A Systematic Review of Computational and Data-Driven Approaches for Energy-Efficient Battery Management in Electrified Vehicles

Abstract

The dynamic growth of the electrified vehicle (xEV) market, including both electric and hybrid vehicles, has increased the demand for advanced Battery Management Systems (BMS). From an energy-systems perspective, xEV batteries act as distributed energy storage units that strongly interact with power grids, renewable generation, and charging infrastructure, making their efficient control a key element of low-carbon energy systems. Traditional BMS methods face challenges in accurately estimating key battery states and parameters, especially under dynamic operating conditions. This review systematically analyzes the progress in applying artificial intelligence, machine learning, and other advanced computational and data-driven algorithms to improve the performance of xEV battery management with a particular focus on energy efficiency, safe utilization of stored electrochemical energy, and the interaction between vehicles and the power system. The literature analysis covers key research trends from 2020 to 2025. This review covers a wide range of applications, including State of Charge (SOC) estimation, State of Health (SOH) prediction, and thermal management. We examine the use of various methods, such as deep learning, neural networks, genetic algorithms, regression, and also filtering algorithms, to solve these complex problems. This review also classifies the research by geographical distribution and document types, providing insight into the global landscape of this rapidly evolving field. By explicitly linking BMS functions with energy-system indicators such as charging load profiles, peak-load reduction, self-consumption of photovoltaic generation, and lifetime-aware energy use, this synthesis of contemporary research serves as a valuable resource for scientists and engineers who wish to understand the latest achievements and future directions in data-driven battery management and its role in modern energy systems.

1. Introduction

Electromobility is evolving rapidly, and the growing number of electric vehicles has become an important element of emission-reduction strategies. Research on charging infrastructure shows that the increasing number of fast-charging stations can significantly burden distribution grids; however, vehicle-to-grid (V2G) concepts and proper energy management can reduce peak loads by over 4% and lower energy costs by up to 40% [1]. Long-term tests have shown that a small city car charged from a 3 kWp photovoltaic carport can cover 30,000 km, requiring about 900 top-ups at 50% State of Charge [2], which underscores the importance of intelligent Battery Management Systems capable of planning charging and monitoring battery status in the context of the power grid. These examples illustrate that xEV batteries should be considered not only as components of the drivetrain but also as flexible energy storage resources embedded in future smart grids and local energy communities. In this sense, advances in BMS directly translate into classic energy-system objectives, including peak-load shaving, increased utilization of renewable generation, and improved overall energy efficiency.

In parallel, electric-vehicle designs are being developed. An analysis of design trends indicates that, despite higher curb weight, manufacturers strive to reduce the mass of BEVs, improve drivetrain efficiency, and increase passenger space. Statistical studies of changes in vehicle dimensions over the last five years show that changes in interior size may determine the competitiveness of EVs relative to conventional cars [3]. Reducing body mass, together with more compact motors and batteries, requires precise temperature and energy management, which sets high demands for BMS.

Electric vehicles are increasingly integrated with photovoltaic installations. Studies presenting algorithms for matching a vehicle to an existing PV system show that analyzing the energy-consumption profile and PV production makes it possible to select a car with an appropriate battery capacity to increase self-consumption and reduce feed-in to the grid [4]. Conversely, sizing a PV system for a specific vehicle requires taking into account the energy demand of both the building and the vehicle, as well as assessing possible panel locations, roof, ground, or carport. Using Metalog probabilistic distributions makes it possible to verify the correctness of PV sizing [5]. Experimental results from a long-term test of a small EV charged from a PV carport confirm that 3 kWp is sufficient for daily vehicle needs [2].

Another important issue is thermal management. Hybrid PVT heat exchangers with minichannels, originally designed for autonomous vehicles, combine solar-energy generation with active cooling. Studies show that such systems maintain PV panel temperatures in the 19.6–22.4 °C range and improve the energy efficiency and reliability of autonomous systems [6]. Ensuring stable thermal conditions is equally important in battery diagnostics; innovative methods for assessing battery condition in small electric vehicles indicate that, after 4.5 years of operation and 30,000 km of mileage, it is possible to effectively determine cell degradation using specialized diagnostic tools [7].

Advances in machine learning are also driving BMS development. A residual convolutional network proposed for State of Charge estimation achieved a mean absolute error of 1.26% and an RMS error of 0.998% [8]. In a study on a hybrid CNN–WNN–WLSTM model, it was emphasized that using wavelet and recurrent networks significantly increases SOH prediction accuracy compared with traditional physics-based methods [8]. Meanwhile, employing a CNN–LSTM network with an attention layer and weight updating (FVIM) enables identification of heterogeneous cell aging and achieves high prediction accuracy [9].

The above examples show that effective battery management in electric vehicles requires a systemic approach, from the proper design of charging infrastructure, through integration with renewable energy sources, to advanced algorithms for state estimation and thermal management. The common denominator is the growing role of artificial intelligence and machine learning techniques, which enable real-time data analysis and decision-making to optimize battery lifespan and safety. In the following sections of this work, we present detailed reviews of SOC and SOH estimation methods, cooling techniques, and optimization algorithms, as well as outline perspectives for the development of intelligent energy management systems.

The main objective of this review is to provide a comprehensive and reproducible synthesis of the current state of research on the application of artificial intelligence, machine learning, and computational methods in electric vehicle Battery Management Systems. The motivation behind this work was the increasing number of fragmented publications with inconsistent scopes and non-comparable metrics, which hinder the assessment of method effectiveness and the transfer of results into engineering practice. Unlike previous studies, this review is based on a fully reproducible procedure, from a clearly defined query in the Scopus database through a five-dimensional classification framework to explicit inclusion and coding criteria. As a result, a coherent picture of the three main BMS tasks, State of Charge estimation, State of Health estimation, and thermal management, has been obtained and presented within a structure that enables cross-comparisons and trend analysis. This framework, described in detail in the methodological Section, ensures the verifiability of results and makes this review more transparent and reliable than previous narrative-based publications.

The structure of this article follows the logic of a systematic review. After the introduction, Section 2 (Materials and Methods) provides a detailed description of the adopted research methodology, data sources, search strategy, and qualification criteria. Section 3 (State of the Art) presents the results of the literature analysis divided into three BMS tasks, highlighting the dominant methods, data types, and evaluation metrics. Section 4 (Statistical Overview) includes quantitative summaries showing the distribution of publications by method classes, countries of affiliation, document types, and methodological approaches. Section 5 (Discussion) synthesizes the results and answers the formulated research questions, while Section 6 (Conclusions) summarizes the findings, practical implications, and directions for further research. This structure allows a smooth transition from the thematic background to the research procedure and then to conclusions and recommendations, ensuring narrative consistency and a logical flow of presentation.

From a broader energy-system perspective, the methods reviewed in this paper influence key indicators such as grid load profiles, efficiency of renewable energy utilization, and lifetime-aware scheduling of charging and discharging, which are central topics for the energy research community.

2. Materials and Methods

This chapter presents a complete and reproducible research procedure, from the design of a systematic review with a bibliometric layer, using Scopus as the data source and searching within the Title, Abstract, and Keywords fields, to a detailed query strategy, keyword normalization, and documentation of applied filters. The need for this review was justified, the objectives were defined, and simple research questions were formulated linking the tasks of Battery Management Systems with method classes and computational costs. The inclusion and exclusion criteria were defined, a two-stage screening process with independent assessment and coded reasons was described, and the record flow was presented in a PRISMA diagram.

The central element of this chapter is a five-dimensional classification scheme, including the BMS task, AI method class, geography of affiliations, document type, and methodological approach, together with labeling rules, terminology normalization, and the hierarchy of dispute resolution. This chapter also describes data extraction and the set of comparative measures for State of Charge estimation, State of Health estimation, and thermal management, including quality metrics, computational costs, and reproducibility elements. Comparative summaries are presented in Tables in Section 4, and accompanying materials have been made available in an open repository under a DOI number (identifier 10.5281/zenodo.17319858). Finally, the limitations related to a single data source, query construction, task and metric heterogeneity, and the influence of the bibliometric layer are discussed, along with mitigating actions and directions for further improvements of the procedure.

2.1. Research Design and Data Sources

This study was designed as a systematic review with a bibliometric analysis component. The aim was to identify and synthesize applications of artificial intelligence and machine learning methods in the management of electric vehicle batteries. Particular emphasis was placed on State of Charge estimation, State of Health estimation, and thermal management. The only controlled data source was the Scopus database. The search was conducted within combined fields of Title, Abstract, and Keywords. The time range was limited to the years 2020 to 2025, the language to English, and the subject areas to Engineering, Computer Science, and Energy. These parameters ensured metadata consistency and procedural reproducibility. The initial query, built around terms related to vehicle batteries and artificial intelligence or machine learning, returned 126 records. After narrowing the search with algorithmic class keywords, including Kalman filters, 105 publications were obtained. According to the predefined inclusion and exclusion criteria, 10 records outside the scope were removed. The final corpus consisted of 95 publications selected for further analysis. The full query text with filters and a description of the limitations was published in an open repository, enabling faithful replication of the search results.

The selection process was two-staged. First, topical relevance was assessed based on metadata and abstracts. Then, the full texts of included and ambiguous papers were analyzed. In parallel, a consistent classification framework was developed. Each publication was coded simultaneously in five interrelated dimensions. The first dimension referred to the BMS task and included State of Charge, State of Health, and thermal management. The second dimension covered the class of artificial intelligence methods and included AI and ML Core, Deep Learning and Neural Networks, and Algorithms and Techniques, which encompassed Genetic Algorithms, Regression Analysis, Support Vector Machines, and Kalman filters. The third dimension captured the country of author affiliation according to Scopus data. The fourth dimension described the document type, including Article, Conference Paper, and Other. The fifth dimension represented the methodological approach derived from the content and included Experiment, Literature Analysis, Case Study, and Conceptual. This design ensured classification comparability, facilitated cross-analysis, and enabled mapping of the research landscape from the perspective of BMS tasks and algorithmic classes.

For transparency and reproducibility, a complete set of supplementary materials was prepared. It includes the exact query text with filters, the list of 95 records with metadata such as authors, affiliations, DOI identifiers and keywords, as well as the input files used for bibliometric visualizations.

Compared with previous works, this review offers an advantage by combining full reproducibility with a multidimensional, consistent classification and bibliometric layer. Instead of a narrative selection of examples, it employs an explicit decision chain from query design to coding, with open access to the exact query, inclusion, and exclusion criteria and the complete record list, eliminating typical verification issues and allowing the results to be reproduced step by step. The unified framework of five interdependent dimensions, including BMS tasks, AI method classes, country of affiliation, document type, and methodological approach, enables cross-comparisons and mapping of the research landscape with a level of precision not achievable in reviews based solely on thematic classifications. The inclusion of thermal management alongside SOC and SOH estimation broadens the scope of analysis to include safety and durability aspects, which are often omitted or treated in a fragmented manner. The application of bibliometric tools for co-occurrence and density analysis structures the conceptual vocabulary and reveals research gaps, while the archived input files for visualization allow immediate verification of the stability of the generated maps. The entire design minimizes selection bias, standardizes metadata, and facilitates the repetition of the procedure, ensuring that the synthetic conclusions in later sections of this paper are based on transparent and repeatable foundations rather than implicit editorial decisions.

2.2. Search Strategy

The search strategy is based solely on Scopus to ensure high metadata quality, consistent filtering, and full reproducibility. Results were limited to English publications from 2020 to 2025, and searches were conducted in the Title, Abstract, and Keywords fields together. The procedure is defined to correspond to the scope of the review, which combines the three tasks of Battery Management Systems, State of Charge estimation, health estimation, and thermal management, with artificial intelligence and machine learning methods.

The search was carried out in two stages. The first step identified vehicle battery entries in relation to AI and ML terminology and then, in the second step, narrowed down the collection using keywords representing method and algorithm classes. The following query is applied to TITLE-ABS-KEY fields, with time, language, and subject area constraints, and a list of SOC, SOH, and thermal task descriptors:

“TITLE-ABS-KEY (“Vehicle Batteries” AND (“Artificial Intelligence” OR “Machine Learning” OR “Learning Systems” OR “Neural Networks” OR “Deep Learning” OR “Long Short-term Memory” OR “Convolutional Neural Network” OR “Genetic Algorithms” OR “Regression Analysis” OR “Support Vector Machines” OR “Mean Square Error”)) AND PUBYEAR > 2019 AND PUBYEAR < 2026 AND (EXCLUDE (SUBJAREA, “CENG”) OR EXCLUDE (SUBJAREA, “DECI”) OR EXCLUDE (SUBJAREA, “ECON”) OR EXCLUDE (SUBJAREA, “ENVI”) OR EXCLUDE (SUBJAREA, “CHEM”) OR EXCLUDE (SUBJAREA, “PHYS”) OR EXCLUDE (SUBJAREA, “SOCI”) OR EXCLUDE (SUBJAREA, “MATE”) OR EXCLUDE (SUBJAREA, “MATH”) OR EXCLUDE (SUBJAREA, “BIOC”) OR EXCLUDE (SUBJAREA, “HEAL”) OR EXCLUDE (SUBJAREA, “DENT”) OR EXCLUDE (SUBJAREA, “EART”) OR EXCLUDE (SUBJAREA, “MULT”)) AND (LIMIT-TO (LANGUAGE, “English”)) AND (LIMIT-TO (EXACTKEYWORD, “State Of Charge”) OR LIMIT-TO (EXACTKEYWORD, “States Of Charges”) OR LIMIT-TO (EXACTKEYWORD, “State-of-charge Estimation”) OR LIMIT-TO (EXACTKEYWORD, “Battery State Of Charge”) OR LIMIT-TO (EXACTKEYWORD, “Battery Health”) OR LIMIT-TO (EXACTKEYWORD, “Battery Degradation”) OR LIMIT-TO (EXACTKEYWORD, “State Of Health”) OR LIMIT-TO (EXACTKEYWORD, “Battery Temperature”) OR LIMIT-TO (EXACTKEYWORD, “Temperature Control”) OR LIMIT-TO (EXACTKEYWORD, “Thermal Management (electronics)”) OR LIMIT-TO (EXACTKEYWORD, “Battery Thermal Managements”) OR LIMIT-TO (EXACTKEYWORD, “Thermal Management Systems”) OR LIMIT-TO (EXACTKEYWORD, “Cooling”))”.

In the first stage of the query, 126 records were obtained, while in the second stage, a refinement filter was applied to the AI and NN method classes, extending the set of keywords to include algorithm families, including LSTM, CNN, SVM, GA, Regression, MSE, and Kalman filters, which narrowed the set down to 105 items.

“AND (LIMIT-TO (EXACTKEYWORD, “Artificial Intelligence”) OR LIMIT-TO (EXACTKEYWORD, “Machine Learning”) OR LIMIT-TO (EXACTKEYWORD, “Learning Systems”) OR LIMIT-TO (EXACTKEYWORD, “Neural Networks”) OR LIMIT-TO (EXACTKEYWORD, “Deep Learning”) OR LIMIT-TO (EXACTKEYWORD, “Long Short-term Memory”) OR LIMIT-TO (EXACTKEYWORD, “Convolutional Neural Network”) OR LIMIT-TO (EXACTKEYWORD, “Genetic Algorithms”) OR LIMIT-TO (EXACTKEYWORD, “Regression Analysis”) OR LIMIT-TO (EXACTKEYWORD, “Support Vector Machines”) OR LIMIT-TO (EXACTKEYWORD, “Mean Square Error”) OR LIMIT-TO (EXACTKEYWORD, “Kalman Filters”))”.

In the second stage, the dataset was narrowed down to 105 records through keyword filtering representing classes of AI and NN methods, including the LSTM, CNN, SVM, GA, Regression, MSE, and Kalman Filter families. A manual verification of scope compliance was then conducted, resulting in the exclusion of 10 records as out of scope, since they did not directly address SOC estimation, SOH estimation or thermal management in vehicles. The excluded publications focused on voltage prediction from the perspective of charging infrastructure, control of charging stations and V2G power flows, fault diagnostics without SOH estimation, active cell balancing, cell sorting for second life, signal simulations, and parameter optimization of models without state estimation. The final corpus selected for extraction and classification consisted of 95 publications.

To reduce bias resulting from terminological differences, a normalization dictionary was developed to merge the most common variants and abbreviations, including SOC and State of Charge, SOH and State of Health, Battery Thermal Management and cooling, as well as algorithmic synonyms such as Extended or Unscented Kalman Filter grouped under the Kalman Filter family, and convolutional neural network grouped under the CNN class.

The data collection and preparation workflow are illustrated in

Table 1. Scopus data collection and preparation workflow for the BMS review, from query to classification framework; final corpus of 95 publications.

Area/Block	Subcategory	Content
Scopus Search	Database	Scopus
	Number of Documents	95
	Years	2020–2025
	Language	English
	Subject Areas	Engineering, Computer Science, Energy
Search Criteria	Title, Abstract, Keywords	Vehicle Batteries AND (Artificial Intelligence OR Machine Learning OR Learning Systems OR Neural Networks OR Deep Learning OR Long Short-term Memory OR Convolutional Neural Network OR Genetic Algorithms OR Regression Analysis OR Support Vector Machines OR Mean Square Error)
Keywords–Battery	Battery/BMS	State Of Charge, States Of Charges, State-of-charge Estimation, Battery State Of Charge, Battery Health, Battery Degradation, State Of Health, Battery Temperature, Temperature Control, Thermal Management (electronics), Battery Thermal Managements, Thermal Management Systems, Cooling
Keywords–AI	AI/Computational Methods	Artificial Intelligence, Machine Learning, Learning Systems, Neural Networks, Deep Learning, Long Short-term Memory, Convolutional Neural Network, Genetic Algorithms, Regression Analysis, Support Vector Machines, Mean Square Error, Kalman Filters
Thematic Categories	Battery Management Systems	State Of Charge, State Of Health, Thermal Management
	Computational Intelligence	AI/ML Core, Neural Networks, Algorithms and Techniques
Dataset Characteristics	Countries	Canada, China, Denmark, France, Germany, India, Malaysia, Morocco, Pakistan, Saudi Arabia, Turkey, United Kingdom, United States, Other
	Document Type	Article, Conference Paper, Other
	Research Methodology	Experiment, Literature Analysis, Case Study, Conceptual

. The top block describes the Scopus search with time, language, and subject constraints and indicates the final number of 95 documents. The middle section shows the set of keywords used in the Title, Abstract, and Keywords fields for the battery domain and for AI and ML method classes. The bottom section organizes the classification framework used in the subsequent analysis, including BMS task categories (State of Charge, State of Health, and thermal management), computational method classes (AI and ML Core, Neural Networks, and Algorithms and Techniques), as well as metadata-based categories from Scopus, such as countries of author affiliation and document types, and methodological approaches derived from the content (Experiment, Literature Analysis, Case Study, and Conceptual). The diagram provides a concise reference to the adopted query rules and classification principles, which facilitates replication of the procedure in future updates of the corpus.

The articles selected for analysis were described and compiled in working files, which, together with the complete metadata and input files for visualization, were archived in the open Zenodo repository under the DOI identifier 10.5281/zenodo.17319858. The repository contains Excel files with Scopus query results and full metadata including titles, authors, affiliations, DOI identifiers, and keywords, as well as the input files used in the bibliometric analyses. Full texts of the articles are not provided there, as they remain available in Scopus and the publishers’ repositories. The dataset prepared in this way ensures transparency, durability, and reproducibility of the results and also enables the corpus to be reused in future updates. The qualified literature corresponds to a corpus of 95 publications that meet the established substantive and formal criteria, identified through the Scopus query, the two-stage screening procedure, and keyword normalization.

Although the TITLE ABS KEY filter ensures high query precision, such a rigorous strategy may omit works that use alternative terminology. To reduce this potential source of bias, a normalization dictionary was constructed. It includes the most common variants and abbreviations, for example, State of Charge and SOC, State of Health and SOH, Battery Thermal Management, thermal management and cooling, and, on the methods side, Extended Kalman Filter and Unscented Kalman Filter as subclasses of Kalman filters, convolutional neural network and CNN, Long Short-Term Memory and LSTM, and Random Forest and RF. Despite the normalization, a few false negatives were identified, for instance, studies predicting remaining useful life under the terms Remaining useful life or Battery health prognosis without explicitly using the SOH acronym, or studies based on message-passing networks described by the acronyms MPNN or GNN without the phrase Graph Neural Networks. Therefore, the Scopus search was complemented with a targeted manual review of bibliographies in leading review papers and high-impact journals. This procedure increases sensitivity while maintaining full transparency regarding the possible limitations resulting from the strict keyword-matching approach.

2.3. Rationale for the Review, Purpose of the Study, and Problems of the Study

The existing literature linking electric vehicle battery management with artificial intelligence methods is fragmented and the thematic scope of individual studies is uneven. Many papers focus on a single task or a specific family of algorithms, often without reference to computational cost, consistent uncertainty reporting or explicit literature selection criteria. This situation makes it difficult to compare results and to transfer methods into engineering practice where both reliable quality metrics and information on computational complexity and hardware requirements are needed. The review presented in this work organizes this field by combining within one reproducible framework three problem axes of BMS, State of Charge estimation, State of Health estimation, and thermal management, with three families of algorithmic approaches, AI and ML Core, Neural Networks, and Algorithms and Techniques. A uniform Scopus search procedure, complete vocabulary normalization, and a five-dimensional classification scheme are used, which enables a cross-sectional review and analysis of the development dynamics of the field.

Preliminary analysis of the corpus shows clear and measurable trends. In the area of BMS tasks, the greatest growth is observed for SOC estimation, where the number of publications increases from 14 to 45 between the periods 2020 to 2022 and 2023 to 2025. A similar trend applies to SOH, which rises from 13 to 33, while works devoted to thermal management remain limited, increasing from 2 to 6. The structure of method classes changes across the entire time horizon, with the number of studies in the AI and ML Core category growing from 12 to 39, those using Neural Networks increasing from 14 to 38 and the Algorithms and Techniques group expanding from 10 to 26. The maturity of the field is also reflected in document types, as journal articles increase from 10 to 46, conference papers grow moderately from 12 to 19, and other types rise from 3 to 5. At the methodological level the share of experimental studies grows from 21 to 59 and conceptual works from 22 to 53, while case studies appear for the first time, increasing from 0 to 8. Geographically, research activity increases most strongly in India from 4 to 22, in China from 7 to 13, and in the United States from 2 to 8, suggesting a concentration of research in several leading centers and an acceleration in the transfer of methods to industrial applications.

The cross-sectional results also make it possible to identify thematic gaps and missing links in the validation chain. Counting matrices show that experimental and conceptual studies are concentrated on SOC with 46 and 46 publications, respectively, while SOH gathers 37 experimental and 36 conceptual works, and the thermal area remains significantly weaker with 8 experimental and 7 conceptual papers. The comparison of method classes with tasks indicates that SOC is dominated by AI and ML Core and Neural Networks, with 31 publications each, while SOH shows a more balanced distribution, with 27 for AI and ML Core, 22 for Neural Networks, and 22 for Algorithms and Techniques. In contrast, thermal studies remain few, with four, seven, and two publications, respectively. This profile indicates that research on heat management is less mature and lacks broad experimental validation and case studies. These findings are consistent with the observed shift in document types toward journal articles, which typically require more mature methodologies and explicit uncertainty and computational cost metrics.

