Energy storage power station fault classification method based on local and global features of relaxation stage

Abstract

The invention discloses an energy storage power station fault classification method based on local and global features of a relaxation stage. The method comprises the steps of 1, collecting terminal voltage, current and temperature data of the relaxation stage after charging and discharging in an actual working condition; 2, extracting local features and global features for fault detection from the collected terminal voltage, current and temperature features; 3, selecting a feature subset finally used for fault detection by using a filtering method and a packaging method in feature selection; and 4, battery cluster fault discrimination is carried out based on a classification algorithm random forest, KNN, logistic regression and an integrated voting strategy.

Description

Technical Field

[0001] The present invention relates to the field of energy storage power station fault diagnosis, and in particular to an energy storage power station fault classification method based on local and global characteristics in a relaxation phase. Background Art

[0002] With the proposal of the dual carbon goals, my country has vigorously developed renewable energy, but due to the instability of renewable energy, it is easy to cause energy waste. The energy storage system has the functions of storing electric energy, peak load regulation, frequency regulation and voltage regulation, which can assist the power system in absorbing renewable energy and improve the overall dispatchability of the power grid system. Electrochemical energy storage systems have become a new driving force for the development of the energy storage market due to their strong tolerance to the geographical environment and high energy conversion efficiency. Among them, the scale of energy storage power stations composed of lithium-ion batteries has increased rapidly and steadily.

[0003] my country's energy storage system based on lithium-ion batteries has developed rapidly, but it also faces some problems. The main component of the energy storage system is the lithium-ion battery, which determines the stable and safe operation of the energy storage system. The failure of lithium-ion batteries may cause thermal explosion or temporary shutdown of the energy storage power station, thereby affecting personnel safety and causing economic losses. Therefore, accurate and timely detection of lithium-ion battery failures is the top priority for the normal operation of the energy storage system.

[0004] The static stage before and after the battery is charged or discharged is also called the relaxation stage. The voltage curve during this period is relatively unaffected by the charging and discharging process, and the voltage and current are relatively easy to collect, and the sampling frequency is relatively high. At the same time, battery failures can be detected before and after charging and discharging, avoiding safety hazards and monitoring the battery status more timely.

[0005] The present invention is based on a data-driven approach. By collecting the voltage, current, and temperature of the battery in the relaxation stage, and extracting local and global features that reflect its changing trend on the voltage, a larger feature set is obtained. Then, a machine learning algorithm is used to achieve real-time monitoring and diagnosis of lithium-ion battery cluster faults in the energy storage system. Summary of the invention

[0006] In order to overcome the deficiencies in the prior art, the present invention provides a fault classification method for energy storage power stations based on local and global characteristics of the relaxation phase, so as to achieve real-time monitoring and diagnosis of lithium battery cluster faults in the energy storage system.

[0007] In order to achieve the above-mentioned invention object, the technical solution adopted to solve the technical problem is as follows:

[0008] A method for classifying energy storage power station faults based on local and global characteristics in a relaxation phase includes the following steps:

[0009] Step 1: Collect the terminal voltage, current and temperature data in the relaxation stage after the end of charging and discharging in actual working conditions;

[0010] Step 2: Extract local features and global features for fault detection based on the collected terminal voltage, current, and temperature features;

[0011] Step 3: Use the filtering and packaging methods in feature selection to select the feature subset that is ultimately used for fault detection;

[0012] Step 4: Battery cluster fault identification is performed based on the classification algorithms random forest, KNN, logistic regression and integrated voting strategy.

[0013] Furthermore, in step 1, the current, voltage, and temperature of the battery cluster in the energy storage system during the relaxation stage are acquired in real time by the BMS, and the corresponding battery status is recorded as a training sample, where the battery status includes normal and faulty. By simulating different working conditions in the laboratory, the current, voltage, and temperature of the battery cluster in different states are acquired as an expansion of the training sample. For the online evaluation link, the voltage, current, and temperature data of the battery cluster during the relaxation stage acquired in real time are acquired from the BMS.

