Health assessment and life extension control method considering inconsistency of large energy storage battery system

Abstract

The invention discloses a health assessment and life extension control method considering the inconsistency of a large energy storage battery system. Aiming at the problems of conservative system available capacity evaluation, unreliable inconsistency evolution prediction and lack of pertinence in life extension control in the prior art, the method comprises the following steps: firstly, cooperatively estimating SOC and SOH of each monomer through data cleaning and an ampere-hour integral method; then, a system-level health assessment model is constructed, the actual available capacity is defined, a multi-dimensional inconsistency feature vector F = [F1, F2, F3] is extracted, F1 represents the capacity difference degree, F2 represents the average state deviation, and F3 represents the boundary unit deviation degree; predicting an inconsistency evolution trend by using a lightweight neural network, and identifying a dominant inconsistency type in combination with SHAP interpretability analysis; and finally, self-adaptively triggering life extension control strategies such as active equalization, power limitation or packet scheduling according to a diagnosis result. According to the method, closed-loop management of perception-evaluation-diagnosis-decision is realized, the evaluation precision of the health state of the system is improved, inconsistent development is effectively inhibited, and the overall service life of the energy storage system is prolonged. The method is suitable for battery life management of new energy power stations, power grid sides and industrial and commercial energy storage systems.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of energy storage and intelligent control. Specifically, it is a health management method for energy storage battery systems that integrates multi-source data processing, system-level health modeling, interpretable machine learning, and closed-loop feedback control. It is particularly suitable for energy storage power station scenarios with high string count, large capacity, and long-term operation. Background Technology

[0002] As the global energy structure transitions towards cleaner and lower-carbon energy, the installed capacity of renewable energy sources such as wind power and photovoltaics continues to grow, highlighting their increasing impact on the volatility and intermittency of the power grid. Large-scale electrochemical energy storage systems, as key infrastructure for smoothing renewable energy output, participating in grid frequency and peak regulation, and improving power supply reliability, are experiencing rapid development. Lithium-ion batteries, with their high energy density, long cycle life, and continuously decreasing cost advantages, have become the preferred technology in the field of electrochemical energy storage. To meet the stringent requirements of grid-scale energy storage for operating voltage and capacity, it is typically necessary to connect thousands or even tens of thousands of battery cells in series and parallel to form large-scale battery systems. However, this large-scale assembly method also brings severe technical challenges. Due to subtle differences in manufacturing processes, uneven temperature distribution in the operating environment, and unbalanced charging and discharging current distribution, there are inherent inconsistencies between battery cells in parameters such as initial capacity, internal resistance, and self-discharge rate. This inconsistency is amplified through a "cumulative effect" during system operation: smaller cells reach their voltage limits more quickly during charging and discharging, resulting in a significantly lower usable system capacity than the average of individual cell capacities; cells with higher internal resistance generate more heat under the same current, accelerating their own aging and affecting surrounding cells; differences in self-discharge rates lead to a gradual divergence of the state of charge, further exacerbating the decline in energy utilization. Studies have shown that for a system composed of hundreds of battery cells, even a 2% difference in capacity between cells can result in a loss of up to 10-15% in the overall usable system capacity.

[0003] Currently, the health status assessment of energy storage battery systems faces the following technical bottlenecks: First, traditional assessment methods are generally based on the "barrel effect" theory, equating the health status of the worst-performing cell in the system with the overall system health status. While this method ensures the safety of the assessment results, it severely underestimates the true usability of the system and cannot quantify the capacity loss caused by inconsistencies. Second, existing methods mostly use fixed thresholds or empirical formulas for equilibrium control, lacking a deep understanding of the evolution of inconsistencies, resulting in lagging and inefficient control strategies. More importantly, existing technologies struggle to accurately identify the dominant inconsistency types that lead to system performance degradation (such as capacity-related, internal resistance-related, or SOC-related inconsistencies), making subsequent life-extending control lack specificity. Regarding the development of battery management system technology, data-driven methods have made significant progress in recent years. By analyzing historical battery operating data, machine learning models can establish a nonlinear mapping relationship between battery parameters and health status. However, existing research mainly focuses on the performance prediction of individual cells, with insufficient research on the dynamic evolution mechanism of system-level inconsistencies; simultaneously, purely data-driven black-box models lack physical mechanism constraints, resulting in poor interpretability and weak extrapolation capabilities in practical engineering applications. Furthermore, existing balancing strategies mostly employ passive balancing or simple active balancing schemes, resulting in low energy transfer efficiency and failing to fundamentally suppress the trend of inconsistency. To address the inconsistency problem in large-scale energy storage battery systems, this invention proposes a health assessment and life extension control method that integrates physical mechanisms and data-driven approaches, balancing assessment accuracy and engineering applicability. This method aims to overcome the shortcomings of existing technologies, such as inaccurate assessment of the system's true usability, difficulty in predicting the evolution trend of inconsistency, and lack of targeted life extension control, thus providing technical support for improving the economy and reliability of energy storage systems. Summary of the Invention

