A method and device for training a classifier model for energy storage system performance evaluation

Abstract

The present invention discloses a method and device for training a classifier model for energy storage system performance evaluation, comprising obtaining a plurality of energy storage system attribute data to be detected, and based on an improved artificial minority class oversampling method, performing data expansion on the attribute data of each energy storage system to be detected, and determining a plurality of target interpolation samples to be detected; inputting each target interpolation sample to be detected into a preset target feature extraction model for feature extraction, and outputting a plurality of feature extraction data to be detected; using a preset target feature dimension reduction model to perform feature dimension reduction on each feature extraction data to be detected, and generating a plurality of feature dimension reduction data to be trained; based on a preset distribution robust optimization strategy, using each feature dimension reduction data to be trained to train an initial energy storage system performance evaluation classifier model, and determining a target energy storage system performance evaluation classifier model; the technical problem that the existing energy storage system performance evaluation technology causes the model to have poor classification accuracy is solved.

Description

Technical Field

[0001] The present invention relates to the technical field of energy storage power grid, and in particular to a method and device for training a classifier model for energy storage system performance evaluation. Background Art

[0002] With the transformation of the global energy structure and the rapid development of new energy technologies, energy storage systems in new energy power facilities play an increasingly important role in energy supply. Energy storage systems, such as battery storage, are one of the key technologies to achieve efficient energy utilization and ensure stable operation of the power grid.

[0003] Energy storage systems are considered an important component of rapid frequency control as a resource with fast regulation speed, flexible power and capacity configuration and wide application range. In the face of the adverse effects of the strong randomness, volatility and intermittency inherent in renewable energy generation on the safe and stable operation of the power system, its performance evaluation has become a necessary step to ensure the efficiency and safety of energy storage systems.

[0004] Existing energy storage system performance evaluation technologies lack the adaptability to nonlinear and complex data patterns in classifier model design, and are not adaptable enough to data deviations and outliers that may be encountered in actual operations, resulting in poor classification accuracy of the model. Summary of the invention

[0005] The present invention provides a method and device for training a classifier model for energy storage system performance evaluation, which are used to solve the technical problem that the classification accuracy of the model is poor due to the existing energy storage system performance evaluation technology.

[0006] A first aspect of the present invention provides a method for training a classifier model for energy storage system performance evaluation, comprising:

[0007] Acquire multiple property data of energy storage systems to be detected, and based on an improved artificial minority class oversampling method, perform data expansion on the property data of each energy storage system to be detected to determine multiple interpolation samples of targets to be detected;

[0008] Inputting each of the interpolated samples of the target to be detected into a preset target feature extraction model for feature extraction, and outputting a plurality of feature extraction data to be detected;

[0009] Using a preset target feature dimension reduction model to perform feature dimension reduction on each of the feature extraction data to be detected, to generate a plurality of feature dimension reduction data to be trained;

[0010] Based on the preset distributed robust optimization strategy, the initial energy storage system performance evaluation classifier model is trained using the feature dimensionality reduction data to be trained to determine the target energy storage system performance evaluation classifier model.

[0011] Optionally, the step of performing data expansion on the attribute data of each energy storage system to be detected based on the improved artificial minority class oversampling method to determine a plurality of interpolation samples of the target to be detected includes:

[0012] Perform characteristic calculations respectively according to the attribute data of each energy storage system to be detected, and output characteristic parameters corresponding to the attribute data of each energy storage system to be detected;

[0013] Using each of the characteristic parameters to perform quantum mapping on the corresponding property data of the energy storage system to be detected, and generating initial quantum state data corresponding to each of the property data of the energy storage system to be detected;

[0014] Performing quantum projection operations on each of the initial quantum state data respectively to generate target quantum state data corresponding to each of the energy storage system attribute data to be detected;

[0015] Performing quantum state measurement on each of the target quantum state data, outputting the probability density distribution value corresponding to each of the energy storage system attribute data to be detected, and comparing each of the probability density distribution values ​​with a preset probability threshold value;

[0016] Taking any to-be-detected energy storage system attribute data corresponding to a probability density distribution value greater than the preset probability threshold as target detection energy storage system attribute data;

[0017] Adaptively performing multiple interpolation on the neighborhood attribute data in the neighborhood of each target detection energy storage system attribute data to generate multiple initial interpolation samples to be detected at the current moment;

[0018] Determine whether the sum of the number of initial interpolation samples to be detected at the current moment and the number of initial interpolation samples to be detected at the historical moments reaches a preset sample number threshold;

[0019] If it is reached, the initial interpolation samples to be detected at the current moment and the initial interpolation samples to be detected at the historical moment are respectively used as the corresponding target interpolation samples to be detected.

