Research Progress on State of Charge Estimation Methods for Power Batteries in New Energy Intelligent Connected Vehicles

Abstract

Accurately estimating the State of Charge (SOC) of power batteries is crucial for the Battery Management Systems (BMS) in new energy intelligent connected vehicles. It directly influences vehicle range, energy management efficiency, and the safety and lifespan of the battery. However, SOC cannot be measured directly with instruments; it needs to be estimated using external parameters such as current, voltage, and internal resistance. Moreover, power batteries represent complex nonlinear time-varying systems, and various uncertainties—like battery aging, fluctuations in ambient temperature, and self-discharge effects—complicate the accuracy of these estimations. This significantly increases the complexity of the estimation process and limits industrial applications. To address these challenges, this study systematically classifies existing SOC estimation algorithms, performs comparative analyses of their computational complexity and accuracy, and identifies the inherent limitations within each category. Additionally, a comprehensive review of SOC estimation technologies utilized in BMS by automotive OEMs globally is conducted. The analysis concludes that advancing multi-fusion estimation frameworks, which offer enhanced universality, robustness, and hard real-time capabilities, represents the primary research trajectory in this field.

1. Introduction

As the global energy crisis intensifies and environmental protection awareness grows, new energy intelligent connected vehicles are gradually becoming significant in the future of transportation [1]. According to Statista, the global market for new energy vehicles is projected to expand at a compound annual growth rate of 5.8% from 2023 to 2029, increasing from USD 769.4 billion to an estimated USD 1.08 trillion [2]. In this developmental context, the importance of various key technologies for new energy intelligent connected vehicles is becoming increasingly prominent, especially regarding power batteries. As the core component of these vehicles, the performance threshold of power batteries directly impacts the iterative process of industrial technology. Among these technologies, the dynamic estimation accuracy of the SOC is a critical technical barrier for BMS. As shown in Figure 1, as the decision-making center of the BMS, the real-time accurate estimation of SOC enables precise control of the battery, thereby optimizing and enhancing battery performance. This is crucial for providing information on remaining battery capacity and ensuring the efficient and safe operation of electric vehicles [3,4].

During the early development stages of intelligent connected electric vehicles, SOC estimation primarily relied on basic current integration and open-circuit voltage methods. These techniques were limited by underdeveloped battery technology, unstable performance, and inadequate sensing capabilities. As a result, they exhibited low accuracy, typically characterized by a Mean Absolute Error (MAE) of 8–15%, and lacked reliability, failing to meet industrial standards [5,6,7]. With advancements in technology, the Kalman Filter was introduced to effectively manage system noise and uncertainties, resulting in a reduction in Gaussian noise errors by 40–60% and achieving a SOC estimation accuracy below 3% Root Mean Square Error (RMSE) [8,9]. Concurrently, emerging methodologies, such as particle filters and Gaussian processes, have demonstrated improved precision (with MAE ranging from 1.5% to 2.8%) and stability in SOC estimation, offering advanced solutions for battery management systems [10,11]. In recent years, there has been a growing interest in data-driven approaches to SOC estimation. By extracting latent features from battery operational data, these methods show strong adaptability to complex operating conditions. However, their generalization capability is still limited by restricted training data coverage (typically covering less than 80% of the SOC range) and high annotation costs (exceeding USD 50/kWh) [12,13]. Although hybrid methods, such as model–data collaborative frameworks, aim to integrate complementary advantages, achieving an optimal balance between algorithmic complexity and real-time performance (with a latency of less than 100 ms) remains a critical and unresolved challenge.

In conclusion, the accuracy of SOC estimation is a key factor in enhancing the performance of new energy intelligent connected vehicles and optimizing user experience. However, most current research tends to focus on the optimization of specific types of algorithms or battery models, resulting in limitations in the comparative analysis and synthesis of these methodologies. This study systematically categorizes existing algorithms by integrating the latest research findings and conducts a comparative analysis of the complexity and accuracy among different algorithms, highlighting the challenges associated with each category. Additionally, it summarizes the applications and research achievements of automotive companies both domestically and internationally in the field of SOC estimation technology. Ultimately, this work identifies future research priorities and developmental directions for SOC estimation.

2. Power Batteries

In the process of energy transition, power batteries serve as the core energy storage devices for new energy smart connected vehicles. Various battery technologies feature distinct characteristics, determining their application scenarios and development prospects.

Lead-acid batteries, as the most traditional and technically mature type of power battery, have the significant advantage of low cost, abundant and easily obtainable lead resources, and well-established production processes that enable large-scale manufacturing within a complete industrial system. Meanwhile, nickel–metal hydride (NiMH) batteries have been widely used in early hybrid vehicles. Compared to nickel–cadmium (NiCd) batteries, NiMH batteries offer higher energy density, superior environmental performance, resistance to overcharging, and maintenance-free sealing. However, NiMH batteries still face challenges such as memory effect and the high costs associated with nickel, which results in elevated battery prices and limits their overall performance compared to lithium-ion batteries. Consequently, their application is largely restricted to specific vehicle models or niche markets. Fuel cells represent another promising technology, directly converting the chemical energy of fuel into electrical energy. They boast high energy conversion efficiency, environmental friendliness (with water as the only byproduct), a wide range of fuel sources, high power density, and rapid startup and response times. Theoretically, these advantages make fuel cells highly promising for new energy smart connected vehicles. However, practical challenges such as hydrogen storage and transportation difficulties, high costs, and concerns over durability and reliability significantly hinder their large-scale commercialization, limiting their use primarily to specific demonstration projects or high-end applications [14,15,16].

As battery technology continues to evolve, the limitations of traditional power batteries have become increasingly apparent, making it difficult to meet the growing demand for efficient energy storage. In this context, lithium-ion batteries have emerged as the most widely applied power batteries, standing out due to their high energy density (ranging from 250 to 350 Wh/kg), long cycle life (exceeding 2000 cycles), and low self-discharge rates (less than 5% per month). As illustrated in Figure 2, the multiscale design simulation framework for lithium-ion batteries integrates information from microscopic electrochemical reactions to macroscopic battery system management. This comprehensive modeling approach not only provides precise data support for battery design and the optimization of structures and materials, but also plays a vital role in battery management, enabling more efficient charge–discharge control and fault detection [17,18]. This advancement signifies that lithium-ion batteries are progressively overcoming the challenges faced by traditional power batteries, while facilitating the development of new energy smart connected vehicles.

