Machine Learning Applications in Energy Consumption Forecasting and Management for Electric Vehicles: A Systematic Review

Abstract

This literature review addresses a major research gap in electromobility by providing a comprehensive synthesis of machine learning (ML) and deep learning (DL) applications for forecasting energy consumption, managing battery state of charge (SoC), and integrating electric vehicles (EVs) with charging infrastructure and smart grids, including vehicle-to-grid (V2G) systems. Despite the rapid increase in publications between 2016 and 2025, few comparative studies systematically evaluate ML/DL approaches, their effectiveness in specific applications, and their limitations under real-world conditions. To bridge this gap, this review analyzes 95 publications, covering methods from ensemble learners (e.g., Random Forest, XGBoost) to advanced hybrids (e.g., LSTM + MPC). Key influencing factors such as driving style, topography, and weather are considered. This review identifies persistent challenges, including the lack of standardized datasets, limited model generalization, and high computational demands. It also outlines research directions, such as adaptive online learning and integration with V2X technologies. By consolidating current knowledge, this review supports engineers, EV system designers, and policymakers in planning effective energy management and charging strategies, thereby contributing to the sustainable development of electromobility.

1. Introduction

Electric vehicles (EVs) have become an integral part of all market segments, from passenger cars to commercial vans. The wide variety of models and price ranges enables buyers to select vehicles that suit their needs [1,2,3]. This growing accessibility is confirmed by the rising number of registered EVs and their steadily increasing market share in both passenger and commercial sectors [4,5,6]. However, large-scale EV adoption requires addressing several key challenges, including efficient energy management, the optimization of charging infrastructure, and integration with smart grids. In this context, data-driven methods, particularly machine learning (ML), deep learning (DL), neural networks (NNs), and statistical approaches, play an increasingly vital role.

The core objective of ML is to process datasets in order to create predictive models. During model training, optimization algorithms are applied to achieve the highest possible accuracy. These models work by adjusting parameters to capture complex relationships in the data. In the EV domain, predictive models are used to estimate energy demand and forecast the battery state of charge (SoC) in real time along a given route. Their purpose is to provide drivers with reliable range predictions and minimize range anxiety. By leveraging large datasets, ML methods can capture nonlinear dependencies between speed profiles, load, road grade, ambient conditions, and actual energy consumption. Forecasting accuracy, however, strongly depends on the quality and diversity of the input data [7,8].

In addition to route-level forecasts, ML techniques also support charging-demand prediction, charging-strategy optimization, and dynamic control of distributed energy resources. These capabilities are essential for efficient EV operation under diverse conditions [9,10,11]. For example, ML models can forecast regional fast-charging demand, enabling more adaptive charging systems tailored to real user needs [7,12,13,14].

The objective of this paper is to systematically review research published between 2016 and 2025 in order to identify trends, challenges, and future directions in the application of data-driven approaches such as ML, DL, NNs, and statistical algorithms for forecasting and managing EV energy consumption. The review focuses on the effectiveness, limitations, and gaps of these methods, particularly in energy and SoC forecasting. It highlights key algorithms such as Random Forest, XGBoost, LSTM, and hybrid models, and examines their ability to incorporate contextual variables such as vehicle speed, road topography, driving style, and weather conditions. The analysis also considers their role in optimizing energy management, including integration with renewable energy sources and hybrid storage systems that combine batteries and supercapacitors.

For clarity, this review adopts the following taxonomy. Machine learning (ML) encompasses broad data-driven techniques, including ensemble methods such as Random Forest and XGBoost. Deep learning (DL) refers to multi-layer neural architectures such as LSTM and CNN. Neural networks (NNs) denote foundational architectures such as ANN and BPNN, which often overlap with DL but are treated separately here for classification consistency. Statistical algorithms include probabilistic methods such as Markov processes. These categories are applied consistently throughout the analysis.

Furthermore, this review aims to highlight the limitations of current approaches and to identify directions for future research. It was conducted to address seven key research questions:

• What machine learning algorithms are most frequently used to forecast energy consumption and battery state of charge (SoC) in electric vehicles, and what are their main applications?
• What factors (e.g., driving style, topography, temperature) are typically considered in ML models for forecasting EV energy consumption?
• What opportunities and challenges are associated with integrating EVs into charging infrastructure and smart grids using ML algorithms?
• What are the most common limitations and challenges related to validating ML/DL models under real-world conditions?
• What are the dominant research directions, and which areas require further investigation?
• Does the distribution of algorithm classes (machine learning, deep learning, neural networks, statistical algorithms) differ significantly between the 2016–2020 and 2021–2025 periods?
• Is publication activity in this area evenly distributed across countries, or is it concentrated in selected hubs?

This review addresses a significant research gap in electromobility by providing a comprehensive synthesis of machine learning applications for predicting energy consumption, managing battery SoC, and integrating EVs with charging infrastructure and smart grids, including vehicle-to-grid (V2G) systems. Despite the growing number of publications between 2016 and 2025, there is still a lack of comparative studies that systematically evaluate different ML/DL approaches, their effectiveness in specific applications, and their limitations under real-world conditions.

To fill this gap, this review analyzes 95 publications covering both ensemble models (e.g., Random Forest, XGBoost) and advanced hybrids (e.g., LSTM + MPC), while considering factors such as driving style, road topography, and weather conditions. The contribution of this review lies in identifying key challenges, including the lack of standardized datasets, limited model generalization, and high computational requirements, as well as in outlining directions for future research. Ultimately, this article supports the development of sustainable electromobility by providing guidance for powertrain engineers, charging-infrastructure designers, and policymakers.

2. Materials and Methods

This paper presents a systematic review of the scientific literature on the application of machine learning, deep learning, neural networks, and statistical methods for forecasting and controlling energy consumption in electric vehicles. The primary goal is to provide a comprehensive analysis of research published between 2016 and 2025 in the field of data-driven methods that support effective energy management, charging infrastructure planning, and the integration of EVs with smart grids. The search was conducted on 10 June 2025. To ensure transparency and replicability, the review followed a structured protocol inspired by the PRISMA guidelines. The publication selection process included four main stages: identification, screening, eligibility assessment, and final inclusion [15] (Supplementary Materials). A flowchart illustrating this process is shown in Figure 1.

2.1. Identification

A literature search was conducted in the Scopus database using a PRISMA-inspired approach. The query included publications from 2016 to 2025 written in English, focusing on the application of ML, DL, and statistical methods in electric vehicles. The search specifically targeted energy consumption forecasting, predictive control, battery management, and integration with the power grid. The following query was used:

“TITLE-ABS-KEY (“Electric Vehicles” AND (“Energy Utilization” OR “Energy Consumption”) AND (“Forecasting” OR “Prediction” OR “Predictive Control”)) AND PUBYEAR > 2015 AND PUBYEAR < 2026 AND (EXCLUDE (SUBJAREA, “DENT”) OR EXCLUDE (SUBJAREA, “NURS”) OR EXCLUDE (SUBJAREA, “IMMU”) OR EXCLUDE (SUBJAREA, “PHAR”) OR EXCLUDE (SUBJAREA, “BUSI”) OR EXCLUDE (SUBJAREA, “SOCI”) OR EXCLUDE (SUBJAREA, “MULT”) OR EXCLUDE (SUBJAREA, “PSYC”) OR EXCLUDE (SUBJAREA, “ARTS”) OR EXCLUDE (SUBJAREA, “AGRI”) OR EXCLUDE (SUBJAREA, “HEAL”) OR EXCLUDE (SUBJAREA, “CENG”) OR EXCLUDE (SUBJAREA, “EART”) OR EXCLUDE (SUBJAREA, “NEUR”) OR EXCLUDE (SUBJAREA, “BIOC”) OR EXCLUDE (SUBJAREA, “CHEM”) OR EXCLUDE (SUBJAREA, “MEDI”) OR EXCLUDE (SUBJAREA, “PHYS”) OR EXCLUDE (SUBJAREA, “ENVI”) OR EXCLUDE (SUBJAREA, “ECON”) OR EXCLUDE (SUBJAREA, “MATE”)) AND (LIMIT-TO (LANGUAGE, “English”)) AND (EXCLUDE (DOCTYPE, “cr”) OR EXCLUDE (DOCTYPE, “re”)) AND (LIMIT-TO (EXACTKEYWORD, “Machine Learning”) OR LIMIT-TO (EXACTKEYWORD, “Machine-learning”) OR LIMIT-TO (EXACTKEYWORD, “Learning Algorithms”) OR LIMIT-TO (EXACTKEYWORD, “Deep Learning”) OR LIMIT-TO (EXACTKEYWORD, “Long Short-term Memory”) OR LIMIT-TO (EXACTKEYWORD, “Neural Networks”) OR LIMIT-TO (EXACTKEYWORD, “Neural-networks”) OR LIMIT-TO (EXACTKEYWORD, “Regression Analysis”) OR LIMIT-TO (EXACTKEYWORD, “Markov Processes”)”.

Following an initial search, 148 papers were retrieved.