[0020] Optionally, it also includes:

[0021] If the sum of the sample quantities of the initial interpolation samples to be detected at the current moment and the initial interpolation samples to be detected at the historical moments does not reach the preset sample quantity threshold, a preset prediction model is used to perform quality assessment on each of the initial interpolation samples to be detected at the current moment, and a sample quality assessment result is output;

[0022] Substituting the sample quality assessment result into a preset loss function and taking a derivative to determine a first model gradient;

[0023] Using the first model gradient to update each of the characteristic parameters, and determining a new characteristic parameter corresponding to each of the energy storage system attribute data to be detected;

[0024] The initial interpolation samples to be detected at the current moment are used as the initial interpolation samples to be detected at the historical moment;

[0025] Based on the new characteristic parameters and the corresponding property data of the energy storage system to be detected, a plurality of initial interpolation samples to be detected at the current moment are determined, until the sum of the number of initial interpolation samples to be detected at the current moment and the number of initial interpolation samples to be detected at the historical moments reaches the preset sample number threshold;

[0026] The initial interpolation samples to be detected at the historical moment, the initial interpolation samples to be detected at the current moment and the initial interpolation samples to be detected at the historical moment determined when the sum of the sample quantities reaches the preset sample quantity threshold are respectively used as the corresponding target interpolation samples to be detected.

[0027] Optionally, the preset target feature dimension reduction model includes a target encoder and a target decoder; the step of using the preset target feature dimension reduction model to perform feature dimension reduction on each of the feature extraction data to be detected to generate a plurality of feature dimension reduction data to be trained includes:

[0028] Using a target encoder to perform low-dimensional conversion on each of the feature extraction data to be detected, so as to generate low-dimensional attribute data corresponding to each of the feature extraction data to be detected;

[0029] The target decoder reconstructs each of the low-dimensional attribute data and outputs a plurality of feature dimensionality reduction data to be trained.

[0030] Optionally, the step of training the initial energy storage system performance evaluation classifier model based on the preset distribution robust optimization strategy using the feature dimension reduction data to be trained to determine the target energy storage system performance evaluation classifier model includes:

[0031] Based on the fractional order parameters of the initial energy storage system performance evaluation classifier model, the initial energy storage system performance evaluation classifier model is used to classify each of the feature dimensionality reduction data to be trained to generate a classification evaluation result to be trained;

[0032] Substituting the evaluation result of the classification to be trained into the preset optimization strategy loss function and deriving it to determine the second model gradient;

[0033] Calculate the momentum term based on the second model gradient at the historical moment;

[0034] Substituting the momentum term into a preset learning rate adjustment formula to calculate the model learning rate;

[0035] The model learning rate and the second model gradient are used to update the model parameters of the initial energy storage system performance evaluation classifier model, determine the intermediate energy storage system performance evaluation classifier model, and count the number of classifier model training times in real time;

[0036] Determine whether the number of training times of the classifier model reaches a preset first training threshold;

[0037] If achieved, the intermediate energy storage system performance evaluation classifier model is used as the trained target energy storage system performance evaluation classifier model.

[0038] Optionally, it also includes:

[0039] If the number of classifier model training times does not reach the preset first training threshold, updating the fractional order parameters to determine new fractional order parameters;

[0040] The intermediate energy storage system performance evaluation classifier model is used as a new initial energy storage system performance evaluation classifier model;

[0041] Based on the new fractional-order parameters, the new initial energy storage system performance evaluation classifier model is used to determine the model learning rate and the second model gradient according to the dimension reduction data of each feature to be trained;

[0042] The model parameters of the new initial energy storage system performance evaluation classifier model are updated using the model learning rate and the second model gradient, the intermediate energy storage system performance evaluation classifier model is determined, and the classifier model training times are counted in real time until the classifier model training times reaches the preset first training threshold;

[0043] The intermediate energy storage system performance evaluation classifier model determined when the number of classifier model training times reaches the preset first training threshold is used as the trained target energy storage system performance evaluation classifier model.

[0044] Optionally, after the step of training the initial energy storage system performance evaluation classifier model based on the preset distribution robust optimization strategy using the feature dimension reduction data to be trained to determine the target energy storage system performance evaluation classifier model, the method further includes:

[0045] When a plurality of feature dimensionality reduction data to be detected are received, each of the feature dimensionality reduction data to be detected is input into the target energy storage system performance evaluation classifier model for classification, and a target classification evaluation result is output.

[0046] Optionally, when a plurality of target interpolation samples to be trained are received, each of the target interpolation samples to be trained is input into an initial feature extraction model for feature extraction, and a plurality of feature extraction data are output;

[0047] Extracting data based on each of the features, calculating a third model gradient;

[0048] Substituting the third model gradient into a preset update amount calculation formula to calculate the model parameter update amount;

[0049] The model parameters of the initial feature extraction model are updated using the model parameter update amount, an intermediate feature extraction model is determined, and the number of feature extraction model training times is counted in real time;

[0050] Determine whether the feature extraction model training times reaches a preset second training threshold;

[0051] If achieved, the intermediate feature extraction model is used as the trained preset target feature extraction model.

