Energy storage system battery cell consistency detection method based on density peak clustering and evidence classification, medium and program product

Abstract

The invention belongs to the technical field of crossing of energy storage systems, data processing and data driving algorithms, and discloses an energy storage system single battery consistency detection method based on density peak clustering and evidence classification, a medium and a program product. Analyzing the association relationship between the historical data of the energy storage system and the related influence factors, and constructing a historical sample library; establishing an energy storage system feature extraction model based on a clustering algorithm, and dividing each battery monomer of the energy storage system into a plurality of categories; and carrying out decreasing mapping on the distance between the real-time operation data and the class center to obtain similarity, completing evidence distribution, calculating and identifying framework evidence, and judging and positioning inconsistent monomers. According to the invention, the accuracy and reliability of consistency detection of the energy storage system can be improved, inconsistent battery monomers in the energy storage system can be accurately positioned, and the reliability and safety of operation of the energy storage system can be improved.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary technical field of energy storage systems, data processing, and data-driven algorithms. It involves data-driven analysis for battery cell consistency evaluation, anomaly identification and localization. Specifically, it is a method, medium and program product for battery cell consistency detection in energy storage systems based on density peaks clustering (DPC) and evidence classification. Background Technology

[0002] With the large-scale grid connection of renewable energy, energy storage systems play a crucial role in modern power systems, primarily for smoothing power fluctuations and improving energy efficiency. Energy storage systems typically consist of numerous battery cells connected in series or parallel to meet voltage and capacity requirements, and are managed by a Battery Management System (BMS) that collects and manages information such as voltage, temperature, current, and State of Charge (SOC). However, due to inherent differences in the production and manufacturing processes of individual battery cells, as well as variations in operating environment factors such as temperature and load, the characteristics of individual battery cells exhibit inconsistencies in internal resistance and SOC. This is reflected in the overall charging and discharging capacity of the energy storage system. During charging, the system is affected by the battery cell that reaches its maximum SOC earliest, and during discharging, it is affected by the battery cell that reaches its minimum SOC earliest. If this bottleneck effect is ignored and all battery cells continue to charge and discharge, it will lead to overcharging, over-discharging, localized overheating, and even serious consequences such as battery explosions and fires.

[0003] Currently, consistency testing methods for individual battery cells in energy storage systems can be broadly categorized into two types: mechanistic models and data-driven methods. While mechanistic models offer strong interpretability, they artificially ignore deviations during actual energy storage operation and the influence of factors such as aging, resulting in lower accuracy. Data-driven algorithms treat the energy storage system as a black box, employing modeling and analysis methods based on clustering and other algorithms to identify outliers, achieving higher accuracy. However, the accuracy of current data-driven methods still needs further improvement. Due to the large volume of data, much of which is repetitive and lacks sufficient information, and the significant non-stationarity influenced by external environments and scheduling strategies, the model's generalization ability, robustness, and online stability face challenges. Furthermore, the scarcity of anomaly sample annotations, significant differences in data distribution between different systems, and the lack of unified and verifiable evaluation metrics can easily lead to difficulties in setting detection thresholds and insufficient interpretability of location results, among other engineering implementation issues.

[0004] In existing technologies, taking a battery consistency detection method disclosed in Chinese patent CN120490817A as an example, it evaluates the battery status by extracting index features and combining a risk accumulation algorithm with dynamic weights. However, this method heavily relies on after-sales feedback data for weight correction, making it difficult to apply effectively in the early stages of energy storage operation where prior labels are lacking. Another example is a battery consistency detection method disclosed in CN120539614A, which identifies abnormal cells using the Local Outlier Factor (LOF) algorithm and optimizes parameters using reinforcement learning to achieve automatic parameter optimization. However, the LOF algorithm has high computational complexity when processing large-scale, high-dimensional data, and the reinforcement learning mechanism still requires manual confirmation of the results as a reward signal, limiting the system's ability to operate fully autonomously in unattended environments. These methods still have limitations in handling multi-source information coupling and unsupervised feature mining.

[0005] In summary, existing methods for detecting the consistency of individual battery cells in energy storage systems still suffer from insufficient online applicability under limited measurement points, inadequate robustness to complex operating conditions, difficulty in balancing false alarms and missed alarms, and insufficient verifiability of location results. Therefore, in energy storage system scenarios characterized by limited measurement points, variable operating conditions, and inconsistent data quality, achieving accurate assessment of individual battery cell consistency and reliable location of inconsistent cells under complex operating conditions is a pressing technical problem that needs to be solved in this field. Summary of the Invention

[0006] (a) Purpose of the invention To address the aforementioned deficiencies and shortcomings of existing technologies, this invention aims to provide a method, medium, and program product for detecting the consistency of battery cells in energy storage systems based on density peak clustering and evidence classification. This method involves cleaning, filtering, and normalizing multi-time-period operational data from multiple energy storage systems. Density peak clustering is used to perform unsupervised partitioning of historical samples, determining cluster centers based on local density and relative distance, and extracting cluster centers and category structures. Furthermore, evidence classification is combined to perform consistency discrimination on real-time samples, identifying abnormal cells that do not belong to existing category distributions and outputting discrimination results that can be used for location and early warning. This achieves accurate assessment of battery cell consistency and reliable detection and location of inconsistent cells, thereby reducing the risk of false alarms and missed alarms and improving the operational safety and reliability of energy storage systems.

