Energy storage fault prediction method and device

Abstract

The present application provides a method and device for predicting energy storage faults, which includes: dividing sample data into positive sample data and negative sample data; classifying the positive sample data, using OVO-SVM classification decision to combine any two categories in pairs, and then inputting the positive sample data to train the SVM model to generate a fault probability classifier; using database language to obtain the type of fault from the negative sample data, filtering the data based on the type of fault, and then using a random forest algorithm to train the negative sample data to generate a fault type classifier; inputting a preset number of data in the sample data into the fault probability classifier and the fault type classifier, respectively, and outputting the predicted fault type and predicted fault probability. The present application improves the accuracy of predicting the failure rate of system operation in the future period and the prediction of the type of fault of system operation in the future period and its corresponding probability of occurrence through bidirectional modeling of positive and negative samples.

Description

Technical Field

[0001] This application seeks to protect a data processing technology, and more particularly relates to a method for predicting energy storage faults. This application also relates to an energy storage fault prediction device. Background Art

[0002] Energy storage technology is an important aspect of grid intelligence. Its safe and reliable operation is related to the peak-shaving and frequency-regulating capabilities of the grid, and is also a prerequisite for the safety of battery energy storage equipment itself.

[0003] In the prior art, safety monitoring of energy storage power stations is usually achieved in the following ways:

[0004] 1. Install multiple fire or smoke sensors inside the energy storage power station to monitor the operation of the energy storage power station;

[0005] 2. Abnormal monitoring of the energy storage system is carried out through the battery management system (BMS) and the bidirectional energy storage converter (PCS).

[0006] However, the safety monitoring system of the energy storage power station has a single function, with only fault alarm and monitoring functions. It is difficult to capture abnormal changes in detection variables in the future, cannot predict the occurrence of faults in advance, and cannot guarantee that faults are discovered in the first time, which means that the energy storage power station still has great safety risks during operation.

[0007] Therefore, in response to the above problems, it is urgent to propose a method for energy storage fault prediction to estimate the operating status of the energy storage power station and nip the safety hazards of the energy storage system in the bud. Summary of the Invention

[0008] In order to solve one or more problems mentioned in the above background technology, the present application proposes a method for predicting energy storage faults. The present application also relates to an energy storage fault prediction device.

[0009] This application provides a method for predicting energy storage faults, including:

[0010] Acquire sample data, and divide the sample data into positive sample data and negative sample data;

[0011] Classify the positive sample data, use OVO-SVM classification decision to combine any two categories, then input the positive sample data to train the SVM model to generate a fault probability classifier;

[0012] Using a database language to obtain the fault type from the negative sample data, filtering the data based on the fault type, and using a random forest algorithm to train the negative sample data to generate a fault type classifier;

[0013] A preset number of data in the sample data is input into the fault probability classifier and the fault type classifier respectively, and the predicted fault type and predicted fault probability are output.

[0014] Optionally, the OVO-SVM classification decision is used to combine any two categories in pairs, and the classification functions of the two categories are as follows:

[0015]

[0016] Wherein, the i represents the first category of the two categories, the j represents the second category of the two categories, the W is the normal vector of the classification hyperplane, and the is the transpose of W, and b is a constant;

[0017] In the above expression, when ≥0, then i gets one vote, otherwise j gets one vote, and then the feature selection sequence table is output.

[0018] Optionally, after the fault type classifier training falls into local convergence, it is optimized by a bionic intelligent algorithm, and the bionic intelligent algorithm includes: particle swarm algorithm, artificial bee colony algorithm, artificial ant colony algorithm, differential search algorithm, cuckoo algorithm, bat algorithm, wolf pack algorithm, whale pack algorithm, salp swarm algorithm or an improved algorithm based on the above algorithms.

[0019] Optionally, the positive sample number has multiple feature sets, each feature set including: total battery voltage, total battery current, SOC, pressure difference, temperature difference, three-phase voltage of PCS, three-phase current, DC voltage, DC current, DC power, grid frequency and load frequency.

[0020] Optionally, the step of outputting the predicted fault type and predicted fault probability includes:

[0021] Calculate the failure rate after counting the number of positive sample features;

[0022] After counting the number of negative sample feature types, the fault probability of each type is calculated, and the fault type is obtained based on the label of each type.