[0052] Optionally, when a plurality of feature extraction data to be trained is received, each of the feature extraction data to be trained is input into an initial feature dimension reduction model for feature dimension reduction, and feature dimension reduction data corresponding to each of the feature extraction data to be trained is output;

[0053] Calculate multiple dimensional probability distributions based on the feature extraction data to be trained and the corresponding feature dimension reduction data;

[0054] Calculating information divergence according to the probability distribution of the multiple dimensions;

[0055] Substituting the information divergence and each of the feature dimension reduction data into a preset inverse reconstruction loss function and taking a derivative to determine a fourth model gradient;

[0056] The fourth model gradient is used to update the model parameters of the initial feature dimensionality reduction model, determine the intermediate feature dimensionality reduction model, and count the number of feature dimensionality reduction model training times in real time;

[0057] Determine whether the number of training times of the feature dimensionality reduction model reaches a preset third training threshold;

[0058] If it is achieved, the intermediate feature dimensionality reduction model is used as the trained preset target feature dimensionality reduction model.

[0059] A second aspect of the present invention provides a device for training a classifier model for energy storage system performance evaluation, comprising:

[0060] An acquisition module is used to acquire multiple property data of energy storage systems to be detected, and based on an improved artificial minority class oversampling method, perform data expansion on the property data of each energy storage system to be detected, and determine multiple interpolation samples of targets to be detected;

[0061] A feature extraction module, used for inputting each of the interpolation samples of the target to be detected into a preset target feature extraction model for feature extraction, and outputting a plurality of feature extraction data to be detected;

[0062] A feature dimension reduction module, used to perform feature dimension reduction on each of the feature extraction data to be detected using a preset target feature dimension reduction model, and generate a plurality of feature dimension reduction data to be trained;

[0063] The training module is used to train the initial energy storage system performance evaluation classifier model based on the preset distributed robust optimization strategy using the dimension reduction data of each feature to be trained, and determine the target energy storage system performance evaluation classifier model.

[0064] It can be seen from the above technical solutions that the present invention has the following advantages:

[0065] The above technical scheme of the present invention provides a method for training a classifier model for energy storage system performance evaluation. First, multiple attribute data of energy storage systems to be detected are obtained, and based on the improved artificial minority class oversampling method, data expansion is performed on the attribute data of each energy storage system to be detected, and multiple interpolation samples of targets to be detected are determined; then, each interpolation sample of the target to be detected is input into a preset target feature extraction model for feature extraction, and multiple feature extraction data to be detected are output; a preset target feature dimensionality reduction model is used to perform feature dimensionality reduction on each feature extraction data to be detected, and multiple feature dimensionality reduction data to be trained are generated; finally, based on the preset classification robust optimization model, the target interpolation sample is input into a preset target feature extraction model for feature extraction, and multiple feature extraction data to be detected are output; ization strategy, using the dimension reduction data of each feature to be trained to train the initial energy storage system performance evaluation classifier model, and determining the target energy storage system performance evaluation classifier model; based on the above scheme, through the preset distribution robust optimization strategy, using the dimension reduction data of the feature to be trained obtained by data expansion, preset target feature extraction model, and preset target feature dimension reduction model to train the initial energy storage system performance evaluation classifier model, and determining the process of the target energy storage system performance evaluation classifier model can improve the model's adaptability to data deviations and outliers that may be encountered in actual operations, and further improve the classification accuracy of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0067] Figure 1 A flowchart of a method for training a classifier model for energy storage system performance evaluation provided in Embodiment 1 of the present invention;

[0068] Figure 2 A flowchart of the steps of the model training process of the preset target feature extraction model provided in the second embodiment of the present invention;

[0069] Figure 3 A schematic diagram of a model training process for a preset target feature extraction model provided in the second embodiment of the present invention;

[0070] Figure 4 A flowchart of the steps of the model training process of the preset target feature dimensionality reduction model provided in the third embodiment of the present invention;

[0071] Figure 5 A schematic diagram of the model training process of the preset target feature dimensionality reduction model provided in the third embodiment of the present invention;

[0072] Figure 6 This is a structural block diagram of a device for training a classifier model for energy storage system performance evaluation provided in Embodiment 4 of the present invention. DETAILED DESCRIPTION

[0073] The embodiment of the present invention provides a method and device for training a classifier model for energy storage system performance evaluation, which are used to solve the technical problem that the classification accuracy of the model is poor due to the existing energy storage system performance evaluation technology.

[0074] In order to make the purpose, features and advantages of the present invention more obvious and easy to understand, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the embodiments described below 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.

[0075] See also Figure 1 , Figure 1 A flowchart of the steps of a method for training a classifier model for energy storage system performance evaluation provided in Example 1 of the present invention.