[0007] (II) Technical Solution To achieve the objective of this invention and solve its technical problems, the present invention adopts the following technical solution: The first objective of this invention is to provide a method for detecting the consistency of individual battery cells in an energy storage system based on density peak clustering and evidence classification, comprising at least the following steps: S100. Historical Data Acquisition and Preprocessing: Collect historical operating data of individual battery cells from multiple energy storage systems at multiple time scales, preprocess the data, and form an effective time-series dataset; S200. Analysis of Influencing Factors and Construction of Historical Sample Library: Collect influencing factor data corresponding to historical operating data, conduct correlation analysis between each influencing factor and the characteristics of the dataset to determine key influencing factors, and construct a historical sample library of energy storage system after merging the key influencing factors with the dataset. S300. Feature extraction based on density peak clustering: The historical sample database is divided into a training set and a test set. Density peak clustering is performed on the training set, including calculating the pairwise distance d between samples. ij And determine the cutoff distance d c Based on d c Calculate the local density ρ of each sample i With relative distance δ i Based on this, cluster centers and the number of categories are determined, historical samples are divided into multiple categories and the core information of each category is extracted. The core information is used to characterize the typical operating mode of a battery cell under different consistency states. S400. Consistency detection based on evidence classification: Collect real-time operating data of each battery cell in the energy storage system, calculate the distance between it and the class center of each category, and obtain the similarity by mapping through a decreasing function. Based on the similarity, assign evidence to each category and calculate the identification frame evidence m(Ω). S500. Inconsistent Cell Judgment and Location: The consistency of battery cells is judged based on the value of the evidence m(Ω) in the identification framework. When m(Ω) meets the preset inconsistency criterion, battery cells that do not belong to any category are extracted and the location result is generated. Otherwise, the consistency evaluation result of the corresponding battery cell is output.

[0008] S600. Online update (preferred step): When the preset data quality and consistency conditions are met, the feature vector corresponding to the real-time running data is written into the historical sample library, and the class core information is incrementally updated according to the preset update cycle; when the feature distribution of the historical sample library is detected to have changed beyond the preset drift threshold relative to the previous update cycle, step S300 is triggered to be re-executed to update the class structure, thereby maintaining the stability of consistency detection under long-term operation.

[0009] The second objective of this invention is to provide a computer program product, including computer instructions, for executing the steps of the above-described energy storage system battery cell consistency detection method based on density peak clustering and evidence classification.

[0010] The third objective of this invention is to provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described method for detecting the consistency of battery cells in an energy storage system based on density peak clustering and evidence classification.

[0011] (III) Technical Effects Compared with existing technologies, the energy storage system battery cell consistency detection method, medium, and program product based on density peak clustering and evidence classification provided by this invention have the following significant technical effects: (1) This invention preprocesses historical operating data at multiple time scales and constructs a historical sample library by combining correlation analysis of influencing factors, so that consistency evaluation can be transformed from a single voltage / temperature index to a multi-dimensional feature characterization. Under the condition of limited measurement point types, it can improve the effective information density, reduce the interference of noise, static sections and repeated segments on the evaluation results, thereby improving the stability and verifiability of consistency detection.

[0012] (2) The present invention uses density peak clustering to determine the cluster center based on local density and relative distance and extract the cluster center information. It can form a category structure of typical operating mode of battery cell without prior category label, reduce the dependence on specific operating conditions and human experience thresholds, enhance the adaptability to different operating stages and different system differences, and provide a stable reference benchmark for subsequent online discrimination.

[0013] (3) In the online detection, the present invention introduces an evidence classification framework, uses the similarity between real-time samples and various types of cores to allocate evidence and calculate the identification framework evidence m(Ω), and explicitly quantifies the unclassifiable uncertainty into an inconsistency measure, so as to realize the reliable identification and location of inconsistent cells; improves the accuracy and reliability of consistency detection of energy storage system, realizes the precise location of inconsistent battery cells in energy storage system, improves the reliability and safety of energy storage system operation, and thus effectively improves the operational reliability, safety and stability of energy system with high renewable energy penetration. Attached Figure Description

[0014] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0015] Figure 1 The flowchart illustrates a method for detecting the consistency of individual battery cells in an energy storage system based on density peak clustering and evidence classification, as provided in an embodiment of the present invention.

[0016] Figure 2This is a flowchart of the feature extraction process based on density peak clustering in this invention.

[0017] Figure 3 This is the density peak clustering decision map in this invention. The blue dots in the map represent 87,436 training samples, and the five points marked with red pentagrams represent the identified cluster centers, which correspond to the five typical operating modes of the battery: standby, low power charging, high power charging, discharging, and abnormal temperature.