Against this background, this review has three objectives. First, to provide a synthetic but comprehensive organization of AI methods in BMS with a clear mapping of tasks to model classes, data types, strategies for embedding physical knowledge, and computational costs in order to compare solutions within a uniform evaluation framework. Second, to conduct a critical analysis of missing elements, particularly uncertainty and computational complexity reporting, and to identify a minimal set of information that should accompany publications to make them useful for design, diagnostics, and maintenance. Third, to identify research priorities and development directions such as strengthening the thermal layer, conducting studies under real conditions with telematics, standardizing features, and developing reproducible workflows.

To translate these observations into specific research tasks, a set of research questions was formulated that are simple, unambiguous, and directly derived from the corpus analysis. The first three questions concern the quality and maturity of the studies, while the last two focus on verifiable statistical trends. The study addresses the following research questions:

• Which model families and hybrid configurations are the most effective in practice for SOC, SOH, and thermal management? That is, which achieve the lowest errors at acceptable computational cost and under which measurement and temperature conditions do they maintain this advantage?
• What data types, including electrical signals, thermal data, telematics, and environmental parameters, and what data preparation and normalization procedures, such as time window selection, alignment, and filtering, dominate in the publications with the best results, and which feature sets most frequently accompany the highest-quality models?
• To what extent do the studies report uncertainty, verification, and validation, as well as computational costs and hardware requirements? That is, do they provide confidence intervals, acceleration factors relative to reference solutions and platform parameters, and which metrics are commonly used as standards?
• How much the number of publications in SOC, SOH, and thermal management categories has increased between 2020 to 2022 and 2023 to 2025, and whether the growth covers all three areas or focuses mainly on SOC and SOH.
• How the shares of method classes, document types, and countries have changed over the same periods, that is, whether the share of studies in the Deep Learning and AI and ML Core families has increased, whether the proportion of journal articles has risen relative to conference papers, and whether there is growing geographic concentration in India, China, and the United States.

These questions structure the subsequent analytical stages. For the substantive part, a taxonomy of tasks and methods, a method-to-task matrix, and evidence cards for representative cases will be prepared, in which for each model the data types, method of embedding physical knowledge, quality metrics, and computational costs will be collected. For the quantitative part density and co-occurrence maps, distributions by document type and country of affiliation and statistical tests of changes in method and publication form shares between the periods will be prepared, allowing the dynamics of the field to be linked with methodological directions and application areas. Such a structure combined with full reproducibility of the search and classification enables not only the formulation of conclusions about the state of research but also the development of recommendations for the design of BMS workflows, the selection of algorithm classes, data curation, and standards for reporting uncertainty and computational costs.

2.4. Eligibility Criteria

The literature selection process was designed to be transparent and reproducible. A two-stage screening procedure was applied. First, titles, abstracts, and keywords were assessed, and in borderline cases, full texts were analyzed. The construction of the criteria follows good practices for reporting reviews in MDPI journals and the logic of the PRISMA diagram while strictly reflecting the parameters of the query adopted in this study, namely the years 2020 to 2025, the English language, the subject areas Engineering, Computer Science, and Energy, the combined search of Title, Abstract, and Keywords fields, and keyword filters describing BMS tasks and AI method classes. After manual verification of thematic relevance, the final corpus consisted of 95 publications.

Only works meeting both substantive and formal requirements were included in the analysis. A direct connection with Battery Management System tasks in at least one of three categories, State of Charge estimation, State of Health estimation, and thermal management, was required, as well as the application of data-driven methods belonging to one of three families, AI and ML Core, Neural Networks, or Algorithms and Techniques, which had to be confirmed by the metadata or content of the article. Publications in English from the years 2020 to 2025 classified in Scopus within the specified subject areas were accepted. Journal articles, conference papers, and items marked as Other, such as book chapters and review articles, were included provided that they presented a coherent methodological contribution or a synthesis that allowed for unambiguous classification. Access to the full text and a complete set of basic metadata, including title, authors, affiliations, DOI identifier, and keywords, was required.

Works that did not meet any of these conditions were excluded from the corpus. Publications without an ML or NN component, studies not directly related to BMS problems, items outside the accepted years, language or subject areas, incomplete records, and duplicates identified by DOI or title were eliminated. Non-technical materials were excluded if they did not contain a computational component or data related to signal models, equivalent circuit models or thermal analyses, as well as works with inadequate reporting, without data description, quality metrics or information enabling assessment reproduction. In cases where the full text did not allow the paper to be assigned to any category, the record was marked as out of scope.

The decision chain corresponds to the stages of the Scopus query. The first query in the Title, Abstract, and Keywords fields with time, language, and subject restrictions returned 126 records. Adding a second set of keywords representing method and algorithm classes, including LSTM, CNN, Support Vector Machines, Genetic Algorithms, and Kalman Filter families, yielded 105 publications. Manual verification of thematic compliance resulted in the exclusion of ten items, setting the corpus size at 95. For each case, the rationale for inclusion or exclusion was recorded. Multiple assignments of a single publication to more than one classification category was allowed, reflecting the multidimensionality of the topic and explaining why totals in cross tables and heat maps may exceed the overall number of records in the corpus. Consistency of the procedure with the adopted protocol and MDPI good practices ensures comparability and reproducibility of the selection in subsequent updates of the review.

2.5. Selection Procedure and Screening

Screening was conducted in two stages. In the first stage, titles, abstracts, and keywords were evaluated, and in the second stage, the full texts of included and ambiguous papers were analyzed. Before the actual screening, a short calibration was carried out on a random sample of records in order to standardize the interpretation of criteria and terminology related to BMS tasks and data methods. Two reviewers independently assessed the publications, and decisions were recorded in a form with three possible outcomes: include, exclude or unclear. Discrepancies were resolved by consensus and, if necessary, with the participation of a third person. For transparency, each decision was assigned a reason code, such as lack of ML or NN component, out of BMS scope, inappropriate document type, incomplete metadata or absence of an assessable method. This approach allowed later summarization of exclusion categories in the final report.

During the title and abstract screening, a minimal set of decision questions was applied. First, whether the publication directly concerns BMS tasks in at least one of the three categories, State of Charge, State of Health or Thermal Management. Second, whether it uses data-driven methods, understood as AI and ML Core, Neural Networks or Algorithms and Techniques. If the answer was yes, the record was directed to full-text analysis. If the answer was no, it was excluded. If the answer was unclear, the record was also directed to full-text evaluation. At this stage, multiple assignments of topics and methods was allowed, reflecting the complexity of the tasks. Terminological inconsistencies were minimized by normalizing terms to a reference list, for example, merging State of Charge and SOC, State of Health and SOH, Battery Thermal Management, thermal management and cooling, and, on the methods side, merging neural networks and neural networks with hyphenation, Long Short-Term Memory and LSTM, convolutional neural network and CNN, and Extended and Unscented Kalman Filter into the Kalman Filters family.

Duplicates were technically confirmed based on DOI identifiers and titles. In cases of twin publications, that is, conference and journal versions of the same work, the methodologically more complete version was preferred. Full-text evaluation served to verify whether the publication met all substantive and formal criteria, particularly whether it actually contained a data-based method component applied in the context of BMS tasks, whether it reported input data and quality metrics relevant to the task, and whether the method description allowed for unambiguous classification. For review and conceptual papers, a consistent taxonomy or methodological conclusions related to the defined classification axes was required. Lack of access to the full text, incomplete metadata or inconsistencies between title, keywords, and content resulted in exclusion with an assigned reason code. All decisions were recorded in a selection log, and disputed classifications were corrected by consensus.

The flow of records between stages, together with the counts, is presented in the PRISMA diagram, Figure 1, which includes identification, screening, eligibility assessment, and final inclusion. At the identification stage, after applying the Scopus query in the Title, Abstract, and Keywords fields with time, language, and subject limitations, 126 records were obtained. Narrowing the list with keywords representing method and algorithm classes, including LSTM, CNN, Support Vector Machines, Genetic Algorithms, and Kalman Filter families, resulted in 105 records directed to full-text assessment. At the eligibility stage, full texts of 105 publications were analyzed, of which 10 were excluded for substantive reasons, setting the final corpus size at 95 studies. Stage metrics confirm the effectiveness of the procedure. Retention after screening was 83.3 percent (105 out of 126). Full-text exclusions accounted for 9.5 percent of the analyzed reports (10 out of 105). The share of included publications relative to identified records was 75.4 percent (95 out of 126). The dominant reason for exclusions at the preliminary stage was refinement of the keyword list on the methods side, confirming that thematic narrowing was conducted at the metadata level before content evaluation. This ensured that further classification and quantitative analyses were based on a technically homogeneous corpus.

The applied selection procedure meets the requirements of transparency and reproducibility while ensuring a technically homogeneous corpus for further analyses. The two-stage screening process, with independent evaluation by two reviewers, calibration prior to the start of the work, vocabulary normalization, and a complete decision log, made it possible to reduce bias and clearly justify exclusions. The record flow presented in Figure 2 confirms high retention after screening and a low percentage of exclusions following full-text review, which demonstrates the appropriateness of the selection criteria and their consistency with the thematic scope of the study. The final set of 95 publications provides a representative basis for building a taxonomy of tasks and methods, cross tables linking methods to tasks, trend analyses, and significance tests, which are presented in the subsequent sections of this paper.

The analysis included articles numbered from 11 to 105, which were accepted for review according to the established inclusion criteria. During the analysis, it was observed that the selected studies show multiple connections in terms of the analyzed categories, allowing for the identification of common trends and convergent conclusions within the examined subject area. These relationships indicate thematic consistency among the reviewed publications and enable a more in-depth interpretation of the results within the adopted research framework.

2.6. Classification Scheme

The classification scheme was designed to unambiguously and consistently code the content of each publication and to enable cross comparisons between BMS tasks, method classes, and bibliometric metadata. For all dimensions, normalization dictionaries and rules for assigning main and auxiliary labels were defined. The same record could receive multiple labels within a single dimension, but for tables and heat maps only one main label was indicated according to the priority rule. The task intent takes precedence over the choice of tool, the data source takes precedence over the data format, and the publication type takes precedence over the manuscript stage. The following dimensions and classification rules apply:

• Dimension 1. BMS Task. The task category is assigned as a continuous description and covers three areas of practice, which are treated jointly to capture their interdependencies. State of Charge estimation refers to determining the current SOC value based on electrical signals and auxiliary information such as voltage, current, temperature, and driving profile, and the output may be a continuous value or a distribution with uncertainty, including hybrid solutions using Kalman filters or circuit models. State of Health estimation includes the diagnosis of condition and aging processes, that is, capacity loss, resistance increase, anomaly detection, and remaining useful life prediction. Both continuous and threshold classification forms are allowed, and typical sources of information are derived features such as incremental capacity curves. Thermal management concerns the modeling and control of cell and pack temperature and the design and regulation of BTMS systems, where the task is to predict temperature distributions and limit the risk of thermal runaway at minimal energy cost. All spelling variants and synonyms are normalized to reference lists. For example, State of Charge, States of Charges, State of Charge estimation, and battery State of Charge are mapped to SOC; State of Health, battery health, and battery degradation to SOH; and battery temperature, thermal management systems, Battery Thermal Management, and cooling to thermal management. If a paper addresses more than one of these tasks, the main label corresponds to the task that governs the evaluation protocol and key result, and the remaining tasks are recorded as auxiliary labels.
• Dimension 2. AI Method Class. Three main families reflect the level of complexity and the nature of the approaches. AI and ML Core includes general frameworks and machine learning or artificial intelligence techniques, such as machine learning, learning systems, and regression and classification methods not based on networks. Neural Networks comprises neural networks in general and their sequential and memory variants, including artificial neural networks, recurrent neural networks, Long Short-Term Memory, and deep learning. Algorithms and Techniques includes learning related techniques, such as Genetic Algorithms, Support Vector Machines, Regression Analysis, and Kalman Filter families. All spelling variants are merged into parent classes, for example, neural networks and neural networks with hyphenation, convolutional neural network and convolutional neural networks, and Long Short-Term memory and LSTM.
• Dimension 3. Geography of Results. The country label is derived from the authors’ affiliations according to Scopus data and is used to analyze the distribution of research activity. In multinational publications, all relevant labels are assigned, allowing measurement of international collaboration. If the first author lists several affiliations, the country of the first listed institution is used.
• Dimension 4. Document Type. The publication form is categorized according to the Scopus classification, Article for journal papers, Conference Paper for conference proceedings, and Other for non-standard forms such as review articles and book chapters, provided they meet substantive and formal criteria. In cases of conference and journal versions of the same work, the journal version is considered the primary one, while the conference version is recorded as auxiliary.
• Dimension 5. Methodological Approach. This category reflects the way the research was conducted and the maturity of the solution. Experiment refers to studies with validation on real or laboratory data, Literature Analysis to critical analyses and bibliometric studies within the adopted axes, Case Study to complete case studies with data and evaluation protocols, and Conceptual to methodological or workflow proposals without full experimental validation. In hybrid approaches, the main label corresponds to the mode that determines the learning objective function or the inference method at the implementation stage.

This five-dimensional framework guarantees unambiguous coding despite terminological differences and enables reliable comparisons between BMS tasks, method classes, and publication metadata, providing a stable foundation for building cross tables, heat maps, and trend analyses in the following sections of the paper.

Since the classification scheme operates on keywords, an additional step was the preparation of supporting visualizations. A keyword density map and a term co-occurrence network were generated from the same dictionary used in the coding process. Together, these visualizations highlight the main topical concentrations in the corpus, the relative strength of links between concepts, and the presence of less explored areas, thereby confirming the dominance of SOC and SOH topics, the strong grounding of methods in ML and NN families, and the comparatively weaker representation of thermal aspects. The maps are based on frequencies of term co-occurrences in the entire dataset rather than on the weight of terms in individual papers, so their interpretation is always linked back to the task and methodological classifications.

The visualizations serve a dual function. First, they offer a quick overview of the dominant themes and relationships between concepts, consistent with the established dictionary and classification scheme. Second, they provide a qualitative reference point for the quantitative analyses in later sections, where distributions by BMS tasks, method classes, document types, and countries of affiliation are compared with the density and network structures to verify whether observed conceptual clusters correspond to actual patterns of algorithm and data selection.

In addition, to align the five-dimensional classification scheme with the visualizations and to ensure full reproducibility of coding, uniform labeling rules and a fixed hierarchy for resolving disagreements were adopted. The same record may have multiple labels within a dimension, but for cross tables only the main label is indicated. The following rules apply:

• BMS Task. The label is assigned based on the research objective expressed in the title and abstract. When a publication addresses multiple tasks, the main label corresponds to the task that governs the evaluation protocol and the key quality outcome, while the remaining tasks are recorded as auxiliary labels.
• AI Method Class. The label corresponds to the family of methods dominant in the learning or inference process. Lexical variants and abbreviations are normalized to parent classes, and in cases of coexistence of network and classical methods, priority is given to the family that determines the objective function and model decision.
• Data Source. Experimental and simulated datasets are distinguished. The main label is assigned to the dominant source, understood as more than 50 percent of the data used for training and evaluation. When the share is comparable, a dual label is applied and disclosed in the supplementary materials.
• Country of Affiliation. The country label is derived from the first affiliation of the first author. In multinational publications, all countries of co-authors are included, and in cases where the first author has multiple affiliations, the country of the first listed institution is used.
• Document Type. The label reflects the publication venue: Article for journal papers, Conference Paper for conference proceedings, and Other for non-standard forms that meet the criteria. If a preprint is later published in a journal, the main label is Article, while the preprint remains auxiliary.
• Methodological Approach. The label reflects the main research mode: Experiment, Literature Analysis, Case Study or Conceptual. In hybrid solutions, the main label corresponds to the mode that determines the learning objective function or inference process at the implementation stage.
• Dispute Resolution and Quality Control. Disputes are resolved according to a hierarchy where task intent takes precedence over tool choice, data source takes precedence over data format, and publication venue takes precedence over manuscript stage. Two reviewers code independently, discrepancies are resolved through discussion, and consistency is verified by re-coding a random 10 percent sample of records after two weeks, achieving an agreement level above 0.9 according to Cohen’s kappa coefficient.

The divisions and label assignments across the five dimensions are presented in the tables in Section 4, and the supplementary set of input data and label descriptions is included in the accompanying materials in the open repository.

2.7. Data Extraction and Benchmarks

From each publication, a standardized set of fields was extracted to obtain a comparable description of the studies and their results. The BMS task was recorded according to the adopted framework, including State of Charge estimation, State of Health estimation, and thermal management, together with the corresponding descriptors used by the authors, for example, State of Charge, State of Health, battery health, battery temperature, and thermal management. In parallel, the class of data method was recorded according to the three families, AI and ML Core, Neural Networks, and Algorithms and Techniques, with full normalization of lexical variants, for example, machine learning and learning systems to the machine learning class, neural networks and neural networks with hyphenation to the neural networks class, and convolutional neural network and convolutional neural networks to the CNN class. For clarity, the document type, country of author affiliation, and methodological approach, experiment, literature analysis, case study or concept were also recorded in accordance with the classification rules.

The set of substantive fields was complemented with technical information essential for interpreting quality and computational cost. These included the type of signals and sensors, voltage, current, temperature, sampling frequency, input window length, temperature range and C rates, cell chemistry and format, pack parameters, dataset size, division into training, validation and test sets, and cross-validation procedures. The architecture and elements of the loss function were recorded, for example, coupling with Kalman filters and circuit models, physical loss terms and boundary conditions in the objective function, feature selection strategies, signal normalization, and filtering. At the resource level, the computational cost and hardware limitations were noted, including training and inference time, hardware type, processor, graphics unit, programmable device or microcontroller, memory footprint, and number of parameters when reported, as well as reproducibility status, availability of code and data, and DOI identifiers.

Comparative metrics were selected according to the nature of the tasks and then standardized within each category to enable direct comparison. For State of Charge estimation, RMSE and MAE in percentage points were primarily used, along with MAPE and the coefficient of determination, and for probabilistic approaches, interval coverage and interval width were additionally applied. For State of Health estimation, relative and absolute error in the assessment of capacity and resistance were used, together with RMSE and MAE in percentages, and in classification scenarios, accuracy parameters and F measures were used, taking into account quality curves as a function of aging stage. For thermal management, RMSE and maximum error in degrees Celsius were used, as well as time to reach the temperature threshold and energy cost of cooling control, while for surrogate models used in the thermal and diagnostic parts, acceleration relative to reference solutions, training time, and inference time were reported. In studies embedding physical knowledge, energy consistency and satisfaction of boundary conditions were recorded. Where authors reported metrics in different units or horizons, results were normalized to the most frequently used measures in the given category and referred to test sets outside the training data, which allowed quality comparisons to be made on a common scale.

The synthesis of results was conducted in two complementary ways. First, a thematic synthesis was performed in three task categories and three method families, which made it possible to link tasks with model types and strategies for embedding physical knowledge, including boundary conditions in the loss function, energy constraints, and filtering components. Second, a quantitative aggregation was carried out, covering geographical distributions, document types, and methodological approaches, with the construction of cross tables, for example, method class by BMS task, and with analysis of popularity and trends in the corpus.

All field summaries, metrics, and comparisons are presented in the Tables in Section 4, including results after normalization and ranking principles, while the complete set of input materials for visualization and the descriptions of extraction and normalization rules are provided in the open repository, ensuring transparency of the procedure and facilitating the update of the review in future editions.

2.8. Limitations

The scope of this review is defined by a single source, the Scopus database, the English language, and the years 2020 to 2025, as well as the restriction to the subject areas Engineering, Computer Science, and Energy. This choice ensures consistent metadata and uniform selection criteria, while also carrying the risk of omitting publications indexed only in other databases and papers written in other languages. Indexing delays should be considered, as they may lead to the underestimation of the most recent publications at the end of the study period.

Further limitations arise from the query design. The use of EXACTKEYWORD descriptors for both BMS task terminology and data method classes increases precision but may exclude works using less common synonyms or alternative naming conventions. This risk was mitigated by combined searching of the Title, Abstract, and Keywords fields and by normalization of lexical variants, although some false negatives remain possible, as well as the inclusion of borderline cases with a broad thematic scope.

The selection procedure included independent double screening and consensus decisions. Despite these safeguards, the classification across five dimensions, BMS task, method class, country of affiliation, document type, and methodological approach, involves an element of expert judgment. Subjectivity particularly affects multidisciplinary papers and publications with brief descriptions of data and metrics. Therefore, a normalization dictionary was used and justifications for decisions were recorded, although some individual ambiguities may persist.

A major challenge for comparison lies in the heterogeneity of tasks and metrics. The analyzed works differ in operating conditions and temperatures, geometric and pack representations, and validation protocols. Authors use different error measures, such as RMSE, MAE, MSE, and MAPE, and surrogate models report computational acceleration under various hardware configurations. Open benchmarks and consistent descriptions of computational cost are rare, which makes a formal meta-analysis of effects impossible. As a result, comparative conclusions are descriptive even after normalization to the most frequently reported metrics within each task category.

The results depend on the inclusion thresholds for terms, the rules for merging synonyms, and the chosen clustering algorithm. Therefore, interpretations of clusters and concept centrality should be related to the substantive context of the review rather than treated as causal indicators.

At the external level, publication and selection bias must be considered. In the technical literature, positive results dominate, while reports of failures are rare, which may overestimate expected performance measures. This review does not include gray literature and preprints outside Scopus, such as industrial reports and internal materials, which often contain information on implementations and operational constraints.

The generalizability of conclusions is limited by environmental and computational differences. The application of AI and NN methods in BMS tasks depends on data quality and size, sampling frequency, temperature and C rate ranges, cell chemistry and format, hardware configuration, and simplifying assumptions, such as replacing three-dimensional models with two-dimensional representations. Many studies rely on synthetic data or limited experiments, and results in industrial conditions may differ from those presented. Some publications do not report prediction uncertainty or full computational costs, which hinders the assessment of risk and scalability.

These limitations do not invalidate the conclusions of the review but define the boundaries of their applicability. In future iterations, the query will be extended to additional databases, including IEEE Xplore and Web of Science, and selected non-English languages will be considered. The synonym dictionary will be refined, and classification agreement coefficients will be reported. The promotion of open benchmarks, standardized evaluation protocols, and transparent reporting of computational costs and uncertainty is recommended, as this will enable more rigorous comparisons in the future. The literature coverage extends to 2025, but due to indexing delays in Scopus, some of the most recent publications may not have been included at the time of the query. This limitation was mitigated by a targeted manual review of citations and the inclusion of 2025 publications identified through journal websites and cross-citation analysis.

3. State of the Art

This chapter contains only the analysis of the literature review divided into categories of Battery Management System tasks, namely State of Charge estimation, State of Health estimation, and thermal management. The aim of this chapter is to present the categorical results in an organized manner, without trend analyses or bibliometric visualizations, so that the reader can easily follow the dominant approaches, data types, and the most frequently used evaluation metrics within each category.

In the section devoted to State of Charge estimation, the methods and model configurations are summarized, along with typical input data streams, signal preparation and normalization procedures, and the reported quality metrics and limitations. An analogous structure is used for State of Health estimation, where differential features and degradation indicators are additionally emphasized, and for thermal management, where models and algorithms used for temperature prediction and cooling system control are discussed. Each subsection presents a concise description of the group of studies assigned to the given category, then discusses representative families of methods with typical data and metrics, and finally identifies gaps and limitations that are of practical relevance. This chapter concludes with a summary section that integrates the observations from the subsections and formulates answers to the research questions, maintaining a categorical character of analysis without extending beyond the scope of the review.

