Electrochemical energy storage technology has been widely applied in various fields, such as mining equipment, new energy vehicles, rail transit, energy storage power stations, and telecommunications. The energy storage battery is the core component of a power station, enabling critical functions like load compensation and peak shaving. Its safety and performance are paramount for the stable and reliable operation of the entire system. Li-ion batteries, known for their environmental friendliness, high energy density, and long cycle life, have garnered significant attention and are extensively deployed globally in energy storage applications, becoming a crucial technological pillar for constructing new power systems.
As the usage time of a Li-ion battery increases, irreversible physical and chemical changes gradually occur inside the cell. These changes directly impact the overall performance of the energy storage system, manifesting as capacity fade and functional abnormalities. In severe cases, they can lead to safety incidents such as fires or explosions. To ensure the long-term and safe operation of energy storage batteries, it is essential to delve into the intrinsic mechanisms of capacity degradation and establish accurate lifetime prediction models. A systematic analysis of battery aging mechanisms not only aids in building a more reliable battery state evaluation system but also significantly improves the prediction accuracy for State-of-Health (SOH) and Remaining Useful Life (RUL), thereby providing vital support for the safe operation and maintenance of energy storage systems.
This article systematically reviews the capacity degradation mechanisms of lithium iron phosphate (LiFePO4)/graphite system energy storage Li-ion batteries, discusses lifetime prediction models, and proposes model selection strategies based on degradation mechanisms. First, the primary aging modes and their underlying mechanisms are analyzed. Subsequently, mainstream lifetime prediction models are introduced, and their advantages, disadvantages, and applicable scenarios are examined in conjunction with practical case studies. By synthesizing research on capacity degradation mechanisms and prediction methods, future research directions are suggested to provide theoretical support for related studies in the energy storage battery field.
Capacity Degradation Mechanisms in Energy Storage Li-Ion Batteries
As of early 2025, LiFePO4 Li-ion batteries constitute a dominant share, over 93%, of newly installed capacity in stationary energy storage projects, making them the mainstream technology in this sector. Therefore, this review focuses on the degradation mechanisms and lifetime prediction methods for the LiFePO4/graphite system. The direct causes of capacity fade and reduced cycle life can be attributed to internal physical and chemical changes. The influencing factors are complex and variable. From a source perspective, the capacity degradation mechanisms can be categorized into Loss of Lithium Inventory (LLI), Loss of Active Material (LAM), Electrolyte Loss (EL), and increased Internal Resistance (IR).
Loss of Lithium Inventory (LLI)
LLI is a primary factor contributing to capacity fade during the cycling of Li-ion batteries. It refers to the permanent loss of cyclable lithium ions due to side reactions, preventing them from being fully intercalated into the anode or cathode materials during charge/discharge cycles. Research on LiFePO4 cells subjected to mild overcharge cycling, using techniques like Incremental Capacity Analysis (ICA) and Electrochemical Impedance Spectroscopy (EIS), has shed light on this behavior. Studies indicate that during charging, a fraction of Li+ ions do not intercalate into the anode lattice but instead combine with electrons and are reduced to metallic lithium, plating onto the anode surface. This leads to permanent lithium loss. The deposited lithium, often with dendritic morphology and high chemical reactivity, promotes further parasitic reactions, exacerbating the loss of active lithium.
Loss of Active Material (LAM)
LAM results from the loss of electrochemical activity or the physical detachment of active materials from the electrodes, caused by issues such as microstructural dislocations, particle cracking, and surface passivation. It is generally divided into cathode LAM and anode LAM.