The characteristic properties of lithium-ion batteries originate from their unique structural design, as illustrated in Figure 3. This architecture establishes the fundamental framework for ordered lithium-ion migration during charge/discharge processes. During charging, under external power supply, lithium ions in the cathode material undergo deintercalation reactions to escape the crystal lattice, migrating through the electrolyte to the anode. Concurrently, electrons are driven by the electric field through the external circuit to the anode. The anode, typically graphite-based, receives lithium ions that combine with incoming electrons to form intercalation compounds within graphite interlayers. During discharge, the battery functions as a power source. Lithium ions deintercalate from the anode composite, traverse the electrolyte to the cathode, while electrons flow from anode to cathode via the external circuit, delivering energy to the load. At the cathode, lithium ions recombine with electrons and reintercalate into the cathode lattice, completing the full charge–discharge cycle [19].

Taking lithium cobalt oxide (LCO) batteries as an example, the electrochemical reactions during cycling are expressed as follows.

Cathode:

$${\mathrm{LiCoO}}_2-{\mathrm{xLi}}^{+}-{\mathrm{xe}}^{-}\stackrel{\mathrm{Charging}}{\to }{\mathrm{Li}}_{1-\mathrm{x}}{\mathrm{CoO}}_2,$$

(1)

$${\mathrm{LI}}_{1-\mathrm{x}}{\mathrm{CoO}}_2+{\mathrm{xLi}}^{+}+{\mathrm{xe}}^{-}\stackrel{\mathrm{Discharging}}{\to }{\mathrm{LiCoO}}_2.$$

(2)

Anode:

$${\mathrm{xLi}}^{+}+{\mathrm{xe}}^{-}+6\mathrm{C}\stackrel{\mathrm{Charging}}{\to }{\mathrm{Li}}_{\mathrm{x}}{\mathrm{C}}_6,$$

(3)

$${\mathrm{Li}}_{\mathrm{x}}{\mathrm{C}}_6-{\mathrm{xLi}}^{+}-{\mathrm{xe}}^{-}\stackrel{\mathrm{Discharging}}{\to }6\mathrm{C}.$$

(4)

Despite the significant advantages of lithium-ion batteries, accurately measuring their SOC presents numerous challenges. These challenges primarily arise from the complex interplay of electrochemical and physical processes during charge and discharge cycles, as well as the influence of environmental variables such as temperature fluctuations (ranging from −30 °C to 60 °C), variations in C-rate (from 0.1 to 5C), and changes in pressure (between 80 and 120 kPa). Among these factors, temperature variations can affect the rate of internal chemical reactions and conductivity within the battery, thereby impacting charging and discharging efficiency. Furthermore, fluctuations in the C-rate directly influence the battery’s discharge characteristics and internal resistance, leading to a decline in the accuracy of SOC estimations. Additionally, pressure changes may affect the flow of the electrolyte and the contact area of the electrodes, further exacerbating SOC estimation errors. Given that SOC is a crucial indicator of remaining energy capacity, the precise estimation of SOC is essential for the optimal utilization and management of lithium-ion batteries. Consequently, the development of advanced SOC estimation methodologies has become a key focus of research, with current efforts aiming to achieve a MAE of less than 2% under real-world operating conditions.

3. SOC Estimation Methodologies for Power Batteries

State of Charge is defined as the ratio of remaining capacity to rated capacity under specified conditions [20,21]. Essentially, it represents the percentage of available energy relative to the maximum storable capacity. The United States Advanced Battery Consortium (USABC) established the following in 2009: SOC is the proportion of residual capacity to rated capacity under defined discharge rate conditions [22]. The mathematical formulation is expressed as follows:

$$\mathrm{SOC}={\displaystyle \frac{{\mathrm{Q}}_{\tau }}{{\mathrm{Q}}_{\mathrm{n}}}}\times 100\%.$$

(5)

For instance, consider a power battery with the rated capacity ${\mathrm{Q}}_n=100Ah$. If the measured and calculated remaining capacity is determined to be ${\mathrm{Q}}_{\tau }=60Ah$, the SOC value of this battery is calculated as follows:

$$\mathrm{SOC}={\displaystyle \frac{60}{100}}\times 100\%=60\%$$

(6)

In new energy intelligent connected vehicles, the accurate acquisition of SOC enables precise remaining range estimation by vehicle control systems. This prediction integrates historical energy consumption patterns (e.g., 5 km per 1% SOC consumption) with real-time SOC values. For instance, at 40% SOC, the estimated remaining range would be 200 km. This predictive capability holds critical significance for drivers in rational route planning and optimal utilization of charging infrastructure. Furthermore, SOC serves as the fundamental parameter for preventing battery overcharge and overdischarge. When approaching 100% SOC, precise current regulation is crucial to prevent hazardous overcharge-induced failures. Electrochemical analysis reveals that for NCM-based lithium-ion batteries (LiNiCoMnO₂ cathode/graphite anode), exceeding a terminal voltage of 4.3 V triggers irreversible electrolyte decomposition, with EC/DMC solvent decomposition rates exceeding 3.2 mmol/h at 45 °C. Thermal propagation modeling indicates that uncontrolled charging above a 1C rate accelerates the cell surface temperature rise beyond 1.5 °C per minute, a critical threshold for initiating thermal runaway. Pressure monitoring data from prismatic cells demonstrate that sustained charging above 95% SOC increases internal pressure by 120–150 kPa due to lithium plating and gas evolution. This situation necessitates the implementation of multi-physics coupled control strategies that integrate feedback from voltage, temperature, and pressure. Furthermore, maintaining SOC above critical thresholds (typically below 20%) through discharge current limitations or circuit interruptions effectively prevents aging mechanisms such as lead sulfate crystallization and the detachment of active materials, potentially extending battery service life by 30–50% [23,24]. This illustrates the vital importance of precise SOC estimation for ensuring operational safety and prolonging battery longevity in intelligent electric vehicles. To achieve enhanced accuracy, researchers have developed a variety of methodologies through ongoing innovation. A systematic review of recent advancements categorizes SOC estimation algorithms into three primary classes: experimental-based, model-based, and data-driven approaches, as depicted in Figure 4.

3.1. Experimental-Based SOC Estimation Methods

3.1.1. Discharge Test Method

The discharge test method is a widely used technique for estimating the SOC of a battery by discharging it to its cutoff voltage. This method involves calculating the remaining capacity of the battery by measuring the current and time during the discharge process [25]. While it provides advantages such as simplicity and high accuracy—typically within ±1% under stable temperature conditions—it also has notable drawbacks, including lengthy testing durations that often exceed four hours for a complete discharge. Furthermore, this method requires taking the battery offline, rendering it unsuitable for online real-time monitoring.

