

Method Mechanism Merits

Physics-based models Based on the underlying electrochemical processes Provide detailed insights into battery operation.
in the battery. Capable of accurately predicting the performance under a wide range of operating
Incorporate physical laws and equations to describe conditions.
battery behavior.
Publication Main content Highlights Challenges
[9] Using a non-linear least squares Tracking individual The use of dynamic bounds and its implications for
method. degradation parameters computational resources.
\Online prediction of remaining using dynamic bounds. Require extensive knowledge of the battery’s physical
useful life (RUL). properties.
Empirical models Based on observed battery behavior, usually Relatively simple to develop and implement.
expressed in the form of mathematical equations. Less computationally demanding compared to physics-based models.
[10] Developing a semi-empirical model Considering It is challenging to generalize fitting parameters to other
that integrates aspects of both computational batteries, and the interplay of temperature effects and
empirical and electrochemical efficiency and physical operational conditions has not been taken into account.
models significance
Equivalent circuit Represent the battery’s behavior using an electrical Simplicity and computational efficiency compared to physics-based models.
models circuit composed of resistors, capacitors, and Suitable for real-time applications.
voltage sources. Can capture key features of battery dynamics.
[11] Propose a 2-RC dynamic equivalent Integration with Balancing complexity and accuracy.
circuit model; temperature effects. Generalization across different batteries
Flexibility for various
cell chemistries.
Comparison with other
models.

mechanisms leading to capacity fade, including SEI layer growth, However, these models exhibit certain limitations, notably a
lithium plating, and particle swelling and cracking [15]. Such mod- lack of flexibility and parametrization difficulties [17]. Although
els hold the capacity to simulate a cell’s entire lifespan under speci- they can achieve a remarkable fit with available data, they cannot
fic operating conditions. Theoretically, these models can serve for account for every conceivable eventuality. Even with numerous
prognostics by parametrizing them with early-life data for a partic- degradation sub-models incorporated, some aspects of the experi-
ular cell and subsequently simulating the remaining life until the mental dataset might not be accurately fitted [15]. Parametrizing
cell’s capacity reduces to 80%. This approach offers the advantage these models for a fresh cell presents a challenge, requiring spe-
of failure cause prediction, enabling preemptive remediation cialized testing equipment and cell teardown. The determination
strategies such as safety limit adjustments in the battery manage- of degradation parameters, such as SEI kinetic parameters, through
ment system. inverse modeling is even more complex, mainly due to the sub-

718
X. Qu, D. Shi, J. Zhao et al. Journal of Energy Chemistry 94 (2024) 716–739

stantial computational time required to simulate the cell’s entire regression and filtering algorithms. Current research focuses on
lifetime [18]. refining these models, addressing issues in PF algorithms, and
To foster wider adoption of physics-based models for lifetime exploring novel techniques to enhance prediction accuracy. Fur-
estimation, reducing this computational time is critical. This reduc- ther progress in empirical models and associated algorithms will
tion could be accomplished either through the development of undoubtedly lead to improved lifetime prediction and optimized
reduced-order models or the enhancement of data-driven meth- maintenance in battery prognostics.
ods. Advancing efficient physics-based machine learning models
is among the primary goals of the open-source battery modeling 2.1.3. Equivalent circuit models
framework, PyBaMM [19]. The study of physical and chemical reactions within batteries
has spurred the advancement of ECMs [29]. ECMs amalgamate cir-
2.1.2. Empirical models cuit elements that emulate the electrical behavior of batteries,
The capacity fade curve, which describes the capacity change allowing for the construction of mathematical models that effec-
over time, charge throughput, or equivalent cycle number, is a tively illustrate battery dynamic response and degradation pat-
commonly used measure of battery degradation. An intuitive terns [30]. This methodology applies circuit analysis techniques
approach for lifetime prediction is to create an empirical/semi- to delineate the complex relationship between battery perfor-
empirical model of capacity fade, parameterized by factors such mance and degradation. A critical facet of battery degradation is
as time and charge throughput [20]. These models often exhibit a the steady increase in internal resistance resulting from the growth
square-root-of-time dependence due to diffusion-limited SEI for- of the SEI layer [31]. Electrochemical impedance spectroscopy (EIS)
mation, and Arrhenius kinetics for temperature dependence [21]. data can serve to identify the ECM parameters, and regression
Improvements have been made to consider factors such as C- models can be utilized for predicting the battery’s RUL [32].
rate, average state of charge (SOC), depth-of-discharge range, and Numerous internal resistance growth models have been proposed,
voltage. Modern models segregate capacity fade into calendar and extended Kalman filter (EKF) and PF algorithms have been
aging [22], dependent on time, and cycle aging [23], reliant on implemented for parameter updating by researchers [33]. While
charge throughput. Despite their widespread industrial use, these EIS tests provide valuable insights, they can be time-consuming
models require large amounts of data, struggle with cell-to-cell and necessitate specialized equipment. Recent research has delved
variations, overlook interactions between aging mechanisms, and into alternative methods, such as employing fractional-order ECMs
often fail to capture changes in degradation rate. [34] and reconstructing EIS spectra using input current and output
To address these limitations, battery prognostics can focus on terminal voltage data. These strategies enable the estimation of
individual raw time series data (e.g., voltage, current, and temper- model parameters without the need for labor-intensive EIS tests.
ature) of cells during cycling. Empirical models can be developed The reconstructed EIS data can then be utilized to develop a regres-
by analyzing correlations in large datasets and incorporating bat- sion model that depicts the growth in internal resistance.
tery degradation behavior. Regression models or suitable empirical In summary, enhanced circuit models present a potent frame-
formulas can then predict the degradation state. Specific aging work for battery prognostics, capturing the complex dynamics of
models for RUL prediction include the weighted Ah aging model battery performance and degradation. The integration of circuit
and the event-oriented aging model [24]. A study abandoned pre- analysis methods with regression models and filtering algorithms
defined invariant battery degradation models in favor of construct- in these models allows for an accurate prediction of battery RUL.
ing an empirical degradation model based on experimental data Current research is focused on further refining these models and
through time series analysis [10]. The predictive model comprises exploring alternative methodologies that decrease dependence
an autoregressive model and a survival regression model based on on time-consuming EIS tests. The development of advanced prog-
segment length. The model employs a parametric bootstrap nostic techniques using enhanced circuit models carries immense
method to generate empirical bootstrap distributions, which in potential for real-time battery health monitoring and the optimiza-
turn are used to establish confidence intervals for battery lifespan, tion of maintenance strategies.
thereby predicting the health status of the battery with enhanced
accuracy and reliability. 2.2. Data-driven prognostics
Under low-current conditions, quadratic polynomial regression
models can capture the degradation trend of nickel-metal hydride In recent years, data-driven models have garnered significant
batteries [25]. Empirical internal resistance models have also been attention in the field of battery health prognostics [35]. Data-
established for calendar and cycle-life prediction over a wide tem- driven approaches rely on machine learning algorithms and statis-
perature range [26]. Models incorporating various health indica- tical techniques to analyze large volumes of battery data [7,36,37].
tors (HIs) such as fully discharged voltage, internal resistance, These methods aim to identify patterns, correlations, and trends in
rate of temperature change, and battery discharge curve, have also the data to make accurate predictions about battery health and
been developed. performance degradation. In particular, machine learning enables
Although simple, empirical models effectively establish correla- exploration of new functionalities in density functional theory cal-
tions between RUL and battery characteristics based on historical culations and new potentials in molecular dynamics simulations
data. However, these correlations evolve due to battery aging and from a microscale perspective. Additionally, machine learning
changing working conditions. Therefore, empirical models often exhibits significant advantages in strategic optimization, such as
blend with filtering algorithms to update parameters with the lat- in the development of fast-charging protocols [38]. By combining
est data. For example, Kalman filters (KF) and particle filters (PF) these models with machine learning techniques, researchers can
are commonly adopted for parameter estimation of batteries make accurate predictions without prior knowledge of the under-
[27]. PFs, in particular, provide tools for battery state estimation lying system dynamics (Fig. 2). Machine learning techniques,
due to their ability to handle nonlinear and non-Gaussian noise including neural networks, support vector machines (SVM), ran-
[28]. dom forests (RF), and regression models (RM), have been success-
In conclusion, empirical models offer a viable and efficient fully applied to predict important battery parameters such as SOC,
approach to predicting battery lifetime, capturing capacity fade SOH, and RUL [39]. In this section, we discuss the potential of data-
behavior under different operating conditions. Enhancements can driven, machine learning-based approaches in addressing the need
include incorporating raw time series data and employing various for continuous innovation in battery research.
719
X. Qu, D. Shi, J. Zhao et al. Journal of Energy Chemistry 94 (2024) 716–739

Fig. 2. Data-driven machine learning techniques for battery lifetime prognostics (abbreviations: LR, linear regression; SVM, supported vector machine; RF, random forest,
Bayes., Bayesian estimation; RNN, recurrent neural networks; PINN, physics-informed neural network). (a) Data-driven models collect and process vast amounts of historical
data; (b) steps for constructing feature engineering based on traditional machine learning; (c) automatic construction of feature engineering and model selection based on
deep learning; (d) selection of model evaluation metrics and output of model results.

2.2.1. Conventional machine learning techniques [49], relevant vector regression (RVR) [50], and Gaussian process
In recent times, methods utilizing machine learning have regression (GPR) [51] are often used. Table 2 details the character-
emerged as effective tools for estimating the SOH of batteries istics of each machine learning technique. After the model is
[40]. These methods are particularly appealing for their capability trained with a dataset, it can be utilized for SOH estimation on a
to capture nonlinear associations and operate independently of testing dataset, offering a measure of the model’s accuracy and
underlying physical models. Such model-free approaches offer sig- reliability. A study delineated a four-step methodology for the esti-
nificant advantages in situations where a balance between model mation of battery SOH [52]. The initial step, feature extraction,
complexity and accuracy poses a challenge. Despite their effective- involves the acquisition and processing of voltage, current, and
ness, these machine learning methods may not provide physical capacity data, inclusive of the decomposition of internal resistance
interpretability, which is a crucial factor for understanding the (IR) curves. The second step, model training, employs datasets
underlying degradation mechanisms. In this section, we’ll investi- from various charging cycles—specifically, the first 50%, 60%, and
gate the various processes and strategies associated with machine a randomly selected 60% of cycles—as the training set, utilizing
learning-based battery health estimation. the SVR technique, which comprises a regression function and a
The general pipeline for machine learning-based battery SOH kernel function, resulting in the generation of three distinct SVR
estimation includes data collection and preprocessing, input models. The third step is the estimation of SOH, which utilizes test
preparation, model training, and testing [41]. Machine learning data and a feature dataset as inputs to the trained SVR models,
algorithms for SOH estimation can be broadly classified into facilitating the output of estimated SOH values and the computa-
feature-based and feature-free methods, based on the types of tion of estimation errors. The final step, error analysis, involves
inputs utilized [42]. While the outcome—SOH—remains constant the application of mean absolute error (MAE) and root mean
for both methods, the preparation of inputs differentiates the two. square error (RMSE) as metrics to quantify the performance of
Feature-based methods involve using HIs—manually extracted the models, with the prediction errors of all three models remain-
features that give insights into a battery’s aging status. HIs can ing within a 2% margin. In research concerning the prediction of
be obtained directly from measurable data such as current, voltage, battery health using RF machine learning, the incremental capacity
and temperature, or indirectly computed from derived data such as analysis method was selected [48]. By extracting features from the
incremental capacity (IC) and differential temperature (DT) curves charging voltage and capacity data of batteries, these extracted fea-
[43,44]. The filter-based method, a popular approach, uses correla- tures were utilized as training inputs for the RF model. This model
tion coefficient analysis to select HIs as inputs for machine learning is capable of predicting battery health based directly on the raw
[45]. However, this approach may result in redundant HIs, some of extracted data, without the necessity for data preprocessing.
which may need to be omitted. An alternative is the wrapper Experimental results indicate that the predictive error of the model
method that optimizes the selection of HIs using advanced search is below 1.3%, suggesting a promising application for online battery
methods [46]. capacity prediction.
Once the inputs are prepared, model training follows. Consider- While machine learning-based methods provide a practical
ing that HIs often display high linear correlations with battery approach to SOH estimation, their lack of physical interpretability
capacity, linear regression (LR) and multi-linear regression (MLR) can pose challenges in deciphering the underlying degradation
methods can be deployed. Other nonlinear mapping algorithms processes. Nonetheless, they offer valuable insights into battery
such as support vector regression (SVR) [47] and RF [48] have also health evaluation and contribute significantly to ongoing innova-
proven effective. For probabilistic estimation, Bayesian estimation tions in battery research. In the subsequent section, we will exam-

720
X. Qu, D. Shi, J. Zhao et al. Journal of Energy Chemistry 94 (2024) 716–739

Table 2
Characteristics of some common machine learning techniques used for battery prognostics.

Machine learning Description Advantages

techniques
Linear regression Linear regression assumes a linear relationship between variables. Easy to understand and implement.
It is often used to predict the battery’s SOH. Computationally efficient.
Publication Main content Highlights Challenges
[53] SOH modelling using MLR. Efficient data handling. Generalization of fitting
The feature reduction process is conducted in two steps: (1) Advanced feature selection parameters.
principal component analysis (PCA); (2) Stepwise selection techniques. Handling temperature effects
combined with L1 regularization. High prediction accuracy and operational conditions.
Support vector SVM constructs a hyperplane in high-dimensional space for classification Effective with smaller datasets and in high-dimensional spaces.
machines or regression tasks, relying on a subset of training data, called support Robust to outliers.
vectors.
[54] Leveraging partial charging data, the battery’s SOH is The prediction of SOH can be Not well-suited for large
predicted online using an SVM model completed using a 15-minute datasets due to increased
charging curve. computational demand.
Requires careful choice of
kernel function and
parameters.
Random forests RF is an ensemble method that builds multiple decision trees from Effective with larger datasets and numerous input variables.
various sub-samples of the dataset, reducing overfitting by averaging the Handles missing values and maintains accuracy even when a
results. large proportion of the data are missing.
[48] By employing incremental capacity analysis for feature There is no need for data May not provide highly
selection, the health of the battery is predicted based on the preprocessing. precise outputs for
RF model. Predictions are made directly continuous numerical
using raw data. prediction.
Less interpretable than
simpler models such as
decision trees.
Bayesian estimation Bayesian Estimation uses Bayes’ theorem to update the probability of Provides a measure of uncertainty, giving a probabilistic
hypotheses as additional evidence becomes available. This approach approach to predictions.
enables updating predictions with new data. Incorporates prior knowledge.
[49] Feature engineering is constructed using charging data. The SOH is estimated and the Computationally demanding
The SOH is predicted based on a Dynamic Bayesian RUL is predicted utilizing a with complex models or large
Network model. particle filtering algorithm. datasets.
Tests were conducted on two Requires appropriate choice
different battery materials. of a prior.
Relevance vector RVR is a Bayesian alternative to SVMs for regression tasks. It provides Delivers a sparser model than SVMs.
regression probabilistic outputs and is more robust to outliers. Provides measure of uncertainty.
[55] A hybrid approach for predicting SOH and RUL has been Multiple kernel relevance vector Computationally intensive.
proposed, employing an Optimized Multi-Kernel Relevance machine is adopted to predict Requires careful selection of
Vector Machine combined with UPF. error trend. kernel function and
The initial estimation is parameters.
corrected by predicted error
trend.
Gaussian process GPR is a Bayesian non-parametric method that provides distribution- Provides a measure of uncertainty (variance) in predictions.
regression based predictions instead of single point estimates. Incorporates prior knowledge.
[56] A predictive model for the SOH of batteries has been The feature correlation was Computationally intensive
developed using GPR. analyzed using the Grey with larger datasets.
Relational Analysis method. Requires careful choice of
A battery with dynamic profile is kernel function and
used to verify the robustness of parameters.
this method.

ine deep neural network techniques—a specialized subset of informed neural networks (PINNs), self-attention transformer
machine learning methods, exploring their applications, advan- models and transfer learning, significantly advance battery prog-
tages, and constraints in the field of battery health prognostics. nostics. CNNs are particularly effective in identifying temporal pat-
terns in battery data, making them valuable for detecting features
2.2.2. Deep neural networks techniques indicative of degradation and predicting future battery health [60].
Deep neural networks (DNNs) have become a powerful tool in Autoencoders, by learning compressed, efficient representations of
battery health prognostics due to their ability to automatically battery data, facilitate anomaly detection and health assessment
learn intricate patterns and representations from extensive data- by reconstructing input data and identifying deviations indicative
sets [57,58]. These networks are comprised of interconnected arti- of battery degradation [61]. RNNs excel in processing time-series
ficial neurons organized into multiple layers, enabling them to data, enabling them to learn from historical battery performance
capture complex nonlinear relationships within high-dimensional data to predict future states accurately, thus providing critical
battery data [59]. DNNs have considerably improved battery insights into battery lifespan and health [62]. Physics-informed
parameter predictions such as SOC, SOH, and RUL, using voltage, neural networks merge domain expertise with data-driven model-
current, and temperature inputs. Six impactful deep learning tech- ing for accurate battery behavior prediction [63]. Self-attention
niques (Table 3), including convolutional neural networks (CNNs), transformer models use a self-attention mechanism to capture
autoencoders, recurrent neural networks (RNNs), physics- complex sequential data dependencies, enhancing prognostics
721
X. Qu, D. Shi, J. Zhao et al. Journal of Energy Chemistry 94 (2024) 716–739

Table 3
Characteristics of PINN, CNNs, transformer, and transfer learning.