Cathode LAM involves the structural degradation, particle fracture, or electrochemical deactivation of the cathode material (e.g., LiFePO4), diminishing its ability to store and release Li+. LiFePO4 undergoes a two-phase transformation during cycling. The lattice mismatch between the LiFePO4 (olivine) and FePO4 (rhombohedral) phases creates strain gradients at the interface. Under high temperature or voltage conditions, trace amounts of Fe2+ may dissolve and migrate, destabilizing the Solid Electrolyte Interphase (SEI) on the anode. Repeated lattice expansion/contraction can lead to micro-crack propagation within particles, potentially causing fragmentation or detachment from the current collector, thereby reducing active material and capacity. Furthermore, a relatively thick Cathode Electrolyte Interphase (CEI) layer can form on the LiFePO4 surface. Volume changes during cycling cause this CEI layer to crack and reform, accelerating electrolyte decomposition side reactions, consuming active lithium, and increasing cell resistance.
Anode LAM is primarily manifested by the destruction of the graphite structure, particle pulverization, or surface passivation, reducing its lithium intercalation capability. During Li+ intercalation/deintercalation, graphite experiences volume changes of about 10%. Long-term cycling under this repeated stress can induce micro-cracks within the particles. Scanning Electron Microscopy (SEM) observations of cycled graphite anodes reveal increased surface roughness, partial delamination of graphite layers, and even complete detachment from the copper current collector, leading to active material loss. The SEI layer on the graphite anode can fracture due to volume changes, exposing fresh graphite surfaces to the electrolyte and resulting in the formation of a thicker, more resistive SEI layer. This process not only consumes active lithium but also increases interfacial impedance.
Electrolyte Loss (EL)
Capacity degradation in Li-ion batteries is closely linked to electrolyte degradation and the associated active lithium loss. The organic electrolyte (e.g., LiPF6/carbonate ester systems) undergoes reduction reactions at the anode surface during cycling, forming an SEI layer composed of compounds like Li2CO3, ROCO2Li, and LiF. Research shows that at elevated temperatures, the decomposition of LiPF6 accelerates, producing LiF and corrosive HF. The HF can further dissolve cathode active materials, aggravating lithium inventory loss. Concurrently, gases generated from electrolyte decomposition (e.g., CO2, C2H4) consume the electrolyte volume. Quantitative analysis indicates that electrolyte loss primarily manifests as an increase in internal resistance, while active lithium loss directly reduces the number of cyclable Li+ ions. These two factors jointly contribute to battery capacity fade.
Internal Resistance Increase (IR) Induced by Side Reactions
Internal resistance impedes ion migration within the battery and is determined by multiple factors, including the electrodes, electrolyte, and interphases. For the cathode material, changes in particle contact, structural alterations, and surface oxidation over long-term use can slow down electron and ion transport, increasing resistance. Increased internal resistance reduces the mobility of Li+ within the active electrode materials and lowers the ionic conductivity of the electrolyte, leading to capacity loss. Studies have found that with increasing depth of discharge, the overall resistance for Li+ to deintercalate, migrate, and intercalate rises. Some Li+ ions accumulate at the electrode surface, forming metallic lithium or other species, resulting in capacity loss. Additionally, graphite particles detaching from the copper foil and becoming suspended in the electrolyte or migrating to the separator can increase cell impedance and accelerate capacity fade.
Lifetime Prediction Methods for Energy Storage Li-Ion Batteries
Under normal operating conditions, battery aging is a slow process that can take several years to reach its end of life. Therefore, predicting the RUL relies on scientific analysis of capacity fade mechanisms and robust prediction methodologies. Based on the methodological framework, lifetime prediction approaches can be classified into three categories: model-based methods, data-driven methods, and hybrid methods. Model-based methods rely on models such as the Arrhenius model, mechanistic models, and empirical models.
Model-Based Methods
Constructing a battery lifetime prediction model requires understanding the differences in external characteristics under various aging states and analyzing the corresponding internal mechanistic changes. Thus, the research foundation is capturing the dominant aging mechanisms or empirical aging patterns.