In laboratory settings, the discharge test method is commonly employed for battery testing and parameter modeling [26], as illustrated in Figure 5. In these controlled environments, researchers can accurately assess a battery’s characteristics and validate mathematical models that predict its performance. However, the operational discontinuity that arises from this method poses significant limitations in practical applications that require an uninterrupted power supply, such as electric vehicles, where in situ SOC estimation is essential. Therefore, although the discharge test method has substantial value for battery evaluation, integrating it with other methods is crucial for the effective monitoring and management of battery status in real-world applications.

3.1.2. Coulomb Counting Method

Coulomb counting, an evolution of the discharge test method, is particularly well-suited for real-time SOC monitoring in applications that require continuous operation, such as BMS for electric vehicles. This method estimates SOC by calculating the accumulated charge over a specified time interval and dividing it by the total battery capacity to determine the variation in SOC. This delta value is then added to the initial SOC to derive the final estimation [27], as illustrated in Figure 6. However, the Coulomb counting method is inherently prone to errors, accumulating inaccuracies of over 5% after just 10 charge–discharge cycles due to prolonged data integration and noise from current sensors (±0.5% full scale). Its accuracy is further limited by the necessity of knowing the battery’s nominal capacity, which can deviate by more than 3% from the actual capacity after 100 cycles. This creates a significant dependence on the initial SOC accuracy, with an error propagation ratio of 1:1.2 for each hour of operation, necessitating frequent calibration against reference methods [28].

To address these challenges, Wang et al. [29] developed an enhanced SOC estimation framework that integrates Coulomb counting with two key improvements: (1) dynamic capacity compensation through experimental characterization of how capacity depends on discharge rate, temperature, and cycle life; and (2) cumulative error correction via open-circuit voltage (OCV) recalibration during battery quiescent periods. This hybrid approach reduced the estimation drift by 68% compared to conventional methods. Building on this research, Zhang et al. [30] implemented a Long Short-Term Memory (LSTM) network to dynamically track capacity degradation patterns. This machine learning model adaptively updates capacity values during operation, thus eliminating the fixed-capacity assumption typical of traditional Coulomb counting. Experimental validation in real-world driving conditions demonstrated SOC estimation errors consistently below 10%, with 95% of data points showing less than an 8.2% deviation.

3.1.3. Internal Resistance Measurement Method

The internal resistance measurement method estimates the SOC by leveraging the deterministic functional relationship between a battery’s internal resistance parameters and its SOC. As illustrated in Figure 7, this technique focuses on analyzing the correlations between ohmic resistance (RO), polarization resistance (RP), and SOC, as well as the temperature-dependent behavior of RO [31]. Experimental observations reveal distinct characteristics: RO exhibits minimal fluctuations (±2% variation) across the entire SOC range, demonstrating superior stability compared to RP, which undergoes significant nonlinear variations (±15–20% amplitude shifts) during SOC transitions. Notably, RO maintains a monotonic relationship with SOC within a defined temperature range, where incremental SOC changes induce unidirectional RO drift (0.5–3 mΩ per 10% SOC variation). This monotonicity provides a robust foundation for utilizing RO as a reliable SOC indicator.

This methodology involves the systematic experimental acquisition of impedance data across varying SOC levels [32], followed by the construction of predictive models through polynomial fitting or machine learning algorithms (e.g., Gaussian process regression). In practical applications, real-time resistance measurements are input into these models to estimate SOC, with adaptive corrections for temperature and other dynamic operating conditions via embedded thermoelectric coupling equations. While offering advantages including non-invasive measurement capability and rapid response (<300 ms latency), this methodology inherently faces two key limitations: susceptibility to multivariate interference, as validated by the resistance drift exceeding 15% after 500 cycles, and the implementation cost constraints. Whereas traditional DC internal resistance measurement systems—such as configurations leveraging the Keysight B1500A (Keysight Technologies, Inc., Santa Rosa, CA, USA)—require capital outlays approaching USD 20,000, modern embedded solutions integrating Hybrid Pulse Power Characterization (HPPC) testing with Kalman filtering algorithms achieve equivalent measurement precision at less than 35% of this cost benchmark.

3.1.4. Open-Circuit Voltage Method

The open-circuit voltage (OCV) method estimates SOC by utilizing the deterministic relationship between OCV and SOC [33,34], as illustrated in Figure 8. The core principle of this approach involves measuring the stabilized OCV after a sufficient resting period (typically more than 2 h) to infer SOC values. While this method offers operational simplicity and quick estimation capabilities, it has inherent limitations, including temperature dependence (with a ±3% SOC error for every 10 °C deviation), capacity fade effects (leading to over 5% SOC drift after 300 cycles), and current hysteresis interference [35].

To address these limitations, hybrid approaches that combine OCV with Coulomb counting and Kalman filtering have become standard practice. However, these implementations face challenges such as slow convergence (taking more than 30 min to reach 95% confidence) and high computational complexity, with floating-point operations often exceeding 10^6 per second. Ling et al. [36] developed a predictive OCV (PSOCV) method that integrates OCV theory with a second-order RC network model. Their analysis compared linear interpolation (LI) and least squares (LS) parameter fitting techniques, with experimental results under HPPC and Urban Dynamometer Driving Schedule (UDDS) profiles showing that the LI-based PSOCV achieved superior accuracy (mean absolute error below 1% compared to 1.8% for the LS method). Building on this work, Wang et al. [37] proposed an OCV-slope-compensated SOC estimation framework. By correlating SOC errors with the rate of change in OCV (d(OCV)/dt) and applying adaptive compensation factors to the outputs of extended and unscented Kalman filters (EKF/UKF), their method reduced voltage-induced SOC errors by 42% under high-current (3C) pulsed conditions.

3.2. Model-Based SOC Estimation Methods

Model-based SOC estimation methods calculate and estimate battery SOC by constructing a specific battery equivalent model. The core of this approach lies in utilizing carefully designed equivalent circuit models or physical models to precisely simulate the battery’s actual operating conditions, thereby achieving accurate SOC estimation. Specifically, the accuracy of SOC estimation improves as the designed equivalent model more realistically reflects the battery’s dynamic characteristics and electrochemical behavior under actual operating conditions in terms of both structure and parameters [38]. This paper categorizes model-based methods in the current field into three types: electrical models, thermal models, and mathematical models. Electrical models focus on simulating the battery’s electrical characteristics. These estimate SOC by establishing relationships between electrical properties and SOC through equivalent circuits or electrochemical principles [39]. Thermal models emphasize simulating the battery’s thermal behavior during charge/discharge processes, considering internal heat generation, transfer, dissipation, and external environmental influences on battery temperature [40]. Mathematical models, represented by the Kalman filter family, are filtering algorithms based on state-space models. They recursively estimate SOC by combining system state equations (e.g., SOC variation with time and current) and observation equations (e.g., relationship between terminal voltage and SOC), while accounting for process noise and measurement noise [41]. These methods integrate multiple information sources rather than relying solely on either electrical or thermal characteristics, enabling more accurate estimations.