Method Mechanism Merits

Convolutional Utilizes a hierarchical structure of convolutional layers, Efficient feature extraction.
Neural Networks pooling layers, and fully connected layers to automatically Strong generalization.
learn spatial hierarchical features from data. Scalability and adaptability.
Publication Main content Highlights Challenges
[66] Leveraging early cycle data, a CNN model has This model automatically extracts features from Computationally
been proposed to predict the SOH. aging data. intensive.
Reducing the required volume of test data. Risk of overfitting and
model interpretability.
Autoencoders The encoder compresses the input data into a low- Automatically extract important features from battery data.
dimensional representation, and the decoder reconstructs Capable of effectively processing high-dimensional data.
the original input data from this latent space. Able to identify data that does not conform to normal data patterns (anomaly
detection).
[67] Combining data interpolation and end-to-end Propose an enhanced end-to-end hybrid model Too complex networks
model approaches, using convolutional prediction method. can lead to overfitting.
autoencoders and bidirectional LSTM models Use autoencoders to extract key integrated time- Highly dependent on the
to predict SOH. series features. quality and
Validate model effectiveness with multiple datasets. representativeness of the
training data.
Recurrent neural Uses recurrent connections to maintain internal state Capable of processing time series data.
networks (memory). Utilizes historical information to make more accurate predictions.
Retaining memory information affects subsequent outputs. Can handle input sequences of varying lengths.
[68] Use the random forest method to extract data Track the time series dependency characteristics of Traditional RNNs are
features, and propose a hybrid model degrading capacity without using memory units. susceptible to vanishing
prediction of capacity degradation using GRU- Analyze model predictive performance on different or exploding gradients.
RNN. datasets. Difficult to capture long-
term dependencies.
Physics-informed Integrate known physical laws into the training process of Provide more accurate and interpretable results.
neural networks deep learning models. Generalize well to different tasks and unseen data.
Potentially require fewer data due to integration of physical laws.
[63] A PINNs has been constructed for battery This framework simulates the dynamic behavior of Complexity in
health management, integrating physical battery degradation. implementation.
dynamics models with data-driven models. An adaptive weighting method is used to balance Require careful tuning
learning tasks. and optimization.
Limited applicability
where underlying
physical laws are not
well understood.
Self-attention Use attention mechanisms to better understand sequence Excel in tasks involving sequence data, such as natural language processing and
transformers data. time-series forecasting.
Look at entire sequences at once to capture long-term Handle large datasets effectively.
dependencies. Inherently parallelizable, facilitating efficient training.
[69] Considering electrochemical information, a Feature extraction incorporates electrochemical High computational cost,
Transformer model has been established to knowledge. especially for longer
estimate the SOH. The attention mechanism is used to capture relevant sequences.
features. Often require large
amounts of training data.
Risk of overfitting
without proper
regularization.
Transfer learning Leverage a pre-trained model on a source task as the starting Can improve model performance, especially when the target task has limited
point for a related target task. data.
Speed up training time by preventing the model from learning everything from
scratch.
[41] A hybrid model combining Long short-term This model is adaptable to the SOH estimation of The source and target
memory (LSTM) networks and transfer various types of batteries. Reducing the volume of tasks need to be related
learning has been developed for estimating data required for training and enhancing for effectiveness.
the SOH of batteries. computational efficiency. Challenge in choosing
the right source model or
task.
The transfer of features
may not always enhance
performance on the
target task.

accuracy [64]. Transfer learning uses pre-training knowledge on (1) Convolutional neural networks
large datasets to improve performance on target tasks, reducing
large target data dependency and boosting model performance Convolutional neural networks (CNNs) are a potent class of
[65]. By leveraging these techniques, researchers can enhance bat- deep neural networks renowned for their high efficiency in analyz-
tery health predictions, leading to improved battery management ing two-dimensional visual imagery, as well as their capability to
and maintenance planning. process one-dimensional data, such as time-series predictions of

722
X. Qu, D. Shi, J. Zhao et al. Journal of Energy Chemistry 94 (2024) 716–739

battery health and signal recognition [70]. Unlike traditional neural eficial for understanding complex patterns in battery usage data,
networks that process data in a flat manner, CNNs leverage the where identifying key features can signal potential health condi-
hierarchical patterns within data, enabling them to adeptly handle tions of the battery. By employing autoencoders to decipher the
information with spatial and temporal correlations [71]. This attri- complexity of battery data, we can reveal meaningful insights into
bute is particularly advantageous for the analysis and comprehen- the battery’s SOH and RUL [77].
sion of the complex dynamics associated with battery usage data, Autoencoders comprise two main components: the encoder and
where temporal and spatial patterns may indicate underlying the decoder [76]. The encoder compresses input data into a lower-
health conditions. Utilizing this characteristic to process and dimensional representation (latent space), capturing the essential
understand the intricacies of battery usage data holds significant characteristics needed to represent the data [78]. In the context
promise. of battery health monitoring, this may involve condensing data
CNNs incorporate convolutional and pooling layers, with fea- from discharge and charge cycles into a compact format that
ture extraction being a hallmark of this algorithm [72], capable retains critical information regarding the condition of the battery.
of automatically and adaptively learning spatial hierarchies of fea- The decoder then reconstructs the input data from this compressed
tures from input data [73]. These layers use a set of learnable filters form, aiming to match the original input as closely as possible [79].
that capture crucial features such as edges and shapes in visual This reconstruction ability enables autoencoders to learn the most
data or significant trends and patterns in time-series data. For bat- prominent features of the battery data without the need for expli-
tery health prediction, CNNs can extract pivotal features from the cit labels.
discharge and charge cycles, which are indicative of the battery’s To illustrate the outstanding performance of Autoencoders in
condition over time. One of the key advantages of CNNs is their the realm of battery health prediction, we will delve into several
ability to capture local dependencies and scale invariance in data. research cases of interest. One study employed gray relational
In the context of battery health prediction, this means CNNs can analysis to examine the relationship between health indicators
identify local patterns of degradation across different cycles, extracted from charge-discharge data and aging, in conjunction
regardless of their position in the time series. This is crucial for with stacked autoencoders, to derive integrated health indicators
accurately modeling the degradation process, which may not be [80]. A Gaussian mixture regression prediction model was then
linear or uniform across the battery’s lifecycle. used, with these integrated features as inputs, to predict the bat-
Moreover, CNNs can be combined with other neural network tery’s RUL. Test results demonstrated that this model achieved a
architectures to enhance their predictive capabilities. For example, prediction precision with a RMSE of less than 1.05%, a MAPE of less
a hybrid model combining CNNs with LSTMs can leverage CNNs’ than 0.42%, and a MAE of less than 0.62%. Another noteworthy
ability to extract features from time-series data and LSTMs’ capac- study exploits the capability of autoencoders to automatically
ity to remember long-term dependencies, offering a robust frame- extract health features, thus circumventing the additional com-
work for predicting SOH and RUL of lithium-ion batteries. A study plexities involved in handling multidimensional data [61].
introduces a battery health prediction model, named Auto-CNN- Researchers constructed a predictive model based on GPR net-
LSTM, which addresses the shortfall of time-series data in CNN net- works to estimate the SOH and employed ensemble learning meth-
works through auto-encoders [74]. The hybrid CNN-LSTM model ods to enhance the precision of the SOH estimates. This model was
leverages the correlation between adjacent cycles of the battery validated using battery data under four different experimental
(where the CNN extracts deep information for feature mapping, conditions. Test results reveal that the model achieved a RMSE of
and LSTM normalizes these feature mappings in time series) to 1.04% and a MAE of 0.77%.
ultimately predict the RUL. The results indicate that this model The integration of autoencoders into battery health prediction
achieves a mean squared error of 4.8%, demonstrating commend- models represents a significant step forward in the field. By captur-
able predictive accuracy. Notably, another study proposes a data ing and analyzing the complex relationships within battery usage
preprocessing method using Empirical Mode Decomposition data, autoencoders facilitate the development of more accurate,
(EMD) for experimental reconstruction of data, while also consid- reliable, and efficient prognostic tools. This advancement holds
ering the utilization of existing data to form prior knowledge the promise of extending battery life, optimizing battery usage,
[75]. This is integrated with a 1DCNN-LSTM model to predict the and ensuring the safety and reliability of battery-powered systems.
battery’s RUL. The findings exhibit high predictive accuracy and As research progresses, the combination of autoencoders with
robustness. Further, by employing techniques such as transfer other neural network architectures and data-driven approaches is
learning, CNN models pre-trained on large datasets can be fine- poised to further elevate the capabilities in battery health manage-
tuned on specific battery datasets, significantly reducing the need ment and prognostics.
for vast amounts of battery-specific data.
CNNs’ ability to process complex, high-dimensional data and (3) Recurrent neural network and its variation
identify intricate patterns makes them exceptionally suited for
the task of battery health prediction. By employing CNNs, research- RNNs are a distinct variety of artificial neural networks built to
ers can develop predictive models that are not only accurate but handle sequential data [81]. Unlike conventional neural networks,
also capable of handling the dynamic and nonlinear behaviors RNNs don’t treat inputs and outputs as independent entities.
characteristic of lithium-ion batteries. As the field progresses, the Instead, they leverage the inherent sequentiality of data, making
integration of CNNs with other data-driven approaches and them particularly advantageous when working with time-series
domain knowledge is poised to further advance the state-of-the- data such as battery lifecycle patterns [82].
art in battery health management and prognostics, ensuring safer RNNs possess a ‘‘memory” feature that encapsulates informa-
and more reliable use of batteries across various applications. tion from prior computations [83]. While theoretically, RNNs can
use this memory to process any sequence of inputs, their practical-
(2) Autoencoders ity is curbed by the ’vanishing gradient’ issue. This problem implies
a geometric decay of information over time, hindering the net-
Autoencoders, a type of deep neural network renowned for work’s ability to learn long-term dependencies effectively. To
their proficiency in data encoding and decoding, are particularly address these issues, there exist several refined versions of RNNs:
suitable for tasks such as feature learning, dimensionality reduc- (i) LSTM. An LSTM is an evolved form of RNN designed to retain
tion, and anomaly detection [76]. This capability is especially ben- information over extended periods. This capability makes LSTMs
723
X. Qu, D. Shi, J. Zhao et al. Journal of Energy Chemistry 94 (2024) 716–739

particularly potent in dealing with data sequences that span a sig- acting as a catalyst for the introduction of pivotal digital technolo-
nificant duration or possess considerable gaps [84]. (ii) Gated gies, fostering the digital transformation of smart manufacturing
recurrent units (GRUs). GRUs streamline the LSTM architecture and transportation.
by merging the forget and input gates into a singular ‘‘update gate”
and blending the cell state with the hidden state [85]. While GRUs (4) Physics-informed neural networks
have fewer parameters, allowing potentially faster training or
fewer data needed to generalize, they might not possess the same Emerging as an effective method for battery health prediction,
level of representational capacity as LSTMs [86]. PINNs blend the capabilities of deep neural networks with the inte-
In the realm of battery health prognostics, RNNs, and their gral physical laws and constraints of a system. This unified frame-
derivatives prove extremely beneficial. They can represent the work marries data-driven learning and physics-based modeling,
time-bound dependencies in battery utilization data to project improving both prediction accuracy and interpretability [91].
future battery health. For instance, a tripartite data science model In the realm of battery health prediction, PINNs decode the
- encompassing EMD, grey relational analysis (GRA), and deep complex dynamics among various physical phenomena, such as
RNNs - has been proposed for RUL prediction [87]. The model electrochemical reactions, diffusion processes, and degradation
applies EMD and GRA methodologies to distil characteristic fea- mechanisms. By leveraging a combination of experimental data
tures from time-series data, and then various deep RNN variants, and physics-based simulations, they can discern the inherent cor-
including vanilla RNN, GRU, LSTM, and bidirectional LSTM, are relations and forecast key battery health indicators such as SOC,
deployed to prognosticate the SOH and RUL of lithium-ion batter- SOH, and RUL with precision [92].
ies. Despite the criticality of accurate degradation trajectory and A distinct benefit of PINNs is their adeptness in handling scarce
future life predictions for intelligent battery and electrochemical and erratic data, a frequent hurdle in battery health prediction.
energy storage systems, realizing precise forecasts with sparse his- They effectively distribute available information and use physical
torical data poses a significant challenge [62]. In response to this constraints to augment prediction accuracy, even in the face of
issue, a generalized deep learning framework rooted in RNNs has sparse or imperfect data. Moreover, PINNs offer a flexible model
been formulated. This innovative framework integrates future cur- that can incorporate domain-specific knowledge and constraints.
rent strategies and preliminary capacity-voltage data as inputs and By integrating physical equations, conservation laws, and recog-
shows promising potential in anticipating future battery behavior nized degradation mechanisms, PINNs adhere to the fundamental
using RNNs. principles of battery operation, providing predictions with physical
Further groundbreaking research in the field has introduced a significance. This enhances not only the model’s interpretability
stacked bidirectional long short-term memory (SBLSTM) network but also its capacity to generalize to novel scenarios.
coupled with extreme gradient boosting (XGBoost) for lithium- Several studies have corroborated the effectiveness of PINNs in
ion battery RUL prediction [88]. The SBLSTM model is adept at dis- battery health prediction, showing their ability to outperform tra-
cerning long-term degradation patterns in batteries, employing ditional data-driven methods in predicting SOC, SOH, and RUL with
XGBoost to identify key time-domain features for multivariate superior accuracy and robustness. For example, an integration of
inputs. Additionally, Bayesian optimization (BO) is engaged to fine- deep learning approaches into these models has been proposed
tune the model’s hyperparameters, resulting in enhanced overall to estimate internal concentrations and potentials at different spa-
performance. In a related endeavor, a self-adaptive long short- tial positions in the electrodes and electrolyte, with a demon-
term memory (SA-LSTM) method has been devised for accurate strated capability to bridge spatial, temporal, and chemical
prediction of lithium-ion battery nonlinear degradation using complexities, by combining physics-based and machine learning
early-cycle data, significantly reducing testing time and fostering models [93].
faster battery optimization and production [89]. This innovative Heat generation rate (HGR) is another significant parameter
approach involves extracting two features from discharge voltage for these batteries, particularly for formulating thermal safety
curves using a time-series method and then forecasting future strategies [94]. Recent research has proposed PINN approaches
cycles with the SA-LSTM model. This predictive model has demon- for accurate HGR estimation under various operating conditions.
strated notable accuracy, with average online prediction errors of The successful implementation of bidirectional long short-term
6.00% for discharge capacity and 6.74% for end of life, employing memory (BiLSTM) networks into these models, combined with
early-cycle data up to 90% capacity retention. Furthermore, the critical physical knowledge, has been instrumental in achieving
importance of temperature control in prolonging battery life and higher accuracy in HGR estimation. This approach has been
ensuring safety has been underscored by correlating temperature shown to perform well under dynamic stress tests and worldwide
data with degradation indicators. As lithium-ion batteries gain fur- light vehicle test procedure conditions. Battery health prognosis is
ther traction, and concerns regarding battery failures mount, also critical in technologies such as unmanned aircraft vehicles
health management and prognostics have become indispensable (UAVs). Accurate models for predicting the remaining useful life
for assuring their reliable and safe operation [90]. A novel SOH pre- of lithium-ion batteries are challenging to construct due to the
diction methodology using gated recurrent unit neural networks complex electrochemistry involved [95]. Reduced-order models
and a hidden Markov model, while incorporating uncertainty based on Nernst and Butler–Volmer equations, complemented
quantification, has been introduced [90]. This approach employs by data-driven kernels, have been used to represent overall bat-
the GRU neural network to adapt to the long-term global down- tery discharge, while an ensemble of variational Bayesian multi-
ward trend, while a hidden Markov model is harnessed to emulate layer perceptrons has been employed to model battery aging.
local fluctuations and quantify the uncertainty introduced by the To further enhance the health prognosis, time-variant prognosis
capacity recovery phenomenon in battery degradation. frameworks have been proposed, which utilize dynamic physics
The employment of RNNs and their variants in predicting bat- information and are sensitive to the implicit dynamic features
tery health and RUL is a burgeoning research domain, with the during long-term prognoses [96]. An LSTM neural network model
advent of innovative models enhancing the accuracy and reliability has been utilized, along with an adaptive model evolution
of these predictions. Although several methodologies have been method that allows for the model to evolve and improve its gen-
investigated, the accurate prognostication of battery health and eralization ability over time. This method has demonstrated high
lifespan using limited historical data remains an ongoing chal- prognosis accuracy under various charging and discharging
lenge. Nevertheless, the continuous refinement of these models is conditions.
724
X. Qu, D. Shi, J. Zhao et al. Journal of Energy Chemistry 94 (2024) 716–739