[0023] The present application provides an energy storage fault prediction device, comprising:

[0024] A segmentation module, configured to obtain sample data and divide the sample data into positive sample data and negative sample data;

[0025] The first training module classifies the positive sample data, uses OVO-SVM classification decision to combine any two categories, and then inputs the positive sample data to train the SVM model to generate a fault probability classifier;

[0026] A second training module is configured to obtain fault types from the negative sample data using a database language, filter the data based on the fault types, and then train the negative sample data using a random forest algorithm to generate a fault type classifier;

[0027] The execution module is used to input a preset number of data in the sample data into the fault probability classifier and the fault type classifier respectively, and output the predicted fault type and the predicted fault probability.

[0028] Optionally, the OVO-SVM classification decision is used to combine any two categories in pairs, and the classification functions of the two categories are as follows:

[0029]

[0030] Wherein, the i represents the first category of the two categories, the j represents the second category of the two categories, the W is the normal vector of the classification hyperplane, and the is the transpose of W, and b is a constant;

[0031] In the above expression, when ≥0, then i gets one vote, otherwise j gets one vote, and then the feature selection sequence table is output.

[0032] Optionally, after the fault type classifier training falls into local convergence, it is optimized by a bionic intelligent algorithm, and the bionic intelligent algorithm includes: particle swarm algorithm, artificial bee colony algorithm, artificial ant colony algorithm, differential search algorithm, cuckoo algorithm, bat algorithm, wolf pack algorithm, whale pack algorithm, salp swarm algorithm or an improved algorithm based on the above algorithms.

[0033] Optionally, the positive sample number has multiple feature sets, each feature set including: total battery voltage, total battery current, SOC, pressure difference, temperature difference, three-phase voltage of PCS, three-phase current, DC voltage, DC current, DC power, grid frequency and load frequency.

[0034] Optionally, the execution module further includes:

[0035] The first statistical unit calculates the failure rate after counting the number of positive sample features;

[0036] The second statistical unit counts the number of negative sample feature types and calculates the fault probability of each type, and obtains the fault type based on the label of each type.

[0037] The advantages of this application over the prior art are:

[0038] The present application provides a method for predicting energy storage faults, including: obtaining sample data, dividing the sample data into positive sample data and negative sample data; classifying the positive sample data, using OVO-SVM classification decision to combine any two categories in pairs, and then inputting the positive sample data to train the SVM model to generate a fault probability classifier; using database language to obtain the type of fault from the negative sample data, filtering the data based on the type of fault, and then using a random forest algorithm to train the negative sample data to generate a fault type classifier; inputting a preset number of data in the sample data into the fault probability classifier and the fault type classifier respectively, and outputting the predicted fault type and the predicted fault probability. The present application uses bidirectional modeling of positive and negative samples to improve the positive classifier for predicting the failure rate of system operation in the future, and the negative classifier for predicting the type of fault of system operation in the future and its corresponding probability of occurrence. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 This is the energy storage fault prediction flow chart in this application.

[0040] Figure 2 This is a diagram of the fault prediction system in this application.

[0041] Figure 3 This is the negative sample training flow chart in this application.

[0042] Figure 4 This is a schematic diagram of the energy storage fault prediction device in this application. DETAILED DESCRIPTION

[0043] The following contents are all examples of specific implementation processes provided for detailed description of the technical solutions to be protected by this application. However, this application can also be implemented in other ways different from the descriptions here. Those skilled in the art can adopt different technical means to implement this application under the guidance of the concept of this application. Therefore, this application is not limited to the specific embodiments below.

[0044] The present application provides a method for predicting energy storage faults, including: obtaining sample data, dividing the sample data into positive sample data and negative sample data; classifying the positive sample data, using OVO-SVM classification decision to combine any two categories in pairs, and then inputting the positive sample data to train the SVM model to generate a fault probability classifier; using database language to obtain the type of fault from the negative sample data, filtering the data based on the type of fault, and then using a random forest algorithm to train the negative sample data to generate a fault type classifier; inputting a preset number of data in the sample data into the fault probability classifier and the fault type classifier respectively, and outputting the predicted fault type and the predicted fault probability. The present application uses bidirectional modeling of positive and negative samples to improve the positive classifier for predicting the failure rate of system operation in the future, and the negative classifier for predicting the type of fault of system operation in the future and its corresponding probability of occurrence.

[0045] Figure 1 This is the energy storage fault prediction flow chart in this application.

[0046] Please refer to Figure 1 As shown, S101 obtains sample data and divides the sample data into positive sample data and negative sample data.