3.1. Battery Management Systems

Below are selected studies from recent years concerning the estimation of the State of Charge (SOC) and State of Health (SOH) of batteries, covering both machine learning models and hybrid, hardware, and review solutions.

In study [10], a residual convolutional neural network with a bidirectional GRU unit optimized using the orangutan algorithm was proposed. This model achieved very low RMSE and MAE errors for SOC and SOH. Study [11] combines the CatBoost algorithm with metaheuristics (BMO, PSO, GA, WOA) and, based on data from 72 BMW i3 trips, reports the best results for the BMO–CatBoost combination. Review [12] describes an IoT-Fog-Cloud architecture, emphasizing the selection of ML algorithms to improve SOC and SOH prediction. Article [13] employs the TimesNet model with DBSCAN clustering and a Savitzky–Golay filter. Although the main focus was on SOH, the analysis also includes SOC estimation, achieving a MAPE error of 0.39%.

The authors of [14] integrated IoT sensors with ML techniques, which increased battery efficiency by 18.6% and reduced fire risk by 72%. Review [15] classifies SOC estimation methods as simple, mathematically advanced, and data-driven, discussing their advantages and limitations. Study [16] introduced synthetic data generation using TimeGAN and the BERT model to account for variables omitted in previous research. The lightweight network with an attention mechanism described in [17] contains only 1713 parameters and achieved an RMSE of 1.23%. AI-assisted fast charging in [18] reduced charging time from 4.5 h to 1.5 h using a PID controller and artificial intelligence algorithms.

The digital twin developed in [19] combines various algorithms (RBF, RF, CNN, LSTM, SVR, XGBoost) for SOC and SOH prediction and ensures model explainability. In [20], a BP neural network estimated SOC at three temperatures (0, 25, and 45 °C) and achieved good agreement between predicted and actual results. Study [21] shows that the second peak of the differential capacity curve correlates better with SOH, with a neural network achieving an RMSE of 0.00330, and introduces a State of Functioning (SoF) indicator. In [22], linear regression, SVM, Random Forest, and neural networks were compared, with neural networks achieving the best results. The hybrid XGBoost–RF model in [23] improved accuracy to 97.6%, reducing both MSE and MAE. The dynamic Kalman model with a genetic algorithm and SVM (DGKNN) in [24] achieved an error of 0.1529%. A modification of RNN that used SOC from the previous step in [25] doubled estimation accuracy. An ensemble of homogeneous LSTM models in [26] shortened training time by 2.6–3.5 times while maintaining MAE ≈ 1.4%.

The DCRNN model with SVM-RFE feature selection in [27] achieved extremely low errors (RMSE ≈ 0.02%). Review [28] characterizes ML methods and their applications to SOC, SOH, and RUL, highlighting the need for comprehensive data analysis. The feedforward network in [29], tested on BMW i3 driving data, achieved a lower RMSE than the extreme learning machine. The digital twin with gradient boosting and adaptive EKF in [30] improved energy extraction and SOC monitoring. The enhanced CRC-SHEKF Kalman filter in [31] reached an MAE of 0.392%. The FPGA accelerator with LSTM in [32] demonstrated that SOC estimation can be performed efficiently directly on hardware. Comparison of gradient boosting and Random Forest in [33] showed the superiority of Random Forest (MAE < 0.3%). The MHDTCN-GRU model in [34], enriched with SHAP-based interpretability, achieved a MAPE of 0.54%.

The algorithm proposed in [35] combines linear regression and LSTM, reaching R² = 99%, while [36] discusses a battery health system based on LSTM and optimization. The critical review in [37] analyzes trends in the application of AI to BMS. The LSTMNNGA model in [38], optimized genetically, achieved an RMSE of 0.0795. In [39], a NARX network estimated SOC and SOH with RMSE of 0.5% and 0.018%, respectively. Gaussian Kalman filters in [40] reduced estimation error by 35–60%. The combination of Kalman filters and a deep network in [41] reduced RMSE to 0.04%, while parallel BiLSTM networks in [42] improved accuracy by 1.5–3 times. Study [43] developed a digital twin with incremental learning (MSE ≈ 0.022). The ML model in [44] predicted SOC and vehicle range for a rural EV with 95% accuracy, and the deep learning strategy in [45] reduced SOC error to 0.835%. In [46], deep deterministic policy gradient (DDPG) reinforcement learning achieved 98.8% accuracy. Comparison of DNN, ANN, and GRU-RNN in [47] showed the superiority of DNN. The telematics system in [48] predicted SOC and detected faults with accuracy above 97%. The hybrid VAR-LSTM model in [49] outperformed simple LSTM models. Study [50] showed that accounting for hysteresis reduced voltage error to 0.002 V, while the improved NARX training algorithm in [51] lowered MSE. Article [52] introduces the Cloud-BMS concept with anomaly detection. Reviews [53,54] discuss deep learning and ANN/SVM approaches, highlighting the need for model integration and data quality improvement. Study [55] presents a fractional-order model with MIUKF, achieving errors below 1.21% for SOC.

An important research direction is implementation under resource-constrained conditions. In [56], a one-dimensional CNN and GRU were implemented on STM32 microcontrollers; quantization reduced model size while maintaining an RMSE of 2.33%. The deep LSTM network in [57] achieved RMSE values between 0.024 and 0.045 for two Nissan Leaf datasets. Study [58] showed that a cascaded feedforward network was more effective than classical backpropagation in SOC and SOH estimation. In [59], an LSTM-SAE model optimized by the Black Widow algorithm achieved lower SOC errors at 25 °C. Energy management in a microgrid with PV forecasts in [60] increased the total SOC of vehicles by 9.49%. Article [61] combined three open-circuit voltage models optimized by the MOGA algorithm, improving SOC estimation and generating financial benefits in V2G services. In [62], a web application (Streamlit) was proposed that allows users to predict SOC based on input data. Review [63] discusses key SOC technologies for battery packs. The optimized Random Forest in [64], using a differential search algorithm, eliminated the need for preprocessing filters. In [65], an artificial neural network achieved a mean error of 2.65%. The support vector data descriptor technique in [66] allowed SOC estimation without temperature and capacity data. Review [67], devoted to equivalent circuit models, emphasizes their simplicity and the potential for future AI integration. The latest study [68] combines SVR regression, RC circuits, DFFRLS, and ASRUKF filtering, achieving RMSE and MAE below 1.95% even with a large initial error.

The collected results indicate that combining methods, such as hybrid models, digital twins, and integration with IoT and explainable AI techniques, systematically improves SOC and SOH estimation accuracy and robustness. At the same time, increasing attention is given to computationally efficient implementations suitable for deployment on embedded platforms. The overview of key methodological categories in State of Charge estimation for EV batteries is presented in

Table 2. Overview of key methodological categories in State of Charge estimation for EV batteries.

Category/Method	Characteristics	Example Findings	Papers
Deep neural networks	CNN, LSTM, GRU, and DNN models, often optimized via meta-heuristics or attention mechanisms; applied to both SOC and SOH estimation.	RMSE as low as 0.02% (DCRNN); MAE ≈ 0.02% (LSTMNNGA); RMSE 0.0873	[10,17,19,26,27,32,34,57,58]
Ensembles/boosting methods	Combine multiple algorithms (e.g., XGBoost + Random Forest or ensembles of LSTMs) or employ gradient boosting techniques.	Accuracy 97.6% and MAE 1.5%; training time reduced by 2.6–3.5	[23,26,33]
Kalman filters and equivalent circuit models	Use KF/EKF/UKF/MIUKF and equivalent circuit models, sometimes coupled with neural networks.	RMSE 0.04% with UKF + DFNN ([49]); MAE 0.392%; SOC MSE < 1.21%	[24,30,31,41,43,55,67,68]
Meta-heuristics and optimization techniques	Tune model parameters (CatBoost, ANN, RNN) using GA, PSO, BMO, DSA or Black Widow Optimization.	RMSE 6.10 and MAE 4.13; error < 2% in BWO-LSTM-SA; improved prediction through MOGA	[11,23,24,38,55,59,61,64]
Hybrid models and digital twins	Blend various algorithms (RBF, RF, CNN, LSTM, SVR) or build digital representations of the battery, often integrated with cloud platforms.	Improved accuracy and interpretability; Cloud-BMS with anomaly detection	[19,30,43,52]
IoT/telemetry-driven approaches	Integrate sensor or telematics data (voltage, current, temperature, vehicle speed) to feed machine learning models.	Performance + 18.6% and fire risk −72% > 97% accuracy for telematics-based predictions	[12,14,48,59,62]
Literature reviews	Survey existing methods, classify them, identify challenges, and point to future research directions.	Emphasize combining physics-based and data-driven models and improving data quality	[15,28,37,53,54,63,67]
Practical implementations	Cover real-world applications such as web apps, FPGA accelerators, energy management in charging stations, and V2G optimizations.	SOC increase in microgrid by 9.49%; Streamlit web app for SOC prediction; hybrid OCV model yields 10% SOC improvement and extra revenue in V2G services	[32,60,61,62]

.

Digital twins and cloud architectures open a new chapter in battery diagnostics because they enable continuous model updates based on operational data and scalable computation. In study [43], a digital twin of a Battery Management System was constructed. Measurement data from the vehicle were transmitted to the cloud, SoH was estimated using an incremental neural network, and SoC was determined using a Kalman filter, resulting in a mean square error of 0.022 and reduced hardware requirements. This concept was expanded in [52], where a battery cloud was developed in which SoC was calculated using a neural network and SoH was estimated using differential DVA and incremental ICA analyses. Anomaly detection was also introduced through monitoring of aggregate event statistics. The concept of explainable digital twins integrating diverse algorithms was presented in [19], emphasizing the importance of combining models with interpretability mechanisms.

In parallel, semi-analytical and hybrid approaches combining physical knowledge with statistical estimation and optimization are being developed. In [69], a two-stage health-aware energy management system was proposed, where relationships between capacity and internal resistance were first linearized and weights for control were then optimized, reducing operating costs without compromising stability. Study [70] combined a particle filter, quantum genetic algorithm, and GRNN network, achieving a mean SoC error of less than 1 percent and a SoH error below 2 percent. In [55], a fractional dynamic model coupled with a multi-innovation Kalman filter enabled simultaneous estimation of SoC and SoH with errors between 2 and 2.5 percent. The perspective of battery model parameter identification using optimization techniques capable of feeding voltage, current, SoC, and SoH computations was presented in [36]. The consideration of degradation in the electric vehicle routing problem and the adaptation of the algorithm to vehicle dynamics constraints were discussed in [71].

A complementary physically inspired direction within signal-based health diagnosis is polarization-oriented analysis, where degradation is interpreted through changes in overpotential-related signatures rather than treated purely as a black-box mapping. In practical BMS implementations, polarization growth can be tracked indirectly via online identification of equivalent-circuit parameters (e.g., ohmic resistance and RC time constants) and their evolution under controlled excitation, providing compact, interpretable indicators that can be fused with learning layers or filtering frameworks. This perspective is consistent with approaches combining fractional or circuit-based dynamics with Kalman-family estimators and with reviews emphasizing the role of equivalent-circuit modeling as a physically grounded backbone for AI integration [36,55,67]. In this sense, polarization-aware features can be treated analogously to differential and incremental curve descriptors, i.e., as physics-linked inputs that improve robustness and interpretability when coupled with machine learning predictors and cloud/digital-twin pipelines already discussed in this section [19,43,52].

Signal analysis provides another source of information on degradation, as it makes it possible to extract sensitive indicators without building full physical models. In [72], incremental capacity analysis was used, which employs derivatives of voltage and capacity curves, allowing SoH estimation with an error of about 2 percent on BMW i3 cells. The authors of [73] extended this concept through feature selection based on mutual information, reducing signal redundancy and improving SoH prediction accuracy by 9 to 52 percent compared with the baseline set. The integration of ICA and DVA methods with cloud infrastructure was described in [52]. The use of differential features to build indicators of functional state, including the correlation of the second peak of the differential curve with SoH, was presented in [21].

Machine learning-based solutions include both classical and deep models, as well as privacy preserving and federated learning approaches. In [74], an additive regression model was developed using fleet data, and the PSO ELM algorithm predicted capacity with an error below 0.4 percent over a two-month horizon. Reviews [53,54] emphasize the potential of convolutional and recurrent architectures but also point to the need for high data quality, large datasets, and careful hyperparameter tuning. A systematic review of privacy protection techniques in battery health prediction for electric vehicles was presented in [75]. A federated state estimation system with an autoencoder and attention mechanism that accounts for participant contribution was introduced in [76], and the combination of federated and self-supervised learning for first life batteries was described in [77]. A feature selection method based on constrained quantum Boltzmann machines, more effective than classical approaches, was presented in [77]. Operational applications of predictive ML for battery performance optimization were reported in [28], and the second part of a comprehensive review of state estimation methods for electric vehicles was presented in [78].

Deep models have brought further improvements in estimation accuracy and model robustness. In [10], a residual convolutional BiGRU network optimized by the orangutan algorithm achieved very low errors for SoC and SoH, including RMSE of 0.0873 and MAE of 0.0866 for SoC, and accuracy of 90.48 percent, RMSE of 0.1089, and MAE of 0.0952 for SoH. Study [79] employed generative adversarial networks with a triple attention mechanism, where a CNN LSTM generator extracted spatial and temporal features and a modified echo state network discriminator processed spatial, temporal, and contextual attention, reducing MAE, MSE, RMSE, and MAPE and improving R squared relative to comparison methods. Study [80] combined physics-informed neural networks with an equivalent circuit model and LSTM, showing that the serial configuration was more accurate and robust than the parallel version and the standard LSTM. In [81], a regularized box particle filter for SoH estimation reduced errors and computation time compared with SIR PF and box filters. The conceptual IoT Fog Cloud system aimed at selecting ML algorithms for battery data was presented in [12]. The framework using TimesNet with DBSCAN, the Savitzky–Golay filter, and correlation analysis, achieving MAPE of 0.39 percent and MSE of 0.20 percent, was described in [13]. The multiscale feature extraction network MMFEN, combining a representation block and a multi-head convolutional attention mechanism, was presented in [82], achieving RMSE of about 1.21 percent and MAPE of 0.99 percent on NASA data, with 2.78 percent and 2.71 percent on CALCE data and 2.39 percent and 2.08 percent on real EV data, respectively. The DCRNN model with SVM RFE feature selection in [27] achieved very low estimation errors, including RMSE of about 0.02 percent and MAPE of about 0.41 percent. The hybrid one-dimensional CNN and bidirectional GRU with Bayesian optimization was introduced in [83], while the combination of wavelet transform, CNN, and LSTM with an attention mechanism was described in [84]. The hardware implementation perspective with an FPGA accelerator for an LSTM model was presented in [32], achieving RMSE of 0.3438 in training and 0.3681 in validation. Estimation during charging using a NARX network was presented in [39], achieving RMSE of 0.5 percent for SoC and 0.018 percent for SoH.

Comparisons of classical methods show varying accuracy depending on data and operating conditions. In [85], KNN, SVM, decision tree, and Random Forest were evaluated, revealing MAPE of 2.791 × 10⁻² for SVR and 9.957 × 10⁻⁴ for KNN, and MAE of 2.429 × 10⁻³ for KNN. The critical review in [37] presents trends and a roadmap for the development of artificial intelligence in Battery Management Systems. Among deep learning methods, the combination of morphological filtering with a neural network in [86] achieved an R squared close to one and MSE of 0.03. The cascaded feedforward network in [58] outperformed the classical network in SoH estimation. Comparative analysis in [87] demonstrated greater robustness of logistic regression to changes in discharge rate than Random Forest or KNN. The two-layer Gaussian process regression model in [88] reduced the SoH prediction error to 1.3 percent and allowed lifetime prediction with an error of less than two cycles. Classification of studies in the “State Of Health” section by subcategory is presented in

Table 3. Classification of studies in the “State Of Health” section by subcategory.

Category	Representative Methods/Models	Key Metrics (Summary)	Sources (Article Numbers)
Battery Health	Federated learning with CAFS; IoT-Fog-Cloud system; explainable digital twins; BaHeS system; battery health monitoring system	Accuracy improvement of SOH by 20.18%; BaHeS—accuracy 93%, improvement 18%; remaining methods are conceptual frameworks	[12,19,28,36,75,76,89]
Battery Degradation	PINN (parallel/series); EV routing with a degradation model; aging analysis (M-SVM/ANN/regression); TCN network for second life; fast ANN	Series PINN more accurate; M-SVM best for aging; TCN–MSE < 1%; fast ANN–error ≈ 1.7%, time 30 s	[71,80,90,91,92]
State of Health	ResDCBi-GRU + OOA; PSO-ELM; TimesNet + DBSCAN/SG; MMFEN; DCRNN + SVM-RFE; LSTMNN + GA; feature selection with QRBM; estimation from partial charges; NARX; LSTM-TCN; BaHeS	RMSE_SOC 0.0873; MAE_SOC 0.0866; SOH 90.48%; RMSE_SOH 0.1089; error < 0.4%; MAPE 0.39%; MMFEN–RMSE 1.21%; DCRNN–RMSE ≈ 0.02%, MAPE ≈ 0.32%; LSTMNN + GA–RMSE 0.0795, MAE 0.0664; partial-charge estimation–RMSE 0.00330; NARX–RMSE_SOC 0.5%, RMSE_SOH 0.018%; BaHeS–accuracy 93%	[10,13,21,27,43,72,74,76,77,79,81,82,86,88,89,91,92,93,94]

.

Battery Thermal Management requires a close connection between modeling and experimental data in order to predict cell temperature and actively limit overheating in real operating scenarios. In study [95], CFD simulations were combined with a neural network trained using the Levenberg–Marquardt algorithm to analyze the influence of ambient temperature, convective coefficient, and discharge rate on the maximum surface temperature of a single cell in a battery pack. The obtained mean square error was 0.00552 and the coefficient of determination was 0.99, while the spread of maximum temperature predictions was small, with a standard deviation below 0.237 degrees Celsius.

Subsequent research focuses on integrating cooling system design with machine learning to improve control and reduce operating costs. The authors of study [96] proposed a hybrid liquid cooling method that combines a seahorse optimizer with an extended residual convolutional neural network, providing a reliable BTMS capable of maintaining the pack temperature at a safe level. Comparisons with existing techniques indicated error reduction, although no numerical values were provided in the abstract. Complementarily, study [97] combined fast charging with thermal management using reinforcement learning. The agent operated in an electrothermal environment that considered aging, controlling both charging current and coolant flow. For a pack composed of twenty cells, it found strategies in less than one second, while model predictive control required over eighty minutes. As a result, the maximum core temperature was maintained below thirty-three degrees Celsius, compared with about forty degrees Celsius under predictive control, and the expected lifetime after one thousand fast charging cycles was extended by up to two years.

Temperature forecasting also benefits from probabilistic models, which allow uncertainty estimation and improved control safety. In study [98], enhanced quantile convolutional and recurrent neural networks were developed, trained on both fleet and simulated data, with parameter tuning performed using a Bayesian approach. The best model achieved an RMSE of 0.66 degrees Celsius and an R² value of 0.84 for median predictions, while the ninety-ninth quantile covered 98.87 percent of the actual values. From a hardware design perspective, rapid exploration of the design space for heat exchangers is particularly important. Study [99] focused on optimizing liquid cooled plates with pin-fin structures. A dataset was created using coupled three-dimensional models and Latin Hypercube sampling, and deep encoder–decoder models were then trained as surrogate models. In multi-objective optimization using the NSGA-II algorithm, a plate geometry was obtained that reduced the maximum temperature by 4.87 kelvins, the temperature gradient by 5.1 kelvins, representing 22.2 percent, and hydraulic losses by 7.93 pascals, representing 9.0 percent.

In parallel, sequential models capable of learning thermal dynamics directly from onboard data are being developed. In study [100], a two-layer BiLSTM model was built to forecast the average pack temperature under real driving conditions using 2.3 million samples collected by an Internet of Things device. The model achieved a mean absolute error of 2.92 degrees Celsius on the test set and 1.7 degrees Celsius in cross-validation for a ten-minute forecasting horizon. In the area of real-time control, study [101] proposed a thermal management strategy based on a double Q network and a GRU unit. Compared with classical fuzzy control and two reinforcement learning strategies, the new method reduced energy consumption by more than 6.7 percent during aggressive driving.

Cross-domain generalization remains critical, as data distributions differ between simulation, fleet, and environmental conditions. In study [102], quantile convolutional networks were trained on data from multiple domains, including simulations, vehicle fleet data, and weather station measurements. Out of 150 variants, the selected model achieved a median prediction error with MAE equal to 0.27 degrees Celsius, and 47 percent of observations were below the median, indicating correct calibration and practical usefulness in thermal control applications. Regarding thermal management, a summary by thematic subcategories is presented in

Table 4. Thermal Management—summary by thematic subcategories.

Category	Representative Methods/Models	Key Results (Numeric Highlights)	Sources (Article Numbers)
Battery Temperature	CFD + ANN (Levenberg–Marquardt), quantile CNN/RNN (Q*NN), Bi-LSTM, cross-domain quantile CNN	MSE ≈ 0.00552, R2 ≈ 0.99; RMSE ≈ 0.66 °C, R2 ≈ 0.84; MAE ≈ 2.92 °C (test) and ≈1.7 °C (cross-validation); MAE ≈ 0.27 °C, 47% values below median	[95,98,100,102]
Temperature Control	Deep RL for fast charging and thermal management, double DQN + GRU, hybrid SHO-DRCNN	RL simulation < 1 s vs. MPC > 80 min; core temp < 33 °C vs. 40 °C; battery life + 2 years; energy reduction > 6.7% during aggressive driving; SHO-DRCNN–qualitative error reduction (no metrics)	[96,97,101]
Thermal Management (electronics)	Deep RL with electro-thermal aging model, double DQN	Energy reduction > 6.7%; core temperature reduction and life extension	[97,101]
Battery Thermal Managements	Hybrid SHO-DRCNN BTMS, deep RL fast-charging management, double DQN	Improved BTMS reliability; RL keeps core temp < 33 °C and extends life; double DQN reduces energy consumption by > 6.7%	[96,97,101]
Thermal Management Systems	SHO-DRCNN BTMS, DeepEDH-CNN with NSGA-II for cold plates, deep RL for fast charging, double DQN	SHO-DRCNN improves BTMS (qualitative); cold-plate design reduces max temperature by 4.87 K, gradient 5.1 K (22.2%) and cooling energy by 9%; RL results as above	[96,97,99,101]
Cooling	Hybrid SHO-DRCNN liquid cooling, RL-controlled coolant flow, pin-fin cold-plate design	SHO-DRCNN lowers cooling errors; RL maintains core temperature < 33 °C vs. 40 °C; cold-plate optimization reduces max temperature by 4.87 K and gradient by 5.1 K (22.2%)

.

3.2. Computational Intelligence

Progress in battery state estimation increasingly relies on models that combine signal representation, physical knowledge, and attention mechanisms. In study [79], a generative model for health state prediction with a triple attention mechanism was presented, in which the generator combines convolutional neural networks and LSTM to extract features, while the echo state network discriminator uses spatial, temporal, and contextual attention. Across three datasets, this approach outperformed classical methods in terms of MAE, MSE, RMSE, MAPE, and the coefficient of determination. Complementarily, researchers in [80] compared two configurations of a physics-informed neural network that combines an equivalent circuit model with LSTM. The serial configuration, which introduces physical model parameters directly into the LSTM, captured degradation patterns more effectively and proved more accurate and robust than the parallel configuration and the baseline LSTM.