PINNs represent a comprehensive approach to battery health expedites the training process. Testing on the NASA battery dataset
prediction, fusing the strengths of deep neural networks with revealed that the dataset’s various portions resulted in distinct
physics-based constraints. Leveraging the benefits of both data- levels of accuracy [100]. Specifically, when 60% of the dataset
driven learning and physical laws, PINNs provide precise, inter- was used for testing, the MAE was recorded at 0.0027, the Mean
pretable predictions of battery health indicators. This synergy of Absolute Percentage Error (MAPE) at 0.0037, the RMSE at 0.0034,
data-driven insights and physics-based comprehension promises and the coefficient of determination (R2) at 0.9987. With 70% of
considerable progress in battery prognostics, aiding in the evolu- the dataset utilized for testing, the MAE slightly improved to
tion of efficient and reliable energy storage systems. However, 0.0026, the MAPE remained closely consistent at 0.0036, the RMSE
future research could explore improvements in data acquisition decreased to 0.0028, and the R2 slightly reduced to 0.9974.
and address challenges related to the complex electrochemistry Remarkably, when 80% of the data was used, the model’s perfor-
of these batteries. mance further enhanced with an MAE of 0.0018, a MAPE of
0.0025, an RMSE of 0.0027, and an R2 of 0.9902. These outcomes
(5) Self-attention transformer indicate that the model demonstrates significant robustness and
efficacy across different volumes of test data.
The self-attention transformer, a key component of transformer In addition to SOH estimation, predicting parameters of com-
models introduced for natural language processing tasks, has been plex, nonlinear battery systems, such as SOC, poses significant
increasingly adopted for diverse applications including time-series challenges, especially in EV applications, due to diverse aging
prediction tasks [97]. These models possess a unique ability to cap- mechanisms, cell-to-cell variations, and dynamic operating condi-
ture complex temporal relationships within sequences of data, tions [101]. To address these challenges, a recent study proposes a
making them highly applicable to time-series analysis involved cloud-based AI-enhanced framework for the co-estimation of SOC
in battery health prediction. A central characteristic of the self- and SOH during a battery system’s operational lifetime. This frame-
attention mechanism is its capability to weigh the relevance of dif- work leverages the capabilities of self-supervised transformer neu-
ferent data points within a sequence, considering each data point’s ral networks to learn observational data representations at
impact on others. For battery health prognostics, this signifies the multiple levels of abstraction. By combining a cloud-edge comput-
model’s capacity to recognize and learn the significance of various ing framework with deep learning versatility, the study demon-
historical battery usage patterns and how they influence the cur- strates the enhanced predictive ability to exploit long-range
rent state and future health of the battery. spatio-temporal dependencies across multiple scales.
The transformer model employs multiple layers of self- The self-attention transformer offers an innovative approach to
attention, allowing it to extract and learn from higher-level features battery health prognostics, capitalizing on its capacity to identify
within the data. For example, in the context of battery health prog- intricate temporal relationships in time-series data. Although chal-
nostics, lower layers might focus on short-term fluctuations in bat- lenges remain, ongoing research continues to enhance its efficacy
tery usage, while upper layers might capture longer-term trends and applicability, potentially heralding new advancements in bat-
indicative of battery degradation. Additionally, the transformer’s tery health prediction. However, there are challenges to address in
capacity to process input sequences in parallel, unlike RNN struc- the further adoption of these models. For instance, while trans-
tures, allows for more efficient training and prediction [98]. This formers excel at learning from large volumes of data, they may
quality is particularly advantageous in battery health prognostics overfit or struggle with limited or sparse datasets—a common sce-
where large volumes of time-series data are often analyzed. More- nario in battery health prognostics. Moreover, their predictions,
over, transformer models also incorporate positional encoding to while accurate, often lack physical interpretability compared to
maintain information about the order of data points within a physics-informed models.
sequence. Study has shown that by design specialized Transformer In the realm of battery lifetime prognostics, the competition
network architectures, capturing both the step-wise (temporal) and between RNNs and Transformers reveals distinct advantages and
channel-wise (spatial) information provides powerful tools for limitations inherent to their architectures and operational mecha-
learning the evolution of non-linear multiscale and multiphysics nisms (Table 4). RNNs, with their sequential data processing capa-
systems with inhomogeneous degradation behavior [99]. bility, have been traditionally favored for time series forecasting,
The study has shown that designing specialized Transformer including the predictive modeling of battery degradation. Their
network architectures that capture both the step-wise (temporal) architecture, designed to handle data one step at a time and pass
and channel-wise (spatial) information provides powerful tools information through hidden states, suits the temporal nature of
for learning the evolution of non-linear multiscale and multi- battery usage patterns but often grapples with long-term depen-
physics systems with inhomogeneous degradation behavior [14]. dencies due to the vanishing gradient problem. Transformers, on
One advantage of investigating long-time series using Transformer the other hand, introduce a paradigm shift with their self-
models is their crucial role in battery health prognostics. The order attention mechanisms, enabling the model to weigh the impor-
of events, such as charge–discharge cycles, significantly impacts tance of different parts of the input data simultaneously. This not
the battery’s health and lifespan. Several studies have demon- only allows Transformers to capture complex, long-range depen-
strated promising results by applying self-attention Transformer dencies more effectively than RNNs but also facilitates significant
models in battery health prognostics. These models have demon- improvements in parallel processing capabilities, scalability, and
strated accurate predictions of key health indicators such as SOH, potentially, model interpretability through attention maps. While
even in cases of complex and nonlinear battery degradation pat- RNNs require meticulous engineering to encapsulate long-term
terns. For example, in one study, a novel SOH estimation frame- trends effectively, Transformers can model intricate, non-linear
work was suggested, utilizing a CNN-Transformer approach [71]. degradation patterns in battery life without extensive feature engi-
The process begins with data pre-processing, where features highly neering. The choice between RNNs and Transformers for battery
related to SOH are selected using the Pearson correlation coeffi- lifetime prognostics thus hinges on specific project requirements,
cient (PCC). Principal correlation analysis (PCA) is used to minimize including the complexity of degradation patterns, sequence length,
redundant feature information, thereby reducing the computa- and computational constraints, with Transformers offering a
tional burden of the estimation model. All features are then nor- promising avenue for tackling more complex prognostic
malized using the min–max feature scaling method, which challenges.

725
X. Qu, D. Shi, J. Zhao et al. Journal of Energy Chemistry 94 (2024) 716–739

Table 4
Comparative analysis of RNN vs. Transformers for battery lifetime prognostics.

Feature RNNs Transformers

Architecture Utilizes a loop to process sequence data, passing the hidden Based on self-attention mechanisms that weigh the importance of different
state from one step to the next. parts of the input data.
Processing of sequence Processes data sequentially, which can be a bottleneck for Processes entire sequences at once, allowing for more efficient
data parallel processing. parallelization.
Memory and long-term Struggles with long-term dependencies due to vanishing Better at capturing long-term dependencies thanks to global self-attention,
dependencies gradient problem, although LSTMs and GRUs mitigate this. without being affected by the position in the sequence.
Training time and Generally faster to train due to simpler architecture, but More computationally intensive due to self-attention, but often requires
complexity may require more epochs. fewer training epochs.
Performance on time Historically favored for time series forecasting due to their Increasingly popular for time series forecasting, showing superior
series data sequential nature. performance in capturing complex patterns.
Application Effective for modeling time-dependent degradation Potentially more effective in modeling complex, non-linear degradation
patterns, but may require careful engineering to capture patterns and interactions over time without the need for extensive feature
long-term trends. engineering.
Scalability Less scalable due to sequential data processing. Highly scalable due to parallel processing capabilities.
Interpretability Can be challenging to interpret due to the hidden state Offers slightly better interpretability through attention maps, showing how
transitions. different positions in the sequence relate to each other.

(6) Transfer learning health forecasting approach, inclusive of transfer learning, for pre-
dicting the states of different battery types. To ascertain predictive
Transfer learning has become a powerful tool in machine learn- uncertainty, they employed variational inference during the con-
ing [102], especially in situations where data are limited or expen- current prediction of battery end-of-life and the forecasting of
sive to obtain. Essentially, transfer learning is a technique that degradation patterns. The method displayed notable versatility
allows a pre-trained model developed for one task to be reused and robustness, accurately predicting the SOH of lithium-ion bat-
as the starting point for a different, but related task. This can signif- teries, even in the absence of past information, making it particu-
icantly speed up the development of effective models and improve larly suitable for second-hand batteries.
performance when data is scarce. In the context of battery health In a separate research endeavor, it was noted that when training
prognostics, transfer learning can be highly beneficial. It can lever- and testing batteries exhibit different degradation patterns due to
age knowledge from previous datasets or models that have been fluctuating operational conditions, significant errors may arise in
trained to predict battery health indicators under different condi- SOH estimations based on data-driven techniques [105]. To tackle
tions or for different types of batteries [41]. The ability to learn this issue, the study proposed the use of a battery’s early aging data
from other tasks and datasets can significantly enhance the mod- for identifying degradation patterns and improving SOH estimation
el’s predictive accuracy and robustness when confronted with accuracy via transfer learning. Their approach notably enhanced
new and possibly distinct battery health prognostic scenarios. the precision of estimations, reducing MAE and RMSEs to a mere
One primary advantage of transfer learning in battery health 0.94% and 1.13%, respectively.
prognostics is the potential to handle diverse types of batteries The existing body of literature underscores a growing interest in
and operating conditions. For instance, a model trained on one type leveraging the potential of transfer and deep learning methodolo-
of lithium-ion battery may be adapted to predict the SOH or RUL of gies to enhance the accuracy of lithium-ion battery health manage-
another type of lithium-ion battery. This can substantially reduce ment. However, the variability in physical and chemical
the need for expensive and time-consuming data collection pro- characteristics among different battery types and operational con-
cesses for each specific battery type and condition. ditions presents significant challenges. Also, there remains an
Transfer learning also offers a solution to the challenges of data ongoing need for more comprehensive and diverse datasets. The
scarcity and uneven distribution, which are often encountered in issue of source task or model selection for transfer learning is
battery health prognostics. By leveraging pre-trained models, another crucial aspect that requires meticulous evaluation to
transfer learning can enhance the model’s generalizability and pre- ensure its relevance and compatibility with the target task. Hence,
dictive performance even when training data for the target task is these areas warrant further investigation in upcoming research
limited. For example, in one recent study, an expansive dataset was endeavors, aiming to improve the accuracy of SOH estimation
curated, encompassing 77 commercial cells and exceeding 140,000 and RUL prediction. Despite these challenges, transfer learning rep-
charge–discharge cycles, setting a benchmark in its category [103]. resents a significant avenue for advancing battery health prognos-
The research team constructed a transfer learning framework, tics. Contemporary research is delving into effective strategies for
which facilitated real-time personalized prediction of health sta- feature transfer and adaptation, and the development of method-
tuses for previously unseen battery discharge protocols, at any ologies for appropriate source task and model selection. The incor-
charge–discharge cycle [20]. The methodology demonstrated poration of transfer learning in battery health prognostics is set to
impressively low mean testing errors 0.176% for capacity estima- facilitate more efficient, accurate, and robust predictions. This, in
tion and 8.72% for the prediction of the RUL. Notably, the innova- turn, is expected to contribute to the reliability and lifespan of bat-
tive framework was also able to leverage insights from two tery systems, signifying an exciting direction for future advance-
distinct battery datasets with different charge configurations and ments in the field.
battery chemistries, thereby reliably predicting the capacity and
RUL of the cells. This holistic approach empowers end-users to cre- 2.3. Hybrid approaches
ate personalized battery consumption strategies and equips manu-
facturers with valuable insights to refine battery designs. Hybrid methodologies have emerged as a compelling strategy
Another study underscored the significance of precise forecast- in the realm of battery health prognostics [106]. By amalgamating
ing of lithium-ion battery’s SOH and RUL [104]. The researchers the robustness of model-based techniques with the flexibility of
concentrated their efforts on developing a deep-learning-based data-driven approaches, these methodologies aim to harness the
726
X. Qu, D. Shi, J. Zhao et al. Journal of Energy Chemistry 94 (2024) 716–739