[0047] The sample data shown refers to the operating data of the energy storage system, including: total battery voltage, total battery current, SOC (system parameters), pressure difference, temperature difference of the BMS (management system); three-phase voltage, three-phase current, three-phase load rate, DC voltage, DC current, DC power, grid frequency and load frequency of the PCS (process control system).

[0048] The sample data is obtained, and the sample data is divided into positive sample data and negative sample data.

[0049] Figure 2 This is a diagram of the fault prediction system in this application.

[0050] Please refer to Figure 2 As shown, at the start of the energy storage system, sample data is collected and stored. During system initialization, the samples are pre-processed, including: creating a sample data table through the database, where the positive sample table is defined as Pos-sample, the positive sample library is defined as Positive, the negative sample table is defined as Neg-sample, the negative sample library is defined as Negative, and the test sample is defined as Sample.

[0051] When the energy storage system begins operation, it communicates with the PCS, BMS, MPPT (power tracking system), wind turbine controllers, and environmental control equipment to complete data collection and processing. The system monitors changes in the alarm flag in real time. When the alarm flag is 0, indicating no system alarms, samples are stored in the Pos-sample table at a rate of 10 minutes. When the alarm flag is 1, indicating an alarm, samples are stored in the Neg-sample table at a rate of 10 seconds. The second column of the table contains alarm details.

[0052] Count the time when the first alarm occurs for different alarm information in the Neg-sample table to form a time array. Use the time array as a filtering condition to filter the data in the Pos-sample table two hours before each alarm occurs.

[0053] If the interval between different alarms is greater than 6 hours, the data from the Pos-sample table for the next 2 hours will be deleted and inserted into the Neg-sample table. If the interval is less than 2 hours, the data from the Pos-sample table will be copied and inserted into the table before the alarm time in the Neg-sample table without deleting the data from the Pos-sample table.

[0054] Copy all the Neg-sample data to Negative, and copy all the Pos-sample data to Positive. Then clear the Neg-sample and Pos-sample data to return to the initialization state. The Positive library contains positive samples, and the Negative library contains negative samples.

[0055] Please refer to Figure 1 as well as Figure 2 As shown, S102 classifies the positive sample data, uses OVO-SVM classification decision to combine any two categories, and then inputs the positive sample data to train the SVM model to generate a fault probability classifier.

[0056] In the present application, the data in the positive sample library is extracted and classified, and preferably the classification categories are: discharge, charge and rest.

[0057] The feature vector set representation of the sample data is performed based on the classification , wherein is the feature vector set of positive samples, The code for the sample type.

[0058] The data in the positive sample feature set is used as sample data to train the SVM (vector machine) model, where each feature set contains 20 values ​​of data such as the total battery voltage, total battery current, SOC, pressure difference, temperature difference of the BMS, and the three-phase voltage, three-phase current, DC voltage, DC current, DC power, grid frequency, and load frequency of the PCS as alternative features.

[0059] In this application, during the SVM training process, OVO-SVM classification decision is used to describe the data association, including:

[0060] The number of combinations of each two categories, i.e. the number of combinations finally formed, is expressed as follows:

[0061]

[0062] Wherein, C represents the number of combinations, S represents the category, and C is also the number of classifiers used in the OVO-SVM classification decision.

[0063] The process of training the SVM model is as follows:

[0064] Step 1: Express the correlation function of the two classifiers as follows:

[0065]

[0066] Wherein, i and j represent two categories respectively, represents the transpose of w, wherein w represents the classification hyperplane normal vector, b is the threshold constant, and It is expressed as a mapping function from the original space to the feature space.

[0067] Step 2: In the above expression, when If ≥0, add one vote to category i, otherwise add one vote to category j. After traversing the C categories, output the feature selection order table according to the number of votes and the classification accuracy of each feature.

[0068] In the present application, the problems that may arise during the training process include local convergence. When local convergence occurs, preferably, a particle swarm algorithm, an artificial bee colony algorithm, an artificial ant colony algorithm, a differential search algorithm, a cuckoo algorithm, a bat algorithm, a wolf pack algorithm, a whale pack algorithm, a salp swarm algorithm or an improved optimization algorithm based on the above algorithms is used to optimize the association function.

[0069] Please refer to Figure 1 as well as Figure 2As shown, S103 uses a database language to obtain fault types from the negative sample data, filters the data based on the fault types, and uses a random forest algorithm to train the negative sample data to generate a fault type classifier.

[0070] In the present application, in processing the negative sample library, the database language is first used to perform statistics on the types of sending faults in the Negative table, and a fault label is generated according to the type.