Arrhenius Model
The Arrhenius model, based on the relationship between reaction rate and temperature, treats battery aging as a temperature-dependent chemical process, particularly under elevated temperatures. While the Arrhenius model often fits experimental data well, its underlying assumption is that temperature is the sole factor affecting battery life, neglecting other influential parameters. Although the Arrhenius equation has a theoretical basis, it does not comprehensively consider the complex mechanisms of battery aging. Therefore, its application in battery life prediction tends to be more empirical. For more accurate predictions, it is usually combined with other models (e.g., mechanistic or data-driven models) to address its limitations. The Arrhenius formula is expressed as:
$$ Q_{\text{loss}}(T, n) = A \cdot n^z \cdot \exp\left(-\frac{E_a}{RT}\right) $$
where $Q_{\text{loss}}(T, n)$ is the capacity loss, $n$ is the number of cycles (or a time-related parameter), $T$ is the absolute temperature, $A$ is the pre-exponential factor, $E_a$ is the activation energy, $R$ is the universal gas constant, and $z$ is a fitting exponent.
Mechanistic Models
Battery mechanistic models simulate internal processes such as mass transport, diffusion, electrochemical reactions, and thermodynamics. However, due to the complexity and difficulty in obtaining internal parameters, and the high computational cost, researchers often simplify these models to improve efficiency. The trade-off is that increased simplification typically reduces model accuracy. Standalone mechanistic models have limited application value, restricting their use in engineering practice. Typically, mechanistic models are combined with state estimation techniques like Kalman filters, particle filters, H∞ filters, or sliding mode observers to estimate model parameters in real-time, thereby enabling state-of-health and RUL estimation.
For instance, a simplified model focusing on SEI layer growth might describe the overpotential due to the SEI layer $\eta_{\text{SEI}, j}(t)$ as:
$$ \eta_{\text{SEI}, j}(t) = \frac{R_{\text{SEI}, j} \cdot j_{f,j}(t)}{a_{s,j}} $$
where $R_{\text{SEI}, j}$ is the area-specific resistance of the SEI layer, $j_{f,j}$ is the volumetric current density, and $a_{s,j}$ is the specific surface area.
Another model relating capacity loss $Q_{\text{loss}}$ to current density $j$ for lithium deposition might propose a linear relationship: $Q_{\text{loss}} = Q_{\text{ir}} + k \cdot j$, where the slope $k$ is determined by Li+ mobility in the SEI layer, linking a macroscopic observable to a microscopic property.
Empirical Models
For Li-ion battery aging, empirical models are derived from fitting regression models to extensive historical degradation data, establishing an empirical relationship that describes the capacity fade trajectory. While these models can accurately predict battery life by incorporating various coupled and nonlinear aging mechanisms, the underlying capacity fade process is complex and variable. Model calculations can be intricate, and it is often challenging to directly parameterize them using externally measurable data for application analysis.
A classic example is the square root of time model for SEI growth, which can be formulated as:
$$ C = a_1 \cdot \left(1 – \exp\left(-a_2 \cdot \sqrt{t} \cdot \exp\left(-\frac{E_a}{RT}\right)\right)\right) $$
where $C$ represents capacity loss, $t$ is time or equivalent cycles, and $a_1$, $a_2$ are fitting parameters related to SEI growth kinetics and temperature sensitivity.
Data-Driven Methods
Data-driven methods utilize measured parameters such as capacity, current, voltage, and impedance, applying intelligent algorithms like neural networks, support vector machines (SVM), and Gaussian process regression (GPR) for RUL prediction. These methods do not require a deep understanding of the internal mechanisms of the Li-ion battery; they only need externally monitored data, which is relatively easier to acquire and process, allowing for flexible application and generalization. However, data-driven models are highly dependent on the density, quality, and volume of historical data. This can pose a challenge in practical applications where obtaining sufficient high-quality data is difficult, potentially affecting prediction accuracy.
For example, a neural network might be trained using features extracted from charge/discharge curves or EIS spectra. Alternatively, a one-order RC equivalent circuit model is often used for its simplicity and real-time capability. Its transfer function $G(s)$ in the Laplace domain, relating terminal voltage $U(s)$ to current $I(s)$, is:
$$ G(s) = \frac{U(s) – U_{\text{OCV}}(s)}{I(s)} = -R_0 – \frac{R_p}{1 + R_p C_p s} $$
where $U_{\text{OCV}}$ is the open-circuit voltage (a function of SOC), $R_0$ is the ohmic resistance, and $R_p$ and $C_p$ are the polarization resistance and capacitance, respectively. Changes in these parameters over time can be used as health indicators for data-driven RUL models.