2.2. Screening

The remaining 148 records were screened, with 46 being excluded by applying an additional keyword filter:

“AND (LIMIT-TO (EXACTKEYWORD, “Battery”) OR LIMIT-TO (EXACTKEY-WORD, “Secondary Batteries”) OR LIMIT-TO (EXACTKEYWORD, “States Of Charges”)

OR LIMIT-TO (EXACTKEYWORD, “State Of Charge”) OR LIMIT-TO (EXACT-KEYWORD, “Digital Storage”) OR LIMIT-TO (EXACTKEYWORD, “Charging (batter-ies)”)

OR LIMIT-TO (EXACTKEYWORD, “Electric Vehicle Charging”) OR LIMIT-TO (EXACTKEYWORD, “Charging Stations”) OR LIMIT-TO (EXACTKEYWORD, “Charging Station”) OR LIMIT-TO (EXACTKEYWORD, “Battery Management Systems”) OR LIMIT-TO (EXACTKEYWORD, “Lithium-ion Batteries”) OR LIMIT-TO (EXACTKEY-WORD, “Vehicle-to-grid”) OR LIMIT-TO (EXACTKEYWORD, “Smart Power Grids”) OR LIMIT-TO (EXACTKEYWORD, “Electric Power Transmission Networks”) OR LIMIT-TO (EXACTKEYWORD, “Electric Power Distribution”) OR LIMIT-TO (EX-ACTKEYWORD, “Smart Grid”) OR LIMIT-TO (EXACTKEYWORD, “Electric Loads”) OR LIMIT-TO (EXACTKEYWORD, “Demand Response”) OR LIMIT-TO (EXACT-KEYWORD, “Energy”) OR LIMIT-TO (EXACTKEYWORD, “Renewable Energy”) OR LIMIT-TO (EXACTKEYWORD, “Model Predictive Control”) OR LIMIT-TO (EXACT-KEYWORD, “Predictive Control Systems”) OR LIMIT-TO (EXACTKEY-WORD, “Model-predictive Control”) OR LIMIT-TO (EXACTKEYWORD, “Adaptive Control Systems”))”.

This process yielded 102 publications.

2.3. Eligibility

During the eligibility assessment stage, full texts were carefully analyzed. Publications were included if they described the practical or conceptual application of ML, DL, or statistical algorithms in the context of electric vehicles, energy consumption, battery management, charging, or grid integration. Categorization was based on titles, abstracts, keywords, and full-text content. As a result, 95 publications were selected for detailed analysis. Studies from non-technical fields and papers without a clear application of data analysis in EV energy were excluded. Review and conceptual papers were included when they provided methodological value or identified research gaps. The quality of each publication was assessed according to citation count, source type (journal or conference), and ranking (e.g., Q1–Q2 in Scopus or SJR). Particular priority was given to works with strong experimental foundations and well-documented methodologies, even if citation counts were low due to recent publication. Each paper was independently evaluated by two authors, and disagreements were resolved by consensus.

2.4. Included

In this review, 95 scientific publications were analyzed, selected from the Scopus database using a defined query and filtering criteria. The collected bibliographic data, including publication title, authors, year of publication, institutional affiliations, DOI, and keywords, were imported into a PostgreSQL 16.2 relational database. This structure enabled systematic storage of information and the execution of advanced SQL queries to aggregate, filter, and segment the data. For further analysis, the Python 3.12.2 environment was used together with the following libraries: Pandas 2.2.2 for processing tabular data, and Matplotlib 3.8.4 and Seaborn 0.13.2 for creating charts and visualizations. Each paper was classified according to four main analytical axes: (i) the algorithms used, (ii) the area of application and control, (iii) the country of author affiliation, and (iv) the type and research methodology. In the case of complex approaches, publications could be assigned to more than one category. A visual summary of the search process, filtering steps, and classification axes is presented in Figure 2, which also illustrates the thematic scope of the analyzed research.

The first classification axis concerned the algorithms used. Categories were assigned based on keywords in the articles, and four main groups were distinguished: machine learning, deep learning, neural networks, and statistical algorithms. The second classification axis covered the areas of algorithm application, also based on keywords. Four categories were identified:

• Battery: issues related to battery operation, characteristics, state of charge (SoC), and energy storage.
• Charging: charging processes, charging stations, and energy demand.
• Energy networks: supervision and strategies for connecting and transferring energy between EVs and the power grid, including vehicle-to-grid (V2G), smart grids, and energy transmission and distribution.
• Predictive control: advanced energy control strategies, such as model predictive control (MPC).

The third axis referred to the geographical location of the authors, determined using affiliation data from Scopus. For each publication, one or more countries were assigned based on the authors’ institutions. The majority of works originated from China, India, the United States, Germany, South Korea, Canada, the United Kingdom, Spain, France, and Saudi Arabia. The remaining countries were grouped under the category “Other.”

The fourth axis concerned the type of publication, classified using bibliographic data from Scopus into three document types: research papers, conference papers, and other documents (e.g., book chapters and monographs).

The fifth axis described the research method, determined by analyzing article content. Four categories were distinguished:

• Experimental studies: authors conducted their own measurements.
• Literature analyses: reviews and syntheses of existing research.
• Case studies: specific implementations or applications of solutions.

Conceptual works: theoretical, model-based, or focused on developing new strategies and mathematical models. This detailed classification enabled a multifaceted analysis of the current scientific output in the field of forecasting and controlling EV energy consumption.

Figure 3 illustrates the spatial concentration of key terms from the analyzed publications, generated using VOSviewer (v. 1.6.20). The color intensity indicates the frequency of term occurrence. Highly saturated areas (red, yellow) represent high frequency and a strong interrelation of topics, signifying intensive research activity [16]. Analysis of the map highlights dominant areas such as electric vehicles, machine learning, and energy management, while clustered terms reveal strong connections between these concepts.

Figure 4 shows a network of keyword connections from the analyzed publications. Each node represents an individual term, with its size indicating frequency and the thickness of connecting lines reflecting the strength of relationships [16]. The red cluster links forecasting and energy utilization with machine learning. The green cluster connects energy utilization with battery and Markov processes, pointing to intensive research on predicting and managing energy and battery status. The blue cluster associates neural networks with charging stations, highlighting their role in monitoring charging processes. The yellow cluster links energy consumption with model predictive control, illustrating the use of advanced control algorithms.

3. State of Art

As noted earlier, machine learning methods for forecasting energy consumption in electric vehicles have developed rapidly in recent years. The analyzed publications were grouped into three main thematic areas: (1) forecasting EV energy consumption, (2) predictive control in EVs, and (3) vehicle-to-grid and infrastructure integration. This review examines the application of ML in EV energy management across these three areas, although the approaches and tools used differ. Some studies are conceptual, presenting new analytical methods and predictive models that require further experimental validation. Others are literature reviews that identify research trends, technological challenges, and gaps that remain to be addressed. Both types of studies make valuable contributions to advancing knowledge in the field of forecasting and controlling EV energy consumption.

3.1. Forecasting EV Energy Consumption

Forecasting energy consumption and battery state of charge is one of the most common applications of ML algorithms in electromobility, as precise estimations are crucial for effective trip planning and energy management. The most frequently used techniques are ensemble models (Random Forest, XGBoost), recurrent neural networks (LSTM, GRU), and hybrid combinations of neural networks with classical algorithms (e.g., syncretic BP + RBF learning).

Table 1. Types of ML algorithms for forecasting energy consumption and SoC in electric vehicles.

Algorithm Type	Example Algorithm	Application	Model Accuracy (R2, MAE, RMSE)	Key Features	Reference Papers
Classical Models	Linear Regression, Decision Trees	Forecasting EV energy consumption based on route data (speed, grade)	R2 = 0.96 [17], R2 = 0.981 [18]	Simple, fast, interpretable; limited in modeling nonlinearities	[16,17,18]
Ensemble Models	Random Forest, XGBoost, GBM-BO	Estimating EV energy consumption and SoC, charging planning, grid management	R2 = 0.981 [18], 81.11% accuracy [19]	High accuracy, robustness to errors, requires large datasets	[16,19,20,21]
Deep Learning	LSTM, GRU, CNN, ConvLSTM	Forecasting SoC, range, and energy consumption under dynamic conditions (driving style, topography)	R2 = 0.99 [22], MAE = 1.76%, RMSE = 1.99% [23], R2 = 0.91 [24]	High precision for time series, high computational cost	[22,23,24,25,26,27,28,29,30,31,32,33,34,35,36,37]
Hybrid Models	GBM-BO, RFM-BO, BP + RBF, LSTM + NARX, MPC + DRL	Estimating SoC, optimizing energy consumption, energy management in BMS	R2 = 0.9647, MSE = 0.000112 [38], R2 = 0.944 [39], error < 1% [40]	Flexible, handles uncertainty, complex implementation	[20,21,37,38,40,41,42,43,44]
Statistical Methods	Markov Processes	Managing battery energy and temperature, real-time optimization	Consumption reduction of 18% [45], 4% [46]	Dynamic adaptation, limited modeling of complex relationships	[45,46]

illustrates the diversity of ML algorithms used in forecasting EV energy consumption, indicating their specific applications and highlighting their varied effectiveness depending on the operational context.

In the field of EV energy consumption forecasting, the use of classical machine learning algorithms and advanced ensemble methods has grown steadily. By combining multiple base learners, ensembles improve the accuracy and robustness of predictions. Study [16] shows that random-forest and XGBoost models outperform alternative approaches for energy-consumption prediction, which is crucial for optimizing charging schedules. A neural network model in [47] achieved 89% accuracy in forecasting battery energy consumption. Studies [17,18] focused on electric buses, identifying key factors such as road gradient and vehicle load. Their regression models explained over 96% of consumption variability [17] and achieved R² = 0.981 [18], confirming high accuracy for rapid predictions. By contrast, study [20] proposed an ML approach that incorporates environmental conditions to deliver accurate, simultaneous forecasts for multiple vehicles while preserving user privacy.

Extending this direction, several works use hybrid ML models to estimate battery state of charge (SoC). Study [21] reports that a gradient boosting machine with Bayesian optimization (GBM-BO) outperforms other approaches for SoC prediction by leveraging energy-consumption patterns and battery capacity. These findings have practical implications for route planning and mitigating range anxiety, as shown in [19,39], where predictive models reached 81.11% accuracy [19] and R² = 0.944 [39] when topographic data were included.

Deep learning (DL), a branch of machine learning that relies on multi-layer neural networks, has also proved effective for EV energy-related forecasting. In [22], a long short-term memory (LSTM) model for SoC prediction achieved R² = 0.99, a level of accuracy vital for energy management. The results in [25] confirm that DL can capture the influence of driving style and acceleration on energy use. To reduce data requirements, ref. [26] introduced a deep transfer learning (DTL) and distributed-learning approach that improves the remaining-range estimation.