[0004] Against the backdrop of profound changes in the global energy landscape and the ever-expanding demand for clean energy, lithium-ion batteries, with their advantages of high energy density and long cycle life, are widely used in the field of electrochemical energy storage. To meet the high operating voltage requirements, lithium-ion batteries are typically assembled in series to form large battery packs containing dozens to hundreds of battery cells. However, while this series architecture increases the voltage level, it also amplifies the impact of inherent parameter inconsistencies between individual battery cells, posing a severe challenge to the accurate assessment and unified control of the system's health status. Existing health assessment methods for energy storage systems often simply rely on the "weakest link" effect, equating the state of health (SOH) of the smallest cell in the system with the SOH of the entire system. This simplified model fails to fully consider the complex inconsistencies between battery cells, making it difficult to effectively carry out assessment-based system maintenance and causing significant economic losses due to prematurely discarding still valuable battery cells. This invention proposes a health assessment and life-extending control method that takes into account the inconsistencies of large-scale energy storage battery systems. First, by using the ampere-hour integral method based on dynamic operating conditions and open-circuit voltage curve fitting, a high-precision coordinated estimation of the health status and state of charge of each battery cell is achieved. Then, a novel system-level health assessment method was adopted, redefining the actual usable capacity of the battery system as "the sum of the minimum rechargeable capacity and the minimum dischargeable capacity of the battery pack." This theoretically and rigorously characterizes the capacity loss caused by inconsistency, quantifies the degree of system inconsistency, and extracts multi-dimensional inconsistency features. Furthermore, machine learning algorithms were introduced to uncover the deep mapping relationship between inconsistency features and battery degradation modes, and SHAP interpretability analysis was integrated to accurately identify the dominant inconsistency types leading to performance degradation. Finally, based on the identification results, targeted life-extending control strategies such as adaptive trigger equalization and group scheduling were implemented to maximize the utilization of remaining system capacity and extend overall lifespan while ensuring safety.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] A method for assessing the health status and intelligently extending the service life of an in-service large-scale energy storage battery system includes the following steps:

[0007] S11: Acquire the operating data of the energy storage battery system, including time, total voltage, current, cell temperature, and cell voltage, record and store the above operating data; when cleaning the operating data, firstly use the dynamic threshold range based on electrochemical characteristics to filter out abnormal data; then use the K-nearest neighbor interpolation method to fill in missing values ​​to ensure the continuity of time series data; finally use a low-pass filter to smooth the noisy data to ensure the stability and reliability of the data.

[0008] S12: When calibrating the battery SOC, a long-term static operating window is selected to obtain the individual cell voltage and temperature data after sufficient rest; then, the battery SOC is calibrated using a SOC-OCV lookup table; based on this, the variable ampere integration method is used to obtain data for the complete charge and discharge range, and the actual battery capacity is calculated.

[0009]

[0010] in, C It is the calculated battery capacity; I ( t ( ) is the charging / discharging current; t start , t end Indicates the start and end times of the integration; SOC start , SOC end Indicates the start and end points SOC Value. (Select only) SOC start (<6% data segment and complete charge / discharge logs to ensure reliability.)

[0011] S13: Based on the physical laws of battery aging, a two-stage empirical model is used to derive the equivalent cycle capacity. First, the equivalent full cycle number EFC is calculated by the total throughput of the statistical cycle. Then, an empirical model that integrates nonlinear and linear decay characteristics is established, and the parameters are solved by fitting historical data using the least squares method to achieve accurate prediction of the battery capacity decay trajectory.

[0012]

[0013] in, C Represents the capacity of the battery cell; a Represents the initial capacity baseline; b Represents the early rapid decay rate; c This represents the linear decay rate.

[0014] S2: Battery System Health Assessment and Inconsistency Quantification

[0015] S21: Accurate Modeling of Actual Usable Capacity Based on Inconsistency. The actual usable capacity of a series-connected battery pack is limited by the weakest cell in the system, rather than the simple sum of the capacities of all cells. To accurately characterize the capacity loss caused by inconsistency, the actual usable capacity of the system is defined as the sum of the minimum rechargeable capacity and the minimum dischargeable capacity:

[0016]

[0017] in Cpack It refers to the packaging capacity; This is the minimum rechargeable capacity; It is the minimum discharge capacity;

[0018] S22: Quantitative Analysis of Inconsistency Based on Energy-Capacity Space. The definition of this index is based on the following theoretical derivation: In an ideal state, if all battery cells in a system have completely identical characteristics, their energy-capacity relationship will exhibit a perfectly linear characteristic; however, in actual systems, due to the dispersion of parameters between cells, the energy-capacity curve will be distorted, resulting in a reduction in the usable capacity area. Based on this physical phenomenon, the formula for calculating the energy-capacity dispersion index is as follows:

[0019]

[0020] Among them, among them, C max and C min These represent the maximum and minimum available capacity of a single unit in the system, respectively. C pack This represents the actual available capacity of the system. The physical meaning of this index is clear: the first term ( C max + C min -2 C pack The second term reflects the degree of deviation of the actual operating point of the system from that of an ideal uniform system. C max - C min This characterizes the inherent dispersion of the individual components in the system. The product of these two factors constitutes a complete description of the impact of system inconsistencies.