[0014] Furthermore, in step 2, the partial data of the terminal voltage of the battery cluster collected is used to find the optimal k, σ, b, U that reflects the battery health status based on the Nelder-Mead simplex search method. oc Four parameters, also called local features that reflect the local change trend of the terminal voltage; six features of the variance, maximum value, minimum value, average value, skewness, and kurtosis of the terminal voltage from the current moment to the beginning of the relaxation phase are calculated using statistical methods to reflect the global change trend of the battery terminal voltage, also called global features; local features, global features, terminal voltage, current, and temperature together constitute a feature set for fault detection. The mathematical relationship between the four parameters and the terminal voltage is as follows:

[0015]

[0016] Among them, k, σ, and b are three parameters that characterize the terminal voltage change curve, also known as the health characteristics of the battery, U L,t represents the terminal voltage of the battery cluster at time t, U oc Indicates the open circuit voltage of the battery cluster;

[0017] The statistical features representing global changes are calculated as follows:

[0018] 1) Variance s 2 ;

[0019]

[0020] Among them, u t represents the terminal voltage of the battery cluster at time t, represents the average value of the battery cluster terminal voltage in the time period [T0, T];

[0021] 2) Skewness S k ;

[0022]

[0023] Wherein, μ3 represents the third-order central moment of the battery cluster terminal voltage in the [T0, T] period, and σ represents the standard deviation;

[0024] 3) Maximum value max;

[0025]

[0026] 4) minimum value min;

[0027]

[0028] 5) average;

[0029]

[0030] 6) Kurtosis

[0031]

[0032] Where n is the sample size, u t represents the terminal voltage of the battery cluster at time t, It represents the average value of the battery cluster terminal voltage in the time period [T0, T].

[0033] Furthermore, in step 3, the Gini index is first used to select feature subset 1 from the basic feature set, and then the decision tree algorithm is used to select feature subset 2 for final classification from feature subset 1. The selection process of feature subset 2 is as follows:

[0034] (1) Calculate the Gini index of each feature and select the top 10 features with the smallest Gini index. The Gini index calculation method is as follows;

[0035]

[0036] Among them, n represents the number of feature categories, different features have different number of categories, p i Indicates the probability of belonging to the i-th category in the sample, Gini index Gini (D i ) reflects the uncertainty of set D, Gini(D,A) reflects the uncertainty of set D after being divided by feature A, and the smaller the Gini index, the better;

[0037] (2) Feature selection based on encapsulation method;

[0038] The forward search strategy is used to traverse the 10 features sorted from small to large in step (1), and add them to the feature subset one by one. In the initial state, the feature subset is empty; the fault detection accuracy is used as the evaluation index, and the decision tree is used as the battery fault classifier; one feature is selected to be added to the feature subset, and the feature subset is used for classification. If the current classification accuracy is higher than the accuracy without the feature, the feature is retained as a new feature subset. Otherwise, the feature is deleted from the current feature subset; when all 10 features are traversed, the final feature subset for classification is obtained, and the feature selection process ends.

[0039] Furthermore, in step 4, based on the idea of ​​decision-layer fusion, the classification results of different classifiers are integrated to obtain the final battery cluster status classification, and three classification algorithms, random forest, KNN model, and logistic regression, are selected as basic classifiers. Classification is performed according to the feature subset 2 calculated in step 3 to obtain the classification results. The number of votes for the fault state and the normal state in the three classification algorithms is counted, and the state with the highest number of votes is used as the current state of the detected battery cluster.

[0040] Furthermore, when the cumulative number of classified data samples is a multiple of 50, the newly stored samples are mixed with the existing samples, the fault classification model is retrained, and a new feature subset 2 is generated for subsequent state classification of the battery cluster.

[0041] Furthermore, fusion methods are used at the feature layer and decision layer respectively. At the feature layer, features reflecting voltage changes from different angles are fused together as basic features; at the decision layer, the classification results of three classification algorithms, random forest, KNN, and logistic regression, are fused to complete the state classification of the battery cluster.

[0042] Furthermore, the battery cluster is classified into normal or faulty states based on the voltage, current and temperature in the relaxation phase after the battery cluster of the energy storage power station is charged and discharged.