LSTM models are particularly effective for analyzing sequential data. In [23], an LSTM model trained on real driving cycles achieved very low errors (MAE = 1.76%, RMSE = 1.99%), supporting precise energy management. In [27], LSTM networks were used to predict an individual driver’s speed profile, which substantially facilitates range estimation. Furthermore, ref. [28] proposed an LSTM-based route-planning strategy that incorporates traffic data, yielding energy savings of 9.9% and a 40.2% reduction in travel time. Other studies, such as refs. [29,30], also employ LSTM and CNNs to forecast time series, leading to reduced energy consumption and improved safety.

Neural network algorithms play a crucial role in EV development by enabling complex tasks in control, forecasting, and energy management. In [31], a neural network model achieved 89% accuracy in predicting energy consumption. In [32], neural networks were used to forecast road conditions as part of an energy-management strategy, enabling the effective coordination of hybrid power sources. In [33], a convolutional neural network with an attention mechanism accurately estimated battery state of charge and capacity over aging; the model was verified in a hardware-in-the-loop (HIL) environment, confirming its practical utility. Similarly, ref. [34] showed that an artificial neural network (ANN) trained with various back-propagation algorithms produced a high correlation between the predicted and experimental values. In [24], a neural network model that accounts for dynamic driving conditions enabled accurate battery-range forecasting (R² = 0.91), outperforming standard manufacturer solutions. Overall, neural networks demonstrated strong robustness, reducing the total cost by 19.99% to 33.13% in [32], which underscores their value for EV energy-management systems.

Statistical methods, particularly Markov processes, support intelligent management of energy consumption and battery temperature, extending battery life and increasing vehicle energy efficiency while adapting to changing conditions. In [45], a cyber-physical vehicle energy-management system used a Markov decision process to forecast speed and optimize energy consumption in real time. The model learned interacting-system behavior dynamically, reducing energy use by 18% and battery-capacity loss by 12%, and achieving 86% of ideal-control efficiency. In [46], a model based on a Markov-chain probability matrix estimated future energy demand and tracked battery temperature, saving up to 4% of total vehicle energy consumption.

Hybrid ML algorithms combine statistical models, neural networks, and deep reinforcement learning (DRL) to improve forecast accuracy and capture complex data dependencies. Their advantage lies in handling system nonlinearities and temporal dependencies while adapting to changing environmental and operational conditions. In [40], several SoC-estimation methods were compared; a simple three-layer feedforward neural network (FFNN) achieved excellent precision (error < 1%, MSE = 1 × 10⁻⁸), making it suitable for devices with limited computational power. Study [41] introduced a hybrid EDOA-DLP model that couples metaheuristic optimization with deep learning to estimate SoC effectively, improving battery performance. Consistently, ref. [42] reported the high effectiveness of hybrid models, where an LSTM trained on real-world data reached R² ≈ 0.99 for SoC forecasting, supporting energy management with renewable sources.

Hybrid architectures such as NARX neural networks combined with a Kalman filter [43,44] have shown strong performance in predicting energy consumption under extreme weather and in estimating battery output power with errors below 1%. In [38], an optimized deep recurrent LSTM model (GA3P-DLSTM) used a genetic algorithm to fine-tune hyperparameters and, when tested on real-world data, achieved very high accuracy (MSE = 0.000112, R² = 0.96470), outperforming other models and confirming the potential of hybrid solutions.

Deep learning, particularly LSTM and ConvLSTM models, is often combined with other techniques for time-series analysis. In [35], a hybrid deep learning approach combining CNN and LSTM enabled an accurate forecasting of the remaining driving range. Trained on real-world driving data, the model captured dynamic speed variations and achieved high accuracy, with an error of only 3762 km over a 7.5 h test, significantly improving estimate reliability. In [36], a kernel extreme learning machine (KELM) optimized using the Whale Optimization Algorithm (WOA) improved speed-forecasting accuracy, reducing the total operating cost by 6.15% and enhancing real-time efficiency by 97.89%. Study [37] applied syncretic learning (BP + RBF) to optimize vehicle speed trajectories at traffic-light intersections, reducing driving cost by 19.82% and travel time. These approaches combine the strengths of different neural network architectures, delivering both high precision and adaptability.

Hybrid ML algorithms, particularly those that combine predictive models, neural networks (e.g., LSTM, BPNN), and reinforcement learning methods, show strong potential for forecasting EV energy consumption. By integrating spatial, meteorological, and operational data, these approaches enable adaptive energy-management systems that remain resilient under changing conditions and are suitable for real-world deployment. LSTM-based deep-learning methods outperform ensemble models such as Random Forest in time-series forecasting because of their ability to capture temporal dependencies. However, in static contexts with limited data, ensemble methods provide greater robustness at a lower computational cost (e.g., R² = 0.981 vs. potential overfitting in [18,22]).

Table 2. Normalized comparison of forecasting algorithms in EV energy consumption prediction.

Algorithm Type	Used Algorithm	Synthesis of Reported Performance (Observed Ranges)—Key Dataset	Trade-Offs
Classical Models	Linear Regression	Reported R2 ≈ 0.96–0.981 in reviewed studies [17,18]; datasets: route-/lab-based (WLTP = standardized laboratory cycle)	High interpretability and very fast inference, but limited accuracy for nonlinear or time-dependent dynamics compared to DL
Ensemble Models	GBM-BO	Reported R2 ≈ 0.944 [39]; accuracy ≈ 81.1% [19]—values from separate fleet/telemetry studies	Robust to noise and competitive accuracy on tabular data; higher data requirements and longer training times vs. classical models
Deep Learning	LSTM, ConvLSTM	Reported R2 ≈ 0.91–0.99 [22,24]; MAE as low as ~1.76% [23] on sequential/real-driving datasets	Superior sequence/spatio-temporal modeling; risk of overfitting and high computational demands; more difficult to deploy on embedded hardware
Hybrid Models	GA3P-DLSTM, BP + RBF, LSTM + NARX	Reported R2 ≈ 0.944–0.965 [38,43]; error < 1% in some studies [40]—datasets include mixed real-world + synthetic/augmented conditions (e.g., weather)	Often best empirical performance, including lower error than pure DL; but complex implementation, tuning overhead, and reduced reproducibility
Statistical Methods	Markov Processes	Reported energy consumption reduction: ~4–18% in Markov-based studies [45,46]	Efficient adaptation to probabilistic transitions, interpretable and lightweight; limited ability to capture nonlinear, long-memory dynamics

summarizes the performance metrics and trade-offs of key algorithms for EV energy-consumption and state-of-charge (SoC) forecasting, enabling a direct cross-method comparison.

Classical models achieve R² ≈ 0.96–0.981 on WLTP and route data [17,18], offering fast inference but a limited ability to capture nonlinear dynamics compared to advanced methods. Ensemble approaches (e.g., GBM-BO) reach competitive performance (R² ≈ 0.944, accuracy ≈ 81.1% on fleet data [19,39]), providing robustness to noise but requiring larger datasets and longer training. Deep learning variants (LSTM, ConvLSTM) excel in sequential modeling (R² ≈ 0.91–0.99, MAE as low as 1.76% on real-driving data [22,23,24]) but face risks of overfitting and impose significant computational demands for embedded systems. Hybrid models (GA3P-DLSTM, LSTM + NARX) often deliver the strongest results (R² ≈ 0.944–0.965, errors < 1% on mixed datasets [38,40,43]), outperforming pure DL in accuracy but at the cost of greater implementation complexity and tuning challenges. Statistical methods such as Markov processes enable efficient probabilistic adaptation, reducing energy consumption by ≈ 4–18% [45,46], though they are less effective at capturing nonlinear or long-memory dynamics.

3.2. Energy Predictive Control in EVs

Predictive control uses forecasts of future states (e.g., speed, road conditions) to optimize vehicle operation in real time. Its main objective is to minimize energy consumption and operating costs while ensuring driving comfort and safety. Model predictive control (MPC) is a central tool in EV energy management because it allows dynamic adaptation of control strategies to changing vehicle and environmental conditions. A key strength of MPC is its ability to integrate forecasts of future kinematic parameters (e.g., speed) and external factors such as terrain, road conditions, and traffic-light patterns. By doing so, MPC optimizes energy consumption, which directly influences driving range and operating costs.

Table 3. Types of ML algorithms and their application in predictive energy control in electric vehicles.

Algorithm Type	Application	Effects	Key Features	Reference Papers
MPC + Deep Learning (LSTM, DRL, CNN)	Speed forecasting, energy management	Reduced costs and energy consumption	Precise forecasts, real-time optimization	[48,49,50,51,52]
MPC + ANN/BPNN	Speed and thermal management optimization	Improved energy efficiency, reduced forecast error	Integration with V2X/V2I, adaptability	[53,54,55,56,57,58,59]
MPC + BTMS	Adaptive cruise control, energy management	Energy savings, improved system collaboration	Dynamic adjustment to driving conditions	[53,54,55,60]
Hybrid Models (Q-learning, QDP, AF-CFFRLS-AEKF, MOR-MPC)	Optimization under uncertainty	Reduced energy consumption in variable conditions	Modeling of random parameters	[50,61,62]

summarizes the main algorithms and methods applied in predictive control for EVs, highlighting their applications, achieved outcomes, and key features.

A key research area in minimizing EV energy consumption is speed forecasting and the optimization of driving strategies. In [48], a hybrid model combining an LSTM network with MPC was proposed, accurately predicting speed and effectively managing battery state of charge. Other approaches also focus on energy optimization. For example, a model that integrates syncretic learning (BP and RBF neural networks) with a trapezoidal collocation algorithm minimized energy use at intersections, reducing total driving costs by 19.82%. In [53], deep reinforcement learning (DRL) was applied in a cooperative adaptive cruise control (CACC) algorithm that learns to anticipate the behavior of surrounding vehicles, thereby improving both comfort and energy efficiency. Communication technologies also play a significant role. In [63], a hybrid MPC model utilized V2X (Vehicle-to-Everything) communication data, including traffic light information, for energy demand estimation. This approach achieved a significant reduction in energy consumption: 30.4% compared to a standard algorithm and 13.5% relative to an existing MPC-based method, thus confirming the potential for reduced energy consumption when integrating advanced algorithms with infrastructure.