[0021] S23: Precise Extraction of Multi-Dimensional Inconsistency Features. To achieve a comprehensive characterization of battery system inconsistencies, this invention constructs a multi-dimensional feature extraction system within the SOC-SOH coupled state space. This system, based on the distribution characteristics of individual battery cells in the state space, reveals the intrinsic structure and influencing mechanisms of inconsistencies from different dimensions. Based on the coupled SOC-SOH state space diagram, we calculate the capacity difference (…). F 1) Mean state deviation ( F 2) and boundary element offset ( F 3). These three features work together. F 1. F 2 and F The larger values ​​of 3 collectively indicate that, as defined by the mathematical expression below, the inconsistency of the battery pack is mainly caused by capacity degradation.

[0022]

[0023] in, and Let ω represent the system's average SOC and average SOH, respectively. soc and ω SOH These are the weighting coefficients.

[0024] S3: Battery System Inconsistency Prediction and Type Identification:

[0025] S31: Construct a lightweight prediction model guided by physical information. Based on the physical mechanism of battery aging, establish constraint relationships to guide the machine learning model in learning battery aging patterns. Construct a composite loss function and use the Adam optimizer and cosine annealing learning rate scheduling strategy for model training. Introduce label smoothing and gradient pruning to prevent overfitting. Update model parameters through adaptive learning rate and weight decay regularization. Establish a dual evaluation standard of validation loss and physical consistency. When the number of consecutive times the validation loss no longer decreases equals the number of tolerances, save the optimal model parameters.

[0026] S32: Multi-objective Bayesian Optimization. A multi-objective optimization framework is established, simultaneously considering estimation accuracy, computational efficiency, and physical consistency. A Bayesian optimization algorithm based on a tree-structured Parzen estimator is adopted, using a probabilistic model to guide the search process and find the Pareto optimal solution set. A neural architecture search is introduced to automatically design the network structure, and knowledge distillation technology is used to enable large models to guide lightweight network learning.

[0027] S33: Inconsistency Type Identification Based on SHAP Analysis. Building upon the lightweight prediction model constructed in S31-S32, this invention introduces the SHAP interpretability analysis framework to quantify the marginal contribution of each inconsistency feature to the prediction of system performance degradation and accurately identify the dominant inconsistency type. The SHAP value provides a consistent and interpretable contribution allocation for model predictions by calculating the average marginal contribution of a feature across all possible feature subsets. i SHAP value of each feature phi i The calculation is as follows:

[0028]

[0029] in N This represents the complete set of all features used by the model; i The target features that indicate its contribution are being analyzed; N { i} indicates from the complete set N Deleting features i The remaining feature subset; S It means N { iAny subset of} represents the set of features introduced when the target features are introduced. i A specific combination of features that have already been used in the modeling; f ( S ) indicates that only a subset of features is used. S The model's predicted output; while f ( S ∪{ i}) indicates that in the feature i Merge into subset S New predictions from the mid-to-late stage model. Overall SHAP value. phi i By using all possible subsets S This is obtained by performing an appropriate weighted summation, thus accurately quantifying the features. i The average marginal contribution of the model to predictions across different combinations of feature contexts.

[0030] S4: Intelligent life extension control of battery system;

[0031] S41: By establishing a mapping relationship between multi-dimensional features and control strategies, a hierarchical control mechanism, including balanced management, group scheduling, and maintenance intervention, is dynamically triggered. This method effectively suppresses the trend of inconsistency by optimizing the system's operating state in real time, significantly improving the system's available capacity and extending its overall service life while ensuring safe system operation.

[0032] S42: Construct a collaborative optimization framework for lifespan prediction and health management, dynamically adjusting control strategy parameters based on real-time system status assessment and aging trend prediction. By combining preventative maintenance with adaptive scheduling, performance maintenance and lifespan extension are achieved throughout the system's entire lifecycle, maximizing the economy and reliability of the energy storage system.

[0033] The beneficial effects of this invention are as follows:

[0034] This invention proposes a health status assessment method for energy storage battery systems based on inconsistency modeling. This method redefines the actual usable capacity of the system as the sum of the minimum rechargeable capacity and the minimum dischargeable capacity, theoretically and rigorously characterizing the capacity loss caused by inconsistency. This achieves an accurate assessment of the system's true health status and overcomes the limitations of traditional "barrel effect" assessment models.

[0035] This invention proposes a method for quantifying inconsistencies in battery systems based on multidimensional feature extraction. This method systematically characterizes the spatial distribution and structural features of inconsistencies by constructing three feature quantities—capacity difference, average state deviation, and boundary cell offset—in the SOC-SOH coupled state space, providing a complete feature description system for subsequent degradation pattern recognition.