strengths of each, offering a promising avenue for the advance- which includes identifying the range of observations for which
ment of battery health prognostics [107]. A study has developed one machine learning model can be mapped onto another or onto
two hybrid models that are grounded in both physics and machine a physical model. It’s also vital to pinpoint the generalization limit
learning [13]. These models incorporate feedback from the infor- beyond which these models cannot be transformed or calibrated
mation captured by the physics-based model into the machine with respect to each other.
learning model, while also integrating mechanisms of battery To improve the generalization and robustness of predictive
aging. They are capable of predicting the health status of batteries models, a stacking ensemble model has been developed to esti-
across a broad spectrum of rates, thereby demonstrating the pre- mate the SOH of commercial lithium iron phosphate (LFP)/graphite
dictive accuracy of the hybrid models. Another study established cells using daily operating charging data, such as time-resolved
a semi-empirical model and data fusion approach for battery inter- voltage, current, and temperature recorded from real-world vehi-
nal resistance and temperature [108]. Model parameters were cle usage [4]. Noise reduction methods have been suggested as a
identified using a genetic algorithm, and a particle filter was preliminary step in data processing. One two-step approach con-
designed to perform real-time updates on these parameters. This sists of a trajectory piecewise-polynomial model and an exponen-
integrated model accurately predicts battery life and health status. tially weighted moving average model for battery data de-noising
A critical component of hybrid methodologies involves the integra- (citation needed). This enhances the quality of the data, which in
tion of different types of battery models to extract an enriched fea- turn improves the accuracy of the battery health prediction. The
ture set from the available data. These models serve to enhance the study also acknowledges the importance of domain-specific fea-
accuracy of battery prognostics and diminish prediction uncer- tures derived from IC and differential voltage (DV) analyses. These
tainty, particularly at the initial stages. A recent study has paved methods provide physically consistent representations of complex
new avenues for the health estimation of novel battery materials. battery degradation behavior, thereby improving the robustness
The research introduces an interpretable hybrid machine learning and stability of the prognostics. The stacking approach, leverages
model that combines time series decomposition with RNNs a two-layer learning model with four base learners (LR, RFR, GPR,
[109]. This hybrid approach stands out in battery forecasting by and GBR) and a meta-learner to improve prediction accuracy and
providing clearer physical explanations for factors leading to bat- model generalization (citation needed). Studies have confirmed
tery degradation, thereby enabling more accurate and insightful the robustness, stability, and high compatibility of this multi-
predictions. model fusion method across different usage histories for more than
There are primarily two types of fused models that have been 1000 EVs.
developed and applied in this context. One common type involves Hybrid methodologies thus offer a versatile approach to battery
the combination of different models to construct a new, more com- health prognostics, striking a balance between the grounded relia-
prehensive model. For example, In the realm of battery health esti- bility of physical models and the adaptability of data-driven algo-
mation, a noteworthy contribution is the creation and execution of rithms. As research progresses, the refinement and further
a battery health management system that incorporates uncertainty development of these hybrid approaches stand to significantly
prediction [110]. The authors put forward a comprehensive strat- improve the effectiveness of battery health prognostics, fostering
egy that employs an array of four distinct machine learning tech- the longevity and reliability of battery systems. In summary, these
niques: two parametric methods (Bayesian ridge regression and progressions in battery health estimation underscore the impor-
RF) and two non-parametric methods (GPR and an ensemble of tance of incorporating uncertainty prediction, model fusion, and
deep neural networks). These are collectively denoted as the foun- generalizability assessment into the development of accurate and
dational models for the system. Each model’s efficacy is assessed reliable models for battery health management.
based on error values and its capacity to determine generalization
error and uncertainty. This methodology astutely underscores that 2.4. Information fusion
every learning algorithm has unique strengths and limitations,
which can differ based on the specific charging methodologies Information fusion methodologies form a significant part of
applied. For example, the random forest model, trained with boot- modern strategies in battery health prognostics. These techniques
strap duplicates, showed the lowest error for constant current- aim to assimilate data from a wide array of sources such as sensor
constant voltage and fast charge protocols (3.6 C-rate). Conversely, readings, historical data, and expert insights, aiming to generate
Bayesian ridge regression was effective for constant-current proto- predictions about battery health that are both comprehensive
cols. It’s important to note, though, that while random forest can and reliable. A key strength of information fusion lies in its ability
be remarkably effective in certain situations, its calibration can to provide a holistic understanding of battery performance. Con-
be demanding and may produce excessively positive forecasts currently processing information from multiple independent
when evaluating uncertainty. Further, training deep neural net- sources, it allows for a multi-faceted view of the battery state,
works typically involves high computational costs and necessitates greatly aiding the accuracy of prognostic assessments.
intricate formulations and coding. Despite these challenges, deep The core idea is that different data sources provide complemen-
neural networks usually achieve superior calibration scores, and tary insights about the state of a battery system. For instance, real-
their adaptability allows for the inclusion of domain-specific time sensor data can offer a glimpse into the current performance
knowledge into the models. Another intriguing study drew from characteristics of the battery, historical records can provide trends
the advantages of multi-model fusion and employed a random for- and patterns over time, and expert knowledge can contribute
est regression model to merge three base learners: multiple linear heuristic information about the typical behavior of the battery
regression, SVR, and GPR [111]. To delve deeper into battery aging, under different operating conditions. By combining these diverse
the study initially extracted geometric characteristics from voltage pieces of information, Information fusion techniques are able to
and temperature change curves, along with peak values from draw a more nuanced picture of the battery state, enabling more
incremental capacity curves, as health indicators. The findings accurate predictions of key indicators such as health and safety.
showed that the random forest-centric fusion model effectively These techniques can help identify potential faults or anomalies
mitigated overfitting and bolstered the robustness of the predic- early on, allowing for preventive maintenance and potentially
tion models. The fusion algorithm can be seen as a decision tree extending the lifespan of the battery. For example, an early-
classifier working within a bagging framework. Importantly, the prediction model developed from commercial battery packs using
generalizability of these transformations needs to be evaluated, NMC/graphite cell data exemplifies this, tackling challenges in
727
X. Qu, D. Shi, J. Zhao et al. Journal of Energy Chemistry 94 (2024) 716–739

robust and accurate prediction of EV battery failure in real-world ity, sparsity, and varying recording patterns and timestamps
applications [112]. Cloud-based closed-loop frameworks are a [116]. Notably, battery performance indicators such as resistance
prominent trend, utilizing historical EV battery data to generate during acceleration and braking, as well as charging impedance,
digital twins and longitudinal electronic health records in cyber- are significantly influenced by seasonal temperature fluctuations,
space. The study demonstrates the effectiveness of machine learn- underscoring the critical impact of real-world operating condi-
ing models integrated with a multifaceted understanding of tions, particularly temperature, on data quality and battery SOH
electrochemical energy storage systems. Such machine learning assessments. Furthermore, variability in charging rates and SoC
techniques contribute not just too accurate battery failure predic- also impacts the quality of health assessment indicators, necessi-
tion but also to lifelong learning through self-taught and self- tating the separate analysis of charging events at different C-
improved AI techniques [113]. Despite progress, accurately pre- rates to enhance predictive accuracy. However, battery degrada-
dicting the evolution of multiphysics systems remains challenging. tion in real conditions inherently exhibits non-linearity and path
To address this, machine learning techniques were developed dependence [15]. This means the sequence of events and condi-
based on cloud-uploaded data, achieving a 96.3% classification tions a battery undergoes significantly affects its degradation pro-
accuracy and a 7.7% misclassification test error. In conclusion, cess, thereby exacerbating the challenge of making accurate
cloud-based AI technology holds significant potential for robustly predictions based on field data.
predicting battery failure in real-world applications, contributing One study proposed an innovative battery aging assessment
to energy storage and the electrification of the transport method combining incremental capacity analysis (ICA) and a radial
community. basis function neural network (RBFNN) model [117]. Real-world
The fusion of information from various sources is typically operational datasets from electric city transit buses were used to
achieved using advanced statistical techniques and machine learn- illustrate the relationship between battery aging level and its influ-
ing algorithms. These methods enable the efficient integration of encing factors. The authors concluded that an average prediction
multi-modal data and account for potential discrepancies or uncer- error of 4.00% was reached and the derived model presented a con-
tainties among different data sources, ensuring robust and reliable fidential interval of 92% with a prediction accuracy of 90%. This
prognostic predictions. Information Fusion methodologies offer a work opens avenues for battery aging prediction utilizing vast
potent approach to battery health prognostics. By effectively inte- real-time operation data. In a parallel effort, [118] proposed a
grating diverse sources of information, these methodologies can method to estimate battery SOH indicators based solely on vehicle
provide a comprehensive and nuanced understanding of battery operating data. Through the incorporation of an electrical circuit
performance, thus paving the way for more accurate, timely, and model, an optimization algorithm, and a Kalman filter, the authors
reliable prognostic assessments. presented a simple but effective method for on-board estimation of
the state of charge and SOH indicators. The promising results from
this study could provide a better understanding of battery degra-
3. Battery lifetime prediction in real-world applications: Case dation laws and enhance state of charge estimations. Moreover, a
studies study conducted in Beijing, China offered a real-world data-
driven nonlinear estimation model for predicting the remaining
In field applications, particularly in vehicular scenarios, driving range of EVs [119]. It highlighted a significant nonlinear
machine learning techniques such as unsupervised learning can relationship between speed and driving distance per SOC. Notably,
be used to generate representative driving cycles [114]. These the study also uncovered an increasing trend in the economically
cycles then determine the charging and discharging patterns for optimal driving speeds for EVs as temperature increases.
battery systems. However, designing tests that mimic the entire A case study aiming to predict the capacity fade of Lithium-ion
spectrum of real-world scenarios in a laboratory setting poses sig- batteries in EVs, incorporating real-world vehicle driving data,
nificant challenges. Recognizing these limitations, there has been demonstrated promising results [120]. The study involved a com-
notable progress in the field of Li-ion battery modeling. This bination of three Li-ion chemistries, different battery architectures,
advancement is especially evident in the integration of real- and recharge strategies in a total of 220 scenarios. It was found
world operational data [115]. Notably, these improvements that, for most combinations, the battery capacity did not fall below
include the integration of real-world operational data into the 80% within 5 years for a driving profile of up to 1000 km/month. A
modeling process, enhancing both its accuracy and practical appli- recent study employed driving data from EV rentals to predict the
cability. Table 5 summarizes recent examples of application meth- actual battery SOH [121]. This study, when compared to other
ods for battery health prediction in practical settings. The process machine learning models, demonstrated that this model had the
(Fig. 3), typically involves collecting data during daily operations. highest predictive accuracy, characterized by lower mean absolute
This is followed by thorough investigations into battery degrada- percentage error and root mean squared error. The significance of
tion mechanisms and/or employing data-driven methods for eval- this research lies in its implications for accurate battery prediction
uating the performance of specific battery cells/chemistries, even in operational EVs. In the process of forecasting battery aging per-
without prior knowledge. Achieving this generally requires offline formance using data-driven models, the accuracy of predictions is
model training and testing to ensure both accuracy and reliability significantly influenced by the selection of features from battery
or alternatively, utilizing cloud-based digital solutions for model data. As a result, a study has proposed a machine learning frame-
training. Prior to practical applications, small-scale scenario work that employs intelligent feature selection to predict both cal-
debugging and validation are conducted, with a focus on specific endar and cycle degradation in lithium-ion batteries within electric
uses such as EVs or frequency energy storage. This step is crucial vehicles [122]. Utilizing data from 192 electric vehicle battery
before the models can be broadly implemented in their respective packs, the dataset comprises a decade’s worth of degradation data
fields. for lithium manganese oxide (LMO) graphite-based batteries
In conjunction with these efforts, addressing data quality issues across five states in the USA. From this dataset, 33 feature param-
inherent in real-world EV applications becomes a cornerstone for eters were identified and utilized to predict both cycle and calen-
reliable battery health prediction. Data collected in actual scenar- dar degradation. The intelligent feature selection method selects
ios display inconsistencies and noise due to varying driving condi- attributes indicative of battery aging from various perspectives,
tions, temperature changes, and user behaviors, necessitating a effectively managing irrelevant and redundant features to create
robust data preprocessing routine to address dataset heterogene- an optimal subset of features. This study contrasts the performance
728
X. Qu, D. Shi, J. Zhao et al. Journal of Energy Chemistry 94 (2024) 716–739

Table 5
The practice of battery health prognostics for real-world applications.

Prediction task Data generation Feature engineering Model Error Ref.

SOH estimation 18 EVs Incremental capacity features Radial basis function neural Average error: 0.85. [117]
Accumulated mileage of vehicles networks Confidential interval: 92%
Initial charging SOC with the prediction accuracy
Average charging temperature of 90%
Average charging current
Average operating temperature
SOH estimation 10 FCVs Battery current Electrical circuit model NRMSE: 0.016 [118]
Cell voltage Optimization algorithm
Battery module temperature Kalman filter
Constant charge/discharge profiles
Internal resistance
Remaining driving range 1 EV (3 drivers SOC Robust nonlinear regression RMSE: 0.1958 [119]
estimation on different Speed R2: 0.9624
roads) Temperature SSE: 2.8367
SOH estimation and RUL Group A 420 39 features are engineered using Stacking ensemble machine Cells: [4]
prediction cells incremental capacity and differential learning, comprised of four base MAPE: 0.0028
Group B 9 EVs voltage learning models RMSPE: 0.0055
EVs:
MAPE (60 s): 0.0198
RMSPE (60 s): 0.0322
MAPE (90 s): 0.0225
RMSPE (90 s): 0.0366

SOH estimation Group A 47 EVs Historical charging data Ampere hour integral Capa:Pre: -Capa:Test [121]
Capa:Test < 0:04
Group B 100 EVs Temperature data Kalman Filtering
Federated Learning
Arrhenius empirical aging model
Capacity prediction 20 EVs Charging data mean, sum, and LSTM Seq2Seq and GPR Seq2Seq Ⅰ + GPR Ⅰ [125]
standard deviation MAEave : 0.0121
RMSEave : 0.0152

Fig. 3. Workflow for implementing battery lifetime prediction in real-world scenarios. (Initially, key parameters such as current, voltage, and temperature are logged via an
on-site data collection module. These data are cleansed of anomalies, noise, and gaps through preprocessing methods. Subsequently, in the feature engineering phase, critical
health indicators are selected and transformed into a format amenable to machine learning algorithms. A variety of machine learning models, including SVM, GPR, RF, and
ANN, are employed for training to forecast battery capacity degradation. The model architecture incorporates various activation functions and network layers, such as ReLU
and LSTM. Ultimately, the models are applied to predict the capacity decline of different battery units, facilitating applications in vehicle usage, secondary life cycles, and
environmental conservation.).

of various machine learning models in predicting battery aging, in tion (SFS) strategy in enhancing the accuracy of battery degrada-
conjunction with intelligent feature selection. By comparing the tion predictions was assessed. The findings demonstrate that the
MAE of the prediction results from different machine learning SFS method can significantly improve the accuracy of machine
algorithms, the effectiveness of the proposed Smart Feature Selec- learning algorithms in predicting battery degradation, with the