[0043] Due to the adoption of the above technical solution, the present invention has the following advantages and positive effects compared with the prior art:

[0044] The present invention discloses a fault classification method for energy storage power stations based on local and global features in the relaxation stage, which collects the terminal voltage, current and temperature of the battery cluster in the relaxation stage after the battery cluster is charged / discharged. By setting a sliding window and a moving step, four health parameters reflecting the health status of the battery are extracted from the local terminal voltage data as local features, reflecting whether the terminal voltage of the battery cluster at the current moment is changing normally in the local range. The statistical features reflecting the terminal voltage change trend are extracted from the terminal voltage data collected from the current moment to the beginning of the relaxation stage as global features, reflecting whether the terminal voltage of the battery cluster is fluctuating normally in the global range. A feature selection method combining filtering and encapsulation is used on a feature set consisting of global features, local features, terminal voltage, current and temperature, so as to reduce feature redundancy and improve the accuracy of fault detection. The voting strategy is used on the three classification algorithms to obtain the final detection result, and the advantages of the three algorithms are combined to reduce the false detection rate of battery cluster fault detection. When the number of samples collected under actual working conditions reaches a specified value, the feature subset and classification algorithm parameters are re-updated to improve the robustness of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments. 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 work. In the drawings:

[0046] Figure 1 It is a schematic diagram of a framework of a fault classification method for energy storage power stations based on local and global characteristics of the relaxation stage of the present invention;

[0047] Figure 2 It is a flow chart of health parameter feature extraction based on terminal voltage and Nelder-Mead simplex search method in the present invention. DETAILED DESCRIPTION

[0048] The technical solution of the present invention will be described clearly and completely below in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0049] like Figure 1-Figure 2 As shown, this embodiment discloses a method for classifying energy storage power station faults based on local and global characteristics in the relaxation phase, comprising the following steps:

[0050] Step 1: Collect the terminal voltage, current and temperature data in the relaxation stage after the end of charging and discharging in actual working conditions;

[0051] Specifically, in step 1, the current, voltage, and temperature of the battery cluster in the energy storage system during the relaxation stage collected in real time by the BMS are obtained, and the corresponding battery status is recorded as a training sample, where the battery status has two types: normal and faulty. By simulating different working conditions in the laboratory, the current, voltage, and temperature of the battery cluster in different states are collected as an expansion of the training sample. For the online evaluation link, the voltage, current, and temperature data of the battery cluster during the relaxation stage collected in real time are obtained from the BMS.

[0052] In this embodiment, in order to make the fault detection model have high accuracy and strong robustness, it is necessary to train the model before actual use. In the present invention, the samples used for model training come from different working conditions of the actual working environment and laboratory simulation, and the terminal voltage, current, and temperature of the battery cluster in the relaxation stage under the two states are collected. By collecting data of the battery cluster in a laboratory environment, the problem of lack of actual working condition data can be compensated.

[0053] Battery relaxation refers to the sudden drop or rise in voltage at the end of charge / discharge until the voltage approaches stability, which can also be considered as the voltage reaching the open circuit voltage state. The relaxation stage of the battery also exists at the beginning of charge and discharge, but the relaxation phenomenon in this stage is short and difficult to capture. Therefore, only the relaxation stage at the end of charge and discharge is considered in the present invention. The voltage, current and temperature time series of the battery cluster relaxation time period in actual working conditions are collected, and the corresponding battery cluster status label is recorded. The laboratory mainly collects data under the fault state of the battery cluster, mainly overcharge fault and over-discharge fault. The steps for collecting battery cluster data are described as follows.

[0054] In the over-discharge experiment, the battery is charged using the constant current and constant voltage method or the constant current method, and a fixed rate discharge is adopted. The discharge cut-off voltage is not controlled, and the discharge time is controlled to achieve over-discharge of battery clusters with different discharge depths. The terminal voltage, current, and temperature of the battery cluster are recorded from the end of discharge to the time when the battery cluster voltage returns to stability.

[0055] In the overcharge experiment, the battery is fully charged using the constant current and constant voltage method, and then the battery is discharged at a constant current with a fixed rate current, and the SOC of the overcharge amount and the charge cut-off voltage are set. The terminal voltage, current, and temperature of the battery cluster are recorded from the time when the terminal voltage of the battery cluster reaches the cut-off voltage to the time when the terminal voltage becomes stable.

[0056] The above experiments can collect data under different ambient temperatures, and the ambient temperature can be set to 25℃, 35℃, 45℃, etc.