Another important application of MPC is vehicle platooning and interaction with other road users. Study [30] compared three autonomous control systems with V2I (Vehicle-to-Infrastructure) communication, showing their superiority over rule-based controllers in minimizing energy consumption and travel time. In [54], a hybrid MPC-DRL model optimized energy use by developing improved speed strategies, achieving simulated savings of 3.2%. A similar approach was described in [55], where a hybrid MPC-DRL architecture enhanced energy efficiency by combining optimization with adaptability, as confirmed in PGDrive simulations. Other studies, such as [64], focused on predicting energy consumption using extensive real-world driving datasets. This two-stage predictive model, which employed advanced feature engineering, outperformed traditional deep learning methods. Meanwhile, ref. [65] tested hybrid solutions in a CarSim (Mechanical Simulation Corporation, Ann Arbor, MI, USA) environment, showing energy-consumption reductions from 1% on uphill segments to 17% on routes with multiple ascents and descents, confirming their practical applicability.

An additional approach was presented in [66], where a model for estimating EV energy consumption considered the longitudinal wheel slip ratio. The model employed a dynamic performance map based on machine learning to determine energy transfer between the motor and the battery.

The integration of predictive control with prognostic models for state of charge (SoC) and state of energy (SoE) is a key direction in optimizing EV energy management. In [36], a multi-objective MPC strategy (MOR-MPC) was developed that simultaneously minimizes energy consumption and battery degradation. This model, which uses ML-based speed prediction, reduced costs by 6.15% and achieved 98.17% optimization efficiency, comparable to dynamic programming. In [43], a hybrid AFCFFRLS-AEKF model combined an adaptive fuzzy system with a Kalman filter to accurately estimate SoE while accounting for temperature effects. The model achieved an estimation error within 1%, significantly improving safety and reducing range anxiety. Other studies [56,57] also confirm the effectiveness of MPC and DL in optimizing energy use in hybrid and electric vehicles, with forecast errors of only 0.1–0.3 kWh. In [49], an innovative energy-management strategy (EMS) employed a backpropagation neural network (BPNN) to approximate an MPC controller, improving real-time performance. Similarly, ref. [67] introduced an adaptive and cooperative quality-aware control design (ACQUA), which optimized system cooperation, reducing energy consumption by 18% and battery-capacity loss by 12%, proving the practical effectiveness of advanced strategies.

Predictive control is increasingly applied in battery thermal management systems (BTMSs), which lowers energy consumption and reduces the risk of power derating. In [61], an MPC-based BTMS strategy with a Markov speed predictor reduced energy use by 5.4% and battery degradation by 3%. In [68], an ML-based BTMS control model optimized cooling and heating power, cutting energy consumption by 35% compared to standard state controllers and by 60% relative to traditional PID controllers. In [58], a hybrid recurrent neural network (RNN) combined with nonlinear MPC reduced energy use by 6.23% while mitigating power-derating effects. In [59], a CNN trained on simulated and real-driving data predicted battery temperature with high accuracy, achieving a mean absolute error (MAE) of only 0.27 °C. Studies [69,70,71] further confirmed the potential of predictive BTMSs by modeling the end-of-discharge temperatures of lithium-ion batteries using real-world data. Ref. [72] proposed a cloud-enabled predictive-control framework that integrates road and traffic data to optimize both the thermal management of the battery and the prediction of energy demand. In real-world simulations, this approach reduced battery-cooling energy consumption by at least 8.58% and 10.31% compared to benchmarks.

An increasing number of studies have integrated ML methods with predictive control to minimize energy consumption and operating costs. Compared with classical MPC, these approaches allow dynamic adaptation under uncertainty while reducing computational load. In [69], a three-layer artificial neural network trained on data optimized with Pontryagin’s Minimum Principle (PMP) reduced operating costs by 19.99–33.13% compared to conventional strategies. In [73], a generalized regression neural network was applied in a hierarchical energy-management strategy, optimizing consumption and extending system life. Other studies, such as [60], combined MPC with Q-learning to reduce operating costs and battery degradation. In [62], stochastic MPC was applied to adaptive cruise control (ACC), cutting energy use by 13% under stop-and-go conditions. In [50], quadrant dynamic programming (QDP) was used to optimize ACC, reducing consumption by 16.1% while maintaining performance. This approach improved computational efficiency and suitability for real-time applications.

Predictive control in EVs, based on forecasts of future vehicle and environmental states, forms the foundation of modern energy-management strategies. By integrating DL, reinforcement learning, and stochastic models, MPC enables the simultaneous minimization of energy consumption, battery degradation, and operating costs, while improving driving comfort and safety. Numerous studies show that hybrid approaches combining ML algorithms with MPC are among the most promising directions for future development, both in vehicle-level management and in integration with energy infrastructure. MPC hybrids perform particularly well under uncertainty due to predictive optimization, but in low-resource environments they can be less effective than simple ANN models, trading accuracy for adaptability (e.g., 6–33% savings vs. real-time constraints in [36,69]).

Table 4. Synthesis of predictive control algorithms in EV applications.

Algorithm Type	Synthesis of Reported Performance (Observed Ranges)—Key Dataset	Trade-Offs
MPC + Deep Learning (LSTM, DRL, CNN)	Reported energy/cost savings ≈ 3–30% on dynamic simulations reported on V2X-based datasets [48,54,63].	High precision under uncertainty (e.g., ~6% cost reduction vs. ANN) but substantial computational overhead and real-time deployment challenges [49,54].
MPC + ANN/BPNN	Reported error ≈ 0.1–0.3 kWh on hardware-in-the-loop and real driving cycles [56,57,69].	Good real-time performance and low latency; less adaptability than DRL hybrids (some studies report larger savings for hybrids, ≈19–33% [53,69]).
MPC + BTMS	Reported thermal-related savings ≈ 5–35% in BTMS simulation studies [61,68].	Effective thermal/energy management; limited generalization across varying operational/ambient conditions [61].
Hybrid Models	Reported reductions ≈ 6–16% in stochastic/ACC scenarios (e.g., Stop&Go) [50,62].	Handles stochasticity and adapts to traffic dynamics; higher algorithmic complexity, stability/validation, and implementation overhead compared to pure MPC [62].

summarizes the performance and trade-offs of key predictive-control algorithms for EV energy management.

MPC integrated with DL techniques (LSTM, DRL, CNN) delivers substantial energy and cost savings (≈3–30% in V2X simulations [48,54,63]), achieving high precision under uncertainty (e.g., ≈6% cost reduction compared with ANN) but at the expense of significant computational overhead [49,54]. In contrast, MPC combined with ANN/BPNN achieves low forecast errors (0.1–0.3 kWh on HIL and real driving cycles [56,57,69]) and offers good real-time latency, yet shows reduced adaptability compared to DRL hybrids, which provide larger savings (19–33% in selected scenarios [53,69]). MPC-enhanced BTMS strategies yielded targeted thermal savings of 5–35% in simulation studies [61,68], although their generalization across diverse operating conditions remains limited [61]. Hybrid approaches, such as Q-learning or stochastic variants, reduce operating costs by 6–16% in stochastic traffic scenarios (ACC and stop-and-go driving [50,62]), effectively handling randomness but at the cost of higher implementation complexity compared to pure MPC [62]. This comparative analysis highlights the need to balance accuracy improvements with practical constraints when selecting algorithms for EV control applications.

3.3. EV Integration with Charging Infrastructure and the Power Grid

The integration of electric vehicles with smart grids and charging infrastructure is essential for maintaining energy-system stability and ensuring efficient energy transfer. Many studies have examined the use of machine learning algorithms in EV energy-management systems and their integration with the power grid. These methods enable the dynamic, adaptive, and predictive management of energy resources at the level of both individual vehicles and entire power systems. By analyzing large datasets and extracting insights, ML models contribute to significant improvements in energy management, charging optimization, grid integration, and precise vehicle control.

Table 5. ML algorithms for integrating electric vehicles with charging infrastructure and the power grid.

Algorithm Type	Application	Effects	Key Features	Reference Papers
Deep learning	Forecasting charging demand	Reduced MAPE by 21.33%, and RMSE by 18.73%; R2 = 0.92	Modeling spatio-temporal dependencies, integration with RES	[51,52,74,75,76]
Regression models	Forecasting charging profiles and session duration	MAE 1.45 kWh, RMSE 6.68 kWh, R2 = 51.9%	Consideration of weather, traffic, and calendar data	[72,77,78,79,80,81,82,83,84]
Stochastic models	Reserve planning, energy management with PV	Minimized costs, extended battery life	Quantification of uncertainty, two-level optimization	[74,85,86]
Neural Networks	Modeling charging sessions, user behavior	Higher accuracy in predicting vehicle count and energy consumption	Capturing nonlinear patterns, scalability	[82,87,88,89,90,91,92,93,94]
Reinforcement Learning	Scheduling charging in microgrids and V2G	Reduced energy costs, improved grid stability	Adaptability to variable conditions, real-time decisions	[95,96,97]
Other ML	Microgrid optimization, charging monitoring	Reduced peak power, real-time fault detection	Two-level architectures, safety monitoring	[86,98,99,100,101,102,103]

presents the application of ML algorithms in EV integration with charging infrastructure and the power grid, highlighting their effectiveness in demand forecasting, charging-schedule optimization, and microgrid and V2G management.

3.3.1. Forecasting Charging Demand

A key challenge in the development of electromobility is the intelligent management of charging infrastructure. ML models provide tools for spatio-temporal demand forecasting, enabling proactive control of energy flows. Convolutional–recurrent models, such as ConvLSTM, and temporal encoder–decoder architectures are particularly effective for this task. For example, ref. [74] presented a model that combined a Markov chain with a Monte Carlo random algorithm to generate training data for a ConvLSTM network, allowing hourly charging-demand predictions in cities with limited historical data. In [51], a T-LSTM-Enc model trained on data from Chinese charging stations effectively forecast demand, reducing the grid load. Other studies have emphasized integration with renewable energy sources. In [52], an AI-based model achieved R² = 0.92 when forecasting energy production for charging stations powered by wind and solar, significantly outperforming traditional methods. Meanwhile, ref. [75] applied a Local Attention Recurrent Convolutional Neural Network (LA-RCNN), which reduced MAPE by 21.33% and RMSE by 18.73%, improving charging-load forecasts. In [76], a deep supervised-learning model achieved high accuracy in medium-term demand forecasting, enabling more effective management of public charging stations.