3.2.1. Electrical Models

Electrochemical Models (EM): EMs are mathematical models constructed based on fundamental principles such as electrode kinetics, ion transport, and thermodynamic equilibrium. They analyze complex electrochemical processes within batteries by describing the characteristics and behaviors of electrodes, electrolytes, and separators, including the use of the Butler–Volmer equation to characterize the relationship between electrochemical reaction rates and potentials at electrode surfaces, and the Nernst–Planck equation to describe ion transport in electrolytes [42]. These models include porous electrode models and concentration polarization models, among others. While they provide deep insights into battery working mechanisms and theoretical guidance for design and management, their applications in performance prediction, state estimation, and fault analysis are limited by high computational complexity, elevated costs, and deviations from real-world conditions. Current electrochemical models are primarily categorized as: pseudo two-dimensional (P2D) models, single particle (SP) models, and multiple particle (MP) models [43], as shown in Figure 9.

(1)

The P2D model, a classical electrochemical model, accurately reflects complex physicochemical phenomena within batteries by detailing mass transport and charge transfer processes in electrodes and electrolytes [44]. Widely used in lithium-ion battery research, it excels in analyzing ion concentration distributions and potential variations during charge/discharge cycles. However, its reliance on numerous partial differential equations results in high computational costs and stringent resource/time requirements.

(2)

The SP model simplifies by treating electrode particles as single entities, focusing on reaction kinetics while ignoring internal concentration gradients [45]. This reduces computational load, making it effective for preliminary design and rapid evaluation where efficiency outweighs precision. However, its oversimplification limits accurate characterization of multiscale phenomena.

(3)

The MP model addresses SP’s oversimplification by simulating battery operational states more realistically through interactions between multiple particles and electrode inhomogeneity [46]. It captures both electrode reaction kinetics and internal mass transport/ion concentration distribution. The MP model demonstrates unique advantages in studying high-rate charging/discharging, long-term cycling, and performance under complex environmental conditions. While providing effective tools for analyzing capacity fade mechanisms, thermal management issues, and battery durability, its intermediate complexity between P2D and SP models requires balancing computational accuracy and cost.

Equivalent Circuit Models (ECMs): ECMs are mathematical representations that conceptualize batteries as circuit networks composed of fundamental components, such as resistors, capacitors, and inductors. These models effectively simulate the electrical behavior of batteries during charge and discharge cycles by establishing relationships between circuit parameters and electrochemical characteristics [47]. Ohmic resistors are employed to represent internal resistance, while RC parallel circuits are utilized to simulate polarization effects. ECMs offer several advantages, including simple structure, low computational complexity, and efficient modeling of the voltage–current–SOC relationships. Common variants include the Rint, Thevenin, and PNGV models, as detailed in

Table 1. Common Equivalent Circuit Models.

Model	Equivalent Circuit	Advantages	Limitations
Voltage Source Model [48] (VSM)		1. Simple structure 2. Fast calculation	1. Oversimplifies 2. High estimation errors
Internal Resistance Model [49] (Rint Model)		1. Simple structure 2. Easy parameter identification	1. Fails to capture dynamic voltage behavior
Thevenin Equivalent Circuit Model [50,51] (TECM)		1. Accounts for polarization effects 2. Low computation	1. Ignores distributed impedance characteristics
Partnership for a New Generation of Vehicles Model [52,53] (PNGV Model)		1. Accurate performance prediction across conditions	1. Requires complex parameter optimization procedures
Dual Polarization Model [54,55] (DP Model)		1. High simulation accuracy 2. Precise dynamic behavior modeling	1. Assumes linear voltage–SOC relationship 2. Sensitive to variations in temperature and aging
n-th Order Resistor-Capacitor Network Model [56] (n-th RC Model)		1. High fidelity 2. Multi-timescale dynamics	1. Complex parameter identification 2. Significant computational load
Generalized Nonlinear Model [57,58] (GNL Model)		1. Incorporates multi-physics nonlinearities	1. Challenges in parameter estimation and model resolution

. Widely used in BMS for SOC estimation, performance prediction, and state monitoring, ECMs facilitate the analysis of electrical performance. Nevertheless, these models have limitations, including insufficient accuracy in describing complex internal electrochemical processes and constrained simulation accuracy under specific operating conditions.

Fractional-Order Models (FOM): FOMs are sophisticated mathematical constructs based on fractional-order calculus, which extends conventional integer-order derivatives and integrals to fractional orders. This extension facilitates enhanced precision in characterizing complex systems with memory and hereditary properties [59]. In battery modeling, FOMs utilize fractional-order derivatives to effectively capture the non-local and history-dependent characteristics of internal ion diffusion processes [60]. Compared to traditional integer-order models, FOMs demonstrate superior accuracy in representing the dynamic behaviors of batteries, particularly in their charge and discharge characteristics, as well as their impedance properties.

Currently, FOMs are primarily employed to model the intricate impedance characteristics of lithium-ion batteries. By applying principles from fractional calculus, these models adeptly simulate impedance patterns that arise from internal ion diffusion and charge transfer processes. Typical implementations include fractional-order equivalent circuit models and fractional impedance network models [61], as illustrated in Figure 10. Within this framework, the Warburg element is a notable 0.5-order fractional component, crucial for describing the diffusion dynamics of particles within the electrode, particularly in the low-frequency range of electrochemical impedance spectroscopy (EIS). In contrast, the Constant Phase Element (CPE) and charge transfer resistance (Rct) are employed to characterize charge transfer reactions and double-layer effects, predominantly occurring in the mid-frequency region of EIS. The enhanced fidelity of FOMs in capturing battery dynamic behaviors is particularly evident in modeling charge/discharge characteristics and impedance properties. By integrating components such as the Warburg element, CPE, and Rct into fractional-order equivalent circuit models and fractional impedance network models, researchers can more accurately simulate the impedance patterns resulting from internal ion diffusion and charge transfer processes.

Despite their promising advantages, FOMs encounter several challenges, including theoretical complexity, difficulties in parameter identification, and higher computational demands. Nevertheless, their application is steadily gaining traction in battery research and management systems, significantly contributing to improved performance analysis and state estimation.

3.2.2. Mathematical Models

Mathematical models refer to the Kalman filter family, as listed in Table 2. Kalman filters are widely used for battery SOC estimation with distinct characteristics. Based on the Kalman Filter [62], these methods have been continuously refined to adapt to diverse battery system properties and application requirements.