Thermal modeling also plays an important role as it supports control strategy design and operational safety. In study [95], CFD calculations were combined with a neural network trained using the Levenberg–Marquardt algorithm to predict the maximum surface temperature of a cell based on operating conditions. The model achieved a mean square error of approximately 0.00552 and an R² value of 0.99, and the deviation from experimental results did not exceed 0.237 degrees Celsius. A purely data-driven approach to State of Charge estimation was presented in [81], where CatBoost was combined with metaheuristics such as Barnacles Mating Optimizer, PSO, GA, and WOA, tested on data from 72 BMW i3 trips. The BMO–CatBoost configuration achieved an RMSE of 6.1031, an MAE of 4.1303, and an R² of 0.8211. The conceptual IoT–Fog–Cloud system in [11] emphasizes the integration of data sources and algorithm selection frameworks for evaluation. Conversely, study [13] shows that combining IoT sensors with learning algorithms allows simultaneous prediction of performance and reduction in fire risk through analysis of voltage, current, temperature, and State of Charge, achieving 99.4 percent accuracy, a 72 percent reduction in fire risk, and an 18.6 percent increase in efficiency.

To enhance scalability and privacy, study [94] applied feature selection using a quantum restricted Boltzmann machine on a D-Wave computer, yielding feature sets that were more unique and compactly reduced than those obtained with PSO or Pearson correlation. Study [77] combined federated and self-supervised learning, which, in simulations for first-life batteries, reduced the mean square error of SoH estimation by 31 percent compared with baseline models while maintaining data protection. For BMS implementation, a lightweight model with an attention mechanism was developed in [17], achieving an RMSE of approximately 1.23 percent with only 1713 parameters, and [18] described dynamic fast charging with a data-driven PID controller, reducing theoretical charging time from 4.5 h to 1.5 h and decreasing cell damage by at least 50 percent. The framework of an explainable digital twin integrating RBF, RF, FNN, CNN, LSTM, DNN, SVR, SVM, and XGBoost was presented in [19], emphasizing the interpretability of models. Study [21] demonstrated that partial charge data can be used to accurately estimate battery health and functionality, with a neural network achieving an RMSE of 0.00330. The review in [78] synthesizes trends in health state estimation methods that combine data-driven and physics-based approaches.

The digital twin concept was extended to renewable energy applications in [30], where collaborative gradient boosting was combined with an adaptive Kalman filter to estimate SoC and resistance in a wind-powered charging system. Field verification confirmed improved accuracy and adaptability. Hardware feasibility was demonstrated in [32] by implementing a BMS accelerator on an FPGA, where an LSTM model trained in Python was synthesized in Xilinx HLS, achieving an RMSE of 0.3438 in training and 0.3681 in validation. Additionally, the design of heat exchangers was optimized in [99], where DeepEDH CNN models acted as surrogates for costly simulations, and the NSGA-II algorithm optimized the geometry of pin-fin plates, reducing the maximum temperature by 4.87 kelvins, the temperature gradient by 5.1 kelvins, and the cooling energy demand by 9 percent.

Comparisons of classical algorithms show varying sensitivity to data and operating conditions. In study [33], Random Forest achieved an MAE below 0.3 percent and an R² of 0.999, outperforming gradient boosting, which obtained an MAE of approximately 1 percent and an R² of 0.997. In study [35], linear regression was combined with LSTM in a web application, achieving an R² of approximately 0.99 for real-time State of Charge estimation. In study [36], conducted on the Sandia dataset, KNN, SVR, decision trees, and Random Forests were compared. At 15 degrees Celsius, MAPE values were 2.791 × 10⁻² for SVR, 1.607 × 10⁻³ for trees, 9.957 × 10⁻⁴ for KNN, and 4.323 × 10⁻³ for forests, while at 25 degrees Celsius, MAE values were 5.593 × 10⁻², 2.379 × 10⁻³, 2.429 × 10⁻³, and 5.073 × 10⁻³, respectively, indicating the advantage of KNN and decision trees. The hybrid one-dimensional CNN and bidirectional GRU with Bayesian optimization described in [83] achieved, on the NASA dataset, MAE of 2.080 percent, RMSE of 2.516 percent, and a zero end-of-life prediction error, with MSE after 19 iterations equal to 1.2 × 10⁻⁵. A critical review of artificial intelligence in BMS was conducted in [37], summarizing the strengths and limitations of methods and outlining future development paths.

Integrating probabilistic approaches with Kalman filtering improves not only accuracy but also the reliability of confidence intervals. Study [40] proposed GP-UKF and GP-PF for State of Charge estimation, which, through covariance matrix adaptation and sample optimization, improved accuracy and reduced uncertainty in UDDS and HWFET simulations compared with classical counterparts. The concept of a digital twin with incremental learning was demonstrated in [43], where SoH was estimated in the cloud and SoC in the vehicle using a Kalman filter, achieving an MSE of 0.022. Application-oriented operational scenarios were expanded in [44], which combined State of Charge estimation with range prediction for rural electric vehicles, achieving an average SoC accuracy of 95 percent and a range error below 2 percent.

The energy and degradation management perspective was addressed in [69], where two-stage linearization of semi-empirical relationships and mapping of degradation features to lost lifetime reduced aging costs in V2G and PHEV scenarios. The rationale for constructing data-driven frameworks based on fleet data was confirmed in [74], where after identifying aging features using the GAM method and analyzing correlations, the ELM weights were optimized with PSO, and the estimation error for a real battery pack dropped below 0.4 percent. A comparison of architectures for State of Charge estimation was presented in [47], where the DNN achieved an average error below 0.5 percent and a maximum error below 2.4 percent on experimental data, and an average below 0.4 percent and maximum below 2.5 percent on simulated data, outperforming GRU RNN. The specific characteristics of aviation eVTOL applications were discussed in [103], which highlighted the dominant role of discharge features and the superiority of Random Forest and XGBoost, while [90] showed that a modified SVM achieved the lowest degradation prediction error compared with neural networks and linear regression.

Telematics data and temporal feature sequencing are gaining importance. In study [61], CAN and GPS data were used for predictive battery maintenance in shared fleets, identifying potential faults with an accuracy above 97 percent and improving State of Charge estimation when temperature was added as a feature. In [48], LSTM, BiLSTM, VAR, and a hybrid VAR-LSTM model were compared, with the combined model yielding higher accuracy than the individual architectures. Study [73] proposed a preprocessing and feature selection pipeline for predicting cyclic capacity loss, improving accuracy by at least 9 percent for LASSO and by 44, 48, and 52 percent when combined with Random Forest, GPR, and XGBoost, respectively. Study [51] compared NARX training using the Levenberg–Marquardt and scaled conjugate gradient algorithms on CALCE data, with the former achieving a lower MSE of 4.61306 × 10⁻⁶. The concept of a cloud-based BMS was further developed in [52], integrating neural networks for SoC estimation with DVA and ICA analyses for SoH estimation and anomaly detection.

Thermal management and real-time control benefit from sequential learning and uncertainty models. In study [100], more than 2.3 million samples from field tests were collected to train a two-layer BiLSTM model that predicted the average pack temperature with an MAE of 2.92 degrees Celsius on the test set and 1.7 degrees Celsius in cross-validation for a ten-minute forecasting horizon. Review [53] organizes deep learning applications for estimating State of Charge, State of Health, and remaining useful life, highlighting challenges related to feature selection, preprocessing, and hyperparameter tuning. A synthetic overview of neural network, gradient boosting, and SVM methods was presented in [70], emphasizing feature selection and implementation limitations. Study [55] proposed a real-time thermal management strategy based on a double Q network with GRU, which, compared with fuzzy control and two reinforcement learning strategies, reduced energy consumption by more than 6.7 percent during aggressive driving. Miniature networks for microcontroller deployment, including one-dimensional CNN and GRU RNN, were presented in [56], where quantization to eight-bit precision in STM32Cube AI and TFLite Micro reduced flash memory from 10.82 to 2.89 kilobytes and RAM to 1.04 kilobytes while maintaining an RMSE of 2.33 percent and MAE of 1.62 percent.

Scaling to production data and integration with digital infrastructure require robust models and tools. In study [86], a health state prediction model based on machine learning, big data, and IoT was proposed, achieving an R² of 0.9999 in training and 0.9995 in testing, with an MSE of 0.03. Study [58] compared a backpropagation network and a cascaded feedforward neural network (CFNN), with the CFNN providing better mapping of State of Charge and State of Health values on NASA and Panasonic 18650 PF datasets. Study [59] applied an autoencoder-based LSTM optimized using the Black Widow algorithm for State of Charge estimation in hybrid vehicles, achieving an error range from minus 3.5 percent to plus 4.3 percent at 25 degrees Celsius and smaller deviations compared with reference methods. Study [87] compared KNN, logistic regression, and Random Forest for lifetime prediction, indicating advantages of KNN, although no detailed metrics were reported. In study [102], quantile convolutional networks were developed and trained on multi-domain data, including simulated, fleet, and environmental datasets. The best model achieved an MAE of 0.27 degrees Celsius for the median and 47 percent of observations below that median. The second-life battery perspective was presented in [91], where a temporal convolutional network maintained MSE below 1 percent even under dynamic load profiles. A user-oriented tool perspective was demonstrated in [62] in the form of a Streamlit web application for State of Charge prediction based on input parameters. Study [64] optimized a Random Forest using differential search, which improved State of Charge estimation accuracy and eliminated the need for preprocessing filters. A real-time method based on a support vector data descriptor was presented in [66], where State of Charge estimation did not require knowledge of temperature or battery capacity and was successfully validated using simulated data. A thematic summary for AI/ML Core is presented in

Table 5. AI/ML Core—thematic summary.

Category	Representative Methods/Models	Key Results (Numeric Highlights)	Literature
Artificial Intelligence	Battery cloud with advanced algorithms; explainable digital twins; digital-twin model for wind-driven EV charging	Cloud-BMS uses neural networks with DVA/ICA for SOC/SOH estimation and detects early thermal anomalies; explainable twins enhance interpretability of AI-based predictions (no specific metrics); wind-driven digital twin uses gradient boosting and adaptive Kalman filtering to improve SOC estimation	[19,30,52]
Machine Learning	BMO–CatBoost metaheuristic; novel feature selection strategy; NARX with Levenberg–Marquardt; quantized CNN/GRU on microcontrollers; robust SOH model; CFNN vs. FBPNN; IoTAI-SOC with LSTM-SAE and BWO; KNN/logistic regression/Random Forest comparison; differential-search-optimized random forest	BMO–CatBoost: RMSE ≈ 6.1031, MAE ≈ 4.1303, R2 ≈ 0.8211; feature selection improves accuracy by ≥ 9% (LASSO) and by 44–52% for RF/GPR/XGBoost; Levenberg–Marquardt NARX: MSE 4.6 × 10−6; 1D CNN: RMSE 2.33%, MAE 1.62%, memory reduced to 2.89 KB flash; robust SOH model: R2 ≈ 0.9999 (train), 0.9995 (test), MSE = 0.03; CFNN outperforms FBPNN (less overshoot); IoTAI-SOC error range −3.5% to +4.3% at 25 °C; KNN noted as most appropriate without metrics; differential-search-optimized Random Forest improves SOC estimation over other methods	[11,51,58,59,64,73,86,101]
Learning Systems	IoT-Fog-Cloud framework; federated learning with contribution-aware strategy (CAFS); self-supervised + federated learning; explainable digital twins; deep learning review; ANN/GB/SVM review; quantile CNN cross-domain; TCN for second-life SOH; Streamlit-based SOC predictor	IoT-Fog-Cloud lays out a sensor-integrating ML architecture (conceptual); CAFS improves SOH accuracy by 20.18%; self-supervised federated learning reduces MSE by 31%; reviews highlight challenges in feature selection, preprocessing and hyper-parameter tuning; quantile CNN achieves MAE 0.27 °C, with 47% actuals below median; TCN yields MSE < 1% for retired batteries; Streamlit-based app demonstrates SOC prediction but lacks metrics	[12,19,53,54,62,76,77,91,102]

.

Research on battery state estimation increasingly combines signal representation, physical knowledge, and attention mechanisms. In study [79], a generative adversarial model for State of Health prediction was proposed, in which the generator combines convolutional neural networks and LSTM for feature extraction, while the discriminator uses a modified echo state network with a triple attention mechanism. Tests on three datasets showed lower MAE, MSE, RMSE, and MAPE values and higher R² compared with competing solutions. Complementarily, study [80] presented two configurations of a physics-informed neural network for health-state forecasting, a parallel and a serial one, both combining an equivalent circuit model with LSTM. The serial configuration, which feeds physical model parameters directly into the LSTM, proved more accurate and more robust than the parallel setup and the standard LSTM.

Thermal modeling also plays an important role as it supports control strategy design and operational safety. In study [95], the maximum temperature of lithium-ion cells was analyzed by combining CFD simulations with a neural network trained using the Levenberg–Marquardt algorithm. The network predicted the maximum surface temperature based on nominal capacity, ambient conditions, and discharge rate, achieving an MSE of 0.00552, an R² of 0.99, and a forecast standard deviation below 0.237 °C. The IoT–Fog–Cloud conceptual system described in study [12] emphasizes the need for integration of heterogeneous sensors, algorithm selection, and evaluation metrics, although no numerical results were reported. A comprehensive review of State of Charge estimation methods, from lookup tables and ampere-hour counting through electrochemical models with observers to data-driven approaches, was presented in study [15], discussing their advantages and limitations in the context of electric vehicles.

The development of feature and representation methods includes both new architectures and data generation. In study [82], the MMFEN network was proposed for State of Health estimation, combining an enhanced representation block and a multi-head convolutional attention module. It achieved RMSE of 1.21 percent and MAPE of 0.99 percent on the NASA dataset, 2.78 percent and 2.71 percent on CALCE data, respectively, and 2.39 percent and 2.08 percent on real EV data. Study [16] introduced the TimeGAN data synthesis system combined with a BERT model for State of Charge prediction, indicating that the inclusion of operational variables and dataset diversification can improve accuracy, which was presented as a conceptual framework. Scalability and privacy were advanced in study [76], which designed a federated learning framework with a contribution-aware attention mechanism that combines an autoencoder and attention module, improving average State of Health estimation accuracy by 20.18 percent. Study [77] combined self-supervised and federated learning, reducing the SoH prediction MSE by 31 percent under simulation conditions. For practical BMS implementations, a lightweight attention-based SoC estimator was proposed in study [17], which, with only 1713 parameters, achieved an RMSE of 1.23 percent on the LG 18650HG2 cell.

The concept of digital twins and control algorithms is being further developed in subsequent studies. Study [19] described explainable digital twins using RBF, RF, FNN, CNN, LSTM, SVR, SVM, and XGBoost for SoC and SoH prediction, emphasizing interpretability without reporting numerical metrics. In study [20], a BP neural network was built to estimate SoC during charging based on feature engineering and historical data, achieving low errors over a wide temperature range. Study [96] combined a seahorse optimizer with an extended residual convolutional neural network in a hybrid liquid cooling system intended to improve BTMS reliability, although no numerical results were reported. In study [22], ML models were compared using data from Europe and Africa, and neural networks achieved the lowest MSE and highest R². Study [25] introduced an autoregressive modification of RNN for State of Charge estimation, feeding the predicted value from the previous step as input, which doubled accuracy compared with baseline models with only a slight increase in training time. Study [26] used a homogeneous ensemble of LSTM models with meta-learning, reducing training time by 2.6 to 3.5 times and achieving MAE of about 1.4 percent. Study [27] applied a diffusion convolutional recurrent neural network with SVM-RFE feature selection for SoH forecasting, obtaining RMSE of about 0.02 percent, MAE of about 0.015 percent, MSE of about 0.026, and MAPE of about 0.32 percent. The operational framework for applying learning methods to optimize battery performance was presented in study [28], and a comprehensive review of State of Health estimation methods for electric vehicles was provided in study [78]. In the context of operational data, a feedforward neural network for SoC estimation based on 70 BMW i3 driving sessions was presented in study [29], achieving an RMSE 2.87 percent lower than that of the next best ELM model.

Energy, thermal, and hardware management are supported by learning methods and uncertainty models. In study [97], reinforcement learning was applied to simultaneously optimize fast charging and cooling, which reduced computation time from more than 80 min to less than one second for a 20-cell pack, maintained the core temperature below 33 °C compared with about 40 °C for predictive control, and extended battery life by up to two years after 1000 fast charging cycles. In study [98], improved quantile convolutional and recurrent neural networks were developed for temperature forecasting, achieving RMSE of 0.66 °C and R² of approximately 0.84 for the median, with the 0.99 quantile covering 98.87 percent of observations. Study [32] designed a BMS accelerator on FPGA using an LSTM network, achieving RMSE of 0.3438 in training and 0.3681 in validation. In study [33], Random Forest and gradient boosting were compared for State of Charge estimation. Random Forest achieved MAE below 0.3 percent and R² of 0.999, outperforming gradient boosting, which yielded MAE of about 1 percent and R² of 0.997. Study [34] developed an interpretable MHDTCN-GRU model with SHAP analysis, obtaining MAPE of 0.54 percent and RMSPE of 0.84 percent with 374,433 parameters. Study [35] combined linear regression and LSTM in a web application for real-time State of Charge prediction, achieving R² of 99 percent. Study [36] described a conceptual battery monitoring system in which parameter optimization and LSTM are used for simultaneous estimation of State of Charge, State of Health, and remaining useful life, without reporting numerical metrics. Study [83] presented a CNN-BiGRU model for State of Health estimation, where Bayesian optimization with Gaussian process reduced MSE to 1.2 × 10⁻⁵. On the NASA dataset, it achieved MAE of 2.080 percent, RMSE of 2.516 percent, and a zero end-of-life prediction error. Study [38] presented a probabilistic LSTMNN model optimized by a genetic algorithm, which achieved an RMSE of 0.0795, MAE of 0.0664, and MAPE of 45.31 percent, outperforming classical LSTMNN and ANN. Study [84] combined wavelet transform, CNN, LSTM, and an attention mechanism, with the addition of wavelets significantly improving State of Health diagnostic accuracy.

Scaling to the level of cells and packs and ensuring robustness under different operating conditions broadens the spectrum of sequential and hybrid models. Study [42] designed a parallel architecture with multiple BiLSTM layers for estimating the State of Charge of individual cells and packs, achieving accuracy improvements of up to 1.5 to 3 times compared with conventional RNNs, particularly near operating extremes. Study [45] presented an optimized deep learning strategy for State of Charge estimation across different temperatures and C-rates, with an error of 0.835 percent. Study [46] applied deterministic policy gradient reinforcement learning for State of Charge estimation and management, achieving 98.8 percent accuracy and shorter convergence times, which improved lifespan and thermal safety. Study [90] analyzed battery aging under dynamic conditions using neural networks, a modified SVM, and linear regression, with the modified SVM achieving the lowest prediction errors. Study [49] compared LSTM, BiLSTM, VAR, and a hybrid VAR-LSTM model, with the combined model producing the most accurate State of Charge forecasts. Study [50] examined the effect of voltage hysteresis, where the LSTM model incorporating hysteresis reduced voltage error to 0.002 V and improved predictions under variable conditions. Study [51] evaluated NARX training algorithms for State of Charge estimation, with the Levenberg–Marquardt algorithm achieving MSE of 4.61306 × 10⁻⁶ and outperforming the scaled conjugate gradient algorithm.

Cross-domain generalization and real-time operation require rich data streams and efficient implementations. In study [100], more than 2.3 million samples collected by an Internet of Things device were used to predict the battery pack temperature in a BEV using a BiLSTM network. The model achieved an MAE of 2.92 °C on the test set and 1.7 °C in cross-validation for a ten-minute prediction horizon. Study [53] presented a review of deep learning applications for estimating State of Charge, State of Health, and remaining useful life, discussing architectures, feature selection challenges, tuning strategies, and development prospects. Study [101] proposed a real-time thermal management algorithm based on reinforcement learning with a GRU module and a double DQN network, which reduced energy consumption by more than 6.7 percent during aggressive driving compared with fuzzy control and other RL methods. Study [56] described State of Charge estimation on microcontrollers using a one-dimensional CNN and GRU network after quantization in STM32Cube AI and TFLite Micro, achieving RMSE of 2.33 percent and MAE of 1.62 percent while reducing flash memory to 2.89 kilobytes and RAM to 1.04 kilobytes. In study [57], a deep LSTM network was applied to predict State of Charge using data from Nissan Leaf, achieving an RMSE of 0.0239 and MAE of 0.0202 on the first dataset and an RMSE of 0.0449 and MAE of 0.0412 on the second. Study [59] presented an IoTAI-SOC model combining LSTM-SAE with the Black Widow Optimization algorithm, achieving an error range from minus 3.5 percent to plus 4.3 percent at 25 °C. Study [60] described a predictive energy management system for a charging station that integrates LSTM-based photovoltaic power forecasting with a rule-based power flow algorithm, increasing the total state of charge of vehicles by 9.49 percent within a 60 min horizon. Study [102] developed quantile convolutional networks trained on multi-domain data for battery temperature forecasting. The best model achieved an MAE of 0.27 °C for the median, with 47 percent of observations below the median. Study [93] applied an LSTM network to predict remaining useful life using two years of data. After cleaning and approximating the health trajectory with a fifth-degree polynomial, an MSE of about 1 percent was achieved. Study [104] presented LSTM-based prediction of State of Health and degradation in the context of recycling, enabling early decisions on reuse, regeneration or disposal. Study [92] developed a unidirectional network based on equivalent circuit parameters optimized by the Gauss method, achieving MSE of about 1.729 percent with a deviation of 0.147 percent and reducing measurement time to 30 s. Study [63] conducted a review of key State of Charge estimation technologies for fully electric vehicles, discussing factors influencing accuracy and research directions. Study [89] described the BaHeS system for State of Health estimation with aging propagation, in which voltage entropy identifies faulty cells and cell-interaction modeling improves prediction. The LSTM network achieved 93 percent accuracy and an 18 percent improvement compared with baseline methods. Study [65] presented a five-step method for State of Charge estimation based on artificial neural networks, in which the maximum error was 18 percent, the mean error was 2.65 percent, and the overall accuracy was 97.35 percent. For Neural Networks, an overview by thematic subcategories is presented in

Table 6. Neural Networks—overview by thematic subcategories.