Supervised regression methods such as Random Forest, XGBoost, SVR, and GPR have also proved effective for predicting daily and hourly charging profiles, especially when weather, calendar, and traffic data are included as features. For instance, ref. [77] employed ensemble models (Random Forest, SVM, XGBoost) to forecast energy consumption and charging-session duration. By incorporating additional data such as weather forecasts and event information, the models achieved high accuracy, with MAE = 1.45 kWh for energy consumption and 66.5 min for session duration. In [78], XGBoost regression was applied to forecast charging demand for distribution-network management. Trained on seven years of data, the model achieved the best performance among those tested, with RMSE = 6.68 kWh and R² = 0.519. This confirms that historical user consumption is a key predictive factor.

For reserve planning and peak-demand risk assessment, stochastic approaches using Markov chains or probabilistic models are particularly useful. These methods quantify uncertainty and can be integrated with bilevel optimization (e.g., home charging with PV). For example, ref. [85] proposed a stochastic MPC model for managing energy in homes with photovoltaic panels and EVs. The model, which combines a Markov chain with an adaptive neuro-fuzzy inference system (ANFIS), minimizes energy costs and extends PEV battery life by avoiding unnecessary charging and discharging cycles.

Integrating large numbers of EVs requires advanced demand forecasting, intelligent charging scheduling, and coordination with renewable energy sources. ML models such as ConvLSTM and T-LSTM-Enc effectively capture spatio-temporal dependencies, achieving superior short- and medium-term forecasting compared to classical time-series methods. Regression methods (Random Forest, XGBoost, SVR, GPR) maintain high accuracy when enriched with weather, calendar, and traffic data, with errors of MAE = 1.45 kWh and RMSE = 6.68 kWh. Stochastic models using Markov chains and probabilistic approaches provide robust uncertainty quantification and support bilevel optimization, for instance, in PV-powered home charging, thereby reducing costs and extending battery life.

3.3.2. Modeling Charging Sessions and EV’s User Behavior

Accurate modeling of charging session duration, start time, and energy consumption distribution within a single session is a key element for assessing the impact of electric vehicles on the power grid and for optimizing the location and sizing of charging stations. The literature emphasizes that such models enable not only accurate forecasting of grid loads but also better adaptation of infrastructure to the actual usage patterns of EVs in a given area.

To estimate session duration and energy consumption based on transactional data, models based on regression and decision trees are widely used. They are characterized by low computational complexity, which makes them useful in designing charging infrastructure. For example, in [79], four regression models were applied to forecast energy demand, achieving better predictive accuracy than in other studies. In [80], a model combining regression with time series was proposed, which achieved high accuracy in predicting charging time, helping drivers with trip planning. In [81], various regression algorithms were tested, with SGDRegressor showing the highest effectiveness in forecasting energy demand, thereby facilitating network management. In [82], several regression models were also utilized, including Gaussian Process Regression (GPR) and neural networks, to forecast daily energy demand and the number of vehicles being charged. This study showed that GPR performs best in forecasting energy consumption, while neural networks are more effective in predicting the number of vehicles, which indicates the complementary nature of these methods in optimizing energy management.

More advanced approaches for forecasting electric vehicle charging demand use neural networks (ANN, CNN) and sequential models (LSTM, GRU). They are capable of capturing the complex, nonlinear behavioral patterns of users, which is particularly important in conditions of high variability, such as in large cities [87] or on highways [88]. For example, ref. [89] proposed a hybrid model that combines probabilistic methods with neural networks for an accurate forecasting of charging demand, providing crucial information for infrastructure planning. In [90], a model based on an ANN was used to forecast future energy consumption at charging stations in Brazil, demonstrating the economic viability of such investments. In [91], a classifier with a CNN architecture was proposed, which achieved higher accuracy in predicting user behavior than previous ML algorithms. In [92], a TA-SSA-LSTM model was presented that combines an LSTM network with optimization algorithms, enabling accurate hourly forecasting of energy consumption at charging stations. Because of their ability to analyze temporal and contextual dependencies, these models provide more precise forecasts in highly variable scenarios, supporting the reduction of peak energy consumption and the optimization of charging schedules.

Hybrid frameworks, such as the combination of ConvLSTM networks with Markov chains [74] or Monte Carlo simulations, further enable the capture of spatio-temporal dependencies and uncertainties related to user behavior. These approaches increase the scalability and portability of models, allowing them to be adapted to different geographical conditions and network types. They therefore support the strategic, long-term planning of charging infrastructure. For example, ref. [93] proposed an EV charging load forecasting model that accounts for energy consumption during driving by relying on speed prediction using a BP neural network. This model also uses a fuzzy inference system (FIS) to determine the probability of charging and successfully predicts load distribution curves, as confirmed by the Monte Carlo simulation results. In [104], an ensemble model was applied to forecast session duration and charging demand, and Random Forest regression was used to predict energy consumption. The analysis showed that these models, which included additional features such as traffic and weather conditions, achieved high accuracy, with an MAE of 56 min for session duration and 1.36 kWh for energy consumption. Study [83] focused on predicting EV charging demand using machine learning algorithms such as SVM and XGBoost. The proposed models achieved high accuracy, supporting effective charging infrastructure management.

Modeling electric vehicle charging sessions is therefore crucial for forecasting grid loads and planning infrastructure. Many studies demonstrate the use of ML algorithms that consider charging demand, charging patterns, and geographical conditions to predict the number of stations and energy consumption for planning the distribution of charging stations [105,106,107,108].

Simple models like regression and decision trees are effective and computationally efficient, while more advanced neural networks (ANN, CNN, LSTM) and hybrid models are able to capture nonlinear behaviors and spatio-temporal dependencies. Supported by transactional and contextual data, these solutions increase scalability and facilitate strategic network management.

3.3.3. EV Integration with Microgrids and V2G Systems

The integration of electric vehicles with microgrids and the development of vehicle-to-grid (V2G) services are key research areas for effectively utilizing the potential of EVs as distributed energy resources. Achieving this requires the development of algorithms that combine demand forecasting, energy flow optimization, and real-time control.

Some of the most frequently used methods are Reinforcement Learning (RL) and Deep Reinforcement Learning (DRL), which are often combined with model predictive control (MPC) in hybrid architectures. Their goal is to learn policies for scheduling charging and managing energy accounting in highly uncertain environments, such as shared fleet operations or energy aggregators [95,96]. Owing to their ability to adapt to variable conditions, these approaches allow for the development of flexible strategies that respond to fluctuations in energy prices and the changing availability of renewable sources. For example, ref. [98] used a machine learning algorithm to predict the uncertainty of EV energy consumption, which enabled the development of a charging and discharging optimization model for a large fleet. As a result, the costs associated with driving, charging, discharging, and battery wear were reduced.

Bilevel architectures are also increasingly used. They combine decisions made at the local level (individual vehicles as agents) with systemic decisions made by microgrid operators. For example, ref. [99] proposed a two-stage deep learning model aimed at reducing energy costs in smart microgrids (MGs) integrated with electric vehicles. This model, by forecasting the grid state and optimizing the EV charging schedule, effectively reduced overall energy consumption and the fees paid to the external grid. Similarly, ref. [86] presented a model that combines a Markov chain, an Adaptive Neuro-Fuzzy Inference System (ANFIS), and MPC. It was shown that the proposed model effectively optimizes charging, maintains the desired battery energy level, and improves system performance under uncertain conditions.

Various methods are applied in research on the integration of electric vehicles with the power grid. They allow for the simultaneous optimization of operating costs and the reduction of battery degradation while maintaining overall system stability. For example, ref. [100] analyzed a machine learning model based on the Random Forest algorithm to predict key parameters for EV charging in residential settings. Using historical user data, the model achieved moderate accuracy (R² = 0.40–0.48) in forecasting session duration and energy demand. This represents a significant improvement over traditional statistical models but also indicates the highly random nature of user behavior. In [101], an energy management model based on Mixed-Integer Linear Programming (MILP) was proposed to optimally plan energy consumption in residential areas. This model, which considers microgrids, renewable energy sources, and electric vehicles, effectively reduces energy costs and peak demand. Meanwhile, ref. [102] presented an imitation learning model based on a Deep Neural Network (DNN) to economically schedule EV charging in households with smart management systems and photovoltaic installations. In [97], a flexible approach to energy management in smart grids was proposed using deep learning and reinforcement learning. The proposed method optimizes EV charging schedules, improving both energy efficiency and grid stability. In [109], a model for forecasting power grid states for regional energy markets was introduced using deep neural networks. The results showed an improved prediction of energy demand, supporting the integration of EVs with the grid. Similarly, ref. [110] applied machine learning to forecast energy consumption in households with electric vehicles, taking into account various usage parameters. The results support energy management planning and enhance the efficiency of home systems.

To improve the safety of electric vehicles during charging, ref. [103] proposed a model for monitoring charging status and fire detection. This model is based on Convolutional Neural Networks (CNNs) and Bi-directional Gated Recurrent Units (BiGRUs). By using historical charging data, it effectively detects faults in real time and sends warning signals, providing timely, early fire alerts.