[0057] Step 2: Extract local features and global features for fault detection based on the collected terminal voltage, current, and temperature features;

[0058] The present invention uses the Nelder-Mead simplex method to search for k, σ, b, U of the relaxation voltage curve. oc The four parameters in which the voltage curve of the battery cluster during the relaxation period is established are described as follows:

[0059] The voltage drop and rise at the end of charge and discharge are caused by the dynamic changes inside the battery. The dynamic voltage can be obtained by subtracting the open circuit voltage from the terminal voltage of the battery cluster. In the relaxation stage, the dynamic voltage U at time t is d,t as follows:

[0060] U d,t =U L,t -U oc

[0061] Among them, U L,t represents the terminal voltage of the battery cluster at time t, U oc Indicates the open circuit voltage of the battery cluster;

[0062] In the present invention, the time constant τ is used to represent the trend of the speed of dynamic voltage change over time. The calculation method of the time constant τ at time t is as follows:

[0063]

[0064] The change of the time constant is represented by a power function, so as to establish a voltage curve model for the relaxation time period. The corresponding method is as follows:

[0065]

[0066] Among them, k, σ, and b are three parameters that characterize the τ-t curve, also known as the health characteristics of the battery. From the above formula, it can be concluded that the terminal voltage of the battery cluster at time t in the relaxation stage can be obtained by k, σ, b, and U oc The four parameter values ​​are estimated, and the optimal k, σ, b, and U are searched by combining the collected terminal voltage in the relaxation stage and the Nelder-Mead simplex method. oc Parameters, and as part of the features for battery fault detection, the health feature extraction process is as follows Figure 2 As shown, the specific steps include the following:

[0067] (1) Initialize k, σ, b, and U based on historical experience oc Four parameters;

[0068] (2) The terminal voltage data length for calculating the four health parameters is set to T w , which is equivalent to the time window W having a length of T w , set the moving step length to 1, when the length of the terminal voltage data collected from the beginning of the relaxation phase to the current moment is less than T wWhen the window is not moved, the window will not move when the length is greater than T w When , move the window;

[0069] (3) k, σ, b, U oc Substituting the four parameters into the above formula, we get [TT w ,T] terminal voltage prediction curve within the time period, T is the current time;

[0070] (3) Calculation of [TT w ,T] The error between the predicted curve and the actual curve of the terminal voltage in the time period. If the error does not meet the set threshold, the above four parameters are updated by the Nelder-Mead simplex method and the above process is repeated until the error reaches the set threshold;

[0071] (4) Output optimized k, σ, b, U oc Four parameters, as health characteristics at the current time T.

[0072] At this point, the extraction of health features of the battery cluster based on the model and terminal voltage is completed. The above four attributes indirectly characterize the stage characteristics of the terminal voltage in the relaxation state.

[0073] Use mathematical statistics to extract the terminal voltage change trend from the beginning of the relaxation state to the current moment, calculate the six features of variance, skewness, maximum value, minimum value, mean value, and kurtosis, characterize the change trend of the terminal voltage of the battery cluster in the entire relaxation state, and increase the accuracy of battery cluster fault detection. The calculation method of statistical features is shown in the following formula, where T0 represents the moment when the relaxation stage of the battery cluster begins, and T represents the current moment (T>T0):

[0074] (1) Variance s 2

[0075]

[0076] Among them, u t represents the terminal voltage of the battery cluster at time t, represents the average value of the battery cluster terminal voltage in the time period [T0, T];

[0077] (2) Skewness S k

[0078]

[0079] Wherein, μ3 represents the third-order central moment of the battery cluster terminal voltage in the [T0, T] period, and σ represents the standard deviation;

[0080] (3) Maximum value max

[0081]

[0082] (4)Minimum value

[0083]

[0084] (5)Averg

[0085]

[0086] (6) Kurtosis

[0087]

[0088] Where n is the sample size, u t represents the terminal voltage of the battery cluster at time t, It represents the average value of the battery cluster terminal voltage in the time period [T0, T].

[0089] The four health features and six statistical features calculated by the above method are used together with the terminal voltage, current, and temperature as the feature set for battery cluster fault detection. The total number of features in this feature set is 13, which detects whether the battery cluster is abnormal from multiple dimensions and improves the accuracy and stability of detection.