[0071] Secondly, the data in the negative sample library is optimized according to the type or label, and redundant data is deleted. Finally, the model training is performed based on the optimized negative sample library.

[0072] Figure 3 This is the negative sample training flow chart in this application.

[0073] Please refer to Figure 3 As shown, the negative sample set (fault feature sample set) is composed of a training sample set, and the following steps are performed:

[0074] Step 1: The training negative sample set is expressed as: ,in is the sample type, is the negative sample feature set, is the i-th training sample, k is The number of eigenvalues ​​in is the label of the sample type.

[0075] Step 2: Bootstrap sampling is performed on the training negative sample feature set X. After sampling n times, the Bootstrap subsample is obtained. .

[0076] Step 3: Sampling CART to build a decision tree model , traverse each sub-sample once and obtain a classifier consisting of a set of decision trees .

[0077] Step 4: Input the test set data and use the classification decision Conduct training, including is the jth decision tree; represents the characteristic function, which is 1 when the set in the brackets satisfies the expression, otherwise it is 0; y represents the category label;

[0078] Step 5: Count the output with the highest votes.

[0079] Please refer to Figure 1 as well as Figure 2As shown, S104 inputs a preset number of data in the sample data into the fault probability classifier and the fault type classifier respectively, and outputs a predicted fault type and a predicted fault probability.

[0080] In the present application, preferably, the preset number is set to: extract nearly 1000 data items from the Sample table, and then randomly select 60% as a test set.

[0081] Specifically, the test set is input into the positive sample classifier, and the number of features belonging to the positive sample class is counted. ; Add up the counted numbers ;

[0082] Calculate the failure rate, failure rate = .

[0083] Among them, the Represents a positive sample, Represents the total number of positive sample feature categories.

[0084] Input the test set into the negative sample classifier and count the number of samples belonging to each feature classifier. ,m is the fault type;

[0085] Calculate the probability of each fault by classification: .

[0086] Among them, the Represents a negative sample, Represents the total number of negative samples.

[0087] Predict fault types, including: , output i, find the corresponding label output fault type according to the fault label table, and obtain the failure rate of the fault type.

[0088] The present application also provides an energy storage fault prediction device, including: a segmentation module 301 , a first training module 302 , a second training module 303 , and an execution module 304 .

[0089] Figure 4 This is a schematic diagram of the energy storage fault prediction device in this application.

[0090] Please refer to Figure 4 As shown, the segmentation module 401 is used to obtain sample data and divide the sample data into positive sample data and negative sample data.

[0091] The sample data shown refers to the operating data of the energy storage system, including: total battery voltage, total battery current, SOC (system parameters), pressure difference, temperature difference of the BMS (management system); three-phase voltage, three-phase current, three-phase load rate, DC voltage, DC current, DC power, grid frequency and load frequency of the PCS (process control system).

[0092] The sample data is obtained, and the sample data is divided into positive sample data and negative sample data.

[0093] Please refer to Figure 2 As shown, at the start of the energy storage system, sample data is collected and stored. During system initialization, the samples are pre-processed, including: creating a sample data table through the database, where the positive sample table is defined as Pos-sample, the positive sample library is defined as Positive, the negative sample table is defined as Neg-sample, the negative sample library is defined as Negative, and the test sample is defined as Sample.

[0094] When the energy storage system begins operation, it communicates with the PCS, BMS, MPPT (power tracking system), wind turbine controllers, and environmental control equipment to complete data collection and processing. The system monitors changes in the alarm flag in real time. When the alarm flag is 0, indicating no system alarms, samples are stored in the Pos-sample table at a rate of 10 minutes. When the alarm flag is 1, indicating an alarm, samples are stored in the Neg-sample table at a rate of 10 seconds. The second column of the table contains alarm details.

[0095] Count the time when the first alarm occurs for different alarm information in the Neg-sample table to form a time array. Use the time array as a filtering condition to filter the data in the Pos-sample table two hours before each alarm occurs.

[0096] If the interval between different alarms is greater than 6 hours, the data from the Pos-sample table for the next 2 hours will be deleted and inserted into the Neg-sample table. If the interval is less than 2 hours, the data from the Pos-sample table will be copied and inserted into the table before the alarm time in the Neg-sample table without deleting the data from the Pos-sample table.

[0097] Copy all the Neg-sample data to Negative, and copy all the Pos-sample data to Positive. Then clear the Neg-sample and Pos-sample data to return to the initialization state. The Positive library contains positive samples, and the Negative library contains negative samples.