(1)

Kalman Filter (KF): The KF recursively calculates SOC using measurements and state-space models, making it suitable for linear systems. Xu et al. [63] developed a lithium battery SOC estimation method that integrates a Recurrent Cerebellar Model Neural Network (RCMNN) with KF. This approach incorporates recursive units in both associative and weight memory layers. Inputs for the model included voltage, current, and temperature measurements, simulating various charge/discharge scenarios in energy storage systems. Experimental results demonstrated high accuracy and robustness across different conditions.

(2)

Extended Kalman Filter (EKF): The EKF extends the KF framework to nonlinear systems through the local linearization of nonlinear functions. However, this linearization can introduce errors. To address this limitation, Tan et al. [64] proposed a Grey Wolf Optimization (GWO)–optimized EKF algorithm, which showed significant reductions in SOC estimation errors and improved accuracy compared to conventional EKF.

(3)

Unscented Kalman Filter (UKF): The UKF employs sigma-point sampling to directly handle nonlinearities, providing enhanced accuracy for nonlinear systems, albeit at increased computational costs. Tang et al. [65] developed a hybrid method combining UKF with Variational Bayesian Adaptive UKF (VBAUKF) to reduce process and measurement noise, achieving a precise SOC estimation. Validation under Urban Dynamometer Driving Schedule (UDDS) conditions confirmed SOC estimation errors below 1%, demonstrating the method’s effectiveness.

(4)

Cubature Kalman Filter (CKF): The CKF approximates probability density functions using cubature rules, achieving high accuracy and stability for nonlinear systems, though its implementation can be complex. Wu et al. [66] proposed an Improved Maximum Correntropy Adaptive Iterative CKF (IMCC-AICKF) to mitigate instability caused by non-Gaussian noise. Simulations verified that this approach accurately converges to true values with enhanced robustness.

(5)

Particle Filter (PF): The PF utilizes Monte Carlo simulations and Bayesian estimation to represent SOC probability distributions through a set of particles, effectively handling arbitrary nonlinear and non-Gaussian systems [67]. However, it suffers from high computational demands and particle degeneracy. Zhang et al. [68] integrated EKF-PF with a second-order Thevenin model to estimate SOC in retired power lithium batteries. Experimental results indicated mean errors below 1.23% and maximum errors under 3.37%, outperforming standard PF approaches.

Table 2. Classification of Kalman Filter Family.

Kalman Filter Variant	Advantages	Limitations
KF [69]	1. Simple structure 2. High computational efficiency	1. Inapplicable to nonlinear systems
EKF [70]	1. Effective for linear/weakly nonlinear systems 2. Moderate computational complexity	1. Poor performance in strong nonlinear systems 2. Sensitivity to initial conditions
UKF [71]	1. Reduced sensitivity to noise/initial conditions 2. Enhanced stability	1. High computational load 2. Requires optimal sigma-point selection
CKF [72]	1. Accurate state distribution characterization 2. Partial non-Gaussian noise robustness	1. High computational complexity 2. Demanding hardware resources
PF [73]	1. Handles strong nonlinearities/non-Gaussian noise 2. Low model dependency	1. Prohibitive computational cost 2. Particle degeneracy issues

Table 2. Classification of Kalman Filter Family.

Kalman Filter Variant	Advantages	Limitations
KF [69]	1. Simple structure 2. High computational efficiency	1. Inapplicable to nonlinear systems
EKF [70]	1. Effective for linear/weakly nonlinear systems 2. Moderate computational complexity	1. Poor performance in strong nonlinear systems 2. Sensitivity to initial conditions
UKF [71]	1. Reduced sensitivity to noise/initial conditions 2. Enhanced stability	1. High computational load 2. Requires optimal sigma-point selection
CKF [72]	1. Accurate state distribution characterization 2. Partial non-Gaussian noise robustness	1. High computational complexity 2. Demanding hardware resources
PF [73]	1. Handles strong nonlinearities/non-Gaussian noise 2. Low model dependency	1. Prohibitive computational cost 2. Particle degeneracy issues

3.2.3. Thermal Model

Thermal models are generally constructed based on electrochemical-thermodynamic principles, integrating chemical reaction heat, Joule heating, and thermal transfer processes within batteries [74]. These models treat the battery’s SOC as a key state variable, while considering the temperature effects on internal parameters and the feedback of these parameter variations on SOC estimation and thermal behavior, forming an interconnected dynamic model. In thermal model construction and analysis, they are typically classified into two categories based on structural characteristics: lumped parameter models and distributed parameter models.

(1)

Lumped Parameter Model: The lumped parameter model is a simplified approach to characterize thermal performance during heat transfer processes. It represents continuously distributed thermal–physical parameters in the system with concentrated parameters. The model assumes instantaneous uniform temperature distribution within the system, neglecting spatial temperature gradients, and treats the entire system or its components as homogeneous “lumped” units with uniform temperature. Based on the law of energy conservation, thermal balance equations are established by analyzing the heat inflow/outflow of the lumped unit and heat storage/variation within the unit, thereby describing the system’s thermal behavior. Zhao et al. [75] proposed an electro-thermal coupled model incorporating temperature effects, which employs discrete identification and UPF-based online parameter identification methods to estimate multiple parameters of the lumped parameter thermal model, with parameter data fitted as continuous environmental functions. As shown in Figure 11, the advantages of this model lie in its simplicity and computational efficiency, enabling the rapid prediction of the system’s overall thermal response trends. This proves particularly valuable for preliminary thermal analysis and estimation. For instance, in scenarios with low temperature accuracy requirements, relatively simple thermal processes, or small-scale systems, the lumped parameter model can provide approximate temperature variation ranges within short timeframes. This assists engineers in rapidly evaluating design feasibility and provides foundational data and conceptual references for subsequent detailed design.

(2)

Distributed Parameter Model: In contrast to lumped parameter models, the distributed parameter model fully considers the continuous spatial variation in internal temperature within objects. The system is divided into numerous micro-units, with thermal balance equations established for each unit. Heat transfer processes in space and time are described through partial differential equations. This unique modeling approach gives distributed parameter models significant advantages in addressing complex heat transfer problems. In battery research, these advantages prove particularly critical. Chen Dafen et al. [76] proposed a construction method for battery distributed parameter equivalent circuit models, with model parameters identified through battery external characteristics. The research process fully leveraged the distributed parameter model’s capability to meticulously characterize heat transfer details. As shown in Figure 12, this model can accurately reconstruct complex temperature distributions within battery systems and reveal detailed heat flow information at different positions. Consequently, it is specifically applicable to scenarios requiring stringent temperature distribution accuracy, involving complex thermal processes and heterogeneous internal heat transfer characteristics, such as battery packs in new energy intelligent connected vehicles and large-scale energy storage battery systems, providing powerful tools for thermal management and performance optimization in these domains.