[0018] Figure 4 The figure shows the distribution of m(Ω) of the complete evidence for 96 battery cells. The green dots represent 94 battery cells with normal consistency, and their m(Ω) values ​​are all less than the set uncertainty threshold T=0.12. The red diamond marks represent the two inconsistent cells detected, namely cell #23 and cell #67, with m(Ω) values ​​of 0.187 and 0.214, respectively, both exceeding the threshold T=0.12.

[0019] Figure 5 The graph shows the trend of inconsistent monomer detections over a 12-month period. The dots connected by the blue solid line represent the actual detections in each month. The two nodes marked with orange pentagrams are located in the 4th and 9th months, respectively, representing the time points when online updates were performed. Detailed Implementation

[0020] This invention aims to provide a method, medium, and program product for detecting the consistency of battery cells in energy storage systems based on density peak clustering and evidence classification. To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions in the embodiments of this invention will be described in more detail below with reference to the accompanying drawings. The described embodiments are some, but not all, embodiments of this invention, and are exemplary and intended to explain the invention, not to limit it.

[0021] Example 1: Method for Testing the Consistency of Individual Battery Cells in Energy Storage Systems like Figure 1 As shown in the figure, the battery cell consistency detection method for energy storage systems based on density peak clustering and evidence classification provided by the present invention is used to detect battery cell consistency and identify inconsistent battery cells during the operation of energy storage systems with limited measurement points and variable operating conditions. The method mainly includes the following steps when implemented. S100. Historical Data Acquisition and Preprocessing: Historical operational data of individual battery cells from multiple energy storage systems at multiple time scales are collected and preprocessed to form a valid time-series dataset. Specifically, this includes collecting and organizing time-series data on temperature and terminal voltage measurements of each battery cell in the energy storage system; collecting and organizing historical State of Charge (SOC) data from the battery management system, as well as charging and discharging current data and operating status identifiers characterizing operating conditions; filtering the data; detecting and removing missing values ​​or imputing or imputing data at each measurement point; removing outliers; and aligning sampling timestamps to form multivariate samples at the same time. Outliers and abrupt changes exceeding preset physical boundaries are removed, and the remaining data is normalized. Data in a fixed state (i.e., batteries not charged or discharged) in the energy storage system's charging and discharging data are removed; specifically, static sections are identified and removed based on a joint criterion of current amplitude, SOC change rate, and voltage change rate, to form a valid time-series dataset.

[0022] As a preferred approach, historical operating data of individual battery cells in the energy storage system are collected at multiple time scales, including short-term scales based on sliding time windows and long-term scales based on the charging and discharging process. When forming an effective time-series dataset, the historical operating data is segmented according to the time window length and step size, and statistical features including at least the mean, standard deviation, range, slope, and cumulative change are extracted in each segment as input features for subsequent construction of the historical sample library and consistency detection.

[0023] S200. Analysis of Influencing Factors and Construction of Historical Sample Database: Data on influencing factors corresponding to historical operating data is collected. Correlation analysis is performed on each influencing factor and the characteristics of the dataset to identify key influencing factors. These key influencing factors are then integrated with the dataset to construct a historical sample library for the energy storage system. Specifically, this includes collecting and organizing data on the number of charge / discharge cycles of individual battery cells, meteorological factors of the environment (such as ambient temperature and humidity), and other relevant influencing factors such as the state of charge and temperature of the energy storage system. Further information on the installation location or heat dissipation conditions of individual battery cells is also included. All relevant data are normalized, and correlation analysis is used to calculate the correlation index between each influencing factor and the characteristic quantities of historical operating data. Based on the absolute value and significance test results of the correlation index, key influencing factors are extracted, an error evaluation system is constructed, and finally, the characteristic quantities of historical operating data are formed, including at least voltage deviation characteristics, temperature deviation characteristics, and state of charge deviation characteristics.

[0024] A historical sample library for energy storage is constructed by taking key influencing factors from historical data, as well as parameters such as state of charge (SOC) and temperature that directly reflect the consistency of individual battery cells in the energy storage system, as a whole. Preferably, when integrating key influencing factors with the dataset to construct the historical sample library, the following steps are taken: normalizing the historical operating data and influencing factor data separately to map features of different dimensions to a unified numerical range; constructing sample vectors for each battery cell, including at least the temperature, terminal voltage, SOC statistical characteristics of the battery cell, and the statistical characteristics of the key influencing factors; establishing an error evaluation index to measure differences between cells, and then writing the samples into the historical sample library after quality screening based on the error evaluation index.