Hybrid Methods
Hybrid methods combine model-based and data-driven approaches to enhance prediction accuracy and robustness, making them suitable for real-world applications. As discussed, model-based methods rely on aging mechanism analysis or empirical conclusions, updating model parameters recursively with sampled data, but they can be computationally expensive and may not handle complex aging dynamics well. Data-driven strategies model aging trends based on historical and real-time data and observed aging features, requiring accurate modeling over time, efficient computation, and high-throughput data generation. Hybrid methods leverage the strengths of both: smaller error and higher stability from models, and adaptability from data. They are well-suited for real-time application in Battery Management Systems (BMS). Their limitation lies in increased model complexity and the need for collaborative design from multiple domains.
For instance, a hybrid approach might use a Gamma stochastic process combined with a state-space model to describe capacity fade randomness, employ a particle filter for state estimation, and then use a meta-heuristic algorithm like Sparrow Search Algorithm (SSA) to optimize parameters for a Support Vector Regression (SVR) model that refines the predictions.
Another advanced example is an Improved Temporal Fusion Transformer (ITFT) model that integrates parameters extracted from a physics-based Single Particle Model with electrolyte (SPMe) – such as internal resistance and diffusion coefficients – with a data-driven sequence analysis architecture like a Bidirectional Long Short-Term Memory (Bi-LSTM) network within a Transformer framework for joint SOH and RUL prediction.
Practical Applications and Case Analysis
Analyzing practical application cases highlights the scenarios for battery lifetime prediction and its coupling relationship with capacity degradation mechanisms.
Arrhenius Model Application: A study proposed a temperature correction method based on the Arrhenius equation and a recursive least squares algorithm for LiFePO4 cells. The model introduced a compensation factor linking available capacity to average cell temperature, accounting for the effect of temperature on active material kinetics (a form of LAM). Experiments across temperatures from 283.15 K to 313.15 K validated the method’s reliability. However, performance prediction under low-temperature conditions and the model’s limitation to a single dominant factor (temperature) were noted.
Mechanistic Model Application: A multi-physics model was developed to simulate SEI growth and lithium plating side reactions in LiFePO4/graphite cells. The model could simulate the rate of these parasitic reactions to predict capacity fade. By simplifying the complex electrochemical mechanisms, it balanced accuracy and computational cost. Another mechanistic model established a linear relationship between capacity loss and current density, with the slope determined by Li+ mobility in the SEI layer, demonstrating the SEI’s role in homogenizing Li deposition. While accurate, such models often require re-fitting coefficients for different battery types, limiting their generalizability.
Data-Driven Model Application: A data-driven approach used a first-order RC equivalent circuit model to track parameters like ohmic resistance ($R_0$) and polarization resistance ($R_p$). For LiFePO4 cells, results showed that as ambient temperature increased from 283.15 K to 313.15 K, the average ohmic resistance decreased by about 26%, aligning with the IR-dominated aging mechanism at lower temperatures. By accurately predicting internal resistance trends, the model improved RUL estimation. Deep learning models, such as those using IC curve features (related to LLI) as input to a neural network, have also been developed for capacity prediction under complex driving cycles, though their performance in non-linear aging phases can be inconsistent.
Hybrid Method Application: A hybrid framework combined a Gamma process-based state-space model (capturing randomness) with a particle filter and an SSA-optimized SVR model for parameter correction and RUL prediction. This method effectively captured the stochastic nature of battery degradation and showed better accuracy and computational efficiency than traditional methods. Another study created an ITFT model that fused parameters from an SPMe mechanistic model with a Bi-LSTM-Transformer data-driven architecture. Tested on LiFePO4/graphite cells, the predictions aligned with IR-increase-dominated capacity fade, achieving a very low RMSE for RUL prediction, showcasing significant improvements in both precision and stability.