729
X. Qu, D. Shi, J. Zhao et al. Journal of Energy Chemistry 94 (2024) 716–739

RF and XGBoost algorithms showing the best performance in pre- in EVs, the demand for batteries in 2030 is expected to be nearly
dicting both cycle and calendar degradation. 50% higher in the growth scenario compared to the base case
The accuracy of battery SOH estimation models under actual [127]. The concept of extending a battery’s lifecycle through a
operating conditions has long been constrained by limited labora- ‘‘second-life” application - one that follows its primary use - repre-
tory conditions. A recent study critically examined current SOH sents an environmentally friendly and economically sensible
estimation methods, emphasizing the necessity of integrating approach, particularly in the rapidly expanding lithium-ion battery
real-world operating conditions into laboratory testing to enhance market (Fig. 4). The nominal capacity degradation of EV batteries to
the accuracy and reliability of battery predictions [116]. This study 70%–80% signifies their end-of-life, a stage at which their dimin-
analyzed over a year of field data collected from a battery pack in ished capacity directly impacts the EVs’ driving range[127]. These
an Audi e-tron vehicle, comprising 1655 battery signals. It consid- batteries, initially used in EVs [128], still retain a considerable
ered correlating extracted battery data features (charging habits, capacity post-retirement from EVs [129], and can later be transi-
acceleration, braking, and seasonally-related ambient tempera- tioned for use in grid or stationary storage applications [130].
tures) with actual driving characteristics. The study underscored Building on this concept, the current study involves the utilization
the complexity of recording field data and the high experimental of second-life EV batteries for residential fixed storage applications
costs involved. It not only advances understanding of the perfor- [131]. The research incorporates a two-stage method of mixed-
mance of electric vehicle battery packs under real-world condi- integer linear optimization. The first stage contemplates the opti-
tions but also provides avenues for improving future battery SOH mal charging strategy for electric vehicles, representing the batter-
estimation techniques. In another noteworthy study, researchers ies’ initial lifecycle. The second stage evaluates the reuse of these
primarily focused on implementing SOH estimation for the health EV batteries within residential buildings as fixed storage, examin-
management of lithium-ion battery packs in electric vehicles, ing various scenarios in conjunction with demand-side manage-
achieving real-world, online capacity estimation for electric vehicle ment, photovoltaic self-consumption, and Vehicle-to-Grid (V2G)
batteries [123]. This research utilized electric buses in Hefei, Anhui, systems. This research assesses service life, economic savings,
China, for commuting purposes, with battery modules consisting of and the impact of electricity pricing, demonstrating how second-
608 series and parallel-connected cells. Data collection captured life batteries can extend the usability of EV batteries while also
four distinct battery discharge modes during electric vehicle oper- achieving cost savings. Transitioning to a related field, the wide-
ation: acceleration, braking, idling, and constant speed. The battery spread adoption of microgrids has significantly addressed the glo-
SOH degradation model considered factors such as total mileage, bal issue of energy wastage. Another study developed an energy
monthly mileage, and seasonal temperature variations. The model storage system based on the second life of batteries, utilizing
parameters were identified using a Variable Forgetting Factor retired Nissan Leaf battery modules [132]. These modules, main-
Recursive Least Squares (VFFRLS) algorithm, and the estimation taining a SOH of 71%, were tested within a microgrid for one year
results were optimized using an Extended Kalman-Particle Filter to assess the economic and environmental benefits of reusing
approach. This study offers a practical solution for the health man- retired electric vehicle batteries, thereby tackling issues related
agement of electric vehicles. Further extending the scope to to energy storage costs and sustainability. However, challenges
include different battery types, the study employs the ICA method, remain, as accurately determining the RUL of a battery set for a
widely applied in battery health forecasting, to predict SOH of real second-life application is fraught with challenges. The availability
EV batteries [124]. The research focuses on LMO/NMC batteries and reliability of historical data, which captures the battery’s
from BMW i3s, conducting aging tests and real EV tests with vari- first-life usage, is often compromised by commercial confidential-
ous mileages. The ICA technique identifies features related to SOH ity. While initiatives such as the Battery Passport strive to enhance
from the battery’s IC curves. The results indicate effective SOH esti- data-sharing reliability, they often face limitations. The battery’s
mation with an overall RMSE of 2%, providing a practical solution operating conditions can vary dramatically between its first and
for EV battery health management. Single predictive models often second life, exemplified by changes in thermal management sys-
disappoint in real-world forecasting performance, whereas hybrid tems and the reduction in the SOC windows and C-rates. Conse-
models leverage the strengths of two different methodologies, quently, models built on first-life performance data may not be
achieving more satisfying prediction outcomes. A study predicting applicable or accurate for second-life predictions.
the actual capacity of electric vehicle batteries based on charging Evaluating a battery pack’s viability for its second life often
data and data-driven algorithms was conducted using data from involves measuring its SOH, commonly done by assessing the
20 vehicles over approximately 29 months [125]. The research capacity under constant-current discharge at different C-rates [3].
incorporates a Sequence-to-Sequence (Seq2Seq) model and a GPR However, relying solely on SOH might not paint the full picture
model. The Seq2Seq model employs LSTM networks to handle of the battery’s potential for second-life use. The trajectory of
sequential data, improving training efficiency through the adjust- capacity fade can demonstrate variability and path dependence
ment of learning rates and optimizers. The GPR model enhances [133]. Batteries with similar SOH may degrade at varying rates in
prediction accuracy by utilizing specific kernel functions and prob- the future, even under identical operational conditions. Hence, pin-
abilistic baselines to boost model performance. The results demon- pointing whether a battery has reached the knee point in its capac-
strate precise capacity forecasting, with errors lower than 1.6%. ity fade curve is pivotal, as degradation tends to accelerate past this
In conclusion, the integration of data-driven modeling, machine point [134].When historical data is scarce, locating the knee point
learning, and cloud data usage presents a promising path toward poses a significant hurdle. Advanced rate performance tests (RPTs)
improving the accuracy and reliability of Li-ion battery aging pre- could be instrumental in estimating the parameters of a physics-
dictions. These advancements can lead to efficient energy manage- based model with more precision than merely capacity and resis-
ment strategies, and enhance the safety and lifecycle of EVs. tance estimation. Furthermore, Table 6 provides a detailed descrip-
tion of the characteristics and challenges associated with
secondary life prediction of batteries, offering a clear framework
4. Battery lifetime prediction for second life and direction for research in this field. Identifying model parame-
ters that correspond with the knee point could prove valuable in
According to a report by Bloomberg New Energy Finance, it is evaluating the viability of second-life batteries. Techniques such
projected that the number of EVs in operation will reach 100 mil- as sensitivity analysis, Fisher information matrix, electrochemical
lion by 2026 [126]. The report further indicates that, with the rise impedance spectroscopy, and differential voltage analysis have
730
X. Qu, D. Shi, J. Zhao et al. Journal of Energy Chemistry 94 (2024) 716–739

Fig. 4. Maximizing the sustainability and economic viability of second-life batteries. (a) Distribution of global new energy electric vehicle ownership; (b) obstacles and
challenges in the secondary use of power batteries; (c) identifying the knee point in the battery aging process is crucial for secondary use, along with listing methods for
locating the knee point; (d) the transition from automotive use to secondary applications (such as energy storage systems) involves changes in battery thermal management,
SOC range, and charge/discharge rates; (e) application scenarios of battery health status at different stages; (f) the positive environmental impact of battery secondary use.

Table 6
Characteristics and challenges of battery prognostics for second life.

Methods Characteristics Challenges

Machine learning Use historical data to identify patterns in battery degradation. Require large amounts of high-quality data.
methods Can incorporate various features such as charge/discharge cycles, temperature, etc. May not generalize well across different battery types or
May include methods such as SVM, artificial neural networks (ANN), etc. operating conditions.
Lack of physical interpretation.
Physics-based Based on understanding of underlying physical processes. Require detailed understanding of battery chemistry and
methods Can provide more accurate and reliable predictions when the physics are well physics.
understood. Computationally expensive and may not be suitable for
real-time applications.
Complexity of adapting to different battery types.
Hybrid methods Combine multiple techniques to improve accuracy and reliability. Increased complexity due to integration of multiple
Might integrate machine learning with physics-based or prognostics and health techniques.
management (PHM) methods. May require more computational resources.
Challenges in selecting and integrating the appropriate
methods.

shown potential in estimating parameters and pinpointing degra- endeavors should aim to tackle these issues and optimize battery
dation modes associated with knee points. Recent advancements health prognostics for second-life scenarios.
in knee point detection methodologies have been notable, particu- In the domain of EV battery repurposing, the economic viability
larly with the application of deep learning approaches. A study uti- of second-life batteries (SLBs) emerges as a key area of exploration.
lizing deep learning to predict battery capacity degradation The economic feasibility of a battery’s secondary life frequently
introduced a dual-phase method, leveraging a CNN model to depends on factors such as the technical performance of battery
extract temporal features from past and future data for real-time degradation, the costs associated with refurbishing battery packs,
prediction of inflection points. This research enhances battery and the potential income for stakeholders [3]. The entire lifecycle
aging prediction performance in real-world scenarios [135]. How- of lithium-ion batteries, which includes production, usage, recy-
ever, translating these techniques to the pack level continues to be cling, and remanufacturing, represents a complex technical trajec-
a challenge as current assessments focus primarily on individual tory [136]. Through health assessment of retired EV batteries,
cells. Addressing this hurdle and fine-tuning prognostic algorithms owners can gain economic compensation by selling these decom-
to cater to different usage patterns in the second life are key steps missioned batteries. Additionally, repurposing these recycled bat-
toward understanding the health and viability of batteries for teries in areas such as energy storage can significantly enhance
second-life applications. Future research and developmental the efficiency of battery reuse. Consequently, establishing an effec-

731
X. Qu, D. Shi, J. Zhao et al. Journal of Energy Chemistry 94 (2024) 716–739

tive system for cascaded utilization and recycling becomes crucial. due to reproducibility concerns. Moreover, implementing
Precise SOH estimation is highlighted as ensuring optimal usage electrochemical impedance spectroscopy (EIS) analysis for
and extending battery life, which directly impacts the economic on-the-spot battery tests is difficult, especially in working
viability and potential revenue of SLB applications [137]. Accurate conditions characterized primarily by time-resolved direct
SOH assessments are instrumental in making better decisions current operations. Such intricacies make it challenging to
about deploying these batteries in stationary storage applications, apply academic findings in an industrial setting. As a result,
where the health and remaining capacity of the battery are crucial it’s crucial for diagnostic methods to be adaptable and effec-
for maximizing operational efficiency and profitability. This precise tive under a variety of conditions.
estimation underpins the development of more effective business (2) Data quality and availability. Field data introduces additional
models and investment strategies, thereby enhancing the overall complexities, including inconsistent battery performance,
value derived from the utilization of second-life batteries. inaccurate sensor readings, noisy data, high charge rates,
To achieve precise SOH estimation and thereby enhance the limited operating conditions, missing or incorrect data, and
revenue and economic feasibility of SLBs, the development of erratic currents. Lab data, in contrast, offers an in-depth
advanced diagnostic and predictive algorithms is essential. These view through various frequency measurements and impe-
algorithms must accurately determine the battery’s current health dance spectroscopy, while real-world data primarily pro-
status and potential lifespan. This process necessitates the integra- vides time-based readings of voltage and current.
tion of battery usage data, historical performance metrics, and Temperature measurements in the field often lack accuracy,
environmental conditions into predictive models that can forecast further complicated by inconsistencies across the cell sur-
battery degradation with greater accuracy [138]. By improving face. This disparity between detailed academic data and lim-
SOH estimation, businesses can more effectively assess the residual ited field data poses challenges for direct application in the
value of SLBs, tailor their use in the most lucrative applications, field. Additionally, transitioning from half-cell to whole-
and optimize replacement cycles. Such strategic enhancements cell models significantly increases complexity, requiring in-
ultimately maximize the profitability and sustainability of battery depth analysis. Lengthy data gaps, such as resting battery
utilization. A significant study introduces a new refurbishment sys- temperature data, further complicate the analysis. Both
tem known as the Heterogeneous Unified Battery (HUB) [139]. This physics-based and data-driven models face limitations when
system addresses the challenge of inconsistent health states addressing these data issues, leading to uncertainties in
among second-life batteries and provides an economic analysis of extrapolations beyond observed data. Physics-based mecha-
this approach. The focus of the study is on the system’s ability to nistic models, while offering valuable insights into the bat-
unify the health states of battery units, thus improving their per- tery’s multiphysics system, are intricate due to heavy
formance in secondary applications. The research contrasts this parameterization. These models may involve hundreds of
new refurbishment method with traditional reuse approaches, uncertain variables ranging from kinetic aspects like reac-
assessing the economic viability and performance benefits of using tion rate constants, which influence electrochemical reac-
refurbished batteries compared to reusing batteries. Special tion speeds, to thermal aspects such as thermal
emphasis is placed on the implications for the battery lifecycle conductivity, crucial for battery heat management. Material
and applications in energy storage solutions. This novel approach attributes of the anode, cathode, electrolyte, and separator,
indicates a forward step in addressing the complexities associated as well as physical dimensions like electrode porosity and
with repurposing second-life batteries, marking a significant con- tortuosity, must also be considered. Addressing real-world
tribution to the field of sustainable energy storage. battery diagnostics with numerous uncertain parameters
and erratic boundary conditions is daunting. Before employ-
ing mechanistic models for real-time use, it’s essential to
5. Challenges in battery prognostics for real-life applications minimize the number of cell parameters required.
(3) Validation and authenticity of data. Validating real-world bat-
In the current era, as we pursue sustainable energy solutions, tery prognostics poses a formidable challenge, as field data
the importance of battery technology is becoming increasingly often lacks a reliable benchmark for comparison. The
prominent, especially in the fields of EVs and renewable energy absence of controlled reference performance tests (RPTs)
storage. However, despite significant advancements in battery complicates the validation of capacity and end-of-life pre-
technology, accurately predicting battery performance and lifespan dictions. Additionally, determining the optimal frequency
in real-world applications continues to pose numerous challenges. for model retraining adds complexity, necessitating a bal-
Deploying battery health prognostics in practical settings faces ance between sufficient training data and the ability to cap-
several obstacles that hinder the accurate assessment of battery ture incremental changes over time. A significant challenge
health and prediction of RUL. These challenges (Fig. 5), arise from is the heterogeneity of data sources, which can range from
a variety of factors including unpredictable usage scenarios, data high-resolution laboratory tests to varied-quality field data
quality and accessibility issues, the comprehensibility of models, collected from sensors with differing accuracies and calibra-
computational efficiency, and scalability concerns. tions. This disparity requires sophisticated data fusion tech-
niques to integrate disparate data streams into a coherent
(1) Unpredictable usage scenarios. In contrast to controlled labo- dataset for effective modeling and analysis. Another chal-
ratory environments, real-life battery usage is fraught with lenge is the computational demand of processing large vol-
variations and uncertainties. Factors such as inconsistent umes of data and running complex models, which calls for
battery pack conditions, temperature fluctuations, varying efficient algorithms and scalable computing infrastructure.
states of charge (SOC) due to different trip lengths, and (4) Model interpretability and explainability. Batteries are sophis-
unpredictable charging/discharging power levels signifi- ticated systems whose functionality is dictated by numerous
cantly impact battery degradation. This path-dependent factors. These range from charge and ion transportation
phenomenon adds another layer of complexity to accurately across various phases and interfaces to a combination of
modeling and predicting battery health. Using galvanostatic both reversible and irreversible chemical reactions. The effi-
charging/discharging (GCD) profiles to derive features from ciency of a cell can be affected by changes in its components,
lab cycling data can pose challenges in practical applications including electrodes, electrolytes, interfaces, and
732
X. Qu, D. Shi, J. Zhao et al. Journal of Energy Chemistry 94 (2024) 716–739

Fig. 5. Exploring the implementation landscape of battery prognostics for real-life applications.

microstructures, as well as by external variables such as offers a scalable infrastructure for battery health prognos-
environmental conditions and operational patterns. tics. Addressing the balance between computational effi-
Although there have been considerable advancements in ciency, scalability, and accuracy remains a critical
comprehensively modeling battery systems, a significant challenge in the field of battery health prognostics.
challenge remains: the development of cohesive predictive (6) Adaptability to new battery technologies. As technology
models that correlate materials and mechanisms across advances, new battery types with diverse chemistries and
diverse scales to the tangible performance of the cell. This configurations continue to emerge. The variability in batter-
disparity has impeded progress in the field. Another major ies extends beyond different types (e.g., LFP, NCA, NMC) and
challenge lies in the interpretability and explainability of is also prevalent within the same batch due to subtle manu-
the models used for battery health prognostics. Complex facturing differences. These discrepancies lead to varied
models, such as deep learning networks, often act as ‘‘black electrochemical properties, affecting the battery’s perfor-
boxes,” making the relationship between inputs and outputs mance and aging process. Although studies using transfer
not easily understood. This lack of transparency can hinder learning highlight domain knowledge and features transfer-
the model’s acceptance, especially in applications where able across different materials, features derived from GCD or
understanding the reasoning behind predictions is crucial. other high-quality static tests often lack practical value
(5) Computational efficiency and scalability. For practical imple- when applied to field data, which is typically short, stochas-
mentation, models for battery health prognostics should be tic, and noisy. Overcoming this challenge necessitates the
computationally efficient and capable of operating in real- development of adaptive algorithms and models that con-
time. However, some sophisticated machine learning algo- sider these variances, allowing for customized strategies
rithms may demand significant computational resources or for batteries from different materials and batches. Battery
time, which may not be practical for certain applications. health prognostic models need to be adaptable to evolving
Additionally, the model should be scalable to handle varying technologies, which involves understanding various degra-
sizes of datasets, from a single vehicle to a fleet of electric dation behaviors and the ability to accommodate different
vehicles (EVs), without significantly compromising predic- data formats and usage scenarios associated with new
tion accuracy or computation time. Exploring innovative technologies.
methods, such as combining physics-based models with
machine learning techniques to create hybrid models, can To address these challenges, hybrid methodologies that merge
optimize the balance between model complexity and com- data-driven and physics-based methods have been explored
putational efficiency. Edge computing presents a viable solu- [140]. These approaches leverage the strengths of both techniques
tion for real-time battery health prognostication. Developing and offer potential solutions for accurate prognostics in real-world
specialized algorithms that can intelligently select and pro- scenarios. By addressing data quality issues, uncertainties in oper-
cess the most relevant data points from large datasets can ating conditions, and the lack of validation data, hybrid methods
significantly reduce the computational load. Employing can significantly enhance prognostic accuracy. The successful
cloud computing for more resource-intensive computations implementation of battery health prognostics in real-life applica-