[0025] S300. Feature extraction of energy storage systems based on density peak clustering: The historical sample database is divided into a training set and a test set. Density peak clustering is performed on the training set, including calculating the pairwise distance d between samples. ij And determine the cutoff distance d c Based on d c Calculate the local density ρ of each sample i With relative distance δ i Based on this, cluster centers and the number of categories are determined, historical samples are divided into multiple categories, and cluster center information for each category is extracted. This cluster center information is used to characterize the typical operating modes of individual battery cells under different consistency states. Specifically, such as... Figure 2 As shown, feature extraction based on density peak clustering includes at least the following sub-steps during implementation: S301. Training data preparation: Divide the historical sample library in step S200 into training set and test set according to a preset ratio, and perform dimensionless processing on the feature vectors of each sample in the training set. S302. Distance Matrix Construction and Truncated Distance Determination: Using the training data as input, density peak clustering is employed to perform cluster analysis on the historical data. First, the distance d between any two samples in the training set is calculated based on their feature vectors. ij And form a distance matrix for all d ij Sort the data and select the distance corresponding to the preset quantile as the cutoff distance d. c , where d c Used to define the neighborhood range of a sample; S303. Local density calculation: For each sample i in the training set, based on the cutoff distance d c The local density ρ is obtained by counting or weighting neighborhood samples using a truncated kernel or Gaussian kernel function. i , used to characterize the degree of clustering of sample i with other samples in its neighborhood; S304. Relative distance calculation: Calculate the relative distance δ for each sample in the training set. i δi For sample i and all local densities greater than ρ i The minimum distance between samples; when sample i has ρ i When the training set has the maximum local density, let δ i Take the maximum distance from sample i to the remaining samples in the training set; S305. Cluster Center and Number of Categories Determination: Based on γ i =ρ i •δ i Constructing a cluster center discrimination index γ i , according to γ i Sort the samples from largest to smallest and select the top K samples as cluster centers, thus determining the number of categories as K; S306. Class Assignment and Label Generation: Assign the remaining samples outside the cluster centers to their nearest high-density clusters, and assign sample i to the cluster with the smallest distance and a local density higher than ρ. i The category to which the sample belongs is determined, thus obtaining the category label for each sample in the training set; S307. Extraction and output of class center information: For each class, calculate the class center vector based on the feature vector of the samples in that class as the class center information, and establish a mapping relationship between the class center information and the class label and output it for the distance calculation and consistency judgment between the real-time samples and the class centers of each class in the subsequent step S400.

[0026] Peak clustering for finding cluster centers is primarily based on two assumptions: the local density of data points around a cluster center is higher than the sample density of other data points; and the distance between data points at two cluster centers is relatively large. Specifically, the local density ρ... i The metric is and In the formula d c To define the range of local density by cutting off the distance; d ij For data points i , j The Euclidean distance between them. Relative distance. d i It is the benchmark for determining whether a point is a core or a noise point; it is the relationship between a point and its corresponding core. ρ i Maximum distance between points in the dataset Based on local density ρ i Relative distance d c The number of clusters and cluster centers are determined, and the category information of individual battery cells in the energy storage system is analyzed based on the clustering results.

[0027] S400. Consistency Detection Based on Evidence Classification: Real-time operating data of each battery cell in the energy storage system is collected, the distance between the cell and the class center is calculated, and similarity is obtained by mapping using a decreasing function. Evidence is assigned to each class based on similarity, and the identification framework evidence m(Ω) is calculated. The consistency detection based on evidence classification includes at least the following sub-steps: S401. Real-time Sample Construction: Sample and align the real-time operating data of each battery cell during the detection period, and generate real-time feature vectors according to feature construction rules consistent with the historical sample library. x Furthermore, the real-time feature vectors undergo the same dimensionless processing as in step S300 to ensure consistency with the distance metric between the class centers of each category. S402. Distance Calculation and Similarity Mapping: Read the class center information of each category and calculate the real-time feature vector. x Distance d between each category center k And through a monotonically decreasing mapping function, d k Mapping to similarity s k , used to characterize how closely a real-time sample resembles the typical operating pattern of the k-th category; S403. Construction of basic probability assignment: based on the similarity s of each class k As input, construct the basic probability assignment m(C) for evidence classification. k ), such that m(C) k The value increases with increasing sk, and for ∑ k m(C k Normalization constraints are applied; simultaneously, the remaining uncertainty that cannot be fully explained by any category is assigned to the identification framework Ω, resulting in the identification framework evidence m(Ω) = 1 - ∑ k m(C k ); S404. Consistency Evidence Output: The output includes m(C) of each category of evidence. k The evidence vector of the identification framework evidence m(Ω) is used to determine and locate inconsistent individuals based on m(Ω) in subsequent steps.

[0028] Preferably, a monotonically decreasing distance mapping function is used to map the distance between real-time running data and the class centers of each category as similarity, and a Gaussian kernel function is used to calculate the similarity between the real-time running data sample x and the class center c of the kth category. k Similarity S between k The formula for (x) is ,in σ Let ||xc be the bandwidth parameter of the Gaussian kernel function. k || represents the relationship between sample x and class center c. k The Euclidean distance between them; evidence allocation uses the basic probability allocation function, assigning evidence m(C) to the k-th category.k The formula for calculating ) is α is the evidence allocation coefficient, and K is the total number of categories; the formula for calculating the identification framework evidence m(Ω) is: This is used to quantify the degree of uncertainty that a real-time running data sample does not belong to any known category.