Comparison of Lifetime Prediction Models
The following table summarizes and compares the key characteristics of the different lifetime prediction model categories.
| Prediction Method | Key Measurement Parameters | Typical Algorithms/Techniques | Advantages | Disadvantages/Limitations |
|---|---|---|---|---|
| Model-Based | ||||
| Arrhenius Model | Temperature, cycle count/time | Least Squares Fitting | Widely applied, theoretically grounded for temperature effects. | Considers only temperature, limited for multi-stress aging. |
| Mechanistic Model | Model parameters (e.g., diffusivity, reaction rate) | SP/SPMe, DFN models combined with state estimators (KF, PF) | Provides physical/chemical insight, high accuracy for specific mechanisms. | High computational cost, complex parameter identification, low generalizability. |
| Empirical Model | Capacity, impedance, cycle count | Non-linear regression (e.g., exponential, square root of time) | Simple to construct based on test data, directly links observable to aging. | Accuracy limited in practical varied conditions, many parameters, lacks physical insight. |
| Data-Driven | Voltage/current curves, temperature, impedance spectra | Neural Networks (DNN, CNN, LSTM), SVM, GPR | High accuracy with sufficient data, strong generalization, no need for mechanism knowledge. | High demand for data quality/quantity, long training time, poor interpretability. |
| Hybrid Methods | Combination of model parameters and operational data | PF/EM with ML models (SVR, NN), Physics-Informed Neural Networks (PINN) | Balances accuracy and robustness, suitable for BMS real-time application. | High model complexity, requires interdisciplinary design, technology still maturing. |
Conclusion and Outlook
With the continued expansion of the energy storage market, the analysis of degradation mechanisms and lifetime prediction models for energy storage batteries will increasingly move toward systematization and standardization, playing a more critical role in practical applications. This review has summarized the primary capacity degradation mechanisms and lifetime prediction technologies for energy storage Li-ion batteries, which is expected to positively contribute to related research.
Although existing models have made progress in predicting battery capacity fade, certain limitations remain. For instance, some models inadequately describe capacity degradation mechanisms under complex operational conditions or lack in-depth study on the aging characteristics of new battery materials. These shortcomings restrict the accuracy and reliability of models in real-world applications. To further enhance the predictive accuracy and practicality of these models, future research could focus on the following two key directions:
- Conduct More Detailed Electrochemical Experiments: Designing more comprehensive and systematic experimental protocols is crucial to deeply investigate the key factors influencing battery capacity fade, such as temperature, charge/discharge rate, cycle number, and State-of-Charge (SOC) window. Concurrently, dedicated studies on the aging characteristics of batteries incorporating novel materials (e.g., high-nickel cathodes, silicon-based anodes) are needed to determine the critical parameters for lifetime prediction models, thereby improving their accuracy and reliability. Furthermore, combining in-situ/operando characterization techniques (e.g., in-situ XRD, Raman spectroscopy) with multi-scale simulation (e.g., molecular dynamics, finite element analysis) can further reveal the microscopic mechanisms of battery capacity degradation.
- Fuse Multiple Lifetime Prediction Model Technologies: Integrating data-driven models (e.g., neural networks, SVM) with mechanistic models (e.g., electrochemical models) to construct a multi-dimensional, multi-scale battery lifetime prediction framework is a promising avenue. By leveraging the strengths of different models, this approach can more comprehensively account for the complex factors influencing battery aging (e.g., temperature, C-rate, cycling history). Hybrid models, particularly those incorporating physics-informed machine learning, offer significant potential for developing accurate, robust, and interpretable prediction tools suitable for real-world battery management systems.
Advancements in these areas will be instrumental in unlocking the full potential of Li-ion batteries for long-duration, safe, and reliable energy storage, supporting the global transition to sustainable energy systems.