3.3. Data-Driven SOC Estimation Algorithm

3.3.1. Neural Network Method

The neural network method is a technique that utilizes the powerful nonlinear mapping and data fitting capabilities of neural networks to measure the SOC of batteries. Also known as the Artificial Neural Network (ANN) method, this computational model mimics the structure and function of human brain neurons, aiming to simulate the information processing capabilities of biological neural systems for the purpose of learning, analyzing, and predicting complex data [77,78]. As shown in Figure 13, the SOC estimation process typically consists of three core stages. First, multidimensional time-series data, such as voltage, current, and temperature, are collected through sensors during the battery’s operation. Second, a feature matrix that captures temporal correlations is constructed using a sliding window approach. Finally, the processed data are input into a feedforward neural network that has been trained on extensive battery cycling data, enabling end-to-end SOC prediction via nonlinear transformations in the hidden layers. In comparison to traditional electrochemical model-based methods, the neural network approach does not require precise modeling of the internal reaction mechanisms of the battery. Its black-box nature allows for greater robustness against complex nonlinear factors such as battery aging and environmental disturbances [79,80,81]. Experimental studies indicate that, with a three-layer Back Propagation (BP) neural network architecture, SOC estimation errors can be controlled within 2%. Furthermore, by incorporating an Elman network with memory capabilities, the tracking accuracy for dynamic conditions can improve by 37% [82].

Currently, neural networks have entered a phase of vigorous development. Deep learning, as a crucial branch of neural network evolution, has significantly expanded both the application boundaries and research depth of neural networks. Modern neural networks have far surpassed the limitations of early models such as Hopfield networks and BP algorithms. Driven by deep learning, numerous novel neural network architectures have emerged. Deep learning emphasizes constructing deep neural network structures with multiple hidden layers to autonomously learn complex features from data. Specifically, Convolutional Neural Networks (CNN) utilize unique structures like convolutional layers and pooling layers to efficiently extract local features from image data, achieving remarkable success in image recognition and object detection. Recurrent Neural Networks (RNN) excel in processing sequential data, demonstrating outstanding performance in speech recognition and language modeling. Variants derived from RNN, including Long Short-Term Memory (LSTM) networks and Gated Recurrent Units (GRU), effectively address the vanishing gradient and exploding gradient problems encountered when processing long sequential data, exhibiting strong capabilities in natural language processing and time series prediction. These deep learning-based neural networks each possess distinct structural and functional characteristics, demonstrating significant advantages across different application domains as shown in

Table 3. Neural network classifications.

Classification	Advantages	Disadvantages
BP [83]	1. Simple structure, easy to comprehend 2. Relatively fast training (especially for small datasets)	1. Cannot process time-series data 2. High computational cost for complex tasks
Hopfield [84]	1. Solves optimization problems and associative memory 2. Energy function enables stability analysis	1. Limited storage capacity 2. Prone to local minima in large-scale systems
CNN [85]	1. Parameter sharing reduces complexity 2. Translation invariance ensures robustness	1. Requires massive training data (typically >104 samples) 2. Black-box decision mechanism
RNN [86]	1. Sequential data processing capability 2. Memory retention for temporal patterns	1. Training instability (vanishing/exploding gradients) 2. Limited long-term dependency capture
LSTM [87]	1. Gradient flow control prevents vanishing gradients 2. Long-term dependency learning (>1000 steps)	1. Complex gating mechanisms (4× parameters vs. RNN) 2. High risk of overfitting
GRU [88]	1. Simplified architecture (2 gates vs. LSTM’s 3) 2. Efficient short-term dependency modeling	1. Compromised long-term memory (<500 steps) 2. Residual gradient issues in deep networks

.

3.3.2. Support Vector Machine

The support vector machine (SVM) is a supervised machine learning method based on statistical learning theory, designed to address classification and regression problems. By identifying an optimal decision boundary (hyperplane) that maximizes the separation between different classes of data points [89], SVM achieves distinct operational modes. For linearly separable data, it directly determines a maximum-margin separating hyperplane that maximizes the distance between two classes to ensure superior generalization capability. For nonlinear problems, it employs kernel functions to map original data into high-dimensional feature spaces where linearly separable hyperplanes can be constructed [90], as illustrated in Figure 14.

During training, support vector machines primarily focus on a small number of critical data points located at the margins of the decision boundary, known as support vectors. These vectors determine the position and orientation of the hyperplane. Parameters of the support vectors and hyperplane are optimized through an objective function to minimize classification errors and maximize margins. The SVM’s unique mechanism for handling nonlinear problems involves mapping input data into a high-dimensional feature space via specific kernel techniques, where effective linear separation or regression fitting operations can be performed. This characteristic endows SVM with significant advantages in modeling and analyzing complex systems, particularly in battery SOC estimation [91].

With increasing demands on the precision and reliability of BMS, SVM have gained growing applications and explorations in this domain due to their potential in handling complex nonlinear relationships between battery parameters (voltage, current, temperature) and SOC [92]. Li Jiabo et al. [93] proposed an improved least squares SVM method for online SOC estimation in Li-ion batteries, incorporating previous voltage, current, and SOC estimates as feedback along with real-time measurements. Experimental results demonstrated <1% estimation error, confirming its effectiveness. Li Xuelin et al. [94] enhanced LSSVM through an improved fruit fly optimization algorithm with adaptive relaxation terms, achieving global optima with accelerated convergence. The IFOA-optimized LSSVM showed excellent agreement (mean absolute error: 1.02%) between predicted and measured SOC in power lithium batteries, exhibiting strong generalization capability. Wang Yuyuan et al. [95] developed an LSSVM-based SOC estimation model using battery current, voltage, and temperature as input vectors, with SOC as output. Adaptive particle swarm optimization was employed for parameter tuning, yielding a high-precision model validated through constant-current charge/discharge experimental data comparison (estimation error: 1.63%).

3.3.3. Fuzzy Logic Method

Fuzzy Logic is a methodology rooted in fuzzy set theory that processes ambiguous and uncertain information. It overcomes the black-and-white dichotomy of traditional binary logic through membership functions quantifying element affiliation degrees to sets. The process involves fuzzifying precise inputs, conducting inference via fuzzy rules derived from expert knowledge and practical experience, and ultimately defuzzifying the outcomes to obtain precise outputs [96], as shown in Figure 15.