[0090] Step 3: Use the filtering and packaging methods in feature selection to select the feature subset that is ultimately used for fault detection;

[0091] In order to reduce the impact of feature redundancy on the accuracy of battery cluster fault detection and find out some relatively important features for battery cluster fault detection, the present invention combines the filtering method and the encapsulation method to select the feature subset ultimately used for fault detection. Specifically, in step 3, the Gini index is first used to select feature subset 1 from the basic feature set, and then the decision tree algorithm is used to select feature subset 2 ultimately used for classification from feature subset 1. The selection process of feature subset 2 is as follows:

[0092] (1) Calculate the Gini index of each feature and select the top 10 features with the smallest Gini index. The Gini index calculation method is as follows;

[0093]

[0094] Among them, n represents the number of feature categories, different features have different number of categories, p i Indicates the probability of belonging to the i-th category in the sample, Gini index Gini (D i ) reflects the uncertainty of set D, Gini(D,A) reflects the uncertainty of set D after being divided by feature A, and the smaller the Gini index, the better;

[0095] (2) Feature selection based on encapsulation method;

[0096] The forward search strategy is used to traverse the 10 features sorted from small to large in step (1), and add them to the feature subset one by one. In the initial state, the feature subset is empty; the fault detection accuracy is used as the evaluation index, and the decision tree is used as the battery fault classifier; one feature is selected to be added to the feature subset, and the feature subset is used for classification. If the current classification accuracy is higher than the accuracy without the feature, the feature is retained as a new feature subset. Otherwise, the feature is deleted from the current feature subset; when all 10 features are traversed, the final feature subset for classification is obtained, and the feature selection process ends.

[0097] Step 4: Battery cluster fault identification is performed based on the classification algorithms random forest, KNN, logistic regression and integrated voting strategy.

[0098] The current state of the battery cluster is classified according to the feature subset obtained in step 3. In the present invention, the state of the battery cluster is divided into two categories, namely normal and faulty. In order to improve the accuracy of classification and integrate the advantages of different classification algorithms, the present invention, based on the idea of ​​ensemble learning, votes on the results of the three classification algorithms as the final classification result. Random forest, K-nearest neighbors Algorithms (KNN), and logistic regression are used as the three basic classification algorithms.

[0099] First, the existing labeled data samples are classified into training set, validation set, and test set. The feature subsets used for online detection are calculated on the training set and validation set, where the training set is used to select the top 10 features, and feature subset 2 is obtained based on the classification accuracy of the validation set on the decision tree. The test set sample data is extracted according to feature subset 2, and then the data is input into the three classifiers respectively to obtain their respective classification results, and the parameters of the classification algorithm are adjusted according to the classification accuracy.

[0100] In online detection, based on the classification results of the three classifiers, the number of votes for the fault state and the normal state is calculated, and the state with the highest number of votes is selected as the state of the battery cluster being detected at the current moment. The feature data and labels of the detection samples are stored by category. When the cumulative number of stored samples reaches a multiple of 50, the newly stored samples are mixed with the existing samples, and the fault classification model is retrained to generate a new feature subset 2 for subsequent battery cluster state classification.

[0101] The above is only a preferred specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed by the present invention should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims

1. A method for classifying energy storage power station faults based on local and global characteristics of the relaxation phase, characterized in that: The following steps are involved: Step 1: Collect the terminal voltage, current and temperature data in the relaxation stage after the end of charging and discharging in actual working conditions; Step 2: Extract local features and global features for fault detection based on the collected terminal voltage, current, and temperature features; Step 3: Use the filtering and packaging methods in feature selection to select the feature subset that is ultimately used for fault detection; Step 4: Battery cluster fault identification is performed based on the classification algorithms random forest, KNN, logistic regression and integrated voting strategy.

2. A method for classifying energy storage power station faults based on local and global characteristics of relaxation phase according to claim 1, characterized in that: In step 1, the current, voltage, and temperature of the battery cluster in the energy storage system during the relaxation stage are obtained in real time from the BMS, and the corresponding battery status is recorded as a training sample, where the battery status is normal and faulty. By simulating different working conditions in the laboratory, the current, voltage, and temperature of the battery cluster in different states are collected as an expansion of the training sample. For the online evaluation link, the voltage, current, and temperature data of the battery cluster during the relaxation stage collected in real time are obtained from the BMS.