[0036] This invention proposes a method for identifying inconsistencies by integrating physical mechanisms and interpretable machine learning. This method constructs a lightweight prediction model guided by physical information, and combines multi-objective Bayesian optimization and SHAP interpretability analysis to achieve in-depth mining of inconsistency features and degradation patterns. It can accurately identify the dominant inconsistency types leading to performance degradation, thereby improving the reliability of the model and the transparency of decision-making.

[0037] This invention proposes an adaptive lifespan extension control method for energy storage battery systems based on characteristic response. This method establishes a dynamic mapping relationship between multi-dimensional features and control strategies, enabling adaptive triggering of hierarchical control mechanisms such as balanced management, group scheduling, and maintenance intervention. This effectively suppresses the trend of inconsistency, significantly improves the system's available capacity, and extends its overall lifespan.

[0038] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0040] Figure 1 This is an overall flowchart of a health assessment and life extension control method for large-scale energy storage battery systems that takes into account inconsistencies in an embodiment of the present invention.

[0041] Figure 2 This is a schematic diagram of the battery system operation data acquisition and preprocessing process in an embodiment of the present invention;

[0042] Figure 3 This is a flowchart of the battery capacity estimation and single-cell health state (SOH) assessment in an embodiment of the present invention;

[0043] Figure 4 This is a schematic diagram illustrating the distribution and inconsistency features of each cell in the SOC-SOH coupled state space of the battery system in this embodiment of the invention.

[0044] Figure 5 This is a schematic diagram illustrating the training and optimization process of the inconsistency prediction model in an embodiment of the present invention.

[0045] Figure 6 This is a flowchart of the intelligent life extension control strategy triggered by inconsistency type identification results in an embodiment of the present invention.

[0046] exist Figure 1The paper demonstrates a complete closed-loop process, starting from raw data acquisition, proceeding through data cleaning, individual SOC / SOH co-estimation, system-level health assessment, inconsistency quantification and evolution prediction, dominant type identification, and finally outputting adaptive life extension control commands. This reflects the integrated architecture of "perception-assessment-diagnosis-decision" of this invention.

[0047] exist Figure 2 The document details the data preprocessing workflow, which involves acquiring raw data such as voltage, current, and temperature from the BMS, and then sequentially performing dynamic threshold filtering, sliding window anomaly detection, K-nearest neighbor interpolation to fill in missing values, and Butterworth low-pass filtering for noise reduction. This ensures the accuracy and reliability of the data used in subsequent analyses.

[0048] exist Figure 3 In this study, a high-precision SOC estimation is achieved by combining the ampere-hour integral method with OCV curve correction. A two-stage empirical model (early nonlinear + later linear) is used to fit the capacity decay trajectory, calculate the equivalent cycle number EFC, and thus obtain the single-cell health state SOH.

[0049] exist Figure 4 In the diagram, the horizontal axis represents the state of charge (SOC), and the vertical axis represents the state of health (SOH). Each data point represents a single battery cell. This distribution allows for the intuitive identification of the minimum rechargeable capacity (min) and the minimum dischargeable capacity (min) and defines the actual usable capacity of the system. Simultaneously, three feature quantities are extracted: F1 (capacity difference), F2 (average state deviation), and F3 (boundary cell offset) to characterize the inconsistent structural features.

[0050] exist Figure 5 The paper demonstrates the construction process of a lightweight neural network model, including input feature selection, Bayesian hyperparameter optimization, physical constraint loss function design (such as monotonicity penalty term), and iterative training mechanism based on validation set performance feedback, which improves the model's generalization ability and physical rationality.

[0051] exist Figure 6 In this process, the control system adaptively triggers corresponding life extension strategies based on the dominant inconsistency type (such as SOC type, capacity type, or internal resistance type) identified by SHAP interpretability analysis: for SOC inconsistency, it initiates active balancing; for excessively large capacity dispersion, it implements group scheduling; and for severely aged individual units, it generates maintenance early warning, forming a closed-loop optimization mechanism. Detailed Implementation

[0052] like Figure 1 As shown, the present invention mainly includes four steps.

[0053] Step 1: Energy Storage System Operation Data Acquisition and Capacity Estimation. First, based on the energy storage system's BMS, distributed acquisition units monitor the battery pack's operating status in real time. Key parameters acquired include the system-level total voltage, total current, and total power, as well as the voltage, temperature, and internal resistance changes of each individual battery cell. Then, for data quality control, the system employs a real-time anomaly detection algorithm based on a sliding window, the core formula of which is:

[0054]

[0055] in, mu W and sigma W These represent the mean and standard deviation of the data within the sliding window, respectively. This algorithm can effectively identify transient abnormal data, laying a solid foundation for subsequent accurate assessment of health status.