The integration of electric vehicles with microgrids and the development of V2G systems are therefore key research areas. Machine learning algorithms, such as RL, DRL, and their hybrids with MPC, are used to optimize charging and discharging schedules. These methods help to reduce energy costs, minimize battery degradation, and improve efficiency under variable conditions. Models based on Markov chains, ANFIS, DNN, Random Forest, and MILP have demonstrated effectiveness in managing fleets and microgrids, though with varying levels of forecasting accuracy. Deep learning models such as ConvLSTM handle spatio-temporal variability better than regression approaches in demand forecasting, but they underperform in scenarios with uncertain user behavior, where stochastic methods reduce errors but increase complexity (e.g., MAE = 1.45 kWh, as noted in [74,77]). In conclusion, the development of adaptive ML algorithms is essential for the effective integration of EVs with energy infrastructure, although further research into their scalability in real-world applications is still needed.

Table 6. Synthesis of ML methods for charging infrastructure and grid integration.

Algorithm Type	Representative Algorithms	Synthesis of Reported Performance (Observed Ranges)—Key Dataset	Trade-Offs
Deep learning (sequence/spatio-temporal)	LSTM, ConvLSTM, LA-RCNN	MAE often ≤2 kWh in some session/aggregate studies; relative improvements vs. baselines reported (for example, MAPE reduction of ≈21%, and RMSE reduction of ≈19% in [75]). Reported on hourly/station/city datasets [74,75,77].	Best for spatio-temporal patterns; high data and compute needs, risk of overfitting, deployment challenges.
Regression and ensembles	Random Forest, XGBoost, SVR, GPR	MAE ≈ 1.4 kWh—RMSE up to ≈6.7 kWh in session datasets [77,78]; robust on tabular/session data when enriched with features.	Fast, interpretable, easy to deploy; generally lower performance than DL for complex spatio-temporal tasks.
Stochastic/probabilistic	Markov chains, ANFIS, stochastic MPC	Demonstrated improvements in cost/reserve planning in PV-integrated/home scenarios; uncertainty quantification reported in [85].	Good uncertainty handling and reserve planning; requires calibration, probabilistic assumptions, and can be complex/slow.
RL and hybrids/other ML	DRL, Q-learning; ensembles + metaheuristics	Reported cost reductions and adaptive scheduling benefits in microgrid/V2G/pilot studies [95,97,98]; qualitative reports of peak shaving.	Powerful for sequential decision making and optimization; high training cost, sim-to-real transfer issues, and potential over-specialization to infrastructure.

presents a comparative synthesis of algorithm groups primarily used for EV charging-demand forecasting, session modeling, and microgrid/V2G integration, consolidating performance metrics across the reviewed literature.

Deep learning approaches, exemplified by LSTM and ConvLSTM, demonstrate a strong capacity to capture spatio-temporal dependencies. This advantage is reflected in MAE values often below 2 kWh and relative error reductions of up to 21% in MAPE when applied to hourly or station-level datasets [74,75,77]. However, these methods require substantial computational resources and are more susceptible to overfitting compared to simpler models. Conversely, regression and ensemble methods, such as Random Forest and XGBoost, offer higher interpretability and faster processing (MAE ≈ 1.4 kWh, RMSE ≈ 6.7 kWh on session data [77,78]), although their performance decreases in highly complex and dynamic scenarios. Stochastic models, particularly Markov chains and ANFIS, excel at quantifying uncertainty in PV-integrated environments [85], delivering measurable cost improvements over deterministic baselines but requiring complex calibration. Finally, RL and hybrid ensemble architectures enable adaptive scheduling in V2G contexts, producing reductions in peak demand and cost savings [95,97,98]. These advanced paradigms, however, face challenges related to high training overhead and the transferability of simulation results to real-world conditions, which underscores the importance of task-specific algorithm selection and a careful balance between scalability and predictive accuracy.

4. Bibliometric and Statistical Analysis

This section presents a bibliometric and statistical analysis of the research articles selected for this review. The data were collected in four categories: algorithm type, document type, algorithm application area, and research methodology. A quantitative analysis of publications from 2016 to 2025 was carried out to assess research dynamics in the application of machine learning and deep learning for forecasting electric vehicle energy consumption. The distribution of algorithm types, categorized into five-year periods, is shown in Figure 5.

A significant increase in the number of publications was observed in the category of applied algorithms. Publications related to machine learning saw a negligible rise (only 3 papers) from 2016 to 2020. Crucially, the quantity then spiked in the subsequent years, reaching 37 papers. A similar trend was recorded for deep learning algorithms, which increased from 4 to 29 publications (87.9%), and for neural networks, which rose from 6 to 23 publications (79.3%). In contrast, the number of papers in the statistical algorithms category slightly decreased, suggesting a gradual replacement of classical techniques with modern, data-driven approaches.

Next, the application areas of ML algorithms were analyzed. The classification was based on keywords and referred to the domains in which the algorithms were applied. Four main categories were identified: Battery (management and characteristics), Charging (charging processes and demand), Power Grids (EV integration with the grid), and Predictive Energy Flow Control Strategies. Figure 6 presents the number of publications illustrating the application of ML algorithms in these areas.

From 2021 to 2025, there was a significant increase in the number of publications across all the analyzed application areas of ML algorithms in electric vehicles. The largest growth was recorded in the “Charging” category (an increase from 6 to 39 publications) and “Battery” (from 13 to 37), which underscores that these topics are currently the focus of researchers. A significant rise in the number of papers was also observed in the “Energy Networks” and “Predictive Control” areas, indicating the growing importance of EV integration with the grid and control optimization. The overall trend suggests that machine learning is becoming a key tool for solving complex problems related to electromobility.

Next, the types of publications in which machine learning algorithms were used to forecast energy consumption in electric vehicles were analyzed. The classification was based on bibliographic data from the Scopus database, distinguishing three types of works: journal articles, conference papers, and other documents, such as book chapters and monographs. Figure 7 illustrates the number of papers classified according to publication type (journal article, conference proceeding, other).

An analysis of the document types shows that the number of both conference papers and journal articles increased significantly after 2020. The number of conference papers rose from 9 to 29, accounting for 76.3% of the total in this group, while the number of journal articles increased from 9 to 47 (83.9%). Only one publication was classified as “Other” (e.g., a chapter in a monograph), which represents a symbolic 1.05% of the total. The significant growth in the number of publications, especially in scientific journals, attests to the increasing interest of the academic community in the topic of forecasting energy consumption in electric vehicles using machine learning methods.

The next stage of the analysis focused on research methodology, which provided a deeper understanding of the nature of the work being conducted. To this end, the papers were classified into four categories: experimental studies, literature reviews, case studies, and conceptual works. This division illustrates the extent to which researchers are moving from theory and reviews toward practical applications and implementations. The number of papers according to their research methodology, categorized by 5-year periods, is presented in Figure 8.

An analysis of the research methodology indicates that the vast majority of publications were experimental studies, a total of 83 papers, of which 68 were published between 2021 and 2025 (81.9%). Conceptual works also showed strong growth, increasing from 15 to 41 publications (73.2%). The number of case studies rose from 2 to 12, and literature reviews from 0 to 9. These data demonstrate that after 2020, research on ML and DL in electric vehicles became more application-oriented, accompanied by a growing interest in comparative analysis and the conceptualization of new solutions. The emergence of “Literature Analysis” papers further indicates the maturation of this research field.

Table 7. Summary of the number of publications in the selected categories in the 5-year periods.

Category	Name	2016–2020	2021–2025	All Years	Share [%]
	Total	18	77	95	100.0
Document type	Conference paper	9	29	38	40.0
	Article	9	47	56	58.95
	Other	0	1	1	1.05
Algorithms	Machine Learning	3	37	40	42.11
	Deep Learning	4	29	33	34.74
	Neural Networks	6	23	29	30.53
	Statistical algorithms	8	7	15	15.79
Energy and control	Battery	13	37	50	52.63
	Charging	6	39	45	47.37
	Energy networks	3	27	30	31.58
	Predictive control	6	21	27	28.42
Research methodology	Experiment	15	68	83	87.37
	Literature analysis	0	9	9	9.47
	Case study	2	12	14	14.74
	Conceptual	15	41	56	58.95

presents a summary of the number of publications in each category. The percentage shares confirm the dominant position of journal articles (58.95%) over conference papers (40%) and the clear predominance of experimental studies (87.37%) over other research types. In terms of content, the application of ML algorithms is most prevalent in topics related to batteries (52.63%), charging (47.37%), and power grids (31.58%). This reflects the increasing complexity and interdisciplinary character of research on forecasting energy consumption in electric vehicles.

To confirm the significance of the observed differences, a chi-squared (χ²) test was conducted across the four main classification categories. The results indicate statistically significant differences only in the “Algorithms” category (χ² = 16.63, df = 3, p < 0.001), meaning that the structure of the algorithms used changed significantly during the study period. In the other categories (document type, p = 0.58; energy area, p = 0.22; research methodology, p = 0.22), no statistically significant differences were found between the 2016–2020 and 2021–2025 periods, even though the quantitative data show increases. This may suggest uniform development in all areas except for the selection of algorithmic techniques, which are subject to the fastest technological and adaptive changes. To identify the regional research centers active in this area, the number of publications was analyzed based on the authors’ affiliations. The Figure 9 shows the number of publications, divided according to the authors’ countries of affiliation, within the defined five-year periods.

A geographical analysis of research activity indicates a clear and dynamic increase in publications between 2021 and 2025. Most publications originated in China (28 papers), with 23 (82.1%) appearing in the second period, accounting for 29.47% of the total. India follows with 17 publications, all of which were published after 2020, demonstrating strong engagement in this field. The next leading countries are the United States (11 publications), Canada (6), Germany, South Korea, and Spain (with 5 each), as well as Saudi Arabia, the United Kingdom, and France. The share of other countries grouped in the “Other” category is 16.84%, reflecting dispersed international activity.

Overall, 81.1% of all publications (77 out of 95) were produced during the 2021–2025 period (

Table 8. Summary of publications according to authors’ countries of affiliation.