Subcategory	Representative Examples	Key Metrics (Summary)	Literature
Neural Networks	ANN for cell temperature prediction; BP neural network for SOC during charging; feedforward neural network (FFNN) for SOC in BMW i3; MHDTCN-GRU model with SHAP; NARX model with Levenberg–Marquardt training; single-layer ANN for second-life SOH; five-step ANN SOC estimator	ANN model predicted maximum cell temperature with MSE ≈ 0.00552 and R2 ≈ 0.99; FFNN reduced RMSE by 2.87% compared with ELM; MHDTCN-GRU achieved MAPE = 0.54% and RMSPE = 0.84%; NARX trained via Levenberg–Marquardt reached MSE ≈ 4.6 × 10−6; single-layer ANN estimated SOH with MSE ≈ 1.73% in ~30 s; five-step ANN yielded max SOC error 18% and average error 2.65% (accuracy 97.35%)	[20,29,34,51,65,92,95]
Deep Learning	GAN with CNN-LSTM and triple attention; physics-informed neural networks (PINN); MMFEN (CNN with attention); lightweight feature-attention network (1713 params); DCRNN with SVM-RFE; deep RL for fast charging and thermal management; quantile CNN/RNN models; CNN-BiGRU with Bayesian optimization; LSTMNN + GA; SOC estimation via CNN on microcontrollers; BiLSTM architectures for cell SOC; deep learning SOC estimation under varying C-rates; RL-based SOC management; double DQN with GRU for thermal control; deep LSTM forecasting for SOC; LSTM-SAE with Black Widow Optimization; predictive energy management using LSTM; quantile CNN trained on cross-domain data	MMFEN achieved RMSE ≈ 1.21% and MAPE ≈ 0.99% on NASA data; DCRNN + SVM-RFE reached RMSE ≈ 0.02%, MAE ≈ 0.015%, MAPE ≈ 0.32%; deep RL reduced computation time from >80 min to <1 s and maintained core temperature < 33 °C; quantile CNN/RNN obtained RMSE ≈ 0.66 °C and R2 ≈ 0.84; CNN-BiGRU (with Bayesian optimization) gave MSE = 1.2 × 10−5, MAE = 2.08%, RMSE = 2.516%; LSTMNN + GA delivered RMSE = 0.0795 and MAE = 0.0664; microcontroller-oriented 1D-CNN attained RMSE = 2.33% and MAE = 1.62% while reducing flash memory to 2.89 KB; BiLSTM architecture improved SOC accuracy by factor 1.5–3; optimized deep learning SOC model reduced error to 0.835%; RL-based SOC estimator achieved 98.8% accuracy; deep LSTM SOC model delivered RMSE = 0.0239 and MAE = 0.0202 on one dataset; IoTAI-SOC produced error within −3.5% to +4.3%; predictive energy management using LSTM increased total SOC by 9.49%; quantile CNN cross-domain model achieved MAE ≈ 0.27 °C	[17,27,38,45,46,59,60,79,80,82,83,86,97,98,101,102]
Long Short-term Memory	BiLSTM and LSTM variants for SOC/SOH prediction; LSTM ensembles; LSTM on FPGA; regression + LSTM SOC predictor; LSTMNNGA; hybrid CNN-BiGRU; LSTM with wavelet transform and attention; BiLSTM temperature forecasting; LSTM hysteresis model; deep LSTM SOC forecasting; IoTAI-SOC (LSTM-SAE); RUL prediction with LSTM; EoU management with LSTM; BaHeS SOH estimation with LSTM	Lightweight feature-attention LSTM achieved RMSE = 1.23% on LG 18650HG2; ensemble LSTM-RNN reduced training time by up to 3.5× with MAE ≈ 1.4%; FPGA-implemented LSTM produced RMSE = 0.3438 (train) and 0.3681 (val); regression + LSTM model had R2 = 0.99; LSTMNNGA obtained RMSE = 0.0795 and MAE = 0.0664; CNN-BiGRU (with BO) reached MSE = 1.2 × 10−5; LSTM hysteresis model reduced voltage error to 0.002 V; BiLSTM temperature model achieved MAE = 2.92 °C (test) and 1.7 °C (cross-val); deep LSTM SOC model delivered RMSE = 0.0239; MAE = 0.0202; IoTAI-SOC error range −3.5% to 4.3%; LSTM RUL model produced MSE ≈ 1%; LSTM-based end-of-use management emphasized early degradation detection; BaHeS (LSTM) accuracy ≈ 93%, improvement 18%	[17,26,32,35,38,45,49,50,56,57,59,60,69,84,89,93,100,104]
Convolutional Neural Network	CNN–LSTM generator in GAN SoH model; MMFEN feature extractor with convolutional attention; quantile CNN for temperature prediction; CNN-BiGRU hybrid for SOH; wavelet-CNN-LSTM model; 1D CNN on microcontrollers; quantile CNN trained on cross-domain data	MMFEN achieved RMSE 1.21% and MAPE 0.99% on NASA data; quantile CNN/RNN yielded RMSE ≈ 0.66 °C, R2 ≈ 0.84 and covered 98.87% of data within the 99th quantile; CNN-BiGRU model reached MSE 1.2 × 10−5, MAE 2.08%, RMSE 2.516%; wavelet-CNN-LSTM improved accuracy relative to plain LSTM; 1D CNN with quantization achieved RMSE 2.33% and MAE 1.62% while reducing memory footprint to 2.89 KB; cross-domain quantile CNN attained MAE 0.27 °C with 47% of real values below the median prediction	[56,79,82,83,84,97]

.

Progress in State of Charge and State of Health estimation increasingly relies on combining deep learning with domain knowledge and on deliberate feature representation design. Study [10] presented a ResDCBiGRU model optimized with the orangutan algorithm that simultaneously estimates SOC and SOH, achieving for SOC an RMSE of 0.0873 and an MAE of 0.0866 and for SOH an accuracy of 90.48 percent with an RMSE of 0.1089 and an MAE of 0.0952. The generative approach was extended in study [79], where SOH prediction was performed using a generative adversarial network with a CNN LSTM generator and an echo state discriminator with a triple attention mechanism, yielding lower MAE, MSE, RMSE, and MAPE and higher R² than comparative methods across three datasets. Study [13] employed TimesNet with DBSCAN clustering, a Savitzky–Golay filter, and Spearman rank correlation, obtaining a MAPE of 0.39 percent and an MSE of 0.20 percent on four datasets. The precision of further designs is illustrated by the CNN BiGRU model with Bayesian optimization in study [83], which on NASA data achieved an MSE of 1.2 × 10⁻⁵, an MAE of 2.08 percent, and an RMSE of 2.516 percent, with a zero end-of-life error. Very low errors were also obtained in study [27], where DCRNN with SVM RFE feature selection achieved an RMSE of about 0.02 percent and a MAPE of about 0.41 percent. Interpretability and feature importance weighting are provided by the MHDTCN GRU model with SHAP analysis in study [34], which reported a MAPE of 0.54 percent and an RMSPE of 0.84 percent.

Classical and hybrid methods remain strong reference points and a basis for fusion. Study [23] combined XGBoost and Random Forest to estimate SOC with an accuracy of 97.6 percent and an MSE between 1.3 and 1.5 percent. Study [33] compared Random Forest and gradient boosting, with Random Forest achieving an MAE below 0.3 percent and R² of 0.999. Study [78] used a feedforward network to estimate SOC from 70 BMW i3 trips, achieving an RMSE that was 2.87 percent lower than that of an ELM model. Study [35] combined linear regression and LSTM in a web application, achieving R² of 99 percent in real time. Study [22] compared linear regression, SVM, Random Forest, and neural networks on European and African datasets, with neural networks obtaining the best results. A review of SOH estimation methods was identified in study [78], while a review focused on neural networks, gradient boosting, and SVM, along with feature selection issues, was presented in study [54].

A strong group of approaches is formed by semi analytical and filtering methods that link physical models with learning. Study [41] coupled a nonlinear Kalman filter based on a two RC model with a deep feedforward network, achieving an SOC error below 0.5 percent and an RMSE of 0.04 percent. Study [81] proposed a regularized box particle filter for SOH estimation in packs that, by incorporating internal resistance and linear regression, reduced sample degeneracy and computation time compared with SIR PF and box filters. Study [24] presented a dynamic Kalman network with genetic optimization integrated with SVM, achieving a minimum SOC error of 0.1529 percent and an MSE of 0.0604. Study [55] used a fractional-order dynamic model with a multi-innovation Kalman filter for simultaneous SOC and SOH estimation, achieving at 25 degrees Celsius an SOC RMSE below 0.38 percent and an SOH RMSE below 0.002 percent, and below 1.21 percent and 0.007 percent, respectively, for different aging states. Study [49] showed that a VAR LSTM hybrid outperforms single models, and study [51] demonstrated that training NARX with the Levenberg–Marquardt algorithm yields an MSE of 4.61306 × 10⁻⁶ and outperforms the scaled conjugate gradient method. Study [90] found that a modified SVM outperforms neural networks and linear regression in degradation prediction under dynamic conditions. The probabilistic perspective is completed by study [88], where a dual Gaussian process regression achieves maximum SOH errors below 1.3 percent and RUL errors below two cycles with training and testing times up to five seconds. In the broader context, ECM methods and their integration with ML are discussed in review [67].

Thermal modeling and control directly affect battery safety and lifetime. In study [95], CFD was combined with a neural network trained using the Levenberg–Marquardt method to predict the maximum surface temperature of a cell, achieving an MSE of 0.00552, an R² of approximately 0.99, and a standard deviation below 0.237 degrees Celsius. Study [98] improved quantile CNN and RNN models for temperature prediction, achieving an RMSE of 0.66 degrees Celsius and an R² of about 0.84 for the median, while the 99th quantile covered 98.87 percent of the observations. Study [31] introduced the CRC-SHEKF filter for SOC estimation with covariance matrix adaptation and a tuning factor, which at 15 degrees Celsius resulted in an MAE of 0.392 percent, RMSE of 0.716 percent, and a maximum error of 0.945 percent, with a computation time of 4.839 s. Study [71] combined a realistic SOC model with a degradation model in the vehicle routing problem, considering energy use during driving and charging and employing a genetic algorithm. Study [32] presented a BMS accelerator implemented on FPGA with an LSTM network, achieving an RMSE of 0.3438 in training and 0.3681 in validation. Study [97] combined fast charging and thermal management using reinforcement learning, obtaining optimal strategies in less than one second and maintaining a core temperature below 33 degrees Celsius compared with about 40 degrees Celsius for predictive control, while lifetime after 1000 fast-charging cycles increased by up to two years. Study [85] compared KNN, SVM, decision tree, and Random Forest algorithms, with KNN and decision tree achieving the lowest errors.

The concept of a digital twin and implementations under resource-constrained conditions bring solutions closer to practical application. Study [43] proposed a digital twin with incremental learning, in which SOC is estimated in the vehicle using a Kalman filter and SOH in the cloud, achieving an MSE of 0.022 for SOH. Study [56] showed that a quantized one-dimensional CNN implemented on STM32 microcontrollers achieved an RMSE of 2.33 percent and MAE of 1.62 percent while reducing flash memory usage to 2.89 kilobytes. Study [57] developed an LSTM model for the Nissan Leaf, achieving an RMSE of 0.0239 and MAE of 0.0202 on the first dataset and an RMSE of 0.0449 and MAE of 0.0412 on the second. Study [61] developed a hybrid OCV model optimized by a genetic algorithm for vehicle-to-grid services, improving SOC estimation by 10 percent and increasing aggregator profit by 445 USD in voltage regulation and by 45 USD in frequency regulation while maintaining grid stability in the range of 0.9 to 1.0 p.u. Study [92] presented a fast SOH estimator based on a single-layer network using parameters from the Randle model, achieving an MSE of 1.729 percent with a deviation of 0.147 percent and a measurement time of about 30 s. Study [64] introduced a differentially tuned random forest for SOC estimation, which reduced error across the entire range without the need for preprocessing filters. Study [20] built a BP neural network for SOC estimation during charging, maintaining low errors across a wide temperature range through feature engineering and historical data. Study [96] combined a seahorse optimizer with an extended residual convolutional neural network in a hybrid liquid cooling system to improve BTMS reliability. Study [72] used incremental capacity analysis for SOH prediction on BMW i3 data and achieved an RMSE of approximately 2 percent, supporting early degradation detection. Study [73] proposed a feature selection strategy for predicting cyclic capacity loss, improving accuracy by at least 9 percent for LASSO and by 44 to 52 percent for RF, GPR, and XGBoost. Study [70] combined a particle filter, a quantum genetic algorithm, and a GRNN network, demonstrating high accuracy with low computational requirements in SOC estimation. Study [102] applied an LSTM network to predict remaining useful life using two years of data, achieving an MSE of approximately 1 percent. Study [36] described a conceptual monitoring system in which parameter optimization and LSTM are used for simultaneous estimation of SOC, SOH, and RUL. For Algorithms and Techniques, a summary by subcategory is presented in

Table 7. Algorithms and Techniques—summary by subcategory.

Subcategory	Representative Examples	Key Metrics/Results (Short)	Literature
Genetic Algorithms	Dynamic Kalman network + GA for SOC estimation; Quantum GA with particle filter and GRNN; GA-optimized OCV model for V2G	SOC error ≈ 0.1529%, MSE ≈ 0.0604; QGA + PF + GRNN noted for high accuracy and low computation; SOC improvement ≈ 10% and aggregator profit + 445 USD (voltage regulation)	[24,61,70]
Regression Analysis	Feature selection strategy (LASSO, RF, GPR, XGBoost); DSA-optimized Random Forest; Dual Gaussian process regression; XGBoost–Random Forest fusion	Accuracy gain ≥ 9% (LASSO) and 44–52% for RF/GPR/XGBoost; improved SOC estimation across full range without preprocessing; MAE < 1.3%, RUL error < 2 cycles; SOC accuracy 97.6%, MSE 1.3–1.5%	[23,64,73,88]
Support Vector Machines	DCRNN + SVM-RFE for SOH; DKNN + SVM for SOC; aging analysis using M-SVM	RMSE ≈ 0.02%, MAE ≈ 0.015%, MSE ≈ 0.032%; SOC error 0.1529%, MSE 0.0604 (with GA); M-SVM outperforms ANN and linear regression under dynamic discharge	[24,75,90]
Mean Square Error	ResDCBi-GRU model; DKNN + GA; CNN-BiGRU with Bayesian optimization; NARX trained via Levenberg–Marquardt; LSTM RUL prediction	SOC RMSE 0.0873; SOH RMSE 0.1089; MSE 0.0604; MSE 1.2 × 10−5; MSE 4.6 × 10−6; MSE ≈ 1% for RUL	[10,24,51,83,93]
Kalman Filters	Cauchy robust correction Sage–Husa EKF; Unscented Kalman Filter with deep network; Kalman filter in digital twin; multi-innovation UKF for SOC/SOH	MAE 0.392%, RMSE 0.716%, MAX 0.945%; SOC error < 0.5%, RMSE 0.04%; SOH MSE 0.022 with Kalman-based digital twin; RMSE_SOC < 0.38%, RMSE_SOH < 0.002% at 25 °C	[31,41,43,55]

.

3.3. Achievements and Development Prospects of EV Battery Management Systems

In recent years, researchers have significantly improved the accuracy and efficiency of algorithms for managing the batteries of electric vehicles. Many studies have focused on State of Charge (SOC) and State of Health (SOH) estimation using advanced machine learning models. In study [10], a residual deep convolutional network combined with a bidirectional GRU unit and optimized with the orangutan algorithm achieved very low RMSE and MAE errors for both SOC and SOH. Another model employing generative adversarial networks with a triple attention mechanism significantly reduced SOH prediction errors across multiple datasets [79]. A hybrid approach combining distributed recurrent neural networks with SVM-based feature elimination achieved errors of only a few hundredths of a percent [27], while a quantile CNN/RNN network enabled precise temperature prediction with RMSE of approximately 0.66 °C [98].

Advances in physically informed models have also contributed to progress in this field. Parallel and serial configurations of PINNs capture battery dynamics more effectively than classical LSTM models [80], while combining CFD with a neural network enabled accurate prediction of maximum cell temperatures with MSE around 0.0055 and a coefficient of determination of 0.99 [95]. Other studies reported that a regularized box particle filter provided more stable real-time SOH estimation than classical particle filters [5], and that a TimesNet network combined with data filtering techniques achieved MAPE of 0.39 percent and MSE of 0.20 percent in SOH prediction [13].

Researchers have also enhanced the learning algorithms themselves. The combination of self-supervised and federated learning reduced MSE by 31 percent [77], while a fusion of XGBoost and Random Forest models achieved SOC prediction accuracy of 97.6 percent with MSE between 1.3 and 1.5 percent [23]. Further developments include a dynamic Kalman network optimized genetically and a NARX model trained using the Levenberg–Marquardt algorithm, which achieved MSE of 4.6 × 10⁻⁶ [51]. Cascaded and feedback-connected networks have also proven highly valuable, reducing convergence time and improving SOC prediction under variable load conditions [58,86].

Particular attention has been paid to hardware implementations and energy-efficient solutions. An FPGA accelerator with an LSTM implementation achieved an RMSE of about 0.34 during training and 0.37 during validation [32], while a compact one-dimensional CNN model running on a microcontroller achieved an RMSE of 2.33 percent with flash memory reduced to 2.89 kilobytes [56]. Deep LSTM networks for SOC prediction in Nissan Leaf vehicles reduced RMSE to 0.024 and MAE to 0.020 [57], while the IoTAI SOC solution combining an autoencoder and Black Widow Optimizer maintained an error range between −3.5 and +4.3 percent at 25 °C [59]. A predictive energy management system based on LSTM increased the total vehicle State of Charge by 9.49 percent within a 60 min forecast horizon [60], while quantile networks used for temperature forecasting in different vehicles achieved low MAE (approximately 0.27 °C) and covered almost half of the actual values below the median [102].

Other important achievements include advances in filtering and calibration methods. A hybrid model combining an unscented Kalman filter with a deep network achieved an SOC error below 0.5 percent and RMSE of 0.04 percent [41], while a fractional-order dynamic model coupled with a multi-innovation UKF enabled simultaneous SOC and SOH estimation with respective errors below 0.38 percent and 0.002 percent [55]. In addition, studies on battery aging and second-life applications have used TCN networks, fast-converging single-layer networks, and LSTM-based RUL models, which achieved MSE values around 1–2 percent and enabled early degradation detection [91,92,93].

In the area of optimization algorithms, significant progress has been made using genetic and heuristic approaches. A dynamic Kalman model with genetic optimization calculated SOC with an error of approximately 0.15 percent [24], while a genetically optimized OCV model improved SOC estimation accuracy by 10 percent and increased V2G service operator profits [61]. The application of a quantum genetic algorithm combined with a particle filter and GRNN network demonstrated high accuracy at low computational cost [70]. Moreover, the development of lightweight architectures such as quantized one-dimensional CNNs enabled the deployment of SOC algorithms on microcontrollers with flash memory reduced to just a few kilobytes while maintaining RMSE around 2.3 percent [56].

Another major achievement is the use of digital twin and battery-cloud concepts for remote diagnostics and management. Digital twin models combined with incremental learning allow SOC estimation in the vehicle and SOH estimation in the cloud with MSE of only 0.022 [43], while telematics and IoT-based systems enable predictive maintenance with accuracy exceeding 97 percent [48] and remote detection of thermal anomalies [52]. These advances are complemented by modern thermal management techniques: deep reinforcement learning models maintain cell core temperatures below 33 °C and shorten charging times, while optimized cooling plate designs reduce maximum temperature by more than 4 K [97,99,101].

In recent years, methods for predicting remaining useful life and analyzing battery aging have advanced rapidly. LSTM models are capable of estimating battery lifetime with errors of around 1 percent [93], while Gaussian process regression models achieve absolute errors below 1.3 percent for SOH and less than two cycles for RUL, with training times under five seconds [88]. Aging analyses have also shown that modified Support Vector Machines outperform classical neural networks and linear regression in degradation prediction [90].

The analysis of the reviewed studies indicates several key directions for future research. First, the integration of telematics data and cloud technologies with digital twins is becoming increasingly important. Such platforms enable continuous collection of operational data, edge preprocessing, and cloud sharing, which allows rapid diagnostics and personalized charging strategies [48,52]. Future developments should therefore focus on standardized data exchange protocols and privacy-preserving methods such as federated and self-supervised learning, which have reduced MSE by more than 30 percent in tests.

Second, further research on hybrid physics–data models is needed. Combining equivalent circuit models with neural networks improves model generalization and interpretability [54,78], while the inclusion of attention mechanisms and fractional filters has reduced SOC and SOH estimation errors to fractional percentages [55]. Such approaches can also better handle incomplete or noisy data through embedded physical constraints.

Optimization algorithms inspired by nature and quantum computing methods are also promising, as they are already being used to optimize model parameters, for example, the orangutan algorithm [10], GA [24], and QGA [70]. With the advancement of quantum computers, even faster tuning of complex architectures can be expected, facilitating the adaptation of models to new battery chemistries and operating conditions. At the same time, compact deep network architectures and their implementation in embedded systems will continue to develop, enabling real-time SOC and SOH prediction on board vehicles or in portable devices.

From a data quality perspective, intelligent methods for feature selection and automated feature engineering will play a major role. Results indicate that the use of quantum methods and preliminary signal transformations can improve prediction accuracy by several tens of percent [72,73]. Future research should also focus on combining heterogeneous datasets that include information from temperature, voltage, current, and environmental sensors, as well as on the use of transfer learning to adapt models across different battery types.

Another important research direction is the integration of intelligent energy and thermal management strategies with prognostic algorithms. Control models based on deep reinforcement learning already reduce charging time, lower energy consumption, and extend cell lifetime [97,101]. Future systems may integrate these algorithms with degradation prediction, allowing the charging profile to be dynamically adjusted to the current battery condition.

A further perspective concerns the circular economy and battery reuse. More accurate RUL and SOH models will make it possible to determine when a cell should be retired from primary use and repurposed for second-life applications [88,93]. Systems such as BaHeS, which combine anomaly detection with recommendations for secondary deployment of cells, already achieve high effectiveness [89,92]. The development of certification standards and automated cell classification systems will be crucial for the sustainable growth of the battery market.

In summary, future solutions for EV battery management will require the synergy of advanced machine learning algorithms, distributed computing architectures, high-quality data, and the integration of sustainability principles. Leveraging these research directions will enable the creation of safer, more efficient, and more environmentally friendly energy storage systems. For Battery Management Systems, an overview of categories and key achievements is presented in

Table 8. Battery Management Systems—overview of categories and key achievements.

Main Category (Subcategories)	Sample Works	Key Results
State Of Charge (State of Charge, States of Charges, Battery State of Charge, State of Charge Estimation)	[10,23,24,39,51,57]	The ResDCBiGRU model achieves RMSE_SOC ≈ 0.0873 and MAE ≈ 0.0866. The XGBoost–Random Forest fusion attains SOC accuracy of approximately 97.6% with MSE between 1.3% and 1.5%. The dynamic Kalman network with genetic optimization (DKNN + GA) reduces SOC error to 0.1529%. The NARX model during charging achieves RMSE_SOC of 0.5%. The Levenberg–Marquardt-trained NARX model reaches MSE of 4.6 × 10−6. The deep LSTM network for the Nissan Leaf yields RMSE of 0.0239 and MAE of 0.0202.
State Of Health (Battery Health, Battery Degradation, State of Health)	[27,43,49,54,72,79,88,90,91,92,98]	The GAN model with triple attention significantly reduces MAE, MSE, and MAPE. The DCRNN combined with SVM-RFE achieves RMSE ≈ 0.02%, MAE ≈ 0.015%. The quantile CNN/RNN attains RMSE_T ≈ 0.66 °C, with the 99th quantile covering 98.87% of observations. The digital twin with incremental learning reaches MSE_SOH ≈ 0.022. Incremental capacity analysis yields RMSE_SOH ≈ 2%. The modified SVM provides the most accurate aging prediction. The hybrid VAR + LSTM model improves accuracy (no numerical data reported). Feature selection increases accuracy by ≥9% for LASSO and by 44–52% for RF, GPR and XGBoost. The TCN model for second-life batteries achieves MSE < 1%. The fast ANN model records MSE ≈ 1.7% with a measurement time of 30 s. The dual GPR model achieves MAE < 1.3% and RUL error < 2 cycles.
Thermal Management (Battery Temperature, Temperature Control, Thermal Management (electronics), Battery Thermal Managements, Thermal Management Systems, Cooling)	[31,95,97,98,99,100,101,102]	CFD combined with ANN predicts maximum cell temperatures with MSE ≈ 0.00552 and R2 ≈ 0.99. Deep reinforcement learning for charging maintains core temperature below 33 °C and achieves simulation time under 1 s. Quantile CNN/RNN yields RMSE ≈ 0.66 °C and R2 ≈ 0.84. Optimization of liquid cooling plates reduces maximum temperature by 4.87 K and temperature gradient by 22.2%. The CRC-SHEKF filter for SOC estimation achieves MAE 0.392% and RMSE 0.716%. BiLSTM predicts pack temperature with MAE ≈ 2.92 °C on the test set. The double-DQN + GRU algorithm reduces energy consumption by more than 6.7% during aggressive driving. Cross-domain quantile CNN achieves MAE ≈ 0.27 °C.