[0056] In the data preprocessing stage, this invention develops a three-level data cleaning mechanism tailored to the operating characteristics of energy storage power stations. First, a dynamic threshold range is set based on electrochemical characteristics, and a strict hard threshold screening is performed on the individual cell voltage. Simultaneously, an improved statistical method is used to identify abnormal fluctuations in current and temperature. Addressing the common data loss problem caused by communication interruptions in energy storage power stations, the system employs a K-nearest neighbor interpolation algorithm based on spatiotemporal correlation. Its core calculation formula is:

[0057]

[0058] in, x i express i The value of the nearest neighbor sample, d (·) represents the spatiotemporal distance function. epsilon To prevent division by zero errors, a small constant is used. Finally, to address the current noise generated during power regulation in the energy storage power station, a Butterworth low-pass filter is selected for smoothing, and its transfer function is:

[0059]

[0060] in, s c The angular frequency corresponding to the cutoff frequency. n This represents the filter order. This processing flow significantly improves the stability and reliability of the data while preserving the effective signal, providing high-quality data assurance for subsequent accurate capacity calculations.

[0061] The battery capacity is obtained by calculating it over a complete charge-discharge cycle using the variable ampere-hour integral method. The specific calculation formula is as follows.

[0062]

[0063] This integration needs to be performed within a strict SOC range, typically selecting a complete charge-discharge process from below 6% to above 90% to ensure the accuracy of capacity calculation. To more accurately describe the battery aging process, this invention also employs a two-stage empirical model to model the capacity decay trajectory:

[0064]

[0065] in a Represents the initial capacity baseline. b Characterizing the early nonlinear decay rate, c Indicates the linear decay rate in the later stage. EFC This represents the equivalent total number of cycles. This model accurately describes the typical characteristics of a battery's rapid initial degradation followed by linear degradation during its lifespan, providing a theoretical basis for the precise assessment of State of Harmony (SOH).

[0066] In practical applications, the system continuously monitors battery charge and discharge data and dynamically updates the State of Health (SOH) estimate based on the aforementioned model. Especially in the operating environment of energy storage power stations, where charge and discharge cycles are relatively regular, this capacity integral-based SOH calculation method can provide accurate and reliable battery health status assessments, offering key input parameters for subsequent inconsistency analysis and life extension control strategy development.

[0067] Step 2: Based on the high-quality data obtained in Step 1 and the monomeric SOH assessment results, this invention further conducts a system-level health status assessment. (See attached...) Figure 4 Analysis shows that the figure clearly illustrates the distribution characteristics of the energy storage battery pack in the SOC-SOH state space. The SOH of the standard battery cell is mainly distributed in the range of 94.5% to 96.0%, while the overall state of the battery pack is severely constrained by the boundary cells, exhibiting an SOH of <93%. During the charging process, min( E C When a single cell reaches its voltage limit first, the system must stop charging; at this point, that single cell limits the total rechargeable capacity of the entire battery pack. Conversely, during discharge, min( E D When a single cell reaches its lower voltage limit first, the system must stop discharging; at this point, that single cell limits the total discharge capacity of the entire battery pack. The combined effect of these two critical boundary units ultimately determines the actual usable capacity of the battery pack. This visual analysis intuitively reveals the mechanism by which inconsistencies limit system performance—the actual usable capacity of the system depends neither on average performance nor on optimal performance, but rather on the discharge capacity of the weakest cell and the charge acceptance capacity of the strongest cell, thus verifying the correctness of the system usable capacity model proposed in this invention.

[0068] First, addressing the "weakest link" effect unique to series-connected battery packs, a precise model of the system's actual usable capacity is established. Unlike simple summation models, this invention defines the actual usable capacity of the battery pack as the sum of the minimum rechargeable capacity and the minimum dischargeable capacity, with the core expression being:

[0069]

[0070] in C pack It refers to the packaging capacity; This is the minimum rechargeable capacity; This represents the minimum dischargeable capacity. The two minimum values ​​are determined by the limiting capabilities of all individual cells within the system under charging and discharging cutoff conditions. This model theoretically and rigorously characterizes the system-level capacity loss caused by cell inconsistencies, laying the theoretical foundation for subsequent quantitative analysis.

[0071] Based on the completed system health assessment, this invention further constructs a quantitative index system for inconsistency in the energy-capacity space. This system comprehensively reflects the impact of inconsistency on system performance through the energy-capacity dispersion index, and its calculation formula is as follows:

[0072]

[0073] Among them, among them, C max and C min These represent the maximum and minimum available capacity of a single unit in the system, respectively. C pack This represents the actual available capacity of the system. The physical meaning of this index is clear: the first term ( C max + C min -2 C pack The second term reflects the degree of deviation of the actual operating point of the system from that of an ideal uniform system. C max - C min This characterizes the inherent dispersion of the individual components in the system. The product of these two factors constitutes a complete description of the impact of system inconsistencies.

[0074] Furthermore, within the SOC-SOH coupled state space, the capacity difference is calculated ( F 1) Mean state deviation ( F 2) and boundary element offset ( F 3) These three features construct a multidimensional inconsistent feature vector. F =[ F 1, F 2,F 3] T This fully reveals the intrinsic structure and impact mechanism of inconsistency, providing precise input features for subsequent intelligent diagnosis and life extension control.