[0076] The present invention provides a method for training a classifier model for energy storage system performance evaluation, comprising:

[0077] Step 101: Acquire multiple to-be-detected energy storage system attribute data, and based on an improved artificial minority class oversampling method, perform data expansion on the attribute data of each to-be-detected energy storage system to determine multiple to-be-detected target interpolation samples.

[0078] The improved artificial minority oversampling method is the quantum projection embedding SMOTE (Synthetic Minority Over-sampling Technique) algorithm.

[0079] It should be noted that the attribute data of the energy storage system to be tested is the attribute data of the energy storage system after being labeled, and the labeling method is manual labeling. The labeled categories include: normal operation, performance degradation, maintenance required and fault status, a total of 4 categories.

[0080] Furthermore, the data sources of the energy storage system attribute data to be detected mainly include but are not limited to the real-time monitoring system of the energy storage facility, historical maintenance records and environmental monitoring devices, and the acquired data is stored in a structured JSON format; wherein the attributes of the acquired data can be set as needed, and the present invention is not limited thereto. For example, the attributes of the energy storage system attribute data to be detected, that is, the multiple energy storage system attribute data to be detected mainly include but are not limited to the battery voltage V a (Unit: Volt), battery current I a (Unit: Ampere), battery temperature T a (Unit: Celsius), charge and discharge status S a (Status: charging, discharging, standby), battery capacity C a (Unit: kWh), number of cycles N a 、Ambient temperature E Ta (Unit: Celsius), ambient humidity E Ha (Unit: percentage), equipment operating status H a (Status: normal, maintenance, failure), maintenance record R a (Record: date and type), etc.; the above examples of the attributes of the energy storage system attribute data to be detected are only one data format and type of the present invention. In practical applications, the attributes of the data are usually more than 10 attributes, and the number of data attributes may reach dozens or even hundreds.

[0081] Furthermore, the acquisition, labeling and preprocessing of training data are time-consuming and labor-intensive, and insufficient training samples can easily lead to poor generalization ability of the model, while affecting the accuracy of the model. The present invention uses a SMOTE algorithm based on quantum projection embedding to generate samples, thereby achieving data expansion. The SMOTE algorithm based on quantum projection embedding aims at the problem that the traditional SMOTE algorithm may be difficult to solve when processing high-dimensional data. By using the probability amplitude based on the quantum state to determine the interpolation weight, the efficiency and quality of data expansion are improved, thereby effectively avoiding the problem of generating redundant or abnormal data points.

[0082] Specifically, based on the improved artificial minority class oversampling method, data expansion is performed on the attribute data of each energy storage system to be detected, and the step of determining multiple interpolation samples of the target to be detected can be achieved by executing the following sub-steps S11 to S18:

[0083] Step S11, performing characteristic calculations according to the attribute data of each energy storage system to be detected, and outputting characteristic parameters corresponding to the attribute data of each energy storage system to be detected;

[0084] The characteristic parameters include a first characteristic parameter and a second characteristic parameter.

[0085] It should be noted that the first characteristic parameter and the second characteristic parameter calculated by the characteristic of the data sample (property data of the energy storage system to be detected) are obtained by normalizing and angle mapping each characteristic vector; wherein the characteristic parameter calculation process is specifically as follows:

[0086] ;

[0087] in, is the first characteristic parameter corresponding to the i-th energy storage system attribute data to be detected; is the characteristic vector of the attribute data of the i-th energy storage system to be detected; is the maximum value function; is the second characteristic parameter corresponding to the attribute data of the i-th energy storage system to be detected; is the jth component of the characteristic vector of the i-th energy storage system attribute data to be detected; is the Euclidean norm of the characteristic vector of the i-th energy storage system attribute data to be detected; is the L2 norm.

[0088] Step S12, using each characteristic parameter to perform quantum mapping on the corresponding property data of the energy storage system to be detected, and generating initial quantum state data corresponding to the property data of each energy storage system to be detected;

[0089] It should be noted that a quantum bit is configured for each data sample (property data of the energy storage system to be detected), and the characteristic vector of each data sample is mapped to the state of the quantum bit, whose initial state is the ground state. The state of each quantum bit is adjusted through quantum gate operation to simulate the characteristics of the data sample; among them, the processing process of the initial quantum state data (quantum state) can be expressed as:

[0090] ;

[0091] in, is the initial quantum state data corresponding to the attribute data of the i-th energy storage system to be detected; is the first characteristic parameter corresponding to the i-th energy storage system attribute data to be detected; is the second characteristic parameter corresponding to the attribute data of the i-th energy storage system to be detected; is the right vector of the quantum state.