Country	2016–2020	2021–2025	All Years	Share [%]
All countries	18	77	95	100.0
China	5	23	28	29.47
India	0	17	17	17.89
United States	3	8	11	11.58
Canada	0	6	6	6.32
Germany	2	3	5	5.26
South Korea	0	5	5	5.26
Spain	0	5	5	5.26
Saudi Arabia	0	4	4	4.21
United Kingdom	1	3	4	4.21
France	2	1	3	3.16
Other	6	10	16	16.84

). Notably, 7 of the top 11 countries showed activity exclusively in this period, underscoring that the intensification of research on the integration of machine learning and deep learning with electromobility is a relatively recent phenomenon, driven by the development of EV infrastructure and R&D strategies.

According to the results of the chi-squared test (χ² = 19.24, df = 10, p = 0.04), the differences in the number of publications between countries are statistically significant at the α = 0.05 level. This means that the changes in research activity across the analyzed countries are not random but reflect real differences in research intensity and the advancement of implementation work in the field of electric vehicles. The analysis revealed that most of the publications were authored by researchers from China, with a substantial contribution from authors based in India. These findings highlight the significant role of Asian research institutions in advancing the application of machine learning algorithms in electric vehicle technology.

The distribution of publications by algorithm type and across the four main application areas in “energy and control” is shown in Figure 10. The largest number of papers was published in the Battery area, where neural networks (20 papers) and machine learning algorithms (17 papers) were applied. In the Charging category, most research focused on the application of machine learning algorithms (21 publications). In the Energy networks area, machine learning (16 papers) and deep learning (13 papers) algorithms clearly dominated. In the Predictive control category, machine learning and deep learning algorithms were most frequently used (10 publications each).

The results of the chi-squared test (χ² = 9.32, df = 9, p = 0.41) indicate that the relationship between algorithm class and energy area is not statistically significant. This suggests a uniform distribution of research preferences.

An analysis of algorithm type in relation to paper type (Figure 11) shows that the vast majority of experimental studies applied machine learning (35 publications) and deep learning (30 publications) algorithms. Conceptual works most often focused on neural networks (22 publications). Literature reviews most frequently referenced classical machine learning algorithms. Case studies illustrated the practical application of ML methods. The chi-squared test for this matrix (χ² = 8.71, df = 9, p = 0.46) does not confirm statistically significant differences, which indicates that the choice of algorithm is not strongly dependent on the type of research methodology.

A bibliometric and statistical analysis revealed the dynamic development of research on the application of machine learning and deep learning in estimating and forecasting energy consumption in electric vehicles, particularly after 2020. The number of publications more than tripled, and a chi-squared test confirmed a statistically significant shift in interest toward more advanced algorithms, while the share of classical statistical methods remained low. The dominant thematic areas are batteries, charging, and power grids, with experimental studies prevailing. Publication activity is clearly concentrated, and a chi-squared test confirmed significant geographical differentiation, highlighting the dominance of China and India. Despite these trends, the absence of significant dependencies between algorithm type and either application area or research methodology suggests that the choice of research tools remains flexible and is determined by the specifics of the issue being analyzed.

5. Answers to Research Questions

A systematic review of the scientific literature from 2016 to 2025 was conducted to analyze the applications of machine learning algorithms in forecasting energy consumption in electric vehicles. This comprehensive analysis provides an overview of the current state of knowledge, identifies research gaps, and outlines directions for future work.

This review focused on seven key research questions. The first asked the following: What machine learning algorithms are most frequently used to forecast energy consumption and state of charge (SoC) in electric vehicles, and what are their key applications? The literature indicates that the most commonly used ML algorithms for forecasting EV energy consumption and SoC are ensemble models (e.g., Random Forest, XGBoost), deep learning (LSTM, GRU, CNN, ConvLSTM), regression models (SVR, GPR, linear regression), and hybrid models that combine various techniques (e.g., MPC + DRL, GBM-BO) [16,17,18,19,20,21,22,23,24,25,26,27,28,29,30,31,32,33,34,35,36,37,38,39,40,41,42,43,44,45,46,47,111]. Ensemble models achieve high accuracy (R² up to 0.981 in [18]) in static forecasts, especially for fleet charging planning [16,20,21]. Deep learning, particularly LSTM and ConvLSTM, dominates in dynamic scenarios such as forecasting time series for speed and SoC (R² = 0.99 in [22], MAE = 1.76% in [23]), supporting real-time energy management [22,23,24,25,26,27,28,29,30,31,32,33,34,35]. Hybrid models that integrate DRL with MPC are used in predictive control and V2G optimization [48,54,55,97,100,101,102,109,110]. Stochastic approaches based on Markov chains are effective for managing uncertainty [45,46,62,77].

The second question asked the following: What factors (e.g., driving style, topography, temperature) are considered in ML models for forecasting energy consumption in electric vehicles? ML models consider a range of influential factors, including vehicle speed, road grade, load, driving style (e.g., sporty, eco, urban), ambient temperature, weather conditions, and traffic [16,17,18,19,20,21,25,39,47,75,76]. For instance, in [17], a multidimensional regression model that incorporated road grade and stop density explained 96% of the variability in energy consumption for BEB buses. Topographical and meteorological data improve forecast accuracy (R² = 0.944 in [39]). In [78], historical user energy consumption was identified as a key predictor (RMSE = 6.68 kWh). Models such as ConvLSTM and LA-RCNN further integrate spatio-temporal dependencies, increasing precision in complex urban scenarios [51,62].

The third question was the following: What are the opportunities and challenges associated with integrating electric vehicles with charging infrastructure and smart energy grids using ML algorithms? ML methods such as ConvLSTM, T-LSTM-Enc, Random Forest, and MPC + DRL hybrids enable the precise forecasting of charging demand (R² = 0.92 in [75], MAE = 1.45 kWh in [77]) and schedule optimization in V2G systems and microgrids [86,95,96,97,98,99,100,101,102,109,110]. These models support reductions in energy costs (e.g., 19.99–33.13% in [69]), minimize battery degradation [91,105,106,107,108], and facilitate integration with renewable energy sources [74,100]. Challenges include the variability of user behavior (R² = 0.40–0.48 in [97]), limited historical data in new locations [62], and the high computational requirements of DL models [50,51]. Stochastic approaches (e.g., Markov chains in [77]) handle uncertainty effectively but require complex calibration.

The fourth question asked the following: What are the most common limitations and challenges related to validating ML/DL models under real-world conditions? The analysis indicates limitations such as a lack of standardized datasets and benchmarks, which hampers comparability of results [16,22]. Low model generalizability across heterogeneous scenarios (e.g., different driving cycles and weather conditions) and susceptibility to overfitting on small datasets are significant challenges [10,32]. Despite high accuracy (e.g., R² = 0.99 in [22]), DL models often require substantial computational resources, limiting their use in EV BMS systems [23,24]. In addition, drivetrain specifics [84,94,112], randomness in user behavior (e.g., [37]), and the difficulty of modeling extreme conditions (e.g., subarctic environments in [43]) can undermine forecast reliability in real-world applications.

The next question asked the following: What are the dominant research directions and areas requiring further analysis? Dominant directions include the development of hybrid ML models (e.g., MPC + DRL, ConvLSTM + Markov) for predictive control and V2G [48,54,86,95,96,98,99], integration with V2X/V2I technologies [55,63], and the optimization of battery thermal management systems (BTMS) [58,59,61,68]. The areas requiring further analysis include improving model generalization, standardizing evaluation metrics, and developing methods with lower computational complexity [40,69]. Increasing model scalability under limited data availability and integrating these models with renewable energy sources in microgrids are also crucial [74,101].

The sixth question asked the following: Does the distribution of algorithm classes (machine learning, deep learning, neural networks, statistical algorithms) differ significantly between the 2016–2020 and 2021–2025 periods? The analysis revealed statistically significant differences in popularity between the two periods (χ² = 16.63, df = 3, p < 0.001). Machine learning grew dynamically from 3 to 37 publications and is now the dominant category. Deep learning and neural networks also gained importance, confirming growing interest in advanced techniques. At the same time, classical statistical algorithms declined in popularity, suggesting their replacement by newer ML and DL approaches.

The seventh question asked the following: Is publication activity in the analyzed area evenly distributed among countries, or is it concentrated in selected centers? This review found that research on machine learning and electric vehicles is not evenly distributed globally. A geographical analysis confirms that China is the dominant research hub, accounting for nearly 30% of publications, with dynamic growth after 2020. India ranks second with 17 papers published between 2021 and 2025 (Figure 8 and Figure 9). The statistically significant concentration of publications in these countries (χ² = 19.24, p = 0.04) indicates a shift of global research activity toward Asia.

6. Discussion and Conclusions

This literature review reveals the broad and dynamic application of predictive machine learning algorithms in electric vehicle technology. Ensemble methods, such as Gradient Boosting Machine, Support Vector Regression, and Random Forest, dominate predictive tasks, including estimating the state of charge (SoC) and forecasting charging demand. Their advantage lies in their ability to model nonlinear relationships in large datasets, which enables high accuracy in energy consumption forecasting (as shown in [18] with R² = 0.981). In turn, deep learning techniques, including Long Short-Term Memory (LSTM) and Convolutional Neural Networks (CNNs), prevail in the analysis of sequential data, such as the dynamics of charging processes, vehicle speed profiles, and battery thermal analysis. For example, LSTM models achieve SoC estimation precision of R² = 0.99 in [22], which significantly improves real-time energy management. Hybrid configurations, such as LSTM integrated with model predictive control (MPC) or artificial neural networks (ANNs) combined with deep reinforcement learning (DRL), are particularly effective in complex scenarios. Examples include integration with vehicle-to-grid (V2G) systems or optimizing eco-driving strategies, where they can lead to energy consumption reductions of 3.2–30.4% [54,63]. A temporal analysis shows an evolution from simpler ML methods between 2016 and 2020 (for example, linear regression in [17]) to more advanced DL and hybrid models in 2021–2025. This reflects advances in computational power and the greater availability of data from sensors monitoring powertrains, batteries, and vehicle surroundings. This trend underscores the growing importance of real-world EV usage data in developing more adaptive systems. ML and DL algorithms demonstrate higher effectiveness than traditional statistical approaches such as Markov processes [45,46], achieving better results in time-series forecasting (for example, an RMSE improvement of 18.73% in [51]) and optimizing operating costs (for example, savings of 19.99–33.13% in [69]).