[0075]

[0076] in, and Let ω represent the system's average SOC and average SOH, respectively. soc and ω SOH These are the weighting coefficients.

[0077] Step 3: Building a battery pack inconsistency prediction model and identifying inconsistency causes. This model uses the multidimensional inconsistency feature vector extracted in Step 2. F =[ F 1, F 2, F 3] T As core inputs, the network aims to predict the evolution of these features over future operating cycles. By learning a mapping from the current state to future states, the network captures the dynamic evolution of inconsistencies during system operation. Its core mapping relationship can be represented as:

[0078]

[0079] in, F t This represents the inconsistency feature vector at the current moment. Theta These are the network parameters. This model can predict capacity differences at future time points (...). F 1) Mean state deviation ( F 2) and boundary element offset ( F 3), thereby enabling forward-looking judgments on inconsistent development trends.

[0080] To ensure that the prediction results conform to the fundamental physical laws governing the inconsistency evolution of battery systems, a composite loss function incorporating physical constraints was employed during the training process. The physical loss term specifically incorporates a monotonicity constraint. This constraint is based on a key physical prior: during battery aging, inconsistencies caused by individual cell differences tend to accumulate or persist macroscopically, rather than spontaneously and significantly decreasing. This physical loss term is defined as follows.

[0081]

[0082] Here, we use the inconsistent feature vectors. LThe 2-norm is used to macroscopically characterize the degree of overall inconsistency in the system. This loss function penalizes predictions that violate physical realities and whose inconsistency decreases significantly over time. The complete composite loss function is:

[0083]

[0084] in, Loss MSE It is the mean square error between the predicted value and the actual value. α and β These are the weighting coefficients. This design ensures that while pursuing high accuracy in data fitting, the model's output conforms to the macroscopic physical laws governing inconsistent evolution.

[0085] The training process employed a systematic optimization workflow, as shown in the attached document. Figure 5 As shown. Training begins with defining the hyperparameter search space, including key parameters such as the number of network layers, number of neurons, and learning rate. The Bayesian optimizer is set to automatically search for these hyperparameters, aiming to find the optimal configuration that minimizes the composite loss function. After the Bayesian optimizer sets a set of hyperparameters, the training loop begins. The model maps current inconsistent features to future states during forward propagation and calculates the composite loss. Subsequently, backpropagation is performed based on the Adam optimizer to update the network parameters. This process integrates gradient pruning and weight decay strategies to prevent overfitting. The system continuously monitors the loss on the validation set and uses this as a criterion for model saving and optimization iteration. Specifically, the algorithm determines whether the "current loss" is better than the "optimal loss." If so, the optimal loss is updated, the current model parameters are saved, and the "number of unimproved iterations" are reset; otherwise, the "number of unimproved iterations" are accumulated. When the "number of unimproved iterations" reaches a preset patience value (total number of iterations) or the "number of optimization iterations" reaches its maximum value N, the current Bayesian optimization iteration ends, and the optimizer records this set of hyperparameters and the corresponding performance. Repeat this process until the Pareto optimal set of hyperparameters is found, and then output the final "optimal model".

[0086] Based on the trained prediction model, the SHAP interpretability analysis framework is introduced to quantify the contribution of each input feature to inconsistent predictions:

[0087]

[0088] in N This represents the complete set of all features used by the model; i The target features that indicate its contribution are being analyzed; N { i} indicates from the complete set N Deleting features i The remaining feature subset; S It meansN { i Any subset of} represents the set of features introduced when the target features are introduced. i A specific combination of features that have already been used in the modeling; f ( S ) indicates that only a subset of features is used. S The model's predicted output; while f ( S ∪{ i}) indicates that in the feature i Merge into subset S New predictions from the mid-to-late stage model. Overall SHAP value. phi i By using all possible subsets S This is obtained by performing an appropriate weighted summation, thus accurately quantifying the features. i The average marginal contribution of the model to predictions across different combinations of feature contexts.

[0089] Step 4: After completing inconsistency prediction and identifying the dominant causes, the system enters the intelligent life extension control phase. (See attached diagram) Figure 6 As shown, this stage begins with a system status update: the system first updates the historical status files of all individual cells, accurately locates key influencing cells based on SHAP interpretability analysis results, and records the type and degree of inconsistency. The system monitors the consistency level index in real time, and when the index exceeds a preset safety threshold, it adaptively triggers a hierarchical control strategy: for systems dominated by SOC inconsistency, active balancing control is initiated to redistribute the state of charge through energy transfer; for systems dominated by capacity or internal resistance inconsistency, group scheduling control is implemented to allocate appropriate power to battery clusters with different characteristics; and for critical cells that have been severely aged, a replacement warning is generated, and a maintenance intervention plan is planned.