. For Computational Intelligence, an overview of categories and key achievements is presented in

Table 9. Computational Intelligence–overview of categories and key achievements.

Main Category (Subcategories)	Sample Works	Key Results
AI/ML Core (Artificial Intelligence, Machine Learning, Learning Systems)	[12,19,48,52,76,77,89]	Federated learning with CAFS improves SOH estimation accuracy by 20.18%. The IoT–Fog–Cloud platform concept organizes ML model selection and integrates telemetry data. Explainable digital twins combining RBF, RF, FNN, CNN, LSTM, SVR, SVM, and XGBoost enhance model interpretability. The battery cloud enables remote data collection and anomaly detection. Telematics-based systems predict faults with accuracy exceeding 97%. The combination of self-supervised and federated learning reduces MSE by 31%. The BaHeS system achieves 93% accuracy and improves SOH estimation by 18%.
Neural Networks (Neural Networks, Deep Learning, Long Short-term Memory, Convolutional Neural Network)	[10,32,56,57,59,79,82,83,84,95,98,102]	ResDCBiGRU achieves RMSE_SOC ≈ 0.0873 and RMSE_SOH ≈ 0.1089. The GAN model with triple attention reduces SOH prediction errors. The ANN + CFD model predicts temperature with MSE ≈ 0.00552. The MMFEN (CNN + attention) network attains RMSE of 1.21% and MAPE of 0.99%. The FPGA-implemented LSTM reaches RMSE of 0.3438 in training and 0.3681 in validation. The CNN–BiGRU model achieves MSE of 1.2 × 10−5, MAE of 2.08%, and RMSE of 2.516%. The wavelet CNN–LSTM network improves accuracy compared with the baseline LSTM. The quantized 1D CNN achieves RMSE of 2.33% and MAE of 1.62%. The LSTM model for the Nissan Leaf attains RMSE of 0.0239 and MAE of 0.0202. The IoTAI-SOC model (LSTM–SAE with Black Widow Optimization) maintains an error range between −3.5% and +4.3%. The cross-domain quantile CNN achieves MAE ≈ 0.27 °C.
Algorithms and Techniques (Genetic Algorithms, Regression Analysis, Support Vector Machines, Mean Square Error, Kalman Filters)	[10,23,24,41,43,51,55,64,70,73,83,88,90,93,104]	GA-optimized DKNN achieves SOC error ≈ 0.1529% and MSE ≈ 0.0604. The QGA + PF + GRNN approach provides high accuracy with low computational cost. The GA-optimized OCV model improves SOC estimation by ≈ 10% and increases profit by 445 USD. Feature selection enhances accuracy by ≥9% for LASSO and by 44–52% for RF, GPR, and XGBoost. The DSA-RFR model reduces error across the full SOC range. The dual GPR model achieves MAE < 1.3% and RUL error < 2 cycles. The XGBoost–RF fusion attains SOC accuracy of 97.6%. The DCRNN combined with SVM-RFE achieves RMSE ≈ 0.02%. The modified SVM provides the best performance in aging analysis. The ResDCBiGRU model yields SOC RMSE of 0.0873. The CNN–BiGRU model achieves MSE of 1.2 × 10−5. The Levenberg–Marquardt-trained NARX model reaches MSE of 4.6 × 10−6. The LSTM model for RUL prediction attains MSE ≈ 1%. The CRC-SHEKF filter achieves MAE of 0.392% and RMSE of 0.716%. The UKF combined with a deep network achieves SOC error < 0.5% and RMSE of 0.04%. The Kalman-based digital twin reaches MSE of 0.022. The MIUKF model achieves RMSE_SOC < 0.38% and RMSE_SOH < 0.002%.

.

The Tables above show that the development of modern Battery Management Systems for electric vehicles is based on three interrelated areas: State of Charge estimation, State of Health monitoring, and thermal management. In the first area, the application of deep networks, hybrid models, and optimization algorithms allows the prediction error of SOC to be reduced to hundredths of a percent. In the second area, through the use of generative and recurrent networks as well as feature selection techniques, high accuracy of SOH estimation is achieved, which is essential for both diagnostics and second-life battery planning. In the third area, the integration of CFD simulations, reinforcement algorithms, and cooling system design optimization effectively limits cell temperature rise and reduces energy consumption.

At the same time, it is clear that these advances would not have been possible without the use of advanced intelligent computing tools. Federated learning, cloud-based telematics, and digital twin concepts enable the secure integration of data from multiple vehicles and provide up-to-date real-time forecasts. Deep neural networks, including CNN–LSTM hybrids, lightweight microcontroller-based architectures, and various LSTM configurations, deliver high performance in both battery parameter prediction and energy management. These are complemented by algorithms and techniques such as genetic optimization, regression methods, Support Vector Machines, and Kalman filters, which provide the mathematical foundation for the above solutions.

In summary, the effectiveness of modern BMS results from the harmonious combination of three elements: precise SOC and SOH estimation algorithms, efficient thermal management, and advanced computational infrastructure. Such an integrated approach not only enhances the safety and performance of electric vehicles but also lays the groundwork for future applications such as autonomous charging strategies, intelligent load balancing and the circular economy of batteries.

3.4. Reversible Degradation and Recovery-Aware Lifetime Modeling

Most State of Health (SOH) prediction approaches reviewed in this study implicitly assume that battery aging is a predominantly monotonic and irreversible process, typically represented by capacity fade, internal resistance growth, or remaining useful life (RUL) trajectories learned from historical data [28,37,53]. Such formulations are well suited for standard laboratory cycling protocols and long-term degradation assessment; however, they may inadequately reflect the behavior of electrochemical systems operating under highly dynamic load profiles. In practice, batteries may exhibit reversible degradation mechanisms, manifested as temporary performance losses followed by partial recovery after rest periods or changes in operating conditions. These recovery phenomena affect voltage response, apparent capacity, and polarization-related indicators that are frequently used as inputs or targets in data-driven SOH models [21,72,73].

The presence of reversible degradation becomes particularly relevant in grid-interactive operating regimes, including frequent vehicle-to-grid (V2G) charge–discharge cycles and shallow cycling scenarios. Under such conditions, repeated current reversals and relaxation intervals may induce short-term voltage recovery effects that interrupt otherwise monotonic aging trajectories. As a consequence, purely irreversible aging models may misinterpret recovery-driven signal changes either as permanent health improvement or as noise, which can lead to biased lifetime prediction and reduced robustness of SOH estimation in real-world Battery Management Systems [19,30,43].

Recent advances in lifetime modeling within the electrochemical energy systems literature have demonstrated that separating reversible and irreversible aging components improves the representation of complex ageing kinetics. Recovery-aware formulations introduce an explicit distinction between long-term degradation trends and short-term voltage recovery dynamics, thereby avoiding the attribution of reversible phenomena to irreversible health changes. Although most AI-based SOH models reviewed in this work do not explicitly implement such separation, several hybrid and digital-twin-oriented studies already point in this direction by combining data-driven learning with physics-inspired descriptors, incremental or differential voltage analysis, and adaptive filtering techniques [19,21,30,52,72]. These approaches implicitly acknowledge that observable health indicators may deviate from monotonic behavior due to reversible electrochemical processes.

From a data-driven modeling perspective, recovery-aware lifetime concepts can be integrated into existing SOH prediction frameworks without altering their core architecture. For example, sequence models such as LSTM- or GRU-based predictors may be augmented with additional latent states representing relaxation or recovery dynamics, while hybrid physics–AI models may treat recovery-sensitive voltage features separately from irreversible degradation indicators [10,19,30]. Similarly, digital twin frameworks and cloud-based BMS architectures provide a natural environment for continuously updating health models and correcting short-term recovery effects using operational data streams [19,43,52].

In summary, extending SOH prediction beyond strictly irreversible aging assumptions and accounting for reversible degradation mechanisms represents an important refinement of lifetime modeling for modern xEV applications. This perspective is particularly relevant for V2G-enabled and energy-system-integrated operation, where dynamic cycling profiles amplify recovery effects. Incorporating recovery-aware concepts into AI-based SOH estimation improves interpretability, reduces prediction bias, and enhances the reliability of lifetime forecasts under realistic operating conditions.

3.5. Prospects for Cutting-Edge Algorithms

Recent developments indicate that the next generation of Battery Management Systems will increasingly rely on algorithms that do not only estimate battery states but also learn control policies and operate within continuously updated digital representations of the battery, vehicle, and grid context. Two particularly dynamic directions in 2023–2025 are reinforcement learning for adaptive charging and thermal and electrical co-control and digital twins as an architectural paradigm for online model updating, integration, and deployment. These approaches are relevant to the energy-systems perspective adopted in this review because they enable actionable flexibility, including V2G and PV-coupled charging, while explicitly accounting for operational constraints, safety margins, and lifetime impacts.

The first cutting-edge direction is the application of reinforcement learning to dynamic charging strategies, including joint optimization of charging profiles and thermal constraints. Reinforcement learning formulations can learn policies that map measured states, or their estimates, to control actions such as charging current and cooling actuation, thereby addressing nonlinear dynamics and time-varying constraints more flexibly than fixed-rule strategies. The reviewed corpus already reflects this trend, including reinforcement-learning-based solutions that report high decision accuracy or control effectiveness and reinforcement-learning frameworks that significantly reduce runtime relative to classical optimization baselines in fast-charging scenarios while maintaining safe temperature limits [46,97]. From a BMS engineering perspective, the key prospective value of reinforcement learning is the ability to co-optimize competing objectives, charging time, thermal safety, and degradation awareness under realistic disturbances. At the same time, this direction requires careful treatment of constraint satisfaction, safety guarantees, and generalization beyond the training distribution, which remain critical barriers for automotive deployment.

A second major direction is the maturation of digital twin concepts, increasingly used to couple state estimation, diagnostics, and lifecycle-aware decision-making within connected BMS architectures. Digital twins can support continuous updating of SOC and SOH models from operational data, facilitate explainability and monitoring, and shift computationally intensive components to cloud-based pipelines when appropriate. Within the reviewed studies, digital-twin-oriented frameworks and cloud BMS concepts already demonstrate incremental updating, hybrid estimation pipelines, and additional analytics such as anomaly detection, illustrating the feasibility of scalable, continuously evolving state estimation in real operating conditions [19,43,52]. Extending this paradigm toward thermal management is a natural next step. A thermal digital twin can integrate temperature prediction, pack-level thermal behavior, and safety-relevant constraints into a unified supervisory layer, enabling more proactive thermal and electrical co-management rather than reactive threshold control. Moreover, surrogate modeling approaches used for thermal design-space exploration demonstrate how high-fidelity thermal knowledge can be approximated by learned models to accelerate engineering workflows, which can be leveraged as the modeling backbone of thermal twins [99].

In forward-looking BMS architectures, reinforcement learning and digital twins should be viewed as complementary rather than competing paradigms. Digital twins provide a controlled environment for policy learning, validation, and stress testing, while reinforcement learning can serve as an adaptive control layer operating on state estimates and twin-informed constraints. This integration supports the review’s energy-systems framing by enabling dispatchable flexibility, for example, V2G services, with explicit safety and lifetime awareness, but it also highlights open research needs in standardized evaluation protocols, transparent reporting of computational requirements, and robust deployment under connectivity, cybersecurity, and functional safety constraints [19,52]. Overall, reinforcement learning for adaptive control and digital twins for continuous, system-level integration represent two high-impact algorithmic prospects that strengthen the practical relevance and technological foresight of data-driven BMS research in the 2023–2025 horizon [46,97].

3.6. Cross-Task Synergy Analysis and Emerging Model Families for Holistic BMS Design

A holistic BMS perspective requires treating SOC estimation, SOH prediction, and thermal management as coupled tasks rather than independent modules. In real xEV operation, SOC estimation errors propagate into SOH inference through biased cycle counting, misestimated depth of discharge, and miscalibrated operating constraints, while temperature dynamics modulate both electrochemical response and aging kinetics, thereby affecting the stability and transferability of data-driven estimators across duty cycles [21,55,97]. Consequently, cross-task integration is not only beneficial but it is often necessary to avoid inconsistent state trajectories where a locally accurate estimator in one task induces systematic bias or constraint violations in another task, particularly under fast charging, V2G, and highly transient conditions [69,97].

From an algorithmic viewpoint, cross-task synergy most frequently emerges through shared representations and shared constraints. Sequence models used for SOC and SOH can exploit common temporal structure in voltage, current, and temperature streams, but multi-task learning must address potential task conflicts; for example, improving short-window SOC accuracy can increase sensitivity to noise that degrades longer horizon SOH stability, especially when training labels mix irreversible aging with short-term recovery effects [10,21,72]. Hybrid formulations mitigate these conflicts by separating fast electrical dynamics from slower degradation components using physically grounded backbones, for example, Kalman family estimators and equivalent circuit models coupled with learned correction terms, which support consistent inference when states are jointly constrained by voltage response and temperature-dependent parameters [55,67,68]. This also provides a practical route to joint estimation, where SOC and SOH are co-inferred under temperature-dependent constraints rather than inferred sequentially with error accumulation [55,68].

Recent advanced architectures in the reviewed corpus can be positioned in a comparative manner by focusing on what they add beyond classical CNN and LSTM baselines, namely stronger temporal generalization, richer attention-based feature selection, and improved robustness under heterogeneity. Temporal models and clustering-based pipelines, including TimesNet-based frameworks, aim to improve SOH-related pattern extraction under nonstationary signals and partial charge information, which is relevant for fleet operation and irregular usage regimes [13]. Attention-augmented deep models and multiscale feature extractors target improved representation of multi-resolution degradation signatures and can reduce dependence on handcrafted features, as illustrated by multi-head convolutional attention mechanisms and wavelet-enriched hybrid architectures [82,84]. Data augmentation and adversarial learning approaches, including GAN-based designs with attention mechanisms, represent another emerging family that seeks improved robustness when labeled aging data are limited or unevenly distributed across operating conditions [79]. Physics-informed neural networks further extend this landscape by embedding constraints into the learning objective, improving interpretability and extrapolation potential compared with purely black-box predictors, which is particularly relevant when cross-task consistency is required under varying thermal states [80].

At the system level, cross-task integration is increasingly expressed through connected architectures rather than single models. Digital twin and cloud BMS concepts provide an integration layer where SOC and SOH models can be updated incrementally from operational data and where thermal diagnostics and anomaly detection can be co-deployed as part of a unified monitoring pipeline [19,43,52]. This architecture supports cross-task consistency by enabling shared data curation, shared validation logic, and continual recalibration, which is difficult to achieve when each task is implemented as a static, isolated estimator [19,52]. In parallel, reinforcement learning extends cross-task integration from estimation to decision-making by learning policies that jointly manage charging and thermal constraints, thereby operationalizing synergy at the control layer rather than only at the prediction layer, with reported advantages in runtime and safety-relevant temperature outcomes under fast-charging scenarios [46,97].

Finally, the feasibility of cross-model quantitative comparison across emerging architectures is fundamentally limited by heterogeneity in datasets, operating conditions, temperature ranges, and evaluation protocols across studies. Therefore, rather than implying strict rank ordering, this review emphasizes comparative interpretation along consistent engineering axes, including accuracy metrics, computational cost evidence, deployment feasibility, and robustness under non-ideal profiles, while directing readers to the source publications for exact dataset-specific performance values [17,32,56]. Within this framing, the main holistic design trend is clear; future BMS research is moving from single-task estimators toward integrated, multimodal, and lifecycle-aware pipelines in which SOC, SOH, and thermal states are jointly constrained, continuously updated, and increasingly connected to adaptive control strategies [19,52,97].

3.7. Conclusions of the Review

The synthesis of findings from the preceding sections indicates that solutions combining physical knowledge with machine- and deep learning achieve the highest effectiveness in battery state estimation and thermal management. For SOC estimation, hybrid configurations that pair Kalman observers with lightweight LSTM or feedforward networks are superior because they stabilize estimation in the presence of noise and drift and transfer well across driving profiles and load regimes. For SOH estimation, sequential models with attention mechanisms and feature selection perform best, especially when inputs include differential and incremental features from charging cycles together with hysteresis indicators and equivalent circuit parameters. In thermal topics, the combination of CNN- and RNN-based quantile prediction with reinforcement learning control policies is advantageous, enabling maintenance of core temperatures below safety thresholds with short decision times. The highest-quality results are obtained when data streams combine electrical and thermal signals with telematics and environmental parameters and when data preparation includes consistent windowing, channel synchronization, filtering, and normalization. Quality assessment still relies mainly on RMSE and MAE, less often on uncertainty and energy consistency measures, and reporting of computational cost and hardware requirements remains inconsistent.

The first research question concerned model families and hybrid configurations and the conditions under which they maintain an advantage. In SOC estimation, systems that combine Kalman observers with LSTM or feedforward networks are most effective, and their advantage holds in typical operating ranges, with C-rates of about 0.5 to 2 and cell and ambient temperatures of about 15 to 45 degrees Celsius, provided that input signals are consistently normalized and filtered. In SOH estimation, sequential architectures with attention and feature selection dominate, particularly in combination with differential and incremental features from charging cycles and with equivalent circuit parameters. In thermal management, the mix of CNN- and RNN-based quantile regression with reinforcement learning control policies simultaneously limits temperature exceedance risk and computational cost. In resource-constrained environments, compact models, including quantized one-dimensional CNNs and reduced LSTMs, retain an advantage by providing a reasonable tradeoff between error, memory footprint, and latency.

The second research question was related to data types and to preparation and normalization procedures, as well as to feature sets associated with the best results. Dominant solutions combine voltage, current, and temperature with telematics and environmental parameters such as speed, elevation profile, and ambient temperature. In fleet data, window lengths of 30 to 120 s with sampling at 1 to 10 hertz are most common, whereas laboratory frequencies are higher. Preparation procedures include channel alignment and synchronization, denoising with bandpass filtering or a Savitzky–Golay filter, resampling to a uniform grid, and normalization, most often min–max or z-score. For SOC, key features include temporal derivatives and aggregates of voltage and current, charge counters, dV/dt derivatives, and relaxation features after load removal. For SOH, ICA, and DVA features, positions and amplitudes of differential peaks, hysteresis indicators, degradation rates, and equivalent circuit parameters provide an advantage. In thermal tasks, geometry and boundary condition representations, sensor placement, and coolant flow descriptors are important, as are features linking load history with local temperature extremes. The best configurations combine these sets with attention mechanisms and with physics embedded in the loss function, strengthening transferability across vehicles and climates.

The third research question addressed the reporting of uncertainty, verification and validation, and computational costs and hardware requirements, as well as the metrics regarded as standard. The share of studies reporting uncertainty is growing, particularly with probabilistic models and quantile regression, yet point estimates still dominate. Validation most often relies on train, validation and test splits, or on cross-validation, while transfer validation across fleets, vehicles, and climates is less common. Information on computational cost and hardware requirements is reported with varying detail, with more complete descriptions appearing in publications that include implementations on programmable devices and microcontrollers, where the number of parameters, memory footprint, and training and inference times are documented. In practice, RMSE and MAE remain the standard for SOC and SOH, with MAPE and the coefficient of determination used as complements, while in the thermal area RMSE and maximum error in degrees Celsius are most frequently reported; for surrogate models, the acceleration relative to baseline methods and run times are emphasized. There is a need for further standardization of reporting to include task-appropriate error metrics, verification and validation elements, computing platform characteristics, and at least one uncertainty or coverage measure.

Taken together, the conclusions point to a set of practices that are both effective and feasible. For SOC and SOH estimation, hybrid solutions with a physical prior and a learning layer should be preferred, built on carefully prepared data streams. In thermal tasks, a tandem of quantile prediction and reinforcement-learning-based control is particularly valuable. Across all tasks, it is necessary to maintain consistent data preparation procedures, window choices and normalization, and to extract features tied to the underlying physics, differential features for SOH, and relaxation features for SOC, with geometry and boundary conditions accounted for in the thermal component. For comparability and engineering usefulness, unified reporting of uncertainty, validation procedures, platform parameters, and computational costs is essential, as it will enable more reliable comparisons and support deployment planning.

The energy-systems framing adopted in this review (V2G and PV-coupled charging) can be strengthened by explicitly clarifying how AI-enabled BMS functions translate into measurable energy indicators. In grid-interactive operation, the BMS does not modify system-level metrics directly; instead, it enables controllable flexibility by providing reliable state awareness (SOC, SOH, and temperature constraints) and by supporting higher-level charging/discharging decisions that must remain safe and lifetime-aware under dynamic duty cycles [19,30,52]. Consequently, the causal pathway from algorithms to energy outcomes should be interpreted as a multi-stage coupling: state estimation and constraint inference enable feasible dispatch decisions, which in turn affect net-load profiles and renewable energy utilization at the system level [1,4,5].

A representative mechanism for reducing the peak–valley difference emerges in V2G-enabled scheduling. Accurate SOC/SOH estimation provides bounded flexibility and reduces uncertainty in available capacity, which is essential for bidirectional operation and for avoiding constraint violations during dispatch [19,30,52]. On this basis, an energy management layer can select charging and discharging setpoints that shift demand away from peak hours and optionally inject power during high-load periods, while incorporating degradation-aware weighting to prevent short-term grid services from causing disproportionate lifetime cost [61,69]. This sequence of actions yields a flatter net-load profile at the point of common coupling, which corresponds to the reported mechanism of peak–valley reduction through controlled V2G participation rather than through estimation accuracy alone [1].

A second case-level mechanism concerns increasing photovoltaic self-consumption through alignment of charging demand with PV generation. Studies on PV–EV matching and sizing indicate that self-consumption improves when charging is scheduled to coincide with PV surplus rather than being driven solely by user arrival times or tariff signals [4,5]. In this setting, BMS-supported prediction and constraint handling provide the feasibility layer for such alignment by maintaining mobility-related SOC requirements and by avoiding excessive cycling that would undermine lifetime objectives [2]. Microgrid-oriented energy management that incorporates PV-related information has been shown to improve operational battery-state outcomes (e.g., increasing aggregate SOC), which is consistent with shifting charging toward locally generated energy and reducing grid import during non-PV periods [60]. Importantly, the energy indicator improvement arises from the coupling of prediction, constraints, and scheduling, not from any single algorithmic component in isolation [4,60].