733
X. Qu, D. Shi, J. Zhao et al. Journal of Energy Chemistry 94 (2024) 716–739

tions requires overcoming challenges associated with varying 6. Opportunities and potential directions
usage scenarios, data quality, and the absence of clear validation
data. Hybrid methodologies, which combine data-driven and Battery health prognostics for real-world applications hold
physics-based techniques, offer promising solutions to these chal- immense opportunities and potential for advancing the reliability
lenges. In addition to these hybrid approaches, a study introduces a and effectiveness of battery systems, particularly in the context
generalizable, online, time-series-based method tailored for of millions of EVs during their service life. The performance of bat-
laboratory-assembled lithium metal batteries to accelerate teries plays a pivotal role in electrified transportation. By combin-
research into new battery chemistries [141]. This approach does ing deep mechanistic understanding, effective prognostic models,
not rely on traditional, high-quality, uniform, and large-scale data- and integration with vehicle systems, battery health prognostics
sets. Instead, it focuses on individual cells by extracting features can play a transformative role in enhancing the performance, reli-
with physical significance from discharge curves to predict the ability, and sustainability of battery-powered systems, contribut-
capacity degradation trajectory of lithium batteries. Typically, ing to a greener and more sustainable future. In this section, we
laboratory-assembled batteries exhibit data quality lower than present four directions that would influence the green and sustain-
that of commercial batteries. According to the results of this study, able environment we strive for, as shown in Fig. 6.
this methodology is poised to have a milestone impact on the
research of novel battery chemistries. Developing robust prognos- 6.1. Battery digital twins
tic models is crucial for effective battery management, mainte-
nance planning, and maximizing battery lifespan across various The term ’Digital Twins’ (DTs), introduced by general electric,
applications. Furthermore, careful data management and consider- encapsulates the idea of a digital replica of a physical entity, such
ation of uncertainties are pivotal for the successful implementation as a battery [142]. DTs, designed to mimic the behavior of their
of battery health prognostics. Apart from the technical challenges, physical counterparts by assimilating real-time measurements
there are several practical issues to consider, such as the return on into computational models, are revolutionizing several industries.
investment ratio. First, there is a significant cost involved in hiring Leveraging Internet of Things (IoT) technology, it is possible to con-
and training data analytics and AI professionals, in addition to the struct a digital twin for cloud-based battery health prognostic sys-
physicochemical experts that the battery-related industry (e.g., EV, tems, significantly enhancing the real-time capabilities of on-site
consumer electronics, electric grid, and so on) already employs. battery health prediction. DT technology integrates advanced algo-
Second, establishing a cloud-based monitoring system demands rithms and machine learning to facilitate fault prediction across
considerable financial commitment. For manufacturers rolling the entire lifecycle of a battery. This approach allows for the itera-
out millions of EVs, this translates to a large investment in plat- tive refinement of battery designs through the testing of virtual
form infrastructure. Third, the time required to train, validate, models, significantly reducing the need for expensive physical pro-
and test AI and machine learning models— even when combined totypes. DTs amalgamate data from various sources, such as
with physics-based approaches— often stretches project timelines charge–discharge cycles, environmental conditions, and usage pat-
and raises expenses. Finally, start-ups are caught in a quandary: terns, thereby offering a comprehensive overview for more effec-
creating an in-house cloud analysis system can be exorbitantly tive battery management. Leveraging IoT technology, it is
expensive, yet sharing battery data with third-party analytics ser- possible to construct a digital twin for cloud-based battery health
vices might not be appealing either. prognostic systems, significantly enhancing the real-time capabili-

Fig. 6. Opportunities and potential directions to improve practical diagnostics of battery lifetime. (a) Challenges associated with the application of digital twin technology in
actual battery health prediction, such as data quality and computational efficiency; (b) describes data-centric artificial intelligence and benchmarks, including the
establishment of evaluation indicators and benchmark design; (c) functional capabilities of blockchain technology applied across various aspects of the battery health
domain; (d) advantages of employing edge cloud computing in battery health prediction are introduced.

734
X. Qu, D. Shi, J. Zhao et al. Journal of Energy Chemistry 94 (2024) 716–739

ties of on-site battery health prediction [143]. This technology sub- develop advanced computational methods and data processing
stantially enhances the accuracy of battery performance predic- techniques, the robustness and reliability of DTs in battery man-
tions, contributing to the development of more robust and agement systems’ real-world applications will continue to improve
efficient battery technologies. However, before these cutting- [144]. The technology promises enhanced battery performance,
edge technologies can be fully operational in practical applications, safety, and life-cycle management, leading to significant benefits
certain fundamental challenges need to be addressed. for a wide range of industries.

(1) Data quality and availability. In real-world settings, available 6.2. Data-centric AI and benchmarks
observational data can be sparse and noisy, and often come
in various formats, such as images, time series data, histori- Data-driven AI, specifically deep learning, has significantly
cal records, etc. Moreover, certain critical information may transformed fields such as imaging, speech, and natural language
not be directly accessible or measurable. Ensuring sufficient, processing [145–147]. Standard benchmarks have been widely
high-quality data is crucial for the accurate and reliable adopted in these areas to gauge algorithmic progress, validate
operation of a battery’s DT. By installing high-precision result reproducibility, and estimate computational cost. The UCI
and highly reliable sensors, key data on battery performance machine learning repository [148], for instance, hosts numerous
is collected, ensuring accurate and consistent data through- databases frequently utilized to compare algorithmic performance.
out the battery’s entire lifecycle. The data collection fre- While the repository primarily caters to conventional machine
quency is optimized by increasing the sampling rate for learning applications, it also houses datasets related to physical
critical parameters while reducing it for parameters with sciences, potentially suitable for physics-informed deep learning
less variability. Furthermore, data undergoes preprocessing methods. However, battery health prognostics in real-world sce-
using techniques such as filtering, anomaly detection, and narios present unique challenges and requirements that necessi-
data imputation. Finally, data from various sensors are inte- tate careful benchmarking considerations. Here, we highlight
grated, and relevant features for predicting battery health several key areas.
are extracted, thereby ensuring the quality and availability
of battery data. (1) Dataset creation and management. Many applications in phy-
(2) Computational and calibration complexity. Traditional sics and chemistry require full-field data, which are often
physics-based computational models heavily rely on cum- complex to acquire experimentally and computationally
bersome preprocessing and calibration processes, which demanding in terms of time and memory. Therefore, we
could limit their usability in real-time decision-making set- need to focus on how to make such data publicly accessible,
tings. To facilitate real-time prediction and decision-making, properly curated, and inclusive of necessary physical models
it is imperative to rapidly process and analyze data. This and parameters. This entails implementing advanced data
becomes particularly critical when dealing with complex acquisition techniques, such as high-resolution sensors and
models and algorithms, as it imposes heightened computa- real-time monitoring systems, to capture detailed and accu-
tional demands. Furthermore, ensuring the accuracy of mod- rate battery performance data. Additionally, leveraging
els under various conditions necessitates ongoing cloud storage and computing resources can address the chal-
calibration and validation. Moreover, these models typically lenges of storing and processing large volumes of data. This
require precise information about initial and boundary con- will require careful planning and coordination among the
ditions. The computational complexity and resource- research community to generate valuable and comprehen-
intensive nature of these processes often pose challenges sive databases suitable for battery prognostics. Furthermore,
to the real-world deployment of DTs. Developing efficient incorporating data standardization protocols and data clean-
computational and calibration methods is crucial to enhanc- ing methods can enhance the quality and usability of the
ing the feasibility of real-time DT-based prognostics. datasets. The integration of simulation data, derived from
(3) Incomplete knowledge of physical systems. Our understanding advanced battery models, with experimental data, can also
of many complex systems is often partial. Such as the perfor- enrich the database, providing a more holistic view of bat-
mance and aging process of batteries are governed by com- tery behavior under various conditions.
plex electrochemical reactions, such as ion migration, (2) Benchmark design. Designing meaningful benchmarks to
electrode reactions, and material degradation. The intrica- assess the accuracy and speed of physics-informed algo-
cies of these processes are not yet fully understood, leading rithms for battery health prognostics is not straightforward.
to uncertainties in model construction. Moreover, the The challenge is further compounded by the goal of predict-
impact of external conditions such as temperature, humid- ing dynamics, necessitating benchmarks capable of handling
ity, and mechanical stress on battery performance and nonlinearities, bifurcations in dynamical systems, and chao-
degradation is not completely elucidated in current model- tic states. To enhance these benchmarks, integrating diverse
ing efforts. Although we have established conservation laws, datasets that represent a wide range of operating conditions
they cannot form a complete system of equations without and battery types is crucial. This includes data from extreme
hypothesizing appropriate constitutive laws. This incom- conditions, such as high temperatures or rapid charge–dis-
plete knowledge can lead to uncertainties in the models charge cycles, which can significantly impact battery health.
used for DTs, affecting their predictive accuracy and Researchers need to collaborate to develop benchmarks that
reliability. test predictive accuracy and account for real-world com-
plexities effectively.
The field of physics-informed learning, which integrates physi- (3) Evaluation metrics. Traditional metrics used in machine
cal models and data and leverages automatic differentiation to learning might not be wholly suitable for battery health
bypass the need for mesh generation, has the potential to be a sig- prognostics, given their dynamic nature. New metrics, such
nificant enabler in the era of DTs. By addressing these challenges, as ’valid-time-prediction’ introduced recently, offer a
DTs can become more accurate and reliable tools for battery health promising direction. These metrics could effectively quantify
prognostics in real-world applications. The future of battery health the validity period of prognostic predictions, making them
prognostics is set to be transformed by DTs. As we continue to more suitable for evaluating prognostic methods that antic-
735
X. Qu, D. Shi, J. Zhao et al. Journal of Energy Chemistry 94 (2024) 716–739

ipate dynamic battery behavior. Developing composite met- sources, thereby enhancing the robustness and comprehen-
rics that combine accuracy, reliability, and computational siveness of health diagnostics.
efficiency could offer a more holistic evaluation of battery (2) Live diagnostics. With live data recording on a blockchain,
health prognostics systems, especially in operational envi- battery health diagnostics can be performed in real-time,
ronments where quick decision-making is critical. enhancing the speed of maintenance and replacement
(4) Hardware and software considerations. Benchmarks should strategies, and in turn, improving the dependability of the
not only focus on the accuracy of predictions but also con- systems they power. Moreover, the integration of artificial
sider the computational efficiency of the methods. This intelligence and machine learning algorithms with
involves an analysis of both software algorithms and the blockchain-recorded data can provide predictive insights,
hardware on which they are run, thereby providing a more forecasting potential future failures and optimizing mainte-
holistic evaluation. In this context, it’s essential to evaluate nance schedules based on actual usage patterns and perfor-
the scalability of algorithms across different hardware plat- mance metrics.
forms, ranging from high-performance computing systems (3) Data sharing and privacy. Blockchain’s decentralized archi-
to edge devices used in real-world applications. The compat- tecture enables diverse stakeholders such as manufacturers,
ibility and optimization of algorithms for parallel processing service providers, and users to access battery health data
architectures, such as GPUs, can significantly impact their without jeopardizing privacy or control of data. This frame-
performance. Additionally, assessing the energy efficiency work provides a secure and transparent environment for
of these methods when deployed on portable or embedded data exchange, where access rights can be precisely man-
systems is critical, as it directly affects the practicality and aged through cryptographic techniques and smart contracts.
sustainability of battery health monitoring solutions. These smart contracts can automatically enforce privacy
policies and data-sharing agreements, ensuring that each
As we shift towards a more data-centric AI approach for battery stakeholder can only access the information that is relevant
health prognostics, establishing robust benchmarks becomes criti- and permitted. This could stimulate data sharing and collec-
cal. Such efforts will require the combined initiative of the research tive problem-solving. The use of blockchain can foster a
community to develop comprehensive databases, design suitable trust-based ecosystem for collaborative research and devel-
benchmarks, devise relevant evaluation metrics, and take into opment, where anonymized data can be utilized for collec-
account hardware-software efficiency. This will set the stage for tive advancements in battery technology, without
significant advancements in battery health prognostics, leading compromising individual data security or proprietary
to more reliable and efficient real-world applications. information.
(4) Smart contracts for streamlined operations. Another feature of
6.3. Blockchain technology blockchain technology is ’smart contracts,’ software pro-
grams that automatically carry out actions when specific
Blockchain technology, best recognized for its use in cryptocur- conditions are fulfilled. For battery health, these could
rencies such as Bitcoin, is fundamentally a secure and transparent initiate maintenance alerts, replacements, or warranty
way to record transactions across a decentralized network [149]. claims when predefined health thresholds are reached.
However, its potential applications stretch far beyond financial Smart contracts can also facilitate automated data analysis
transactions. Among these, an innovative prospect is its application for health prediction, integrating with IoT devices to contin-
to tackle the complexities of battery health diagnostics in real- uously monitor battery parameters and execute real-time
world scenarios. While there’s a considerable buzz around block- diagnostics. This automated process enhances operational
chain technology, it’s important to fully comprehend its features efficiency, reducing human error and response time in
and apply them appropriately. Data-driven technologies are critical situations.
increasingly gaining popularity in the field of battery health pre-
diction, with data security emerging as a crucial component in Deciding between permissionless and permissioned blockchain
the on-site applications of EVs [150]. The technology relies heavily networks, based on the necessary level of access and control,
on cryptography, using private and public keys to securely carry requires thoughtful decision-making. Dealing with malicious users,
out transactions. Unlike conventional databases, altering data on implementing robust controls, and understanding the limitations
a blockchain is not straightforward. Changes are typically incorpo- of blockchain implementations are also important considerations.
rated as new blocks, preserving a complete alteration history and, Governance-related challenges, especially for permissioned block-
in the context of batteries, creating a ’digital mirror’ of their chain networks, need to be addressed to decide who will manage
lifecycle. and govern the network. In conclusion, while blockchain technol-
The essential structure of blockchain as a distributed ledger sys- ogy is not a universal solution when understood and applied aptly,
tem made of blocks, each block comprising a header with metadata it can serve as a powerful instrument in the realm of battery health
and transaction-related data, can be harnessed for battery health diagnostics. As a continually evolving technology, its application in
diagnostics. Here are some potential outcomes. battery health diagnostics should keep pace with these advance-
ments. The potential advantages in terms of dependability, trans-
(1) Robust records. Blockchain’s inherent resistance to tamper- parency, and efficiency are noteworthy and call for continued
ing delivers a reliable and immutable record of a battery’s research and exploration.
health data throughout its life cycle. This technology can
facilitate secure and transparent data sharing among multi- 6.4. Edge-cloud collaboration
ple stakeholders, such as manufacturers, service providers,
and end-users, enhancing collaborative diagnostics and In the context of Industry 4.0, the integration of cloud-based AI
maintenance. This could enable precise diagnostics, mini- technologies with battery PHM holds immense potential [151].
mize warranty-related fraud, and potentially boost the This convergence paves the way for collaborative endeavors
resale worth of used batteries by providing an irrefutable between battery scientists and AI experts, enabling the systematic
health record. Integrating blockchain with IoT devices facil- fusion of data and models. Moreover, fostering closer collabora-
itates the aggregation of decentralized data from diverse tions between academia and industry is crucial to enhancing our
736
X. Qu, D. Shi, J. Zhao et al. Journal of Energy Chemistry 94 (2024) 716–739