[0090] The system dynamically adjusts control parameters by evaluating the effectiveness of the control strategy in real time, forming a closed-loop optimization mechanism. This adaptive control method based on precise inconsistency diagnosis effectively suppresses the trend of inconsistency development while ensuring the safe operation of the system, maximizes the utilization of the system's remaining capacity, significantly extends the overall lifespan of the battery system, and improves the economy and operational reliability of the energy storage system.

Claims

1. A method for health assessment and life extension control considering inconsistencies in large-scale energy storage battery systems, comprising the following steps: S1: High-precision collaborative estimation of the health status and state of charge of each individual battery cell; S11: Acquire the operating data of the energy storage battery system, including time, total voltage, current, cell temperature, and cell voltage, record and store the above operating data; when cleaning the operating data, firstly use the dynamic threshold range based on electrochemical characteristics to filter out abnormal data; then use the K-nearest neighbor interpolation method to fill in missing values ​​to ensure the continuity of time series data; finally use a low-pass filter to smooth the noisy data to ensure the stability and reliability of the data. S12: When calibrating the battery SOC, select the operating window of the system that has been idle for a long time, and obtain the individual cell voltage and temperature data after sufficient rest; then use the SOC-OCV lookup table to calibrate the battery SOC; based on this, use the variable ampere integration method to obtain the data of the complete charge and discharge range, and calculate the actual capacity of the battery. in, C It is the calculated battery capacity; I ( t ( ) is the charging / discharging current; t start , t end Indicates the start and end times of the integration; SOC start , SOC end Indicates the start and end points SOC Value. (Select only) SOC start (<6% data segment and complete charge / discharge logs to ensure reliability.) S13: Based on the physical laws of battery aging, a two-stage empirical model is used to derive the equivalent cycle capacity. First, the equivalent full cycle number EFC is calculated by the total throughput of the statistical cycle. Then, an empirical model that integrates nonlinear and linear decay characteristics is established, and the parameters are solved by fitting historical data using the least squares method to achieve accurate prediction of the battery capacity decay trajectory. in, C Represents the capacity of the battery cell; a Represents the initial capacity baseline; b Represents the early rapid decay rate; c S2: Represents linear degradation rate; S3: Battery system health assessment and inconsistency quantification; S41: Accurate modeling of the actual usable capacity of the system based on inconsistency. The actual usable capacity of a series-connected battery pack is limited by the weakest cell in the system, rather than the simple sum of the capacities of all cells. To accurately characterize the capacity loss caused by inconsistency, the actual usable capacity of the system is defined as the sum of the minimum rechargeable capacity and the minimum dischargeable capacity. ;in C pack It refers to the packaging capacity; This is the minimum rechargeable capacity; S22: Quantitative analysis of inconsistency in the energy-capacity space. The definition of this index is based on the following theoretical derivation: In an ideal state, if all battery cells in the system have completely identical characteristics, their energy-capacity relationship will exhibit a perfectly linear characteristic; however, in actual systems, due to the dispersion of parameters between cells, the energy-capacity curve will be distorted, resulting in a reduction in the usable capacity area. Based on this physical phenomenon, the formula for calculating the energy-capacity dispersion index is as follows; in, C max and C min These represent the maximum and minimum available capacity of a single unit in the system, respectively. C pack This represents the actual available capacity of the system. The physical meaning of this index is clear: the first term ( C max + C min -2 C pack The second term reflects the degree of deviation of the actual operating point of the system from that of an ideal uniform system. C max - C min This characterizes the inherent dispersion of individual components in the system. The product of the two constitutes a complete description of the impact of system inconsistency; S23: Accurate extraction of multi-dimensional inconsistency features; S3: Battery system inconsistency prediction and type identification; S31: Constructing a lightweight prediction model guided by physical information. Based on the physical mechanism of battery aging, a constraint relationship is constructed, and the physical constraints of the battery are introduced to guide the machine learning model to learn the battery aging law; a composite loss function is constructed, and the Adam optimizer and cosine annealing learning rate scheduling strategy are used for model training. Label smoothing technology and gradient pruning are introduced to prevent overfitting; the model parameters are updated through adaptive learning rate and weight decay regularization, and a dual evaluation standard of validation loss and physical consistency is established. When the number of consecutive times the validation loss no longer decreases is equal to the number of tolerances, the optimal model parameters are saved; S32: Multi-objective Bayesian optimization. A multi-objective optimization framework is established, simultaneously considering estimation accuracy, computational efficiency, and physical consistency. A Bayesian optimization algorithm based on a tree-structured Parzen estimator is adopted, using a probabilistic model to guide the search process and find the Pareto optimal solution set. A neural architecture search is introduced to automatically design the network structure, and knowledge distillation technology is used to enable large models to guide lightweight network learning. S33: Inconsistency type identification based on SHAP analysis; S4: Intelligent life extension control of the battery system; S41: By establishing a mapping relationship between multi-dimensional features and control strategies, a hierarchical control mechanism including equalization management, group scheduling, and maintenance intervention is dynamically triggered. This method effectively suppresses the inconsistency development trend by optimizing the system's operating state in real time, significantly improving the system's available capacity and extending its overall service life while ensuring safe system operation; S42: A collaborative optimization framework for life prediction and health management is constructed, dynamically adjusting control strategy parameters based on real-time system state assessment and aging trend prediction. By combining preventive maintenance and adaptive scheduling, performance maintenance and life extension are achieved throughout the system's entire life cycle, maximizing the economy and reliability of the energy storage system.