[0029] Furthermore, the evidence allocation coefficient α and the bandwidth parameter of the Gaussian kernel function... σ The cross-validation method was used to optimize and determine the value on the test set, including setting the candidate range for α to be 0.85 to 0.99. σ The candidate range is 0.1 to 2.0. Within this range, a grid search method is used to traverse the parameter combinations. For each parameter combination, the classification accuracy and F1 score are calculated on the test set. The parameter combination that optimizes the overall performance is selected as the final α and 0. σ The comprehensive performance index is evaluated by weighted summation of classification accuracy and F1 score, with the weighting coefficients determined according to actual application requirements.

[0030] It should be noted that in evidence classification algorithms, distance metrics are often used to measure the similarity between different samples, and it is assumed that "similar inputs lead to similar outputs," meaning that samples that are close to each other have the same attributes. For the real-time running data samples to be detected... x It can determine its distance from the historical sample database or training set. k nearest neighbor samples And calculate the corresponding distance di=||xc i ||. Sample x and its nearest neighbor sample x i The similarity between them is determined by a distance-dependent decreasing function φ(d). i The constraints satisfy φ(0) = 1 and Its general form is expressed as an exponential decay function. In the formula γ >0 is the discount factor, used to adjust the rate at which distance diminishes the similarity. When the aforementioned Gaussian kernel distance mapping function is used, it is mathematically equivalent to the aforementioned decreasing function, and a consistent expression can be achieved through parameter conversion to ensure a unified approach to similarity calculation.

[0031] Based on the obtained test samples x i Construct evidence m(·|) belonging to each category. x i Let the recognition frame be Ω = {ω1, ω2, ..., ω}. K} represents the set of K categories obtained in step S300, and the nearest neighbor sample x i Category ω q Then for the set of propositions The allocation of evidence satisfies: In the formula: α Evidence is assigned a coefficient, which is a fixed value used to characterize the credibility of standard evidence; m(A | x i The reliability of the assignment to category A; m (Ω∣ x i () is used to identify frame evidence and quantify samples. x Uncertainty that is difficult to be fully explained by any known category. Furthermore, the conditional basic probability assignments of the k nearest neighbors are fused or converged to obtain the comprehensive evidence m(C) for each category of sample x. k Based on this, the identification framework evidence m(Ω) is obtained, which serves as the basis for subsequent steps S500 to determine and locate inconsistent individuals. The larger the value of m(Ω), the greater the probability that sample x does not belong to any known category.

[0032] S500. Identification and Location of Inconsistent Individual Components: Identification framework evidence based on evidence classification m The value of (Ω) is used to analyze the consistency of parameters such as the state of charge of the energy storage system. m The larger the m(Ω) value, the more likely the battery cell is to be an unknown category, i.e., an inconsistent battery cell. When m(Ω) meets the preset inconsistency criterion, battery cells that do not belong to any category are extracted and a location result is generated; otherwise, the consistency evaluation result of the corresponding battery cell is output. The preset inconsistency criterion is determined by the statistical distribution of m(Ω) in the historical sample library, and includes at least one or two of the following: the threshold criterion is that when m(Ω) is greater than or equal to the preset threshold T, the corresponding battery cell is determined to be inconsistent; the relative criterion is that the m(Ω) of all battery cells in the same time window is sorted and the battery cell with the highest sorted value and a value exceeding the preset ratio difference is determined to be inconsistent. Furthermore, when the m(Ω) of the same battery cell exceeds the inconsistency threshold for N consecutive detection periods, the battery cell is determined to be an inconsistent battery cell and a location result is generated. At the same time, a detection record containing the cell identifier, m(Ω) value, and inconsistency level is output.

[0033] Preferably, in steps S300 to S500, consistency detection is performed in parallel at at least two different time scales to obtain the identification framework evidence m(Ω) for the corresponding time scale, and a fusion inconsistency metric is formed by weighted fusion or evidence synthesis rules; in step S500, inconsistent individual is determined and located based on the fusion inconsistency metric to reduce the sensitivity of a single time scale to operating condition fluctuations and improve the reliability of detecting abnormal individuals under complex operating conditions.

[0034] In addition, to adapt to the data distribution changes caused by limited measurement points and variable operating conditions, steps S400~S500 use a sliding time window for online consistency detection. The sliding time window has a preset window length and step size. In each window, feature vectors consistent with the historical sample database are extracted from the real-time running data and then evidence classification is performed. Battery cells whose m(Ω) satisfies the preset inconsistency criterion in multiple consecutive windows are identified as continuously inconsistent battery cells, thereby suppressing false alarms caused by transient disturbances.