These evidence chains clarify how BMS algorithms contribute to energy-system indicators through constraint-aware flexibility provision and dispatch enablement. Making this linkage explicit supports the review’s energy-systems perspective and demonstrates why reporting only estimation error is insufficient when the intended application includes V2G services and PV-integrated charging strategies [1,19,52].

While many studies emphasize accuracy metrics (e.g., MAE/RMSE/MAPE), practical deployment of AI-based BMS solutions also depends on the accuracy–computational cost balance, including model footprint, training effort, inference latency, and hardware feasibility. Representative examples in the reviewed corpus illustrate several recurring engineering trade-offs. First, some works explicitly minimize model size to enable embedded inference; for instance, a lightweight network with an attention mechanism is reported to contain only 1713 parameters while achieving an RMSE of 1.23%, indicating a clear pathway toward microcontroller/ECU-oriented deployment through footprint reduction [17]. Second, training efficiency improvements are addressed by ensemble-type strategies; an ensemble of homogeneous LSTM models is reported to maintain an MAE of approximately 1.4% while shortening training time by about 2.6–3.5 times, suggesting that practical maintainability and retraining cost can be improved without sacrificing accuracy [26]. Third, hardware-aware implementations demonstrate that sequence models can be accelerated on dedicated platforms; an FPGA accelerator for an LSTM-based estimator is reported, with RMSE values of 0.3438 in training and 0.3681 in validation, providing evidence that real-time inference constraints can be addressed through specialized computation rather than purely architectural simplification [32].

A parallel line of work addresses computational feasibility at the system-architecture level. Digital twin and cloud-connected BMS concepts aim to shift computational load off-board and enable model updates based on operational data. For example, a digital-twin-oriented BMS is described where measurement data are transmitted to the cloud, SOH is estimated using incremental learning, and SOC is determined using a Kalman filter-based component; the study reports an MSE of approximately 0.022 and emphasizes reduced local hardware requirements, which illustrates an explicit accuracy–infrastructure trade-off (edge constraints vs. connectivity and off-board processing) [43]. Related cloud-BMS concepts combine SOC/SOH analytics with additional functions such as anomaly detection and signal-based health analyses, further motivating the view that deployment feasibility is not only model-specific but also architecture-dependent [52]. These approaches also align with broader digital twin frameworks discussed in the corpus, where interpretability and continuous updating are treated as part of the engineering value proposition [19].

Embedded feasibility is also demonstrated directly through implementation-oriented studies. A practical example is the deployment of a one-dimensional CNN and GRU on STM32 microcontrollers, where quantization is applied to reduce model size while maintaining an RMSE of 2.33%, explicitly illustrating the performance–compression trade-off required for resource-constrained BMS hardware [56]. In the thermal-management and charging domain, computational cost considerations become even more visible in runtime comparisons: reinforcement-learning-based joint fast-charging and thermal-control strategies are reported to be found in less than one second, whereas model predictive control requires over eighty minutes under the compared setup; the same study reports maintaining a maximum core temperature below 33 °C, thereby linking computational feasibility with safety-relevant thermal constraints in online control [97]. Finally, thermal management research also uses surrogate modeling to reduce computational burden in design-space exploration; deep encoder–decoder surrogate models are employed to support multi-objective optimization of liquid-cooled plate structures, reporting reductions such as a 4.87 K decrease in maximum temperature, a 5.1 K decrease in temperature gradient (22.2%), and a 7.93 Pa decrease in hydraulic losses (9.0%), demonstrating how computational methods can substitute expensive high-fidelity simulations in engineering design workflows [99].

Overall, these examples show that accuracy alone is insufficient for engineering selection of SOC/SOH/thermal solutions. The reviewed literature indicates three practical routes to achieve deployable performance: (i) footprint minimization for embedded inference, (ii) hardware acceleration for real-time computation, and (iii) system-level architectures (digital twins/cloud BMS) that redistribute computational effort. At the same time, cost reporting remains inconsistent across publications, and future work would benefit from more standardized disclosure of platform specifications and runtime indicators to enable fair cross-model comparisons.

4. Statistical Overview

The corpus includes 95 publications, of which 25 fall within the years 2020–2022 and 70 within 2023–2025, corresponding to 26.3% and 73.7%, respectively. The structure of the corpus remains stable between the two periods, as confirmed by independence tests calculated for the categorical distributions: document type, χ² = 5.04, df = 2, p = 0.08; BMS categories, χ² = 0.28, df = 2, p = 0.87; method classes, χ² = 0.24, df = 2, p = 0.89; and methodological approaches, χ² = 3.25, df = 3, p = 0.35. In other words, while the research volume is growing rapidly, the proportions between the main categories do not show statistically significant changes. Publications by year in all categories are presented in

Table 10. Publications by year in all categories.

Name	2020–2022	2023–2025	All Years	Share [%]
Total	25	70	95	100.0
Document Type
Article	10	46	56	58.95
Conference Paper	12	19	31	32.63
Other	3	5	8	8.42
Battery Management Systems
State Of Charge	14	45	59	62.11
State Of Health	13	33	46	48.42
Thermal Management	2	6	8	8.42
Computational Intelligence
AI/ML Core	12	39	51	53.68
Neural Networks	14	38	52	54.74
Algorithms and Techniques	10	26	36	37.89
Research Methodology
Experiment	21	59	80	84.21
Literature Analysis	3	7	10	10.53
Case Study	0	8	8	8.42
Conceptual	22	53	75	78.95

.

The data indicate not only an increase in research volume but also relationships between categories that are significant for method selection and workflow design. The ratio of SOC to SOH studies in the entire corpus is approximately 1.28 to 1, reflecting a greater number of investigations on State of Charge estimation than on State of Health estimation. The thermal topic remains niche and is represented about seven times less frequently than SOC. This structure suggests that SOC and SOH research form the core of BMS practice, while the thermal component requires targeted strengthening, particularly in areas where safety and battery lifespan depend on accurate temperature forecasts.

In terms of document types, journal articles dominate, accounting for 56 publications, or about 59% of the total corpus. Conference papers represent 31 publications, about 33%, and the Other category includes 8 publications, about 8%. Comparing the two periods, the share of journal articles increased from 10 of 25 publications in 2020–2022 (40%) to 46 of 70 in 2023–2025 (65.7%). The share of conference papers decreased from 48% to 27.1%, while the share of the Other category declined from 12% to 7.1%. The interpretation is straightforward: research outputs increasingly culminate in journal publications, while conferences continue to play a significant complementary role. The distribution of document types indicates a maturing publication stream. Journal articles now constitute nearly three fifths of all works, and their share has grown in the more recent subperiod, whereas the share of conference materials has declined. Such a shift typically accompanies a transition from the exploratory phase to the consolidation phase, characterized by methodological refinement and validation of results. The Other category serves as a supplement and does not materially affect the overall picture. Document types in the corpus, including Article, Conference Paper, and Other, for the periods 2020–2022 and 2023–2025 are presented in Figure 2.

In the breakdown by BMS tasks, SOC estimation prevails with 59 publications, representing 62%, followed by SOH estimation with 46 publications, or about 48.4%, while the thermal topic is less represented, with 8 publications, or about 8.4%. Comparing the two periods, SOC increased from 14 of 25 publications (56%) to 45 of 70 (64.3%), SOH from 13 of 25 (52%) to 33 of 70 (47.1%), and thermal management from 2 to 6 publications, corresponding to 8% and 8.6%, respectively. This indicates the persistent dominance of SOC and SOH topics and a continuing shortage of studies focused on thermal management.

Within the classes of Computational Intelligence methods, parallel development can be observed without displacement effects. The shares of AI and ML Core and Neural Networks are similar across the entire corpus, at approximately 54% each, while Algorithms and Techniques remain around 38%. Comparison between periods suggests a mild shift toward AI and ML Core and a slight flattening of the network share; however, the differences are small, reinforcing the conclusion that network-based and classical methods are used concurrently. In practice, this means that hybrid configurations combining observers and physical priors with learning will remain the dominant implementation pattern. Battery Management Systems, State of Charge, State of Health, and thermal management for the periods 2020–2022 and 2023–2025 are presented in Figure 3.

All three method families are growing in parallel. AI and ML Core appear in 51 publications, accounting for approximately 53.7%, Neural Networks in 52 publications, about 54.7%, and Algorithms and Techniques in 36 publications, about 37.9%. In the first period, the shares were 48%, 56%, and 40%, respectively, and in the second period, 55.7%, 54.3%, and 37.1%. The data indicate that research practice does not abandon classical methods in favor of neural networks but rather develops both directions simultaneously, which supports the prevalence of hybrid configurations. Computational Intelligence classes, AI and ML Core, Neural Networks, and Algorithms and Techniques for the periods 2020–2022 and 2023–2025 are presented in Figure 4.

Experimental and conceptual works dominate among methodological approaches. Experiments account for 80 publications, approximately 84.2%, conceptual studies for 75 publications, about 79%, literature analyses for 10 publications, about 10.5%, and case studies for 8 publications, about 8.4%. Between the two periods, the number of experimental works increased from 21 to 59 publications, with shares remaining similar at 84.0% and 84.3%, respectively. Conceptual works increased from 22 to 53 publications, with shares of 88.0% and 75.7%, while case studies appeared only in the 2023–2025 period, accounting for 11.4%.

This profile confirms the simultaneous presence of data-based validation and the conceptual foundation required for further methodological refinements. Experimental and conceptual approaches dominate, which is characteristic of a field in a phase of rapid development. Experiments indicate the growing importance of validation using real data, while conceptual studies provide the framework for subsequent iterations of methods. The emergence of case studies in the more recent subperiod signals the beginning of translation into implementation scenarios. Literature analyses maintain an organizing role and help identify research gaps, including the gap in the thermal domain. Research methodology in the corpus, including Experiment, Literature Analysis, Case Study, and Conceptual for the periods 2020–2022 and 2023–2025, is presented in Figure 5.

In summary, the entire corpus is growing rapidly while the proportions between the main categories remain essentially unchanged. Journal articles have become the dominant form of dissemination, SOC and SOH continue to serve as leading topics, and the thermal component requires further strengthening. The parallel development of neural and classical methods supports hybrid approaches, while full comparability of findings requires further standardization of metric reporting, validation procedures, and information on computational cost and hardware requirements. From the perspective of future research decisions, three additional conclusions are significant. First, the ratio of SOC to SOH remains similar in both periods, which means that when planning comparative reviews and benchmarks, it is advisable to maintain an equal emphasis on both categories while increasing the intensity of thermal management research in dedicated tracks. Second, the parallel growth of AI and ML Core and Neural Networks justifies the continuation of strategies combining physical priors with learning layers, especially in conditions requiring stability under varying temperature and load. Third, the dominance of journal articles and experimental studies supports the introduction of a reporting standard that includes a validation scheme, a set of error metrics, computational cost, and hardware requirements, which will improve comparability and accelerate transfer to practice.

The following section interprets the distribution of publications by country, based on the data summarized in

Table 11. Publications by year in countries.

Country	2020–2022	2023–2025	All Years	Share [%]
All countries	25	70	95	100.0
India	4	22	26	27.37
China	7	13	20	21.05
United States	2	8	10	10.53
Malaysia	3	5	8	8.42
Canada	2	3	5	5.26
Morocco	1	4	5	5.26
Pakistan	2	3	5	5.26
France	1	3	4	4.21
Denmark	0	3	3	3.16
Germany	2	1	3	3.16
Saudi Arabia	0	3	3	3.16
Turkey	0	3	3	3.16
United Kingdom	1	2	3	3.16
Other	3	8	11	11.58

and illustrated in Figure 6. The analysis covers both subperiods, 2020–2022 and 2023–2025, followed by conclusions for the entire corpus.

Across the full dataset of 95 publications, the distribution by country is uneven. India contributes the largest share with 26 papers, representing 27.37%, followed by China with 20 papers, 21.05%, and the United States with 10 papers, 10.53%. Together, these three countries account for nearly 59% of the entire corpus, indicating the presence of strong research hubs. The next group includes Malaysia with 8 publications, 8.42%, and the category of Other Countries with 11 publications, 11.58%, which aggregates smaller but still active research centers. Stable, though smaller, contributions come from Canada, Morocco, and Pakistan, each with 5 publications, 5.26%, followed by France with 4 publications, 4.21%, and Denmark, Germany, Saudi Arabia, Turkey, and the United Kingdom, each with 3 publications, 3.16%.

The results of the Chi-square independence test confirm the stability of the geographical profile between subperiods. The test yielded a χ² value of 10.12 with 13 degrees of freedom and a p-value of 0.68, which provide no grounds for rejecting the hypothesis of independence between country distributions and time periods. This means that although the data show clear increases in publication counts in India and the United States and a relative slowdown of growth in China, these changes fall within the range of random fluctuations for the sample size and do not constitute a statistically significant structural shift. It should also be noted that the totals exceed 95 because some papers have co-authors from multiple countries and are counted multiple times in accordance with the adopted classification rules, which increases the heterogeneity of the table and reduces test power. Nevertheless, the high p-value indicates no evidence of a change in the geographical structure within the analyzed time frame.

Figure 6 allows comparison between the two subperiods. India increased its number of publications from 4 to 22, corresponding to a rise in share from 16% to 31.43%, which means that its growth rate exceeds that of the overall corpus. China increased its volume from 7 to 13 publications, but its share decreased from 28% to 18.57%, suggesting a lower-than-average growth rate. The United States expanded from 2 to 8 publications, with its share rising from 8% to 11.43%, strengthening its position in the more recent period. In Southeast Asia, Malaysia maintained steady activity, growing from 3 to 5 publications. In Europe, France and Turkey show moderate growth, while Denmark appears in the newer subperiod with 3 publications. In the MENA region, Saudi Arabia demonstrates increasing activity, rising from 0 to 3 publications, which broadens the thematic scope. Among previously active countries, Germany shows a slight decline from 2 to 1 publication, whereas the United Kingdom increases from 1 to 2 publications.

The Figure above also reveals a long-tail effect: beyond the three leading countries, the distribution includes a broad group of nations with only one or a few publications. This pattern indicates simultaneous concentration and diffusion, where the main results are produced in a few leading research centers while the topic continues to expand to additional countries. For future benchmarks and collaborations, this implies access to increasingly diverse datasets and operational scenarios, while also requiring consistent reporting practices to ensure that cross-center comparisons remain reliable.

In summary, India’s dominance continues to grow, China maintains a strong yet relatively declining position, the United States is strengthening its share, and Europe and the MENA region are joining with a moderate but steadily increasing contribution. This distribution should be taken into account when planning collaboration networks and developing datasets representative of different climatic conditions and operational profiles.

The next Section interprets the distribution of studies according to Battery Management System tasks and their relationships with artificial intelligence method classes and methodological approaches. The basis for this analysis is

Table 12. Publications by Battery Management Systems in other categories.

Name	State of Charge	State of Health	Thermal Management	Total
Total	59	46	8	95
Computational Intelligence
AI/ML Core	31	27	4	51
Neural Networks	31	22	7	52
Algorithms and Techniques	20	22	2	36
Research Methodology
Experiment	46	37	8	80
Literature Analysis	8	6	0	10
Case Study	3	4	2	8
Conceptual	46	36	7	75

. Publications by Battery Management Systems in other categories and the intensity distributions are illustrated on two heat maps in Figure 7 and Figure 8.

The results of the Chi-square independence tests indicate that the distribution of method classes relative to BMS tasks and the distribution of methodological approaches relative to BMS tasks do not differ significantly in statistical terms. For the Computational Intelligence cross-section, χ² was 3.16 with 4 degrees of freedom and p = 0.53, and for the Research Methodology cross-section, χ² was 3.99 with 6 degrees of freedom and p = 0.68. These results provide no basis for rejecting the null hypothesis of independence. In other words, preferences regarding method families and research modes are similarly distributed across SOC, SOH, and the thermal domain, even though the heat maps suggest practical tendencies such as a stronger presence of neural networks in thermal studies or the high share of experimental approaches across all tasks. The interpretation should be made with caution, as the thermal subset contains small counts and the multi-label classification increases margins and reduces test power. Nevertheless, the obtained p-values indicate that the observed differences are descriptive rather than statistically confirmed relationships between task type and the choice of method family or research approach.

In the cross-section of BMS tasks vs. method classes, two clear patterns emerge. First, State of Charge estimation most frequently coincides with both the AI and ML Core family (31 publications) and Neural Networks (31 publications), followed by the Algorithms and Techniques group (20 publications). This distribution confirms the practical tendency to construct hybrid systems in which the network layer cooperates with classical methods and filtering components. Second, State of Health estimation exhibits a similar distribution, with Neural Networks and classical methods appearing in 22 and 22 publications, respectively, and AI and ML Core in 27 publications, again indicating the concurrent use of both paradigms. In thermal management tasks, the network-based family dominates, with seven out of eight studies employing neural approaches, while classical methods are less frequent (two studies), and AI and ML Core appears in four studies. It should be noted that the method classes are not mutually exclusive at the level of individual publications; therefore, row totals exceed column counts, which is expected and reflects the co-occurrence of method families within single solutions.

In the cross-section of BMS tasks and methodological approaches, experimental and conceptual works dominate. For SOC, out of 59 publications, 46 are experimental and 46 are conceptual, while literature analyses appear eight times and case studies three times. For SOH, out of 46 publications, 37 are experimental and 36 are conceptual, with literature analyses and case studies occurring six and four times, respectively. The thermal section is the smallest in size but strongly grounded in validation: all eight studies include an experimental component, seven are conceptual, and case studies appear twice, while no literature analyses were recorded. It should be noted that experimental and conceptual layers often coexist within a single publication; therefore, the row totals may exceed the number of items in a given column. Figure 8 presents these relationships in the form of a heat map, facilitating quick comparison of intensities between tasks and methodological approaches.

The geographical distribution of affiliations was quantified using the Herfindahl–Hirschman Index (HHI), computed from the country-level publication shares (

Table 13. Geographical concentration of affiliations (HHI).

Period	Total Country Contributions	HHI (0–1)	HHI × 10,000	Interpretation
2020–2022	30	0.118	1178	Low concentration—publications are relatively evenly distributed across countries, without dominance of a single location.
2023–2025	82	0.132	1318	Still low, but clearly higher concentration—a larger share of output comes from a few leading countries (especially India and China).

). The index increased from 0.12 (HHI = 1178) in 2020–2022 to 0.13 (HHI = 1318) in 2023–2025. Both values are well below the conventional threshold for high concentration, indicating that BMS research remains geographically diversified. At the same time, the modest increase in HHI suggests a slight shift towards higher concentration, with a growing share of publications originating from a few leading countries, most notably India and China.

The contingency-based effect sizes reported in

Table 14. Association between key dimensions (Cramér’s V).

Crosstab	N	χ2	df	Cramér’s V	Strength of Association
BMS Task × AI Method Class	166	mar.16	4	0.10	Very weak
BMS Task × Methodological Approach	203	mar.99	6	0.10	Very weak
Period (2020–2022 vs. 2023–2025) × Document Type	95	5.kwi	2	0.23	Small to moderate
Period × BMS Task	113	0.28	2	0.05	Negligible
Period × AI Method Class	139	0.24	2	0.04	Negligible
Period × Methodological Approach	173	mar.25	3	0.14	Weak

show that the structural relationships within the corpus are generally weak. The associations between BMS tasks and AI method classes (Cramér’s V = 0.10) and between BMS tasks and methodological approaches (V = 0.10) are very small, indicating that neither AI method families nor research designs are strongly task-specific; similar techniques and study types are used across State of Charge, State of Health, and thermal management problems. The links between the time period and the distributions of BMS tasks (V = 0.05), AI method classes (V = 0.04), and methodological approaches (V = 0.14) are negligible to weak, which suggests that the internal structure of the field in terms of topics and methods remains relatively stable over time despite the rapid growth in publication volume. The strongest association is observed between period and document type (V = 0.23), reflecting a small-to-moderate shift from conference proceedings towards journal articles in 2023–2025 and supporting the interpretation that the research area is gradually consolidating and maturing.

Three practical conclusions emerge from the presented results. First, for SOC and SOH, studies most often combine two main method families, Neural Networks and AI and ML Core, confirming that hybrid architectures are the dominant design pattern. Second, thermal tasks show a clear affinity with the neural family, which can be associated with the need to model nonlinear spatial and temporal couplings in temperature distributions. Third, the strong presence of the experimental component across all tasks indicates a gradual transition from conceptual propositions to solutions validated on real data, while the small number of case studies highlights potential for further development of implementation-oriented research.

Based on the above analyses it is possible to answer the formulated research questions. Regarding the first question, the number of publications has increased across all three BMS categories. In SOC estimation, an increase from 14 to 45 publications is demonstrated, representing an increase of 31 papers or about 221 percent compared with the first period. In SOH estimation, from 13 to 33 publications, an increase of 20 papers or about 154 percent is seen. In thermal management, from 2 to 6 publications, an increase of 4 papers or about 200 percent is seen. Thus, growth covers all three areas, with the largest absolute increases observed in SOC and SOH, while the thermal domain remains the least developed topic despite its relatively high percentage growth rate.

Regarding the second question, the shares of method classes, document types, and countries changed within a limited range, and the direction of change is consistent. Among method families, the share of AI and ML Core increased from 48.0 percent to 55.7 percent, the number of studies using neural networks rose, but the share of Neural Networks remained approximately stable, from 56.0 percent to 54.3 percent, while Algorithms and Techniques decreased moderately from 40.0 percent to 37.1 percent. In terms of document types, the proportion of journal articles increased from 40.0 percent to 65.7 percent, while the share of conference publications decreased from 48.0 percent to 27.1 percent, indicating a maturing publication stream. Geographically, the concentration of activity within the three main centers, India, China, and the United States, has increased. The share of India grew from 16.0 percent to 31.4 percent, China’s share declined from 28.0 percent to 18.6 percent despite an increase in the absolute number of papers, and the United States expanded from 8.0 percent to 11.4 percent. The combined share of these three countries rose from about 52 percent to about 61 percent, pointing to stronger concentration of research activity while maintaining a long-tail distribution among other countries.

The temporal comparison shows a strong overall expansion of the BMS–AI publication landscape between 2020–2022 and 2023–2025 (

Table 15. Growth in publication counts between 2020–2022 and 2023–2025.

Dimension	Category	2020–2022	2023–2025	Absolute Change	Growth Rate (%)
BMS Task	State of Charge (SoC)	14	45	31	221%
	State of Health (SoH)	13	33	20	154%
	Thermal Management	2	6	4	200%
AI Method Class	AI/ML Core	12	39	27	225%
	Neural Networks	14	38	24	171%
	Algorithms and Techniques	10	26	16	160%
Document Type	Article	10	46	36	360%
	Conference Paper	12	19	7	58%
	Other	3	5	2	67%
Methodological Approach	Experiment	21	59	38	181%
	Literature Analysis	3	7	4	133%
	Case Study	0	8	8	100%
	Conceptual	22	53	31	141%

). Among BMS tasks, State of Charge (SoC) exhibits both the largest absolute increase (+31 records) and the highest growth rate (+221%), followed by State of Health (SoH, +154%) and thermal management (+200% from a much smaller base). This confirms the dominant and still-accelerating position of SoC, while thermal aspects remain underrepresented despite their relatively high relative growth.