2. The health assessment and life extension control method for large-scale energy storage battery systems considering inconsistencies as described in claim 1, characterized in that, The specific method of step S11 is as follows: In the data preprocessing stage, this invention develops a three-level data cleaning mechanism tailored to the operating characteristics of energy storage power stations. First, a dynamic threshold range is set based on electrochemical characteristics, and a strict hard threshold screening is performed on the individual cell voltage. Simultaneously, an improved statistical method is used to identify abnormal fluctuations in current and temperature. Addressing the common data loss problem caused by communication interruptions in energy storage power stations, the system employs a K-nearest neighbor interpolation algorithm based on spatiotemporal correlation. Its core calculation formula is: in, x i express i The value of the nearest neighbor sample, d (·) represents the spatiotemporal distance function. ϵ To prevent division by zero errors, a small constant is used. Finally, to address the current noise generated during power regulation in the energy storage power station, a Butterworth low-pass filter is selected for smoothing, and its transfer function is: in, s c The angular frequency corresponding to the cutoff frequency. n This represents the filter order. This processing flow significantly improves the stability and reliability of the data while preserving the effective signal, providing high-quality data assurance for subsequent accurate capacity calculations.

3. The health assessment and life extension control method for large-scale energy storage battery systems considering inconsistencies as described in claim 1, characterized in that, The specific method of step S23 is as follows: To achieve a comprehensive characterization of battery system inconsistencies, this invention constructs a multi-dimensional feature extraction system within the SOC-SOH coupled state space. This system, based on the distribution characteristics of individual battery cells in the state space, reveals the intrinsic structure and influencing mechanisms of inconsistencies from different dimensions. Based on the coupled SOC-SOH state space diagram, we calculate the capacity difference (…). F 1) Mean state deviation ( F 2) and boundary element offset ( F 3). These three features work together. F 1. F 2 and F The larger values ​​of 3 collectively indicate that, as defined by the mathematical expression below, the inconsistency of the battery pack is mainly caused by capacity degradation; ;in, and Let ω represent the system's average SOC and average SOH, respectively. soc and ω SOH These are the weighting coefficients.

4. The health assessment and life extension control method for large-scale energy storage battery systems considering inconsistencies as described in claim 1, characterized in that, Step S33 introduces the SHAP interpretability analysis framework to quantify the marginal contribution of each inconsistency feature to the prediction of system performance degradation and to accurately identify the dominant inconsistency type. The SHAP value provides a consistent and interpretable contribution allocation for model predictions by calculating the average marginal contribution of a feature across all possible feature subsets. i SHAP value of each feature ϕ i The calculation is as follows: ;in N This represents the complete set of all features used by the model; i The target features that indicate its contribution are being analyzed; N { i } indicates from the complete set N Deleting features i The remaining feature subset; i>S refers to N { i Any subset of} represents the set of features introduced when the target features are introduced. i A specific combination of features that have already been used in the modeling; f ( S ) indicates that only a subset of features is used. S The model's predicted output; while f ( S ∪{ i }) indicates that in the feature i Merge into subset S New predictions from the mid-to-late stage model. Overall SHAP value. ϕ i By using all possible subsets S This is obtained by performing an appropriate weighted summation, thus accurately quantifying the features. i The average marginal contribution of the model to predictions across different combinations of feature contexts.

5. The health assessment and life extension control method for large-scale energy storage battery systems considering inconsistencies as described in claim 1, characterized in that, The calculation formula for step S32 is as follows: in, F t This represents the inconsistency feature vector at the current moment. Θ These are the network parameters. This model can predict capacity differences at future time points (...). F 1) Mean state deviation ( F 2) and boundary element offset ( F 3) This allows for a forward-looking judgment on the development trend of inconsistencies. To ensure that the prediction results conform to the basic physical laws of the evolution of inconsistencies in battery systems, a composite loss function with physical constraints is used in the training process. The physical loss term specifically incorporates a monotonicity constraint. This constraint is based on a key physical prior: during battery aging, inconsistencies caused by differences between individual cells tend to accumulate or remain stable macroscopically, rather than spontaneously and significantly decreasing. The physical loss term is defined as follows: Here, the inconsistent feature vectors are used. L The 2-norm is used to macroscopically characterize the degree of overall inconsistency in the system. This loss function penalizes predictions that violate physical realities and whose inconsistency decreases significantly over time. The complete composite loss function is: in, Loss MSE It is the mean square error between the predicted value and the actual value. α and β These are the weighting coefficients. This design ensures that while pursuing high accuracy in data fitting, the model's output conforms to the macroscopic physical laws governing inconsistent evolution.

Images