[0035] Preferably, steps S200 to S400 further include establishing a working condition segment consistency detection mechanism. The working condition segments are labeled according to the charging / discharging direction, rate range, and ambient temperature range, and databases are built or core information is calculated separately. When classifying evidence in step S400, the core information corresponding to the working condition to which the real-time operating data belongs is selected first for similarity calculation and evidence allocation. This reduces the interference caused by the difference in feature distribution between different working conditions on consistency judgment and improves the stability of consistency detection under cross-seasonal and cross-strategy operating conditions.

[0036] S600. Online update (preferred step): When the preset data quality and consistency conditions are met, the feature vector corresponding to the real-time running data is written into the historical sample library, and the class core information is incrementally updated according to the preset update cycle; when the feature distribution of the historical sample library is detected to have changed beyond the preset drift threshold relative to the previous update cycle, step S300 is re-executed to update the class structure and maintain the consistency detection stability under long-term operation.

[0037] Example 2: Application Case This embodiment 2 uses a grid-side lithium iron phosphate energy storage power station as an example for illustration. Each energy storage cluster of this power station consists of 96 strings of battery cells. The sampling period is 1 minute, and 30 days of operation data are continuously collected, including cell voltage, cell temperature, SOC, cluster current and operation status indicators. At the same time, influencing factors such as ambient temperature and humidity, charge and discharge cycles and battery compartment heat dissipation status indicators are also collected.

[0038] Data and feature settings: For short-term timescales, a sliding window of 30 min and a step size of 5 min are used for segmentation; for long-term timescales, segments are based on a single charge-discharge process. For each segment, the mean, standard deviation, range, slope, and cumulative change of voltage / temperature / SOC are extracted, and voltage deviation, temperature deviation, and SOC deviation features are constructed as inputs for error evaluation. The training and test sets are divided in an 8:2 ratio.

[0039] Clustering and Recognition Framework Construction: Perform density peak clustering on the training set sample segments and set the cutoff distance d. cThe local density ρ is calculated as 1.5% of the average Euclidean distance between samples. i With relative distance δ i Decision value γ i =ρ i ×δ i Used to identify cluster centers. For example... Figure 3 As shown in the decision graph, the five samples with significantly higher γ values ​​than other sample points were identified as cluster centers, corresponding to five typical operating modes: standby, low-power charging, high-power charging, discharging, and abnormal temperature. Distance distributions were calculated within each cluster, and the 95th percentile was used as the distance threshold D for that mode. th,k The prior probability P(Ck) and evidence weight wk of each mode were statistically determined using training set samples. When constructing the recognition framework, the focal element set Θ={C1,C2,C3,C4,C5} and the global set Ω=Θ were used. Multi-scale evidence was fused using combination rules, and an uncertainty threshold T=0.12 was set.

[0040] Consistency Detection and Online Update: The test set sample segment is inspected, and the basic probability allocation m(C) of each battery cell is calculated. k ) and the complete set of evidence m(Ω). For example Figure 4 As shown, 94 out of 96 cells had an m(Ω) < 0.12, indicating normal consistency. Two cells (numbered #23 and #67) had m(Ω) values ​​of 0.187 and 0.214 respectively, exceeding the threshold T, and were therefore identified as inconsistent cells. Further analysis revealed that cell #23 exhibited a persistently high voltage deviation during charging, while cell #67 showed an abnormal temperature deviation during discharging, consistent with the increased internal resistance and thermal management failure observed in actual maintenance records. The detection accuracy reached 97.9%, with a false alarm rate of only 1.3%.

[0041] After six months of continuous operation, an online update mechanism was initiated to address the changes in operating modes introduced by newly added seasonal climate variations and power plant power adjustments. Newly collected sample segments from the past seven days were selected as the update dataset, and incremental cluster analysis was performed, identifying one new cluster center corresponding to the summer high-temperature frequency regulation mode. Following the online update process in step S7, the new mode was incorporated into the identification framework, the focal element set was expanded to Θ'={C1,C2,C3,C4,C5,C6}, and the distance threshold D was recalculated. th,6 =8.73 and prior probability P(C6)=0.078, and update the evidence weights for each mode. After the update, in the subsequent 14 days of running data, the number of inconsistent monomers detected decreased from an average of 3.2 per day before the update to 1.1, and the false alarm rate decreased from 2.7% to 0.8%. Figure 5 As shown, the adaptability of the online update mechanism to environmental changes has been verified.

[0042] The energy storage power station was continuously monitored for 12 months, with a comprehensive consistency test performed monthly. A total of 23 inconsistent cells were detected. After manual review and maintenance verification, 21 were confirmed to be actual faults or performance degradation cells (including abnormal internal resistance, capacity decay, and thermal management failure), while 2 were false alarms (due to temporary sensor malfunctions). The detection sensitivity reached 91.3%, and the specificity reached 98.7%. Two online updates were performed in the 4th and 9th months, respectively. After each update, the false alarm rate decreased by approximately 1.5%, demonstrating the long-term stability and adaptability of the method.