understanding of observational data, which often exhibits signifi- (3) Merging edge cloud and machine learning. The deployment of
cant heterogeneity across various modalities. These include remote a battery health prediction model on the edge cloud, serving
sensor data, in-situ observations, lab tests, historical data, as well a range of IoT devices, can redefine the conventional
as time and frequency domain information. The amalgamation of approach. These devices can transmit battery usage and
battery health prognostics and edge-cloud technology introduces associated data to the edge cloud, which the model can pro-
tremendous prospects to augment the lifespan and reliability of cess to predict the battery’s health status. By integrating
IoT devices. In this context, an innovative end-cloud collaboration machine learning algorithms into this setup, the edge cloud
method is emphasized, integrating end-side empirical models with can not only handle large volumes of data but also learn and
cloud-side data-driven models for the predictive maintenance of adapt over time, improving the accuracy of its predictions.
lithium-ion batteries [152]. This approach leverages ensemble These algorithms can identify subtle patterns and trends in
learning methods for constructing highly accurate cloud-side pre- battery performance, which might be missed by traditional
dictions, and employs a double exponential decay model for real- methods. Utilizing technologies such as network function
time end-side predictions. By incorporating Kalman filtering, it virtualization (NFV) and software defined networking
updates the empirical model’s predictions with cloud-side data, (SDN), the edge cloud platform could efficaciously allocate
aiming for enhanced accuracy and timeliness in battery health esti- resources to process and scrutinize the offloaded data, facil-
mation. This methodology could significantly contribute to the bat- itating precise and swift predictions. With the help of NFV
tery health management systems, ensuring safety and efficiency and SDN, the edge cloud infrastructure can dynamically
throughout the battery’s lifecycle, supported by a cyber hierarchy adjust its computational and network capacities to meet
and interactive network architecture. Transposing these burgeon- the varying demands of battery health analysis, ensuring
ing technologies into practical applications mandates a collective efficient utilization of resources and reducing operational
effort from diverse entities, such as IoT device manufacturers, costs.
edge-cloud service providers, and academicians [153]. By harness- (4) Pioneering advanced models. The performance of such a
ing the power of cloud-edge collaborative computing, this innova- system is significantly influenced by the development of
tive approach offers a more flexible and efficient alternative [154]. sophisticated machine learning or AI models designed
The manufacturing sector is harnessing the synergies of IoT and specifically for battery health prognostics. These models
cloud computing, aiming to supercharge operational efficiencies, should efficiently handle large data sets, adapt to variations
refine product quality, bolster response capabilities to shifting cus- in battery usage patterns, and continuously improve predic-
tomer requirements, and facilitate market expansion [155]. Pillars tion accuracy. To achieve this, the incorporation of deep
such as productivity, quality, safety, and agility in adapting to learning techniques could be particularly beneficial. These
dynamic conditions form the bedrock of the industry’s competitive techniques are capable of extracting complex features and
edge. The compelling strength of edge-cloud technology is gaining temporal patterns from battery data, which are critical for
momentum in this sector, as it marries the processing prowess of accurate health assessment. Furthermore, the use of transfer
the cloud with the proximity and real-time responsiveness of edge learning can expedite the training process by leveraging pre-
computing. The resultant streamlined processes, real-time data trained models on similar tasks, thus enhancing the model’s
analytics, and swift adaptations to changing circumstances set ability to generalize across different types of batteries and
the stage for an industry primed to meet the future’s challenges operational conditions. Additionally, the implementation of
and opportunities. Addressing a sequence of salient questions is reinforcement learning could optimize battery usage and
the stepping stone toward materializing these theoretical strides maintenance decisions based on the prognostic outputs,
into applicable benefits. further extending the battery’s lifespan and operational
efficiency.
(1) Empowering battery health prognostics through IoT and edge- (5) Prioritizing data security and privacy. Implementation of
cloud computing. The fusion of IoT and edge-cloud technol- battery health prognostics on edge cloud necessitates
ogy could transform battery health prognostics radically. addressing issues of data privacy and security. With the like-
Given that most IoT devices run on batteries, prognosticating lihood of sensitive data being offloaded to the edge cloud,
battery life and health is crucial to bolstering their perfor- ensuring secure data transmission and storage is paramount.
mance and extending their usage period. Processing com- This requires the incorporation of stringent security mea-
prehensive battery health data at the edge can drastically sures such as data encryption, secure access controls, and
curtail latency, thereby enabling prompt battery health privacy-preserving computing techniques. To enhance this,
prognostics. By leveraging edge computing, data analysis advanced encryption protocols and technologies, along with
can be performed closer to where it is generated, reducing blockchain technology, can be employed to safeguard data
the need to transmit large volumes of data to centralized during transmission, thereby further augmenting data
cloud servers. This not only minimizes transmission delays security and privacy.
but also enhances data security by limiting exposure.
(2) Proactive measures stemming from real-time predictions. Rapid To sum up, the synergy of battery health prognostics and edge-
predictive capability lays the groundwork for instant pre- cloud computing offers substantial potential. Actualizing this
ventive or remedial actions, enhancing the functionality potential requires a concerted effort to navigate the challenges
and reliability of IoT devices. Battery health prognostics and develop efficacious solutions, thereby transforming these
could find extensive utility across domains such as EVs, promising technologies into real-world applications.
drones, and other battery-reliant IoT devices. By analyzing
patterns and anomalies in battery usage and condition data, 7. Conclusions
these systems can forecast future battery health status,
enabling timely interventions. The real-time aspect of these In this review, we have conducted a comprehensive analysis of
predictions is crucial for dynamic environments where bat- battery health prognostics for real-world applications, highlighting
tery performance directly impacts the overall functionality the challenges and opportunities associated with this field. Battery
of the device. health prognostics have gained significant importance in the con-

737
X. Qu, D. Shi, J. Zhao et al. Journal of Energy Chemistry 94 (2024) 716–739

text of energy storage systems, particularly in EVs and renewable [14] J. Liu, Y. Zhang, J. Bai, L. Zhou, Z. Wang, Electrochim. Acta. 454 (2023) 142362.
[15] V. Sulzer, P. Mohtat, A. Aitio, S. Lee, Y.T. Yeh, F. Steinbacher, M.U. Khan, J.W.
energy sectors, where the durability and dependability of batteries
Lee, J.B. Siegel, Joule 5 (2021) 1934–1955.
are crucial. We have explored various methodologies for battery [16] Z. Chen, D.L. Danilov, R.A. Eichel, P.H. Adv, Energy Mater. 12 (2022) 2201506.
health prognostics, including battery modeling approaches such [17] U. Krewer, F. Röder, E. Harinath, R.D. Braatz, B. Bedürftig, R. Findeisen, J.
as physics-based models, empirical models, and equivalent circuit Electrochem. Soc. 165 (2018) A3656.
[18] J.S. Edge, S. O’Kane, R. Prosser, N.D. Kirkaldy, A.N. Patel, A. Hales, A. Ghosh, W.
models. Additionally, we have delved into data-driven prognostics, Ai, J. Chen, J. Yang, Phys. Chem. Chem. Phys. 23 (2021) 8200–8221.
employing conventional machine learning techniques and [19] Pybamm, Battery Modelling Webinar Series 2021. https://pybamm.org/,
advanced deep neural network approaches. The advantages and (accessed Jan. 2024).
[20] M. Varini, P.E. Campana, G. Lindbergh, J. Energy Storage. 25 (2019) 100819.
limitations of hybrid approaches and the role of information fusion [21] S. Sankarasubramanian, B. Krishnamurthy, Electrochim. Acta. 70 (2012) 248–
in enhancing prognostic accuracy have also been examined. 254.
Through the analysis of real-world case studies, we have demon- [22] E.J. Dufek, T.R. Tanim, B.-R. Chen, S. Kim, Joule 6 (2022) 1363–1367.
[23] J. Zhao, Y. Gao, J. Guo, L. Chu, A.F. Burke, Int. J. Electrochem. Sci. 13 (2018)
strated the practicality and adaptability of these methodologies. 1773–1786.
However, the deployment of battery prognostics in real-life appli- [24] S. Ansari, A. Ayob, M.H. Lipu, A. Hussain, M.H.M. Saad, Energy Rep. 8 (2022)
cations presents several implementation challenges, including 12153–12185.
[25] F. Yang, X. Song, G. Dong, K.-L. Tsui, Energy. 171 (2019) 1173–1182.
scalability issues, complexities in data interpretation, and inconsis- [26] M. Soltani, J. Ronsmans, J. Van Mierlo, J. Energy Storage. 31 (2020) 101659.
tencies in data quality. These challenges need to be addressed to [27] L. Zheng, J. Zhu, G. Wang, D.-D.-C. Lu, T. He, Energy. 158 (2018) 1028–1037.
ensure the successful integration of battery health prognostics in [28] M. Ahwiadi, W. Wang, IEEE Trans. Instrum. Meas. 68 (2018) 923–935.
[29] X. Hu, S. Li, H. Peng, J. Power Sources 198 (2012) 359–367.
real-world scenarios. Nevertheless, we have identified promising
[30] M. Shen, Q. Gao, Int. J. Energy Res. 43 (2019) 5042–5075.
opportunities and potential directions for future advancements in [31] J.-F. Ding, R. Xu, C. Yan, B.-Q. Li, H. Yuan, J.-Q. Huang, J. Energy Chem. 59
battery health prognostics. These include the development of bat- (2021) 306–319.
tery digital twin models, the application of data-centric AI tech- [32] A. Guha, A. Patra, IEEE Trans. Instrum. Meas. 67 (2018) 1836–1849.
[33] B. Khaki, P. Das, J. Energy Storage. 39 (2021) 102587.
niques and the establishment of benchmarks, the integration of [34] J. Guo, W. Liu, L. Chu, J. Zhao, Energies 12 (2019) 4251.
blockchain technology for enhanced data security and trans- [35] J. Zhao, A.F. Burke, Batteries 8 (2022) 142.
parency, and the synergy of edge-cloud collaboration to optimize [36] J. Zhao, X. Feng, Q. Pang, J. Wang, Y. Lian, M. Ouyang, A.F. Burke, J. Power
Sources 581 (2023) 233474.
data computation and analysis. In conclusion, this paper has pro- [37] M.-F. Ng, J. Zhao, Q. Yan, G.J. Conduit, Z.W. Seh, Nat. Mach. Intell. 2 (2020)
vided a comprehensive overview of battery health and lifetime 161–170.
prognostics for real-world applications. By expanding our under- [38] X. Chen, X. Liu, X. Shen, Q. Zhang, Angew. Chem. 133 (2021) 24558–24570.
[39] A. Manoharan, K. Begam, V.R. Aparow, D. Sooriamoorthy, J. Energy Storage 55
standing of existing techniques, addressing challenges, and capital- (2022) 105384.
izing on emerging opportunities, researchers and professionals in [40] X. Sui, S. He, S.B. Vilsen, J. Meng, R. Teodorescu, D.-I. Stroe, Appl. Energy 300
this field can further enhance the reliability, durability, and perfor- (2021) 117346.
[41] X. Shu, J. Shen, G. Li, Y. Zhang, Z. Chen, Y. Liu, IEEE Trans. Transp.
mance of energy storage systems, contributing to the broader Electrification. 7 (2021) 2238–2248.
adoption of EVs and renewable energy technologies. [42] Y. Che, X. Hu, X. Lin, J. Guo, R. Teodorescu, Energy Environ. Sci. (2023).
[43] H. Meng, M. Geng, T. Han, Reliab. Eng. Syst. Saf. 236 (2023) 109288.
[44] S. Yang, C. Zhang, J. Jiang, W. Zhang, L. Zhang, Y. Wang, J. Clean. Prod. 314
Declaration of competing interest (2021) 128015.
[45] Y. Xu, X. Chen, H. Zhang, F. Yang, L. Tong, Y. Yang, D. Yan, A. Yang, M. Yu, Z. Liu,
The authors declare that they have no known competing finan- Int. J. Energy Res. 46 (2022) 19615–19652.
[46] A. Jović, K. Brkić, N. Bogunović, In: 2015 38th International Convention on
cial interests or personal relationships that could have appeared
Information and Communication Technology, Electronics And
to influence the work reported in this paper. Microelectronics (MIPRO), 2015: IEEE, pp. 1200–1205.
[47] M. Lin, C. Yan, J. Meng, W. Wang, J. Wu, Energy 250 (2022) 123829.
[48] Y. Li, C. Zou, M. Berecibar, E. Nanini-Maury, J.-C.-W. Chan, P. Van den Bossche,
Acknowledgments
J. Van Mierlo, N. Omar, Appl. Energy 232 (2018) 197–210.
[49] G. Dong, W. Han, Y. Wang, IEEE Trans. Ind. Electron. 68 (2020) 10949–10958.
This research was funded by the Independent Innovation Pro- [50] C. Qian, N. He, L. He, H. Li, F. Cheng, J. Energy Storage. 68 (2023) 107754.
[51] P. Tagade, K.S. Hariharan, S. Ramachandran, A. Khandelwal, A. Naha, S.M.
jects of the Hubei Longzhong Laboratory (2022ZZ-24) and the Cen-
Kolake, S.H. Han, J. Power Sources. 445 (2020) 227281.
tral Government to Guide Local Science and Technology [52] X. Li, C. Yuan, Z. Wang, Energy. 203 (2020) 117852.
Development fund Projects of Hubei Province (2022BGE267). [53] S.B. Vilsen, D.-I. Stroe, J. Clean. Prod. 290 (2021) 125700.
[54] X. Feng, C. Weng, X. He, X. Han, L. Lu, D. Ren, M. Ouyang, IEEE Trans. Veh.
Technol. 68 (2019) 8583–8592.
References [55] X. Sun, K. Zhong, M. Han, Measurement. 170 (2021) 108679.
[56] D. Yang, X. Zhang, R. Pan, Y. Wang, Z. Chen, J. Power Sources. 384 (2018) 387–
[1] M. Li, J. Lu, Z. Chen, K. Amine, Adv. Mater. 30 (2018) 1800561. 395.
[2] X. Hu, L. Xu, X. Lin, M. Pecht, Joule 4 (2020) 310–346. [57] Y. Zhang, Y.-F. Li, Renew. Sustain. Energy Rev. 161 (2022) 112282.
[3] M. Shahjalal, P.K. Roy, T. Shams, A. Fly, J.I. Chowdhury, M.R. Ahmed, K. Liu, [58] J. Zhao, X. Han, M. Ouyang, A.F. Burke, J. Energy Chem. 87 (2023) 416–438.
Energy 241 (2022) 122881. [59] B. Rezaeianjouybari, Y. Shang, Measurement 163 (2020) 107929.
[4] J. Zhao, H. Ling, J. Liu, J. Wang, A.F. Burke, Y. Lian, eTransportation 15 (2023) [60] C. Qian, B. Xu, L. Chang, B. Sun, Q. Feng, D. Yang, Y. Ren, Z. Wang, Energy 227
100214. (2021) 120333.
[5] J. Zhao, X. Feng, M.-K. Tran, M. Fowler, M. Ouyang, A.F. Burke, J. Power Sources [61] J. Wu, J. Chen, X. Feng, H. Xiang, Q. Zhu, J. Energy Storage 55 (2022) 105708.
598 (2024) 234111. [62] J. Lu, R. Xiong, J. Tian, C. Wang, C.-W. Hsu, N.-T. Tsou, F. Sun, J. Li, En. Stor.
[6] Y. Wang, J. Tian, Z. Sun, L. Wang, R. Xu, M. Li, Z. Chen, Renew. Sustain. Energy Mater. 50 (2022) 139–151.
Rev. 131 (2020) 110015. [63] P. Wen, Z.-S. Ye, Y. Li, S. Chen, P. Xie, S. Zhao, IEEE Trans. Intell. Veh. 9 (2023)
[7] J. Zhao, X. Feng, Q. Pang, M. Fowler, Y. Lian, M. Ouyang, A.F. Burke, Progr. 2276–2289.
Energy Combust. Sci. 102 (2024) 101142. [64] Y. Jiang, Y. Chen, F. Yang, W. Peng, J. Power Sources. 556 (2023) 232466.
[8] X. Han, L. Lu, Y. Zheng, X. Feng, Z. Li, J. Li, M. Ouyang, eTransportation 1 (2019) [65] S. Shen, M. Sadoughi, M. Li, Z. Wang, C. Hu, Appl. Energy 260 (2020) 114296.
100005. [66] G. Lee, D. Kwon, C. Lee, Mech. Syst. Signal Process. 188 (2023) 110004.
[9] A. Downey, Y.-H. Lui, C. Hu, S. Laflamme, S. Hu, Reliab. Eng. Syst. Saf. 182 [67] C. Zhu, M. Gao, Z. He, H. Wu, C. Sun, Z. Zhang, Z. Bao, J. Energy Storage. 65
(2019) 1–12. (2023) 107218.
[10] C.-P. Lin, J. Cabrera, F. Yang, M.-H. Ling, K.-L. Tsui, S.-J. Bae, Appl. Energy. 275 [68] R. Rouhi Ardeshiri, C. Ma, Int. J. Energy Res. 45 (2021) 16633–16648.
(2020) 115338. [69] K. Luo, H. Zheng, Z. Shi, J. Power Sources. 576 (2023) 233139.
[11] S. Amir, M. Gulzar, M.O. Tarar, I.H. Naqvi, N.A. Zaffar, M.G. Pecht, IEEE Access [70] Z. Li, F. Liu, W. Yang, S. Peng, J. Zhou, IEEE Trans. Neural Netw. Learn. Syst.
10 (2022) 18279–18288. (2022) 6999–7019.
[12] M. Guo, G.-H. Kim, R.E. White, J. Power Sources. 240 (2013) 80–94. [71] X. Gu, K. See, P. Li, K. Shan, Y. Wang, L. Zhao, K.C. Lim, N. Zhang, Energy 262
[13] H. Tu, S. Moura, Y. Wang, H. Fang, Appl. Energy. 329 (2023) 120289. (2023) 125501.