The internal electrochemical reactions of batteries involve multiple interdependent parameters with inherent uncertainties that dynamically vary with operating conditions, ambient temperature, and charge/discharge rates [97,98]. These complexities resist precise mathematical modeling. Fuzzy Logic addresses such challenges through fuzzy sets and rules: battery parameters (voltage, current) are mapped to fuzzy values, processed via expert knowledge-based rule bases, and defuzzified to obtain precise SOC estimates [99,100]. Unlike ANN, this method requires minimal training data, primarily relying on expert-derived rules. Consequently, Fuzzy Logic is gaining increasing scholarly attention in SOC estimation. Building on this growing interest, extensive research has been conducted on Fuzzy Logic-based SOC estimation. Jing Ziyuan from Jilin University [101] implemented an improved Fuzzy Logic algorithm for parameter identification in a second-order RC battery model. Experimental results demonstrated high online identification accuracy with reduced computation time. Zhou Bin et al. [102] established a hybrid SOC estimation model combining Fuzzy Control with EKF, Coulomb counting, and open-circuit voltage methods. Case studies revealed a maximum error of 3.12%, confirming high precision and capability to overcome limitations of single algorithms. Wang Yungan et al. [103] proposed a multiple-input multiple-output (MIMO) Fuzzy Control-based adaptive parameter equivalent circuit model, experimentally validating the effects of adaptive parameters on model accuracy and adaptability under varying operating conditions. These studies collectively demonstrate that Fuzzy Logic methods provide not only high SOC estimation accuracy but also adaptability to dynamic operational environments, highlighting their practical value and potential in battery management systems.

3.3.4. Extreme Learning Machine

The Extreme Learning Machine (ELM) is a machine learning algorithm based on feedforward neural networks, characterized by fixed hidden layer node parameters (randomly assigned or manually defined) that require no tuning during training, with learning focused solely on output weight calculation [104]. Compared to traditional backpropagation-based neural networks, ELM eliminates the iterative adjustment of hidden layer parameters, significantly reducing training time and computational load. This enables rapid SOC estimation for lithium-ion batteries with enhanced real-time performance. For SOC estimation, battery state parameters including operating temperature, charge–discharge cycles, state of charge/discharge, voltage and current at specific C-rates are processed as input layer data. The hidden layer transforms inputs through activation functions (e.g., trigonometric, Gaussian, radial basis, or Sigmoid functions). The output layer ultimately provides the estimated SOC value. Building upon the ELM framework, Zhao Xiaobo et al. [105] developed a multiple-input ELM (MI-ELM) method with online parameter identification for high-precision SOC estimation. As depicted in Figure 16, the model’s architecture and data flow are visually demonstrated, providing intuitive insights into its operational mechanisms. Comparative experiments were conducted against EKF and standard ELM methods. The MI-ELM achieved performance improvements up to 60.00% in MAE and 88.83% in RMSE, significantly outperforming conventional approaches.

3.4. Comparative Analysis of SOC Estimation Methods

Based on the analysis above, experimental-based methods, model-based approaches, and data-driven techniques exhibit significant differences in estimating the SOC of batteries. To systematically evaluate the performance of these three methodologies across various application scenarios, this study compares them based on four dimensions: accuracy, real-time capability, cost, and robustness (scoring scale: 1–5, with 5 being the best). The applicability of each method is summarized in conjunction with typical application scenarios, as illustrated in

Table 4. Comparison of Battery SOC Estimation Methods

Method Category	Accuracy	Real-Time Capability	Cost	Robustness	Typical Application Scenarios	Core Limitations
Experimental-based [106]	4	2	3	4	Laboratory parameter calibration, offline validation	Relies on offline testing, unable to track dynamically in real time
Model-based [107]	3	5	4	3	Real-time control in BMS, dynamic condition response	Ignores battery nonlinearities and aging dynamics
Data-driven [108,109]	5	4	2	5	Adaptation in complex environments, multi-factor coupling scenarios	Strong dependence on data quality, high training costs

.

Experimental-based methods demonstrate exceptional accuracy (±1% MAE) in controlled environments (e.g., constant temperature and low noise). These methods typically involve conducting charging and discharging tests under specific conditions while measuring the battery’s voltage and current to accurately calculate SOC. However, this approach requires the battery to remain stationary or undergo offline testing, resulting in poor real-time performance (with delays exceeding 2 h). Consequently, experimental methods are more suitable for laboratory calibration rather than online monitoring scenarios. In practical applications, they are often employed for the performance validation of new batteries and during the research and development phase to ensure the accuracy and reliability of the BMS.

Model-based approaches achieve efficient computation (with delays of less than 100 milliseconds) by simplifying dynamic equations, thus adapting well to real-time monitoring requirements. While model-driven methods can predict SOC to some extent, they often overlook internal nonlinear factors of the battery (such as concentration polarization and solid electrolyte interphase film growth). These nonlinear phenomena gradually manifest during the long-term use of batteries, leading to limitations in accuracy (±3% RMSE). This approach is suitable for rapid SOC estimation, particularly within energy management systems of electric vehicles, where it effectively aids in optimizing energy distribution and enhancing driving range.

Data-driven techniques efficiently capture complex nonlinear relationships through data mining, providing the highest accuracy (±1.5% MAE) under dynamic conditions. Data-driven methods rely on a substantial amount of high-quality labeled data (costing over USD 50/kWh) and sufficient computational resources (such as GPU acceleration) for model training and optimization. With the advancement of big data technologies and increased computational capabilities, these methods are increasingly showing potential in practical applications, particularly in battery state monitoring and fault prediction, effectively enhancing battery lifespan and safety.

In summary, no single method can adequately meet the comprehensive demands of New Energy Intelligent Connected Vehicles for high accuracy, strong real-time capability, and low cost. Therefore, future advancements need to focus on the integration of multiple methodologies to overcome existing technical bottlenecks. Within this framework, it is possible to leverage the high accuracy of experimental methods, the real-time capabilities of model-driven approaches, and the adaptability of data-driven techniques, thereby achieving an optimal SOC estimation performance across various application scenarios and significantly enhancing the intelligence and operational efficiency of new energy vehicles.

4. Industry Frontier Algorithms

With the rapid advancement of the new energy intelligent connected vehicle (NEV-ICV) industry, SOC estimation technology has emerged as a pivotal factor influencing vehicle performance and safety. The theoretical framework encompasses a diverse array of SOC estimation algorithms, including Coulomb counting, open-circuit voltage, equivalent circuit models, Kalman filters, neural networks, support vector machines, fuzzy logic, and extreme learning machines. These algorithms exhibit significant variations in model construction, data processing techniques, computational complexity, and estimation accuracy, attributable to their distinct theoretical foundations and methodologies.