[0092] Step S13, performing quantum projection operations on each initial quantum state data to generate target quantum state data corresponding to each energy storage system attribute data to be detected;

[0093] It should be noted that the quantum bit is operated through a specific quantum logic gate to reach a super-position state, which represents the probability distribution of the original data point. The present invention uses Hadamard gates and controlled NOT gates to operate the quantum bit (initial quantum state data) to obtain the converted quantum state (target quantum state data); wherein the processing process of the target quantum state data can be expressed as:

[0094] ;

[0095] in, is the target quantum state data corresponding to the attribute data of the i-th energy storage system to be detected; For Hadamard gate operation; is the initial quantum state data corresponding to the attribute data of the i-th energy storage system to be detected; To control the CNOT operation.

[0096] Step S14, performing quantum state measurement on each target quantum state data, outputting the probability density distribution value corresponding to each energy storage system attribute data to be detected, and comparing each probability density distribution value with a preset probability threshold value;

[0097] It should be noted that when measuring the quantum bit, data points are selected for interpolation according to the probability distribution of the measurement result. The state with a higher probability of the measurement result is more likely to be selected as the starting point of the interpolation. That is, the probability density distribution value corresponding to each energy storage system attribute data to be detected is calculated, and the energy storage system attribute data to be detected corresponding to the probability density distribution value greater than the preset probability threshold is selected as the target detection energy storage system attribute data; wherein, the calculation process of the probability density distribution value can be expressed as:

[0098] ;

[0099] in, is the probability density function corresponding to the attribute data of the i-th energy storage system to be detected; is the state result corresponding to the target quantum state data after the i-th quantum state measurement, that is, the probability density distribution value corresponding to the i-th energy storage system attribute data to be detected.

[0100] Step S15, taking any energy storage system attribute data to be detected corresponding to a probability density distribution value greater than a preset probability threshold as target detection energy storage system attribute data;

[0101] Step S16, performing adaptive multiple interpolation on the neighborhood attribute data in the neighborhood of each target detection energy storage system attribute data, to generate multiple initial interpolation samples to be detected at the current moment;

[0102] It should be noted that, according to the selected data points, that is, the selected target detection energy storage system attribute data, multiple interpolation is performed using the neighborhood data points (neighborhood attribute data) in the neighborhood of each target detection energy storage system attribute data, wherein the interpolation method and weight are determined by the local density and attribute difference of the data points. For example, for data points in sparse areas, a wider interpolation range and a larger weight are used, that is, the generation process of the initial interpolation sample to be detected can be expressed as:

[0103] ;

[0104] in, is the initial interpolation sample to be detected; The weight is calculated based on the local density and attribute difference of the target detection energy storage system attribute data; Detecting the jth neighborhood attribute data in the neighborhood of the target energy storage system attribute data; is the parameter that controls the weight decay rate; Detecting energy storage system attribute data for the i-th target; is the average value of the distance between all data points in the neighborhood and is used to adjust the width of the Gaussian kernel to adapt to data areas of different densities; kre is the total number of all data points in the neighborhood.

[0105] It is worth mentioning that the multiple initial interpolation samples to be detected at the current moment are interpolation samples generated after the current adaptive multiple interpolation operation is performed.

[0106] Step S17, determining whether the sum of the number of the initial interpolation samples to be detected at the current moment and the number of the initial interpolation samples to be detected at the historical moments reaches a preset sample number threshold;

[0107] It should be noted that the sum of the sample numbers of all the initial interpolation samples to be detected at the current moment and the initial interpolation samples to be detected at all historical moments is judged, that is, during the data expansion process, the sum of the sample numbers of all the initial interpolation samples to be detected accumulated historically and all the initial interpolation samples to be detected currently generated reaches a preset sample number threshold; wherein the preset sample number threshold can be set as needed, for example, it can be set to 5000, and the present invention is not limited to this.

[0108] Step S18: If reached, the initial interpolation samples to be detected at the current moment and the initial interpolation samples to be detected at the historical moment are respectively used as the corresponding target interpolation samples to be detected.

[0109] It should be noted that if the sum of the number of historically accumulated initial interpolation samples to be detected and the number of all currently generated initial interpolation samples to be detected reaches a preset sample number threshold, it indicates that the number of samples can meet the fitting performance of the present invention, and sample generation is stopped, and all historically accumulated initial interpolation samples to be detected are used as corresponding target interpolation samples to be detected, and all currently generated initial interpolation samples to be detected are used as corresponding target interpolation samples to be detected, thereby obtaining multiple target interpolation samples to be detected.