[0098] Please refer to Figure 4 As shown, the first training module 402 classifies the positive sample data, uses OVO-SVM classification decision to combine any two categories, and then inputs the positive sample data to train the SVM model to generate a fault probability classifier.

[0099] In the present application, the data in the positive sample library is extracted and classified, and preferably the classification categories are: discharge, charge and rest.

[0100] The feature vector set representation of the sample data is performed based on the classification , wherein is the feature vector set of positive samples, The code for the sample type.

[0101] The data in the positive sample feature set is used as sample data to train the SVM (vector machine) model, where each feature set contains 20 values ​​of data such as the total battery voltage, total battery current, SOC, pressure difference, temperature difference of the BMS, and the three-phase voltage, three-phase current, DC voltage, DC current, DC power, grid frequency, and load frequency of the PCS as alternative features.

[0102] In this application, during the SVM training process, OVO-SVM classification decision is used to describe the data association, including:

[0103] The number of combinations of each two categories, i.e. the number of combinations finally formed, is expressed as follows:

[0104]

[0105] Wherein, C represents the number of combinations, S represents the category, and C is also the number of classifiers used in the OVO-SVM classification decision.

[0106] The process of training the SVM model is as follows:

[0107] Step 1: Express the correlation function of the two classifiers as follows:

[0108]

[0109] Wherein, i and j represent two categories respectively, represents the transpose of w, wherein w represents the classification hyperplane normal vector, b is the threshold constant, and It is expressed as a mapping function from the original space to the feature space.

[0110] Step 2: In the above expression, when If ≥0, add one vote to category i, otherwise add one vote to category j. After traversing the C categories, output the feature selection order table according to the number of votes and the classification accuracy of each feature.

[0111] In the present application, the problems that may arise during the training process include local convergence. When local convergence occurs, preferably, a particle swarm algorithm, an artificial bee colony algorithm, an artificial ant colony algorithm, a differential search algorithm, a cuckoo algorithm, a bat algorithm, a wolf pack algorithm, a whale pack algorithm, a salp swarm algorithm or an improved optimization algorithm based on the above algorithms is used to optimize the association function.

[0112] Please refer to Figure 4 As shown, the second training module 403 is used to obtain fault types by statistically analyzing the negative sample data using a database language, filter the data based on the fault types, and then use a random forest algorithm to train the negative sample data to generate a fault type classifier.

[0113] In the present application, in processing the negative sample library, the database language is first used to perform statistics on the types of sending faults in the Negative table, and a fault label is generated according to the type.

[0114] Secondly, the data in the negative sample library is optimized according to the type or label, and redundant data is deleted. Finally, the model training is performed based on the optimized negative sample library.

[0115] Figure 3 This is the negative sample training flow chart in this application.

[0116] Please refer to Figure 3 As shown, the negative sample set (fault feature sample set) is composed of a training sample set, and the following steps are performed:

[0117] Step 1: The training negative sample set is expressed as: ,in is the sample type, is the negative sample feature set, is the i-th training sample, k is The number of eigenvalues ​​in is the label of the sample type.

[0118] Step 2: Bootstrap sampling is performed on the training negative sample feature set X. After sampling n times, the Bootstrap subsample is obtained. .

[0119] Step 3: Sampling CART to build a decision tree model , traverse each sub-sample once and obtain a classifier consisting of a set of decision trees .

[0120] Step 4: Input the test set data and use the classification decision Conduct training, including is the jth decision tree; represents the characteristic function, which is 1 when the set in the brackets satisfies the expression, otherwise it is 0; y represents the category label;

[0121] Step 5: Count the output with the highest votes.

[0122] Please refer to Figure 4 As shown, the execution module 404 is used to input the preset quantity data in the sample data into the fault probability classifier and the fault type classifier respectively, and output the predicted fault type and the predicted fault probability.

[0123] In the present application, preferably, the preset number is set to: extract nearly 1000 data items from the Sample table, and then randomly select 60% as a test set.

[0124] Specifically, the test set is input into the positive sample classifier, and the number of features belonging to the positive sample class is counted. ; Add up the counted numbers ;

[0125] Calculate the failure rate, failure rate = .

[0126] Among them, the Represents a positive sample, Represents the total number of positive sample feature categories.

[0127] Input the test set into the negative sample classifier and count the number of samples belonging to each feature classifier. ,m is the fault type;

[0128] Calculate the probability of each fault by classification: .