Despite the solid groundwork established by theoretical studies, challenges persist in the practical adaptation of these methods for industrial applications. Globally, automakers are actively exploring and implementing various SOC estimation algorithms within their NEV-ICV development processes, guided by their technological expertise, research and development strategies, and market positioning. Analyzing these real-world implementations yields critical insights into the performance of algorithms under actual operating conditions, providing both technical references and practical experiences that can expedite technological innovation within the NEV-ICV sector.

This study compiles case studies of SOC estimation algorithms employed by leading global automakers. As summarized in

Table 5. Industrial SOC Estimation Algorithms: Global Practices.

Automaker	Algorithm	Key Technical Features	Advantages	Disadvantages	Vehicle Models
NIO [110]	Adaptive EKF	Real-time noise covariance updates	Superior Dynamic performance	Complex parameter tuning	ES8, ET7
XPeng [111]	Particle Filter	Monte Carlo probability simulation	Extreme-condition accuracy	High computational load	G9, P7
BYD [112]	Dual Kalman Filter	Stepwise SOC/SOH estimation	Robust in complex conditions	Needs high-end BMS chips	Han EV, Tang DM-i
CATL [113]	Coulomb + OCV	OCV error segmentation correction	Low cost	Low-temperature recalibration	NIO ES6
Chery [114]	ECM + Online Parameter Identification	1st-order RC model with temperature compensation	Low hardware cost	High-rate parameter lag	Tiggo e, Arrizo 5e
Volkswagen [115]	Fuzzy Logic + MPC	Fuzzy rules + multi-objective optimization	Effective for aged batteries	Extensive rule base calibration	ID.4, ID.6
Toyota [116]	ECM + Sliding Window	2nd-order RC error suppression	Long-term consistency	Slow response	Prius, bZ4X
Honda [117]	Regression- Based SOC Estimation	Pseudo-SOC value acquisition, correlation analysis	Improved accuracy post -depolarization	Complexity in correlation classification	Honda e
Audi [118]	Electrochemical + Observer	Physicochemical modeling with observer	Anti-aging precision	High ECU requirements	e-tron series
Volvo [119]	Voltage Level-Based SOC Estimation	Voltage level detection	Efficient SOC estimation in hybrid systems	Dependency on voltage levels	XC40 Recharge, C40 Recharge

Note: Industry implementation cases were collected from automakers’ technical disclosures and third-party industry analyses between 2020 and 2025. Given the proprietary nature of automotive R&D, specific validation datasets are not publicly accessible.

, a comparative analysis of the characteristics, advantages, disadvantages, and applicable scenarios of these algorithms offers actionable solutions for the NEV-ICV industry and lays a robust foundation for future technological breakthroughs.

5. Conclusions and Perspectives

5.1. Conclusions

SOC estimation plays a critical role in battery management. With the continuous advancement of battery technology, SOC estimation algorithms have made significant progress in both theoretical and practical applications, particularly in the BMS of new energy intelligent connected vehicles. However, several key challenges remain in the field of SOC estimation:

• Trade-off between Accuracy and Real-time Performance: SOC estimation requires rapid provision of high-accuracy results to facilitate real-time adjustment of charging strategies. While high-accuracy algorithms, such as Kalman filtering and particle filtering, can deliver precise estimates, they typically demand substantial computational resources, thus restricting their application in resource-constrained embedded systems.
• Balancing Data Quality and Reliability: SOC estimation relies on data from multiple sensors (such as voltage, current, and temperature), which can be significantly affected by external environmental factors. Extreme conditions may degrade sensor performance, resulting in estimation errors. Although multi-sensor collaboration can enhance accuracy, it also adds complexity and cost to the system.
• Adaptability of Models and Algorithms: In practical applications, the diversity of battery types, operating environments, and dynamic states necessitate that SOC estimation models and algorithms exhibit strong adaptability. However, existing SOC estimation solutions each have their strengths and weaknesses, making it challenging to meet all complex requirements.

Despite these challenges, modern solutions such as edge computing and lightweight model compression offer fresh perspectives for SOC estimation. Edge computing effectively reduces latency and enhances responsiveness by processing data closer to the source, making real-time SOC estimation feasible. Lightweight models decrease computational resource requirements through reduced parameter counts and structural complexity, allowing efficient SOC estimation within resource-limited embedded systems. Together, these technologies provide a diverse array of optimization pathways for battery management systems.

5.2. Perspectives

Future research in SOC estimation can evolve by integrating the strengths of existing methods while addressing their challenges, focusing on the following directions:

• Development of Multi-Fusion Estimation Frameworks: Experimental-based methods demonstrate practical value in specific scenarios but struggle to track SOC dynamics in real time. Model-driven approaches excel in real-time responsiveness, enabling rapid reactions to battery operating states, yet inherently fail to comprehensively reflect internal battery characteristics. Data-driven methods exhibit strong adaptability by automatically adjusting estimation models according to battery properties and operating conditions, but they face challenges such as data collection difficulties, transmission delays, and high storage costs. This highlights the inadequacy of single-method solutions in meeting increasingly stringent BMS requirements. Therefore, creating a multi-fusion estimation framework that enhances universality, robustness, and real-time capability will be a crucial direction for the future.
• Research on Adaptive Algorithms: Investigating adaptive algorithms that can dynamically adjust under varying operating conditions will be vital for addressing the challenges posed by battery performance degradation and environmental changes. Incorporating advanced technologies such as reinforcement learning can further enhance the flexibility and intelligence of these algorithms.
• Optimizing Computational Efficiency:

  (a)

  Parameter Quantization: Reducing the precision of model parameters can decrease storage and computational demands, thereby improving efficiency.

  (b)

  Lightweight Model Compression: Techniques such as pruning and knowledge distillation can simplify model structures and reduce complexity.

  (c)

  Hardware Acceleration: Utilizing specialized hardware accelerators like FPGAs and GPUs can enhance computational performance, ensuring efficient SOC estimation.

• Standardization and Modular Design: Promoting the standardization and modular design of SOC estimation algorithms and models will facilitate flexible deployment across different systems and applications, simplifying the integration process.
• Forward-looking Research and Innovation: Strengthening collaboration between academia and industry to conduct forward-looking research will be essential for exploring new materials, structures, and technologies in battery management. Such innovations could lead to significant advancements in the future of battery management systems.

By integrating the strengths of existing methods and addressing current challenges, the SOC estimation field is poised to achieve more precise and efficient solutions. This advancement will not only meet the increasingly stringent requirements for SOC estimation in new energy intelligent connected vehicles and other battery-operated devices but also drive progress toward sustainable transportation.