[0110] Further, if the sum of the number of samples of the initial interpolation samples to be detected at the previous moment and the number of samples of the initial interpolation samples to be detected at the historical moment does not reach the preset sample number threshold, it indicates that the number of samples cannot meet the fitting performance of the present invention, and it is necessary to continue to generate samples. This process can be implemented by executing the following steps S01 to S06:

[0111] Step S01: If the sum of the number of initial interpolation samples to be detected at the current moment and the number of initial interpolation samples to be detected at the historical moments does not reach the preset sample number threshold, a preset prediction model is used to perform quality assessment on each initial interpolation sample to be detected at the current moment, and a sample quality assessment result is output;

[0112] Step S02: Substitute the sample quality assessment result into a preset loss function and derive it to determine the first model gradient;

[0113] Step S03: using the first model gradient to update each characteristic parameter, and determining a new characteristic parameter corresponding to each property data of the energy storage system to be detected;

[0114] Step S04: using the initial interpolation sample to be detected at the current moment as the initial interpolation sample to be detected at the historical moment;

[0115] Step S05: based on the new characteristic parameters and the corresponding property data of the energy storage system to be detected, determine a plurality of initial interpolation samples to be detected at the current moment, until the sum of the number of initial interpolation samples to be detected at the current moment and the number of initial interpolation samples to be detected at the historical moments reaches a preset sample number threshold;

[0116] Step S06: The initial interpolation samples to be detected at the historical moment, the initial interpolation samples to be detected at the current moment, and the initial interpolation samples to be detected at the historical moment determined when the sum of the sample quantities reaches a preset sample quantity threshold are respectively used as corresponding target interpolation samples to be detected.

[0117] It should be noted that the preset prediction model is a prediction model built based on new data points (samples generated by interpolation). The network structure used by the prediction model is a logistic regression model. The preset prediction model is used to evaluate the quality of the newly generated data points (the initial interpolation samples to be detected at the current moment). The preset loss function is used to adjust the quality feedback of the new data points output. and Parameters are used to optimize the data generation process; the preset generation loss function can be expressed as:

[0118] ;

[0119] in, Generate loss function for preset; is the target label; To preset prediction models; is the i-th initial interpolation sample to be detected; is the rth model parameter in the set of model parameters; is the generator loss regularization coefficient; A collection of model parameters for a preset prediction model.

[0120] Further, each characteristic parameter is updated by using the first model gradient, and a new characteristic parameter corresponding to each property data of the energy storage system to be detected is determined, and the process jumps to execute step S12 until the sum of the number of samples of the initial interpolation samples to be detected at the current moment and the initial interpolation samples to be detected at the historical moments reaches a preset sample number threshold, and the accumulated initial interpolation samples to be detected at all historical moments are used as the corresponding target interpolation samples to be detected, and the initial interpolation samples to be detected at the current moment determined when the sum of the number of samples of the initial interpolation samples to be detected at the current moment and the initial interpolation samples to be detected at the historical moments reaches the preset sample number threshold is used as the corresponding target interpolation sample to be detected.

[0121] In this embodiment, multiple property data of energy storage systems to be detected are obtained, and based on the improved artificial minority class oversampling method, data expansion is performed on the property data of each energy storage system to be detected to determine multiple interpolation samples of targets to be detected.

[0122] Step 102: input each interpolation sample of the target to be detected into a preset target feature extraction model for feature extraction, and output a plurality of feature extraction data to be detected.

[0123] It should be noted that the preset target feature extraction model is a 6-layer fully connected neural network, and its network structure consists of an input layer, a hidden layer and an output layer, specifically: a-input layer: receiving and transmitting sample data after sample expansion; b-hidden layer: the network contains four hidden layers, each layer is composed of a certain number of neurons, b1-first hidden layer: with 128 neurons, for preliminary feature extraction, b2-second hidden layer: with 256 neurons, to further abstract features, b3-third hidden layer: with 512 neurons, to continue to deepen feature extraction, b4-fourth hidden layer: with 1024 neurons, to continue to deepen feature extraction; b5-output layer: with 512 neurons.

[0124] Furthermore, after the input data (the interpolated sample of the target to be detected) is received by the preset target feature extraction model, it passes through each hidden layer in the preset target feature extraction model, and each hidden layer is subjected to linear transformation and nonlinear activation function processing, and finally the feature representation of the output layer is obtained. The output calculation method of each hidden layer, that is, the feature extraction process of the preset target feature extraction model, can be expressed as:

[0125] ;

[0126] in, is the activation output of the i-th hidden layer in the preset target feature extraction model; is the adaptive ReLU activation function; is the weight matrix of the i-th hidden layer in the preset target feature extraction model; The activation output of the i-1th hidden layer in the preset target feature extraction model; is the bias vector of the i-th hidden layer in the preset target feature extraction model.

[0127] Furthermore, the calculation method of the adaptive ReLU activation function proposed in the present invention can be expressed as:

[0128] ;

[0129] in, is the adaptive ReLU activation function; Interpolate samples for the target to be detected; is the positive adaptive coefficient, and its value is 0.3; is the maximum value function; is the negative adaptive coefficient, and its value is 0.7; is the minimum value function.

[0130] In this embodiment, each interpolation sample of the target to be detected is input into a preset target feature extraction model for feature extraction, and a plurality of feature extraction data to be detected is output.