[0043] The objectives of this invention have been fully and effectively achieved through the above embodiments. Those skilled in the art will understand that this invention includes, but is not limited to, the contents described in the accompanying drawings and the specific embodiments described above. Although the invention has been described with reference to what is currently considered the most practical and preferred embodiments, it should be understood that the invention is not limited to the disclosed embodiments, and any modifications that do not depart from the functional and structural principles of the invention will be included within the scope of the claims.

Claims

1. A method for detecting the consistency of individual battery cells in an energy storage system based on density peak clustering and evidence classification, characterized in that, At least the following steps are included: S100. Collect historical operating data of individual battery cells from multiple energy storage systems at multiple time scales, preprocess the data, and form an effective time-series dataset; S200. Collect influencing factor data corresponding to historical operating data, conduct correlation analysis between each influencing factor and the characteristics of the dataset to determine key influencing factors, and construct a historical sample library of energy storage system after merging the key influencing factors with the dataset. S300. Divide the historical sample database into a training set and a test set, and perform density peak clustering on the training set, including calculating the pairwise distance d between samples. ij And determine the cutoff distance d c Based on d c Calculate the local density ρ of each sample i With relative distance δ i Based on this, the cluster centers and the number of categories are determined, historical samples are divided into multiple categories, and the cluster center information of each category is extracted; S400. Collect real-time operating data of each battery cell in the energy storage system, calculate the distance between it and the class center of each category, and obtain the similarity by mapping through a decreasing function. Based on the similarity, assign evidence to each category and calculate the identification frame evidence m(Ω). S500. Determine the consistency of battery cells based on the value of the evidence m(Ω) in the identification framework. When m(Ω) meets the preset inconsistency criterion, extract battery cells that do not belong to any category and generate the location result; otherwise, output the consistency evaluation result of the corresponding battery cell.

2. The method according to claim 1, characterized in that, In step S100, the historical operating data includes at least the temperature, terminal voltage, and state of charge (SOC) time-series data of each battery cell in the energy storage system, as well as the charging and discharging current data and operating status identifier that characterize the operating conditions. The preprocessing of the collected data includes at least the following: detecting and removing or imputing missing values ​​in the data of each measuring point, and aligning the sampling timestamps to form multivariate samples at the same time. Abnormal points and abrupt changes that exceed the preset physical boundaries are removed, and the remaining data are normalized. Stagnant sections are identified and their data are removed based on a joint criterion of current amplitude, SOC change rate and voltage change rate.

3. The method according to claim 1 or 2, characterized in that, In step S100, the multiple time scales include a short time scale based on a sliding time window and a long time scale based on the charging and discharging process. When forming an effective time series dataset, the historical running data is segmented according to the time window length and step size, and statistical features including at least the mean, standard deviation, range, slope and cumulative change are extracted in each segment as input features for subsequent construction of the historical sample library and consistency detection.

4. The method according to claim 1, characterized in that, In step S200, the influencing factor data includes at least the number of charge and discharge cycles of the battery cell and meteorological factor data of the environment. The meteorological factor data includes at least the ambient temperature and ambient humidity, and further includes information on the installation location or heat dissipation conditions of the battery cell. The correlation analysis includes calculating the correlation index between the characteristic quantities of each influencing factor and the historical operating data, and determining the key influencing factors based on the absolute value of the correlation index and the significance test results. The characteristic quantities of the historical operating data include at least voltage deviation characteristics, temperature deviation characteristics, and state of charge deviation characteristics.

5. The method according to claim 1 or 4, characterized in that, In step S200, when the key influencing factors are integrated with the dataset to construct a historical sample library, the following steps are included: normalizing the historical operating data and the influencing factor data respectively, so that features of different dimensions are mapped to a unified numerical range; constructing sample vectors with battery cells as objects, including at least the temperature, terminal voltage, SOC statistical characteristics of battery cells and the statistical characteristics of key influencing factors; establishing an error evaluation index to measure the differences between cells, and writing the samples into the historical sample library after quality screening based on the error evaluation index.

6. The method according to claim 1, characterized in that, In step S300, feature extraction based on density peak clustering includes at least the following sub-steps: S301. Training data preparation: Divide the historical sample library into training set and test set according to the preset ratio, and perform dimensionless processing on the feature vectors of each sample in the training set. S302. Distance Matrix Construction and Truncation Distance Determination: Calculate the distance d between samples based on the feature vectors of any two samples in the training set. ij And form a distance matrix for all d ij Sort the data and select the distance corresponding to the preset quantile as the cutoff distance d. c , used to define the neighborhood range of a sample; S303. Local density calculation: For each sample i in the training set, based on the cutoff distance d c The local density ρ is obtained by counting or weighting neighborhood samples using a truncated kernel or Gaussian kernel function. i , used to characterize the degree of clustering of sample i with other samples in its neighborhood; S304. Relative distance calculation: Calculate the relative distance δ for each sample i in the training set. i δ i For sample i and all local densities greater than ρ i The minimum distance between samples; When sample i's ρ i When the training set has the maximum local density, let δ i Take the maximum distance from sample i to the remaining samples in the training set; S305. Cluster Center and Number of Categories Determination: Based on γ i =ρ i •δ i Constructing a cluster center discrimination index γ i , according to γ i Sort the samples from largest to smallest and select the top K samples as cluster centers, thus determining the number of categories as K; S306. Class Assignment and Label Generation: Assign the remaining samples outside the cluster centers to their nearest high-density clusters, and assign sample i to the cluster with the smallest distance and a local density higher than ρ. i The category to which the sample belongs is determined, thus obtaining the category label for each sample in the training set; S307. Extraction and Output of Category Core Information: For each category, calculate the category core vector based on the feature vector of the samples within that category as the category core information, and establish a mapping relationship between the category core information and the category label and output it.