All three AI method families grew substantially, with AI/ML core methods showing the highest growth rate (+225%), but neural networks and algorithmic techniques have also more than doubled their counts (+171% and +160%, respectively). This pattern suggests that the field is not moving away from any particular family but is instead expanding its methodological portfolio in parallel.

The strongest structural change is observed in document types: journal articles increase by +360% (from 10 to 46), whereas conference papers and other formats grow only moderately (+58% and +67%). This indicates a clear shift from conference-proceedings-driven output towards journal-based dissemination, consistent with the maturation and consolidation of the research area. Methodological approaches follow a similar trend: experimental studies almost tripled (+181%), while conceptual work also grew markedly (+141%). Case studies emerged as a new, previously unused category (0 to 8 records), and literature analyses more than doubled (+133%). Taken together, these dynamics point to a field that is rapidly expanding in volume, increasingly anchored in journal publications and empirical designs, while still maintaining a diversified mix of tasks and AI methods.

To complement the descriptive trend analysis in Table 15 with a region-oriented perspective, a macroregional aggregation of affiliation contributions was derived from the country-level counts reported in Table 11, with countries grouped into Asia, North America, Europe, MENA and Africa, and the residual Other category. In 2020–2022, Asia accounts for 16 contributions, 57.1% of the aggregated country contributions, followed by Europe and North America with 4 contributions each, 14.3% each, the Other group with 3 contributions, 10.7%, and MENA and Africa with 1 contribution, 3.6%. In 2023–2025, the ordering remains unchanged; Asia contributes 43, 53.1%, Europe 12, 14.8%, North America 11, 13.6%, Other 8, 9.9%, and MENA and Africa 7, 8.6%. In relative terms, all macroregions expand in volume, while the fastest growth is observed in MENA and Africa, from 1 to 7 contributions, which corresponds to 600% growth from a small base. The Chi-square test for macroregion by period indicates no statistically significant structural shift, χ² = 0.82 and df = 4, p = 0.94, with a very small association size, Cramér’s V = 0.087. Therefore, the regional profile is stable in proportional terms, while the corpus grows rapidly and gradually broadens its long tail of emerging contributors.

From the thematic perspective, Table 15 can be interpreted as a two-stage evolution timeline. The 2020–2022 subperiod reflects an initial consolidation of SOC and SOH estimation as the dominant application axes, with research outputs still strongly represented in conference proceedings. The 2023–2025 subperiod corresponds to a maturation phase, with journal articles increasing markedly and case studies emerging, which suggests a gradual transition from predominantly methodological demonstrations to implementation-oriented scenarios. In parallel, the substantial growth of both AI and ML core techniques and neural architectures indicates that the dominant trajectory is not methodological replacement but hybridization, where learning-based estimators are increasingly combined with observers, physical priors, and multi-source data streams, including cloud and digital twin settings and grid-aware workflows [19,43,61].

Thermal management remains underrepresented in the corpus, with eight publications in total, despite a 200% growth from a small base. Two structural factors can be inferred from the internal characteristics of the thermal subset and the reporting practices observed across the full corpus. First, thermal studies are disproportionately experiment-anchored; all thermal items include an experimental layer, and they are dominated by network-based families, which is consistent with the need to learn nonlinear spatiotemporal heat propagation under coupled boundary conditions from sparse measurement layouts. Second, compared to SOC and SOH, thermal research more often requires pack-level instrumentation, safety-constrained experiments, and multi-physics integration, which reduces the availability of open benchmarks and makes cross-paper quantitative synthesis difficult. Works in the corpus illustrate that progress in this area is increasingly obtained through multi-domain data fusion, including environmental signals, and through coupling of prediction with control policies, for instance, in reinforcement-learning-based charging and cooling [97,98,102]. Consequently, while the macroregional structure of contributions can be statistically described, a rigorous verification of regional performance differences is not methodologically defensible within the present corpus because studies rely on heterogeneous datasets, chemistries, operating envelopes, and evaluation protocols and often report results using non-uniform metrics. This gap further supports the need for cross-region benchmark initiatives spanning diverse climates and usage profiles, which would enable future updates of this review to test regional performance variation under consistent protocols.

In summary, the increase in the number of publications is widespread across all BMS categories, but in absolute terms it is concentrated in SOC and SOH. In the methodological dimension, there is a moderate increase in the share of AI and ML Core with stable importance of neural networks, supporting the prevalence of hybrid configurations. The publication structure is shifting toward journal articles, confirming the maturity of the field. Geographically, India’s role is strengthening, followed by a smaller increase in the United States, while China maintains a high contribution despite a relative decline in percentage share. Overall, research activity is becoming more concentrated in a few leading centers, yet still remains broadly distributed across a wide range of countries.

5. Discussion

At the outset, it is useful to relate the overall picture of the corpus to the directions identified in previous reviews, which emphasize the growing importance of learning-based methods for State of Charge and State of Health tasks, along with the increasing organization of algorithm families and their roles within BMS architectures. Our findings are consistent with these trends: applications for SOC and SOH dominate, and the highest-quality results are achieved by systems that combine learning with physical knowledge. In practice, the most effective configurations are hybrid designs, such as Kalman filters coupled with deep networks for SOC, sequential hybrids with feature selection for SOH, and approaches integrating equivalent circuit models with LSTM or using SVM-RFE feature selection. This is supported by individual studies showing RMSE reduction through nonlinear filtering combined with deep networks, high accuracy achieved by DCRNN with feature elimination, the effectiveness of XGBoost and Random Forest fusion, and cloud–onboard integrations in which SOC is estimated in the vehicle and SOH in the cloud using incremental learning [10,23,27,28,41,43].

A second recurring theme is the practice of data and feature selection. Studies reporting the best results combine voltage, current, and temperature with telematics and environmental parameters, while their data preparation pipelines include channel synchronization, filtering, and normalization. For SOH, key features are differential and incremental measures from charge cycles, the second peak of the incremental capacity curve, and high mutual-information-based feature selection, which improve resolution with respect to nonlinear aging. In thermal tasks, the use of quantile regression and reinforcement-learning-based control is increasing, as these methods provide both forecasts and coverage metrics and operate within time horizons suitable for cooling regulation. Fleet and telematics applications confirm the usefulness of multimodal data streams, while evidence of the benefits of feature selection and signal engineering is consistent across independent studies [21,48,72,73,98,101].

The third observation concerns implementability and reliability. Alongside improvements in predictive accuracy, there is growing evidence of implementations on resource-constrained platforms and the first consistent reports of uncertainty quantification. FPGA accelerators for LSTM and quantized networks on microcontrollers demonstrate that estimation can be performed within the memory and energy budgets typical for BMS, effectively linking quality metrics with computational cost. Quantile regression for temperature and Gaussian process filters with uncertainty modeling point toward a standardization trend in reporting, where confidence measures and platform characteristics are provided in addition to RMSE and MAE. In domains requiring scalability and data protection, federated learning is increasingly being considered, enabling model training in distributed environments without centralizing data, which aligns with the operational needs of large fleets [32,40,76,101].

The results of our review confirm the effectiveness of hybrid solutions in which machine learning and deep neural networks are coupled with physical knowledge and Bayesian filtering, while also clarifying the conditions under which these solutions retain their advantage. In SOC estimation, configurations combining Kalman observers with sequential, feedforward or LSTM layers maintain low errors under varying load profiles and within typical temperature ranges. This is consistent with the results obtained for filtering–network hybrids and with improvements reported in EKF and UKF variants in the literature [41], as well as with generalization effects in PINN, where the serial configuration, which injects ECM parameters into LSTM, outperforms both the parallel architecture and the baseline LSTM [80]. In SOH estimation, the best results are achieved by sequential architectures with targeted feature selection, for example, DCRNN with SVM-RFE exhibiting very low errors, as well as multiscale feature extraction models. Together, these findings indicate that differential feature selection and circuit-based priors remain essential for distinguishing nonlinear degradation trajectories [27,82]. In the thermal domain, quantile CNN and RNN models perform best, providing not only point predictions but also coverage information, along with CFD–network hybrids that predict maximum temperatures more efficiently, which is crucial for cooling control under variable load conditions [95,98]. These observations are consistent with the effectiveness of classical and learning-based fusions, such as XGBoost combined with Random Forest for SOC estimation, and with digital twin solutions in which SOC is estimated in the vehicle using a Kalman filter and SOH in the cloud using incremental learning [23,43].

Answering the first research question, namely which model families and hybrid configurations are most effective in practice and under what conditions they maintain their advantage, the results indicate three stable patterns. First, in SOC estimation, observer plus sequential network systems dominate, effectively suppressing noise and drift while maintaining quality within typical operating ranges. This is confirmed by designs combining Kalman filters with deep networks and variants coupled with ECM [41,80]. Second, in SOH estimation, sequential architectures with selection mechanisms prevail, with DCRNN combined with SVM-RFE achieving errors on the order of hundredths of a percent, confirming that feature filtering prior to learning remains essential for prediction stability [27]. Third, in thermal management, quantile CNN and RNN models validated on fleet and simulated data are effective, as they enable control design with uncertainty consideration, while CFD hybrids reduce computational cost in maximum temperature prediction [95,97]. These results are consistent with the observed effectiveness of deep configurations optimized through metaheuristics for combined SOC and SOH tasks, such as the ResDCBi-GRU model with parametric fine-tuning, which achieves low RMSE in both tasks [10].

Answering the second research question, namely what types of data and preprocessing and normalization procedures dominate in the best-performing studies and which feature sets most often accompany them, the corpus is dominated by data streams combining voltage, current, and temperature with telematics and environmental context, following consistent channel alignment, filtering, and normalization. The effectiveness of such prepared data is confirmed by studies using fleet and road datasets, for example, for the BMW i3, where feedforward networks outperform ELM when datasets are properly partitioned and normalized, as well as by studies combining multiple driving sessions with iterative hyperparameter tuning [23,29]. In the thermal domain, cross-domain approaches and Bayesian hyperparameter optimization improve RMSE median values and high-quantile coverage, illustrating the usefulness of integrating fleet, simulation, and weather station data into a single processing pipeline [98]. In probabilistic and Bayesian GP filtering, significant reductions in uncertainty and accuracy gains over reference filters demonstrate that explicit uncertainty modeling within the estimator is a practical advantage when working with noisy data [40]. Finally, the digital twin paradigm divides computational load between onboard and cloud systems, facilitating scalability, with SOC estimated in the vehicle and SOH in the cloud, while incremental learning improves adaptation to data drift [42].

Answering the third research question, namely the extent to which uncertainty, verification and validation, as well as computational costs and hardware requirements, are reported, clear yet uneven practices can be observed. Quantile models in the thermal domain report coverage and interval width measures, supporting engineering decisions under risk [98]. GP-UKF and GP-PF filters incorporate uncertainty directly into the estimator, reporting variance reduction and improved accuracy over reference filters, which can be regarded as a good standard for UQ reporting in SOC estimation [40]. At the same time, reports on computational cost and hardware requirements remain limited, although studies with hardware implementations quantify memory footprint, parameter count, and inference times after quantization, confirming feasibility on microcontrollers with errors of a few percent of RMSE [101]. In studies focused on achieving high accuracy for both battery states, low RMSE and MAE values within a single model are also reported, indicating consolidation of the processing pipeline and the potential for joint optimization targeting SOC and SOH estimation [10].

To complement the discussion with two statistical questions, we first note that the number of publications increased in all three BMS categories. In SOC estimation, a jump from 14 in 2020–2022 to 45 in 2023–2025 was seen, representing an increase of 31 papers. In SOH estimation, moving from 13 to 33, an increase of 20 papers was seen. In thermal management, moving from two to six, an increase of four papers was seen. Thus, growth covers all three areas, but in absolute terms it is concentrated in SOC and SOH, while the thermal component remains the smallest in volume despite relatively high growth dynamics.

Regarding the changes in the shares of method classes, document types, and countries, we observe an increase in the share of AI and ML Core from about 48.0 percent to 55.7 percent, with a stable level of network-based solutions, from about 56.0 percent to 54.3 percent, and a moderate decrease in the Algorithms and Techniques group, from about 40.0 percent to 37.1 percent. In the structure of document types, the share of journal articles increased from about 40.0 percent to 65.7 percent, while the share of conference publications decreased from about 48.0 percent to 27.1 percent, with a smaller share for the remaining category. In the geographical distribution, concentration increased among the three leading centers, with India’s share growing from about 16.0 percent to 31.4 percent, China’s share decreasing from about 28.0 percent to 18.6 percent despite an increase in the number of papers, and the United States expanding from about 8.0 percent to 11.4 percent. The Chi-square independence test for country distribution indicates no significant structural change between the subperiods, meaning that the observed shifts reflect a quantitative growth trend without statistically confirmed restructuring of the geographical profile.

The limitations of this review arise primarily from the selection of sources and query parameters. The analysis relied exclusively on the Scopus database, the English language, and the years 2020–2025. Searches were conducted in the Title, Abstract, and Keywords fields using EXACTKEYWORD filters, which increases metadata consistency and reproducibility but may exclude papers outside the index, in other languages or those using different terminology. Indexing delays may also reduce the number of the most recent entries. The multi-label classification and two-stage screening across five dimensions introduce an element of expert judgment. In borderline cases, decisions depend on the interpretation of research objectives and abstract quality. Multi-labeling violates independence assumptions in Chi-square tests and reduces test power, especially in low-count categories such as the thermal domain. The bibliometric layer is descriptive; term distributions depend on inclusion thresholds and synonym merging rules, and geographical or typological aggregation simplifies reality. Country assignment was based on the first affiliation of the first author, with additional labels added for co-authors, which may slightly overestimate the shares of countries with high international collaboration.

Another barrier to comparison is the heterogeneity of tasks, datasets, and evaluation protocols. Different error metrics are used, including RMSE, MAE, MAPE, and R², while in the thermal section maximum errors and coverage measures for quantile models are also reported. Authors present results in different units and time horizons, rarely reporting full computational costs, training and inference times, parameter counts or memory footprint. The absence of common benchmarks limits the possibility of formal quantitative synthesis, making comparative conclusions descriptive in nature. The generalizability of results is constrained by environmental and computational differences such as signal quality and sampling frequency, temperature and C-rate ranges, cell chemistry, pack size, and hardware configuration. Some results are based on synthetic data or limited experiments, creating risks of validation leakage and sensitivity to hyperparameter selection. These factors require cautious interpretation and emphasize the need for standardized reporting of uncertainty, computational costs, and transferability tests across fleets and climatic conditions.

Despite rapid progress in data-driven SOC and SOH estimation and thermal prediction, translation into production BMS is constrained by engineering deployment barriers that are often underreported in academic studies. A major constraint is compliance with automotive functional safety requirements, where an AI-enabled estimator must be treated as part of a safety-related item and supported by a structured safety case, hazard analysis, and verification evidence across the full development lifecycle. In practice, this implies clearly defined operational boundaries, monitoring for out-of-distribution conditions and sensor faults, diagnostic coverage, and safe degraded modes when estimator confidence deteriorates. Approaches that combine model-based backbones with learning components or include explicit validation and anomaly-detection layers are more naturally aligned with such safety argumentation than purely black-box inference because they provide mechanisms for constraint checking and fault containment [52,55,68].

A second barrier concerns cybersecurity risks in connected and cloud-assisted BMS architectures. Cloud-BMS and digital-twin pipelines enable continuous updating and fleet-scale analytics, but they also introduce additional attack surfaces through communication links, remote updates, and aggregation of operational telemetry. Deployment-oriented solutions therefore require secure data transport, authentication and authorization, integrity protection, robust update governance, and continuous monitoring for abnormal behavior. The reviewed corpus includes cloud-connected concepts and telematics-oriented monitoring that highlight both the opportunity and the need for system-level safeguards, particularly when estimation and diagnostics are coupled with off-board computation [19,48,52]. In practical terms, a robust split architecture is required, where safety-critical inference remains credible on-board while cloud components support non-safety-critical optimization, long-horizon analytics, and supervised recalibration under controlled conditions [19,43,52].

A third deployment challenge is adaptation to emerging and diverse battery chemistries and formats. Data-driven models trained on limited cell types and operating envelopes can exhibit domain shift when applied to different chemistries, temperature regimes, aging modes, or pack configurations, which undermines generalization and complicates validation. Mitigation requires explicit reporting and enforcement of validity domains, chemistry-aware feature design, transfer and incremental learning strategies, and hybrid physics-guided constraints that reduce reliance on purely empirical correlations. Studies employing incremental updating, signal-based diagnostic descriptors, or hybrid estimation structures indicate practical pathways to increase robustness under heterogeneous real-world conditions, particularly when operational data streams are used for controlled recalibration rather than unverified extrapolation [43,52,72].

Finally, resource-constrained execution at the edge remains a practical bottleneck. Many high-accuracy neural architectures are difficult to deploy on embedded automotive hardware without compression, quantization, or hardware acceleration and without demonstrating predictable latency and memory usage. The corpus contains representative directions that map the accuracy–efficiency trade space, including lightweight model designs with explicitly reduced parameter counts, microcontroller implementations supported by quantization, and dedicated acceleration platforms [17,32,56]. These examples motivate a practical reporting recommendation for future studies; accuracy metrics should be accompanied by platform-specific runtime indicators, memory footprint evidence, and an explicit statement of whether inference is feasible on-board, accelerated, or off-board.

6. Conclusions

The present review, based on a homogeneous corpus of 95 publications and a five-dimensional classification framework, provides a synthetic description of the current state of research on BMS in three key tasks: State of Charge estimation, State of Health estimation, and thermal management. Quantitative results indicate a rapid increase in publication volume after 2022, accompanied by stability in the proportions between categories, which facilitates inter-period comparison. Across the entire corpus, SOC and SOH dominate, while the thermal component remains comparatively weaker, although it shows a steady upward trend. Within the method classes, parallel development of AI and ML Core and network-based solutions can be observed, while the share of journal articles is increasing, confirming the maturation of the field.

Summarizing the research questions, the most effective configurations in practice are hybrid systems that combine physical priors and filtering with a learning layer. Specifically, Kalman observers combined with sequential networks for SOC and sequential architectures with targeted feature selection for SOH deliver the best results, while in thermal tasks the most effective solutions combine quantile prediction with reinforcement-learning-based control policies. The highest accuracy is achieved using data streams that integrate voltage, current, and temperature with telematics and environmental parameters, with consistent preprocessing procedures including synchronization, filtering, resampling, and normalization. Effective feature sets include temporal derivatives and aggregates for SOC, ICA, and DVA features and hysteresis indicators for SOH and geometry and boundary condition representations for thermal tasks. Reporting of uncertainty and computational costs is not yet standardized, with RMSE and MAE remaining dominant, while MAPE and R² appear less frequently. In thermal studies, RMSE and maximum error are most common, and surrogate models report execution times and speed-up factors. A unified minimal reporting package is needed, including at least one measure of uncertainty, an explicit validation scheme, and a description of the computational platform.

The methodological contribution of this study lies in ensuring full reproducibility of the procedure, from the Scopus query through vocabulary normalization and two-stage screening to five-dimensional coding and cross-sectional tabulation. This design made it possible to map relationships between tasks, method classes, geography, and publication types and then coherently link them to qualitative conclusions regarding the most effective configurations and data preparation practices. The resulting picture reflects a mature yet heterogeneous research field, while also identifying areas requiring further investment, particularly the thermal domain and standardized reporting of uncertainty and computational costs.

The implications for engineering practice are direct. For SOC and SOH, hybrid configurations with a physical prior and observer, as well as sequential architectures supported by feature selection, are recommended. For the thermal domain, the combination of quantile prediction and reinforcement learning control is most effective. In resource-constrained environments, compact quantized models with controlled memory footprints are preferable. Data workflows should incorporate telematics and environmental conditions, and the preparation pipeline should be adapted to the decision horizon, window length, sampling frequency, and normalization scheme. For evaluation, it is recommended to report task-specific error metrics and verification and validation elements, including transferability tests, as well as platform parameters and computational times.

In future work, an important direction for improving multi-state co-estimation in BMS is the explicit inclusion of novel sensing modalities, particularly non-destructive testing (NDT) technologies, as complementary information channels to conventional electrical measurements. While the reviewed literature already highlights sensor fusion and thermal–electrical co-management as key trends, current pipelines are still predominantly driven by voltage–current features, which may limit observability under fast transients and delay the detection of safety-critical internal changes. Integrating non-invasive sensing (e.g., ultrasound-based observations treated as sparse or event-triggered measurements) could enhance joint SOC–temperature inference, strengthen state-consistency checks, and support early-warning capabilities relevant to thermal runaway prevention, while aligning with the engineering objective of reducing sensing cost by enabling targeted measurements rather than continuous high-bandwidth instrumentation. This perspective is consistent with the growing emphasis on connected BMS architectures and advanced control approaches that jointly consider charging and thermal constraints, indicating a clear pathway for future research on practical, deployment-oriented multimodal state estimation [12,97,102].

From the perspective of further development, three directions are most important. First, strengthening the thermal component through standardized datasets and cross-domain validation, which will facilitate comparability and transfer to implementation. Second, standardizing the reporting of uncertainty, computational costs, and resource requirements, which will enable more accurate planning of computational and energy budgets and support risk assessment. Third, developing open benchmarks and shared validation and verification protocols, together with public access to visualization and classification input files, which will ensure long-term transparency and replicability of future review editions. Such a framework supports the development of solutions that are both computationally feasible and qualitatively effective, while accelerating the transition of methods into BMS practice.

The authors are planning a subsequent stage of work in which the procedure will be extended to include additional bibliographic sources and language variants. The query will be repeated and harmonized across at least two independent databases, IEEE Xplore and Web of Science, with consideration of including ACM Digital Library and arXiv for preprints. For this purpose, a shared normalization dictionary and metadata field mapping will be developed, along with a unified set of subject filters and a deduplication procedure based on DOI identifiers and title–author–year matching. The updated PRISMA diagram will include multi-source paths, and the resulting datasets will be subjected to sensitivity analysis to assess the impact of database, language, and keyword scope on corpus composition and the stability of conclusions.

Future work should strengthen the thermal-management perspective by explicitly addressing extreme operating scenarios and pack-level thermal behavior, which are currently underrepresented relative to SOC and SOH topics. In particular, fast charging, low-temperature operation, and high-ambient conditions can amplify thermal gradients and accelerate safety-relevant transients, while pack-level propagation mechanisms require models and controls that remain valid beyond single-cell settings [97,100]. In addition, a deployment-oriented synthesis should report not only temperature prediction accuracy but also energy–efficiency trade-offs of thermal control strategies using consistent metrics, for example, parasitic cooling power or energy overhead, net impact on driving range, temperature reduction per unit of auxiliary energy, and joint constraints on maximum temperature and spatial gradients in the pack [99]. Such an extension would align thermal–electrical co-management with engineering decision-making by making explicit how improved thermal safety and robustness are achieved at an acceptable energy cost, particularly in integrated electrothermal control settings already reflected in recent studies [97,98].

This future-oriented perspective also includes registering the review protocol, publishing a replication package, and iteratively refining the classification framework. Plans include refining the synonym dictionary, incorporating terminology variants used outside Scopus, introducing transferability tests for results derived from multiple databases, and recalibrating labeling rules across the five dimensions. Such an extended review will allow verification of the robustness of the observed patterns to source selection, fill thematic gaps resulting from single-database limitations, and enhance the reliability of recommendations for BMS practice.