[0131] Step 103: Use a preset target feature dimension reduction model to perform feature dimension reduction on each feature extraction data to be detected, and generate a plurality of feature dimension reduction data to be trained.

[0132] It should be noted that the present invention adopts an autoencoder algorithm based on inverse reconstruction as a preset target feature dimensionality reduction model, which includes two parts: a target encoder and a target decoder. The encoder is responsible for compressing the input data into a low-dimensional representation, and the decoder reconstructs the data from this low-dimensional representation. In order to enhance the feature dimensionality reduction adaptability of the model.

[0133] Further, step 103 can be implemented by executing the following sub-steps S31 to S32:

[0134] Step S31, using a target encoder to perform low-dimensional conversion on each feature extraction data to be detected, to generate low-dimensional attribute data corresponding to each feature extraction data to be detected;

[0135] Step S32: reconstruct each low-dimensional attribute data through the target decoder, and output a plurality of feature dimensionality reduction data to be trained.

[0136] It should be noted that the input feature extracted data (feature extraction data to be detected) is converted into a low-dimensional space representation through the encoder, and the low-dimensional representation (low-dimensional attribute data) is reconstructed into an output with the same dimension as the original data through the decoder. In the forward propagation process, the data processing process of the target encoder can be expressed as:

[0137] ;

[0138] in, is the low-dimensional feature representation of the encoding (low-dimensional attribute data); is the Sigmoid activation function; is the weight of the target encoder; Extract data for the features to be detected; is the offset of the target encoder.

[0139] Furthermore, the reconstructed output of the target decoder, that is, the data processing process of the target decoder, can be expressed as:

[0140] ;

[0141] in, is the low-dimensional feature representation of the encoding (low-dimensional attribute data); is the Sigmoid activation function; is the weight of the target decoder; Reduce the dimension of the data for the features to be trained; is the bias of the target decoder.

[0142] In this embodiment, a preset target feature dimension reduction model is used to perform feature dimension reduction on each feature extraction data to be detected, so as to generate a plurality of feature dimension reduction data to be trained.

[0143] Step 104: Based on the preset distributed robust optimization strategy, the initial energy storage system performance evaluation classifier model is trained using the dimension reduction data of each feature to be trained to determine the target energy storage system performance evaluation classifier model.

[0144] It should be noted that the reduced-dimensional data (feature reduced-dimensional data to be trained) is input into the classifier, that is, the initial energy storage system performance evaluation classifier model is used to train the classifier model. The present invention adopts a fractional-order neural network algorithm based on fractional robust optimization as the initial energy storage system performance evaluation classifier model for classification. The network structure of the model adopts a multi-layer neural network framework, and each layer enhances the learning ability and generalization of the model through fractional-order derivatives. The present invention uses fractional-order derivatives instead of integer-order derivatives, which improves the adaptability of the model to nonlinear and complex data patterns, making it more robust to data deviations and outliers that may be encountered in actual operations, thereby improving classification accuracy.

[0145] Specifically, step 104 can be implemented by executing the following sub-steps S41 to S47:

[0146] Step S41, based on the fractional order parameters of the initial energy storage system performance evaluation classifier model, the initial energy storage system performance evaluation classifier model is used to classify the feature dimensionality reduction data to be trained, and generate a classification evaluation result to be trained;

[0147] It should be noted that the model parameters of the initial energy storage system performance evaluation classifier model include the weights and biases of the fractional-order neural network model, and its initialization method can be expressed as:

[0148] ;

[0149] in, is the initial value of the weight of the fractional-order neural network model; is to generate normally distributed random numbers; drf is the dimension of the feature reduction data to be trained, indicating the dimension of the input feature; mrf is the dimension of the output layer; nrg is the number of samples input in the current batch; is the initial value of the bias of the fractional-order neural network model; To generate a zero vector.

[0150] Furthermore, the input feature-reduced data (feature-reduced data to be trained) is forward propagated through the initial energy storage system performance evaluation classifier model. The activation function of each layer of the model uses a fractional-order nonlinear function to capture the nonlinear relationship in the data. The output of the last layer of the model is the classification evaluation result to be trained. The forward propagation process can be expressed as:

[0151] ;

[0152] in, is the linear transformation result of the lth layer; is the weight of the lth layer; is the output of the lth layer; is the bias of the lth layer; is the fractional-order ReLU function; is the maximum value function; is a fractional parameter, between 0.5 and 1; is the output of the l-1th layer.