[0129] Among them, the Represents a negative sample, Represents the total number of negative samples.

[0130] Predict fault types, including: , output i, find the corresponding label output fault type according to the fault label table, and obtain the failure rate of the fault type.

Claims

1. A method for predicting energy storage failure, characterized in that: include: Acquire sample data, and divide the sample data into positive sample data and negative sample data; The positive sample data is classified, and any two categories are combined in pairs using OVO-SVM classification decision. Then, the positive sample data is input to train the SVM model to generate a fault probability classifier. The classification functions of the two categories are as follows: Wherein, i represents the first category of the two categories, j represents the second category of the two categories, W is the normal vector of the classification hyperplane, and w T is the transpose of W, and b is a constant; In the expression, when f(x)≥0, i gets one vote, otherwise j gets one vote, and then the feature selection order table is output; Using a database language to obtain the fault type from the negative sample data, filtering the data based on the fault type, and using a random forest algorithm to train the negative sample data to generate a fault type classifier; Inputting a preset number of data in the sample data into the fault probability classifier and the fault type classifier respectively, and outputting a predicted fault type and a predicted fault probability; The sample data refers to the operating data of the energy storage system, including: total battery voltage, total battery current, SOC (system parameters), pressure difference, temperature difference of BMS (management system), three-phase voltage, three-phase current, three-phase load rate, DC voltage, DC current, DC power, grid frequency and load frequency of PCS (process control).

2. The energy storage fault prediction method according to claim 1, characterized in that: After the fault type classifier training falls into local convergence, it is optimized by a bionic intelligent algorithm, and the bionic intelligent algorithm includes: particle swarm algorithm, artificial bee colony algorithm, artificial ant colony algorithm, differential search algorithm, cuckoo algorithm, bat algorithm, wolf group algorithm, whale group algorithm, salp group algorithm or an improved algorithm based on the above algorithms.

3. The energy storage fault prediction method according to claim 1, characterized in that: The positive sample number has multiple feature sets, each feature set including: total battery voltage, total battery current, SOC, pressure difference, temperature difference, three-phase voltage of PCS, three-phase current, DC voltage, DC current, DC power, grid frequency and load frequency.

4. The energy storage fault prediction method according to claim 1, characterized in that: The step of outputting the predicted fault type and predicted fault probability comprises: Calculate the failure rate after counting the number of positive sample features; After counting the number of negative sample feature types, the fault probability of each type is calculated, and the fault type is obtained based on the label of each type.

5. A device for predicting energy storage failure, characterized in that: include: A segmentation module, configured to obtain sample data and divide the sample data into positive sample data and negative sample data; The first training module classifies the positive sample data, uses OVO-SVM classification decision to combine any two categories, and then inputs the positive sample data to train the SVM model to generate a fault probability classifier; the classification functions of the two categories are as follows: Wherein, i represents the first category of the two categories, j represents the second category of the two categories, W is the normal vector of the classification hyperplane, and w T is the transpose of W, and b is a constant; In the expression, when f(x)≥0, i gets one vote, otherwise j gets one vote, and then the feature selection order table is output; A second training module is configured to obtain fault types from the negative sample data using a database language, filter the data based on the fault types, and then train the negative sample data using a random forest algorithm to generate a fault type classifier; The execution module is used to input a preset number of data in the sample data into the fault probability classifier and the fault type classifier respectively, and output the predicted fault type and the predicted fault probability.

6. The energy storage fault prediction device according to claim 5, characterized in that: After the fault type classifier training falls into local convergence, it is optimized by a bionic intelligent algorithm, and the bionic intelligent algorithm includes: particle swarm algorithm, artificial bee colony algorithm, artificial ant colony algorithm, differential search algorithm, cuckoo algorithm, bat algorithm, wolf group algorithm, whale group algorithm, salp group algorithm or an improved algorithm based on the above algorithms.

7. The energy storage fault prediction device according to claim 5, characterized in that: The positive sample number has multiple feature sets, each feature set including: total battery voltage, total battery current, SOC, pressure difference, temperature difference, three-phase voltage of PCS, three-phase current, DC voltage, DC current, DC power, grid frequency and load frequency.

8. The energy storage fault prediction device according to claim 5, characterized in that: The execution module also includes: The first statistical unit calculates the failure rate after counting the number of positive sample features; The second statistical unit counts the number of negative sample feature types and calculates the fault probability of each type, and obtains the fault type based on the label of each type.

Images