3. The energy storage power station fault classification method based on local and global characteristics of the relaxation stage according to claim 1 is characterized in that: In step 2, the collected terminal voltage data of the battery cluster is used to find the optimal k, σ, b, and U that reflect the battery health status based on the Nelder-Mead simplex search method. oc Four parameters, also called local characteristics reflecting the local variation trend of terminal voltage; The six characteristics of the terminal voltage from the current moment to the beginning of the relaxation phase are calculated using statistical methods, namely, the variance, maximum value, minimum value, average value, skewness, and kurtosis, which reflect the global change trend of the battery terminal voltage, also known as the global characteristics. The local characteristics, global characteristics, terminal voltage, current, and temperature together constitute the feature set for fault detection. The mathematical relationship between the four parameters and the terminal voltage is as follows: Among them, k, σ, and b are three parameters that characterize the terminal voltage change curve, also known as the health characteristics of the battery, U L,t represents the terminal voltage of the battery cluster at time t, U oc Indicates the open circuit voltage of the battery cluster; The statistical features representing global changes are calculated as follows: 1) Variance s 2 ; Among them, u t represents the terminal voltage of the battery cluster at time t, represents the average value of the battery cluster terminal voltage in the [T0, T] period; 2) Skewness S k ; Wherein, μ3 represents the third-order central moment of the battery cluster terminal voltage in the [T0, T] period, and σ represents the standard deviation; 3) Maximum value max; 4) minimum value min; 5) average; 6) Kurtosis Where n is the sample size, u t represents the terminal voltage of the battery cluster at time t, It represents the average value of the battery cluster terminal voltage in the time period [T0, T].

4. A method for classifying energy storage power station faults based on local and global characteristics of relaxation phase according to claim 1, characterized in that: In step 3, the Gini index is first used to select feature subset 1 from the basic feature set, and then the decision tree algorithm is used to select feature subset 2 for classification from feature subset 1. The selection process of feature subset 2 is as follows: (1) Calculate the Gini index of each feature and select the top 10 features with the smallest Gini index. The Gini index calculation method is as follows; Among them, n represents the number of feature categories, different features have different number of categories, p i Indicates the probability of belonging to the i-th category in the sample, Gini index Gini (D i ) reflects the uncertainty of set D, Gini(D,A) reflects the uncertainty of set D after being divided by feature A, and the smaller the Gini index, the better; (2) Feature selection based on encapsulation method; The forward search strategy is used to traverse the 10 features sorted from small to large in step (1), and add them to the feature subset one by one. In the initial state, the feature subset is empty; the fault detection accuracy is used as the evaluation index, and the decision tree is used as the battery fault classifier; one feature is selected to be added to the feature subset, and the feature subset is used for classification. If the current classification accuracy is higher than the accuracy without the feature, the feature is retained as a new feature subset. Otherwise, the feature is deleted from the current feature subset; when all 10 features are traversed, the final feature subset for classification is obtained, and the feature selection process ends.

5. A method for classifying energy storage power station faults based on local and global characteristics of relaxation phase according to claim 4, characterized in that: In step 4, based on the idea of ​​decision-making layer fusion, the classification results of different classifiers are integrated to obtain the final battery cluster status classification. Random forest, KNN model, and logistic regression are selected as basic classifiers. The classification is performed according to the feature subset 2 calculated in step 3 to obtain the classification results. The number of votes for the fault state and the normal state in the three classification algorithms is counted, and the state with the highest number of votes is taken as the current state of the detected battery cluster.

6. A method for classifying energy storage power station faults based on local and global characteristics of relaxation phase according to claim 5, characterized in that: When the accumulated number of classified data samples is a multiple of 50, the newly stored samples are mixed with the existing samples, the fault classification model is retrained, and a new feature subset 2 is generated for subsequent battery cluster status classification.

7. A method for classifying energy storage power station faults based on local and global characteristics of relaxation phase according to claim 6, characterized in that: Fusion methods are used at the feature layer and decision layer respectively. At the feature layer, features reflecting voltage changes from different angles are fused together as basic features. At the decision layer, the classification results of the three classification algorithms, random forest, KNN, and logistic regression, are integrated to complete the status classification of the battery cluster.

8. A method for classifying energy storage power station faults based on local and global characteristics of relaxation phase according to claim 7, characterized in that: The battery cluster is classified into normal or faulty states based on the voltage, current and temperature during the relaxation phase after the battery cluster of the energy storage power station is charged and discharged.

Images