738
X. Qu, D. Shi, J. Zhao et al. Journal of Energy Chemistry 94 (2024) 716–739

[72] K. Luo, X. Chen, H. Zheng, Z. Shi, Energy Chem 74 (2022) 159–173. [117] C. She, Z. Wang, F. Sun, P. Liu, L. Zhang, IEEE Trans. Ind. Inform. 16 (2019)
[73] Q. Zhang, M. Zhang, T. Chen, Z. Sun, Y. Ma, B. Yu, Neurocomputing 323 (2019) 3345–3354.
37–51. [118] L. Vichard, A. Ravey, P. Venet, F. Harel, S. Pelissier, D. Hissel, Energy 225
[74] L. Ren, J. Dong, X. Wang, Z. Meng, L. Zhao, M.J. Deen, IEEE Trans. Ind. Inform. (2021) 120235.
17 (2020) 3478–3487. [119] J. Bi, Y. Wang, Q. Sai, C. Ding, Energy 169 (2019) 833–843.
[75] Q. Xie, R. Liu, J. Huang, J. Su, Energy 281 (2023) 128232. [120] M. De Gennaro, E. Paffumi, G. Martini, A. Giallonardo, S. Pedroso, A. Case Stud.
[76] W.H.L. Pinaya, S. Vieira, R. Garcia-Dias, A. Mechelli, Autoencoders, In: Transp. Policy. 8 (2020) 517–534.
Machine Learning, Elsevier, Amsterdam, 2020, pp. 193–208. [121] K. Li, P. Zhou, Y. Lu, X. Han, X. Li, Y. Zheng, J. Power Sources 468 (2020)
[77] J. Obregon, Y.-R. Han, C.W. Ho, D. Mouraliraman, C.W. Lee, J.-Y. Jung, J. Energy 228192.
Storage 60 (2023) 106680. [122] H. Rauf, M. Khalid, N. Arshad, J. Energy Storage 68 (2023) 107577.
[78] K. Pawar, V.Z. Attar, Assessment of Autoencoder Architectures for Data [123] J. Tian, X. Liu, S. Li, Z. Wei, X. Zhang, G. Xiao, P. Wang, Energy 270 (2023)
Representation, Deep Learning: Concepts and Architectures, Springer, Berlin, 126855.
2020, pp. 101–132. [124] A. Gismero, K. Nørregaard, B. Johnsen, L. Stenhøj, D.-I. Stroe, E. Schaltz, J.
[79] L. Ren, L. Zhao, S. Hong, S. Zhao, H. Wang, L. Zhang, IEEE Access. 6 (2018) Energy Storage 64 (2023) 107110.
50587–50598. [125] Z. Deng, L. Xu, H. Liu, X. Hu, Z. Duan, Y. Xu, Appl. Energy 339 (2023)
[80] M. Wei, M. Ye, Q. Wang, J.P. Twajamahoro, J. Energy Storage. 47 (2022) 120954.
103558. [126] BloombergNEF, Electric Vehicle Fleet Set to Hit 100 Million by 2026, but
[81] H. Salehinejad, S. Sankar, J. Barfett, E. Colak, S. Valaee, arXiv preprint Stronger Push Needed to Stay on Track for Net Zero. https://about.bnef.com/
arXiv:1801.01078. (2017). blog/electric-vehicle-fleet-set-to-hit-100-million-by-2026-but-stronger-
[82] H. Hewamalage, C. Bergmeir, K. Bandara, Int. J. Forecast. 37 (2021) 388–427. push-needed-to-stay-on-track-for-net-zero/, 2023 (accessed March 2024).
[83] A. Sherstinsky, Phys. D. 404 (2020) 132306. [127] BloombergNEF, Battery bloat could backfire on electric vehicle
[84] Y. Yu, X. Si, C. Hu, J. Zhang, Neural Comput. 31 (2019) 1235–1270. manufacturers. https://about.bnef.com/blog/battery-bloat-could-backfire-
[85] H. Sun, D. Yang, J. Du, P. Li, K. Wang, Energy Rep. 8 (2022) 442–449. on-electric-vehicle-manufacturers/, 2023 (accessed March 2024).
[86] L. Alzubaidi, J. Zhang, A.J. Humaidi, A. Al-Dujaili, Y. Duan, O. Al-Shamma, J. [128] M.F. Börner, M.H. Frieges, B. Späth, K. Spütz, H.H. Heimes, D.U. Sauer, W. Li,
Santamaría, M.A. Fadhel, M. AI-Amidie, L. Farhan, J. Big Data. 8 (2021) 1–74. Cell Rep. Phys. Sci. 3 (2022) 101095.
[87] J.C. Chen, T.-L. Chen, W.-J. Liu, C. Cheng, M.-G. Li, Adv. Eng. Inform. 50 (2021) [129] E. Hossain, D. Murtaugh, J. Mody, H.M.R. Faruque, M.S.H. Sunny, N.
101405. Mohammad, IEEE Access 7 (2019) 73215–73252.
[88] R.R. Ardeshiri, M. Liu, C. Ma, Reliab. Eng. Syst. Saf. 224 (2022) 108481. [130] X. Hu, X. Deng, F. Wang, Z. Deng, X. Lin, R. Teodorescu, M.G. Pecht, Proc. IEEE.
[89] C. Han, Y.C. Gao, X. Chen, X. Liu, N. Yao, L. Yu, L. Kong, Q. Zhang, InfoMat. 110 (2022) 735–753.
(2024) e12521. [131] J. Thakur, C.M.L. de Almeida, A.G. Baskar, J. Clean. Prod. 375 (2022) 134066.
[90] M. Lin, Y. You, W. Wang, J. Wu, Reliab. Eng. Syst. Saf. 230 (2023) 108978. [132] J. Lacap, J.W. Park, L. Beslow, J. Energy Storage. 41 (2021) 102837.
[91] G.E. Karniadakis, I.G. Kevrekidis, L. Lu, P. Perdikaris, S. Wang, L. Yang, Nat. Rev. [133] T. Raj, A.A. Wang, C.W. Monroe, D.A. Howey, Batteries & Supercaps. 3 (2020)
Phys. 3 (2021) 422–440. 1377–1385.
[92] J. Shi, A. Rivera, D. Wu, Mech. Syst. Signal Process. 179 (2022) 109347. [134] P.M. Attia, A. Bills, F.B. Planella, P. Dechent, G. Dos Reis, M. Dubarry, P.
[93] W. Li, J. Zhang, F. Ringbeck, D. Jöst, L. Zhang, Z. Wei, D.U. Sauer, J. Power Gasper, R. Gilchrist, S. Greenbank, D. Howey, J. Electrochem. Soc. 169
Sources. 506 (2021) 230034. (2022) 060517.
[94] H. Pang, L. Wu, J. Liu, X. Liu, K. Liu, J. Energy Chem. 78 (2023) 1–12. [135] S. Sohn, H.-E. Byun, J.H. Lee, Appl. Energy. 328 (2022) 120204.
[95] R.G. Nascimento, M. Corbetta, C.S. Kulkarni, F.A. Viana, J. Power Sources. 513 [136] M. Chen, X. Ma, B. Chen, R. Arsenault, P. Karlson, N. Simon, Y. Wang, Joule. 3
(2021) 230526. (2019) 2622–2646.
[96] B. Sun, J. Pan, Z. Wu, Q. Xia, Z. Wang, Y. Ren, D. Yang, X. Guo, Q. Feng, J. Power [137] H. Iqbal, S. Sarwar, D. Kirli, J.K. Shek, A.E. Kiprakis, Carbon Neutrality 2 (2023)
Sources. 556 (2023) 232432. 8.
[97] S. Ahmed, I.E. Nielsen, A. Tripathi, S. Siddiqui, R.P. Ramachandran, G. Rasool, [138] X. Gu, H. Bai, X. Cui, J. Zhu, W. Zhuang, Z. Li, X. Hu, Z. Song, Renew. Sustain.
Circuits Syst. Signal Process. 42 (2023) 7433–7466. Energy Rev. 192 (2024) 114191.
[98] M. Dehghani, S. Gouws, O. Vinyals, J. Uszkoreit, Ł. Kaiser, arXiv preprint [139] N. Horesh, C. Quinn, H. Wang, R. Zane, M. Ferry, S. Tong, J.C. Quinn, Appl.
arXiv:1807.03819. (2018). Energy. 295 (2021) 117007.
[99] D. Shi, J. Zhao, Z. Wang, H. Zhao, J. Wang, Y. Lian, A.F. Burke, Electronics 12 [140] L. Xu, Z. Deng, Y. Xie, X. Lin, X. Hu, IEEE Trans. Transp. Electrification (2022).
(2023) 2598. [141] X. Liu, X.-Q. Zhang, X. Chen, G.-L. Zhu, C. Yan, J.-Q. Huang, H.-J. Peng, J. Energy
[100] NASA, Li-ion Battery Aging Datasets. https://data.nasa.gov/dataset/Li-ion- Chem. 68 (2022) 548–555.
Battery-Aging-Datasets/uj5r-zjdb/data, (accessed March 2024). [142] B. Wu, W.D. Widanage, S. Yang, X. Liu, Energy AI. 1 (2020) 100016.
[101] D. Shi, J. Zhao, Z. Wang, H. Zhao, C. Eze, J. Wang, Y. Lian, A.F. Burke, Energies. [143] W. Li, M. Rentemeister, J. Badeda, D. Jöst, D. Schulte, D.U. Sauer, J. Energy
16 (2023) 3855. Storage. 30 (2020) 101557.
[102] K. Weiss, T.M. Khoshgoftaar, D. Wang, J. Big Data. 3 (2016) 1–40. [144] W. Wang, J. Wang, J. Tian, J. Lu, R. Xiong, Chin. J. Mech. Eng. 34 (2021) 1–19.
[103] G. Ma, S. Xu, B. Jiang, C. Cheng, X. Yang, Y. Shen, T. Yang, Y. Huang, H. Ding, Y. [145] A. Torfi, R.A. Shirvani, Y. Keneshloo, N. Tavaf, E.A. Fox, arXiv preprint arXiv:
Yuan, Energy Environ. Sci. 15 (2022) 4083–4094. 2003.01200. (2020).
[104] S. Kim, Y.Y. Choi, K.J. Kim, J.-I. Choi, J. Energy Storage. 41 (2021) 102893. [146] J. Zhao, W. Zhao, B. Deng, Z. Wang, F. Zhang, W. Zheng, W. Cao, J. Nan, Y. Lian,
[105] Z. Deng, X. Lin, J. Cai, X. Hu, J. Power Sources. 525 (2022) 231027. A.F. Burke, Expert Syst. Appl. (2023) 122836.
[106] B. Gou, Y. Xu, X. Feng, IEEE Trans. Veh. Technol. 69 (2020) 10854–10867. [147] Y.-T. Zhuang, F. Wu, C. Chen, Y.-H. Pan, Front. Inf. Technol Electron. Eng. 18
[107] M.A. Chao, C. Kulkarni, K. Goebel, O. Fink, Reliab. Eng. Syst. Saf. 217 (2022) (2017) 3–14.
107961. [148] D. Newman, S. Hettich, C. Blake, C. Merz., UCI Repository of machine learning
[108] X. Han, Z. Wang, Z. Wei, Appl. Energy. 302 (2021) 117511. databases. https://www.bibsonomy.org/bibtex/
[109] X. Liu, H.J. Peng, B.Q. Li, X. Chen, Z. Li, J.Q. Huang, Q. Zhang, Angew. Chem. 134 e1b40ce6b507248ba38ba27eb0b735a9, 2007 (accessed Jan. 2024).
(2022) e202214037. [149] J. Yli-Huumo, D. Ko, S. Choi, S. Park, K. Smolander, PLoS ONE 11 (2016)
[110] D. Roman, S. Saxena, V. Robu, M. Pecht, D. Flynn, Nat. Mach. Intell. 3 (2021) e0163477.
447–456. [150] T. Kim, J. Ochoa, T. Faika, H.A. Mantooth, J. Di, Q. Li, Y. Lee, Sel. Top. Power
[111] M. Lin, D. Wu, J. Meng, J. Wu, H. Wu, J. Power Sources. 518 (2022) 230774. Electron. 10 (2020) 1270–1281.
[112] J. Zhao, H. Ling, J. Wang, A.F. Burke, Y. Lian, Iscience 25 (2022) 104172. [151] J. Zhao, A.F. Burke, J. Energy Chem. 84 (2023) 30–33.
[113] J. Zhao, J. Nan, J. Wang, H. Ling, Y. Lian, A. Burke, In: 2022 IEEE Vehicle Power [152] B. Ma, L. Zhang, H. Yu, B. Zou, W. Wang, C. Zhang, S. Yang, X. Liu, J. Energy
and Propulsion Conference (VPPC), 2022: IEEE, pp. 1–6. Chem. 82 (2023) 1–17.
[114] J. Zhao, Y. Gao, J. Guo, L. Chu, Cities Soc. 42 (2018) 301–313. [153] G. Rong, Y. Xu, X. Tong, H. Fan, J. Cloud Comput. 10 (2021) 1–14.
[115] Z. Wang, D. Shi, J. Zhao, Z. Chu, D. Guo, C. Eze, X. Qu, Y. Lian, A.F. Burke, [154] D. Shi, J. Zhao, C. Eze, Z. Wang, J. Wang, Y. Lian, A.F. Burke, Energies 16 (2023)
eTransportation (2023) 100309. 4403.
[116] G. Pozzato, A. Allam, L. Pulvirenti, G.A. Negoita, W.A. Paxton, S. Onori, Joule 7 [155] D. Georgakopoulos, P.P. Jayaraman, M. Fazia, M. Villari, R. Ranjan, IEEE Cloud
(2023) 2035–2053. Comput. 3 (2016) 66–73.

739