7. The method according to claim 1, characterized in that, In step S400, the consistency check based on evidence classification includes at least the following sub-steps during implementation: S401. Real-time sample construction: Sample and align the real-time operating data of each battery cell during the detection period, generate real-time feature vectors according to the feature construction rules consistent with the historical sample library, and perform the same dimensionless processing on the real-time feature vectors as in step S300. S402. Distance Calculation and Similarity Mapping: Read the class center information of each category and calculate the distance d between the real-time feature vector and the class center of each category. k And through a monotonically decreasing mapping function, d k Mapping to similarity s k , used to characterize how closely a real-time sample resembles the typical operating pattern of the k-th category; S403. Construction of basic probability assignment: based on the similarity s of each class k As input, construct the basic probability assignment m(C) for evidence classification. k ), such that m(C) k ) with s k Increases and increases, and for ∑ k m(C k Normalization constraints are applied; simultaneously, the remaining uncertainty that cannot be fully explained by any category is assigned to the identification framework Ω, resulting in the identification framework evidence m(Ω) = 1 - ∑ k m(C k ); S404. Consistency Evidence Output: The output includes m(C) of each category of evidence. k The evidence vector of the identification framework evidence m(Ω) is used to determine and locate inconsistent individuals based on m(Ω) in subsequent steps.

8. The method according to claim 1 or 7, characterized in that, In step S400, the decreasing function is a monotonically decreasing distance mapping function, and a Gaussian kernel function is used to calculate the distance between the real-time running data sample x and the k-th class center c. k Similarity S between k The formula for (x) is ,in σ Let ||xc be the bandwidth parameter of the Gaussian kernel function. k || represents the relationship between sample x and class center c. k The Euclidean distance between them; evidence allocation uses the basic probability allocation function, assigning evidence m(C) to the k-th category. k The formula for calculating ) is α is the evidence allocation coefficient, and K is the total number of categories; the formula for calculating the identification framework evidence m(Ω) is: This is used to quantify the degree of uncertainty that a real-time running data sample does not belong to any known category.

9. The method according to claim 8, characterized in that, In step S400, the evidence allocation coefficient α and the Gaussian kernel function bandwidth parameter σ The cross-validation method was used to optimize and determine the value on the test set, including setting the candidate range for α to be 0.85 to 0.

99. σ The candidate range is 0.1 to 2.

0. Within this range, a grid search method is used to traverse the parameter combinations. For each parameter combination, the classification accuracy and F1 score are calculated on the test set. The parameter combination that optimizes the overall performance is selected as the final α and 0. σ The comprehensive performance index is evaluated by weighted summation of classification accuracy and F1 score, with the weighting coefficients determined according to actual application requirements.

10. The method according to claim 1, characterized in that, In step S500, the preset inconsistency criterion is determined by the statistical distribution of m(Ω) in the historical sample library, and includes at least one or two of the following: threshold criterion and relative criterion. The threshold criterion is that when m(Ω) is greater than or equal to the preset threshold T, the corresponding battery cell is determined to be inconsistent. The relative criterion is that the m(Ω) of all battery cells are sorted within the same time window, and the battery cell with the highest sorted value and the difference exceeding the preset ratio is determined to be inconsistent. When the m(Ω) of the same battery cell exceeds the inconsistency threshold for N consecutive detection periods, the battery cell is determined to be an inconsistent cell and a positioning result is generated. At the same time, a detection record containing the cell identifier, m(Ω) value and inconsistency level is output.

11. The method according to claim 1 or 10, characterized in that, In steps S300 to S500, consistency detection is performed in parallel at at least two different time scales to obtain the identification framework evidence m(Ω) for the corresponding time scale, and a fusion inconsistency metric is formed through weighted fusion or evidence synthesis rules; in step S500, inconsistent individuals are determined and located based on the fusion inconsistency metric.

12. A computer program product comprising computer instructions, characterized in that, The computer instructions are used to execute the steps of the energy storage system battery cell consistency detection method based on density peak clustering and evidence classification as described in any one of claims 1 to 11.

13. A computer-readable storage medium having a computer program stored thereon, characterized in that, When a computer program is executed by a processor, it implements the steps of the energy storage system battery cell consistency detection method based on density peak clustering and evidence classification as described in any one of claims 1 to 11.

Images