Discrepancies that remain an active research area concern the optimal algorithmic configurations for specific applications in EV energy management. For example, ensemble models are well suited to static forecasts of charging demand [75,76], whereas DL handles dynamic V2X data better [63]. This differentiation points to a key principle: selecting the appropriate algorithm depends on the specifics of the task, including the nature of the data (for example, sequential versus static), accuracy requirements (for example, MAE < 1.45 kWh in [77]), available computational resources, and specific optimization goals, such as minimizing battery degradation [61,68] or reducing grid load [50,51]. There is no universal solution, which underscores the need to adapt the methodology to the problem context, taking into account external factors such as topography, driving style, and weather conditions [17,39].

Most of the papers selected for this review were experimental studies or case studies. This enabled an in-depth understanding of many aspects of the proposed algorithms in controlled research environments, such as GT-Suite simulations [47] or CarSim tests [65]. However, the following limitations were observed:

• Most models rely on simulation or laboratory data, which means they have not been validated under real-world conditions such as variable traffic density or extreme weather. This reduces their reliability and applicability in practical EV systems. Examples of real-world applications include fleet trials in urban buses ([17,18], reducing consumption variability) and commercial V2G implementations ([95,96], improving grid stability), although broader industrial case studies, such as Tesla’s Autopilot integrations, remain underexplored in the literature;
• DL algorithms such as LSTM and CNNs are computationally complex and require significant hardware resources, which makes it difficult to implement them in low-power, onboard EV systems such as battery management system (BMS) modules;
• Results obtained in one study are difficult to compare directly with those from other publications because of discrepancies in testing protocols (for example, different driving conditions) and a lack of uniform datasets. This significantly hinders the reproducibility of research and meta-analyses.

Hybrid approaches (for example, ANN + DRL in [54]) show promise by offering adaptive, real-time control. However, their effectiveness for wider implementation, such as in commercial fleets, requires further empirical verification. Despite extensive research, the literature lacks comprehensive analyses of the impact of user behavioral data (for example, driving style and charging schedules) on predictive models. The effectiveness of these methods in less-developed regions, where EV infrastructure is limited (for example, in developing countries outside China and India), is also not well documented. There is a shortage of empirical data from field tests in diverse climates and usage scenarios, which hinders the generalization and transferability of results across regions.

Despite significant progress in applying machine learning algorithms to estimate and forecast energy consumption and state of charge in electric vehicles, an analysis of the literature from 2016 to 2025 reveals numerous research gaps. The studies analyzed highlight data-, model-, and implementation-related challenges that limit the practical application of ML in electromobility. This literature review uncovered several key research gaps:

• Lack of model generalization and challenges with limited data: Despite high accuracy (R² = 0.99), models are often tested under very specific conditions, which prevents their application in diverse scenarios (e.g., different vehicle types, driving styles, or weather conditions). There is a clear need for research on transfer learning and federated learning to create models that work effectively across heterogeneous vehicle fleets.
• Limited use of benchmark datasets: Commonly used datasets include WLTP-based datasets, which are standardized for laboratory testing but underrepresent real-world variability. Real-world driving datasets such as those described in [64] capture dynamic factors but are limited by privacy concerns and a lack of standardization. These differences highlight the need for open and shared repositories to enhance comparability.
• High computational requirements and lack of real-time scalability: Advanced models such as LSTM and ConvLSTM require significant computational resources. This limits their implementation in low-power, onboard battery management systems. There is a need to develop simplified, optimized models that can operate effectively in real time.
• Data uncertainty and insufficient consideration of external factors: Forecasting models often overlook dynamic, nonlinear dependencies resulting from weather conditions, traffic density, or user preferences. There is a need to integrate data from V2X and IoT systems. This would allow dynamic model adaptation to changing conditions and increase forecast accuracy, especially in the case of unpredictable events.
• Lack of standard evaluation metrics and the problem of interpretability. The diversity of metrics (R², MAE, RMSE) and heterogeneous datasets makes it difficult to compare algorithm effectiveness objectively. Additionally, the complex nature of advanced DL models makes it challenging to interpret which factors have the greatest impact on energy consumption.

The literature review demonstrated that ML algorithms, including ensemble models, deep learning, regression models, and hybrid models, play a key role in forecasting EV energy consumption and SoC, as well as in predictive control and integration with charging infrastructure and smart grids. Models such as Random Forest, XGBoost, LSTM, ConvLSTM, and MPC + DRL hybrids achieve high accuracy (R² up to 0.99, MAE up to 1.45 kWh) in various applications, from route planning to V2G optimization [16,17,18,19,20,21,22,23,24,25,26,27,28,29,30,31,32,33,34,35,36,37,38,39,40,41,42,43,44,45,46,47,48,49,50,51,52,53,54,55,56,57,58,59,60,61,62,63,64,65,66,67,68,69,70,71,73,74,75,76,77,78,79,80,81,82,83,85,87,88,89,90,91,92,93,104,105,106,107,108,111]. Factors such as driving style, topography, temperature, and traffic are also effectively considered, which improves forecast precision [17,39,76]. The integration of EVs with microgrids and V2G, supported by RL, DRL, and stochastic approaches, enables cost and battery-degradation reductions, although further optimization is needed [83,91,92,93,104,105,106,107,108]. The main challenges are limited generalization, high computational requirements, and the randomness of user behavior. This indicates a need for further research into scalable and adaptive models. The development of data standards, simpler architectures, and integration with renewable energy sources (RESs) and safety technologies are key directions for accelerating the adoption of sustainable electromobility.

Future research should focus on several key areas to accelerate the development of effective energy-management systems in electric vehicles. The priority is to develop adaptive online models that enable dynamic, real-time control, adjusting to changing environmental and behavioral conditions. Equally important is integration with V2X and Internet of Things (IoT) technologies, which will allow better energy management and optimization of communication between vehicles and infrastructure, as in smart grids. Attention should also be paid to precise battery-degradation forecasting that accounts for user behavioral data and variable operating conditions, including extreme temperatures and charging cycles.

Standardizing validation is also crucial for progress. This can be achieved by introducing uniform testing protocols, such as benchmarks based on real-world datasets from EV fleets, and by conducting reliable real-world tests, including long-term field studies. Furthermore, the development of edge AI and federated learning methods could address data-privacy and computational-resource issues, allowing secure sharing of models among vehicles without data centralization. The integration of EVs with microgrids and renewable energy sources requires further research on probabilistic hybrid models that combine Markov chains with DRL to balance energy in real time more effectively. Implementing these recommendations could contribute to increased energy efficiency, higher safety, and the sustainable development of electromobility on a global scale.

The results of this review are significant for engineers, researchers, and specialists working with electric drivetrains, related technologies, and the broader topic of charging. ML and DL models can support the optimization of charging infrastructure, reduce operating costs (e.g., by forecasting demand peaks), and facilitate integration with power grids. This is particularly relevant for V2G systems, where these models can help reduce battery degradation and improve grid stability. The high prediction accuracy reported for SoC and charging-demand tasks (e.g., battery-temperature prediction MAE = 0.27 °C in [59]) suggests significant commercial potential. However, implementation will require improving scalability, for example, through less complex hybrid models. Policymakers can use these findings to plan sustainable charging infrastructure, especially in regions with a high concentration of electric vehicles.

The main limitation of this review is its focus on selected algorithm classes (ML, DL, hybrids), which omits other approaches such as heuristic or evolutionary methods. This review was also limited to a single database (Scopus) and a specific time period (2016–June 2025). This may have excluded earlier foundational work or sources from other platforms, such as Web of Science. The selection of literature may have influenced the conclusions by limiting the analysis to publications indexed in that database. While focusing on this period reflects contemporary trends, it might have excluded older research with a long-term impact. Additionally, the classification of topics was challenging because of varied terminology (e.g., “energy management” vs. “charging optimization”), which required subjective interpretation and could introduce bias into the analysis. The authors are aware that this is a very broad and interdisciplinary field. Limiting the search to only technical domains could affect the conclusions reached.

Despite the aforementioned limitations, this review concludes that ML and DL methods play a crucial role in the development of electromobility, offering effective solutions for EV energy management. Their potential is significant, but it requires overcoming challenges related to validation, scalability, and data availability. This review indicates a dynamic evolution of these approaches and the need to adapt algorithms to specific applications, paving the way for future innovations in this field. To broaden the perspective, the authors plan to review literature from other databases, including Google Scholar and Web of Science, in future work.

In conclusion, this article provides a synthetic overview of current achievements and serves as a key starting point for further work aimed at the efficient and safe use of electric vehicles. The analysis presented is significant for several reasons:

• A rapid increase in interest in data-driven methods in electromobility was observed between 2016 and 2025, particularly in the fields of machine learning, deep learning, and neural networks. This highlights the dynamic development and crucial importance of this field for the future of transport.
• It was noted that hybrid models (such as LSTM + MPC, ANN + DRL) offer the highest effectiveness in forecasting and optimizing energy consumption, which is essential for the efficiency and reliability of electric vehicles.
• The main barriers and challenges in the field were identified. These include the limited availability of real-world data, the problem of model transferability, the lack of standardized validation procedures, and data-privacy concerns. Recognizing these challenges is the first step toward solving them and making further progress.
• Future research should prioritize the following: (i) adaptive online models validated on real-world fleets to address data scarcity (see the gaps identified in Section 4); (ii) V2X integration using probabilistic hybrid controllers, with potential for measurable cost reductions in specific deployment studies (cf. [63,74]); (iii) edge AI and model compression for a low-power BMS, with a recommended research target of reducing SoC error to below 1% under extreme conditions (a proposed target for future validation); and (iv) the development of standardized benchmarks and shared metrics for cross-study comparisons.

This work indicates that implementing the recommendations outlined here will contribute to improved energy efficiency, enhanced safety, and the sustainable development of electromobility worldwide.