[0153] Step S42: Substitute the evaluation result of the classification to be trained into the preset optimization strategy loss function and derive it to determine the second model gradient;

[0154] It should be noted that the loss function is calculated according to the distributionally robust optimization strategy, so that the model maintains performance when facing data distribution deviation. The calculation method of the loss function of the preset optimization strategy can be expressed as:

[0155] ;

[0156] in, is the preset optimization strategy loss function, which represents the loss function of the fractional-order neural network; is the true label; is the evaluation result of the classification to be trained, indicating the model prediction; is the regularization coefficient, and its value is 0.3; is the weight of the output layer; mgf is the number of neurons in the output layer; N is the total number of samples, that is, the number of samples of feature dimensionality reduction data to be trained; To adjust the coefficient of the influence of the distribution sensitive loss term, the value is taken as 0.1; is the Kullback-Leibler divergence between the true distribution and the predicted distribution of the i-th sample (the feature reduction data to be trained); is the weight of the Lth layer; is the output of the L-1th layer; is the bias of the Lth layer; L is the total number of layers of the initial energy storage system performance evaluation classifier model.

[0157] Step S43, calculating the momentum term according to the second model gradient at the historical moment;

[0158] Specifically, the second model gradient at the historical moment (the second model gradient at the previous moment, the second model gradient at the previous moment) and the momentum term at the historical moment (the momentum term at the previous moment) are used to calculate the momentum term at the current moment. The processing process of the momentum term can be expressed as:

[0159] ;

[0160] in, is the momentum term of the tth iteration, that is, the momentum term at the current moment; is the learning rate decay coefficient; is the momentum term of the t-1th iteration, that is, the momentum term of the previous moment; is the gradient change rate of the t-1th iteration, that is, the gradient change rate of the previous moment; is the gradient change at the t-1th iteration; is the second model gradient of the t-2th iteration, that is, the second model gradient of the previous moment; is the second model gradient of the t-1th iteration, that is, the second model gradient of the previous moment.

[0161] Step S44, substituting the momentum term into the preset learning rate adjustment formula to calculate the model learning rate;

[0162] It should be noted that the present invention adopts an adaptive learning rate adjustment strategy to adjust the model learning rate, and presets a learning rate adjustment formula, which is specifically:

[0163] ;

[0164] in, is the model learning rate of the tth iteration, that is, the model learning rate at the current moment; is the learning rate attenuation coefficient, which is set to 0.9; t is the number of iterations, i.e. the number of times the classifier model is trained; is the momentum term of the tth iteration, that is, the momentum term at the current moment; is the initial learning rate.

[0165] Step S45, using the model learning rate and the second model gradient to update the model parameters of the initial energy storage system performance evaluation classifier model, determine the intermediate energy storage system performance evaluation classifier model, and count the number of classifier model training times in real time;

[0166] The second model gradient includes the classifier model weight gradient and the classifier model bias gradient.

[0167] The model parameters of the initial energy storage system performance evaluation classifier model include weights and biases of the fractional-order neural network.

[0168] It should be noted that the fractional gradient descent algorithm is used for back propagation, that is, based on the fractional gradient descent algorithm, the model learning rate and the second model gradient are used to update the model parameters of the initial energy storage system performance evaluation classifier model, and the intermediate energy storage system performance evaluation classifier model is determined to adjust the flow and update rate of the gradient to achieve more detailed parameter adjustment; wherein, the update process of the model parameters of the energy storage system performance evaluation classifier model can be expressed as:

[0169] ;

[0170] in, is the weight gradient of the classifier model; The partial derivative of the preset optimization strategy loss function with respect to the classifier model weight output; is the weight of the energy storage system performance evaluation classifier model of the t-th iteration, that is, the weight of the fractional-order neural network in the model parameters of the intermediate energy storage system performance evaluation classifier model; is the weight of the energy storage system performance evaluation classifier model of the t-1th iteration, that is, the weight of the fractional-order neural network in the model parameters of the initial energy storage system performance evaluation classifier model; is the model learning rate of the tth iteration, that is, the model learning rate at the current moment; Bias gradient for the classifier model; The partial derivative of the preset optimization strategy loss function with respect to the bias output of the classifier model; is the bias of the energy storage system performance evaluation classifier model of the t-th iteration, that is, the bias of the fractional-order neural network in the model parameters of the intermediate energy storage system performance evaluation classifier model; It is the bias of the energy storage system performance evaluation classifier model of the t-1th iteration, that is, the bias of the fractional-order neural network in the model parameters of the initial energy storage system performance evaluation classifier model.

[0171] Step S46, determining whether the number of classifier model training times reaches a preset first training threshold;

[0172] Step S47: If reached, the intermediate energy storage system performance evaluation classifier model is used as the trained target energy storage system performance evaluation classifier model.

[0173] It should be noted that if the number of classifier model training times reaches the preset first training threshold, it indicates that the model training is completed, and the intermediate energy storage system performance evaluation classifier model is used as the trained target energy storage system performance evaluation classifier model; wherein, the preset first training threshold can be set as needed, for example, it can be set to 1000 times, and the present invention is not limited thereto.

[0174] Furthermore, if the number of classifier model training times does not reach the preset first training threshold, the surface model has not yet completed training and needs to continue to adjust the model parameters. This process can be achieved by executing the following steps S001 to S005:

[0175] S001. If the number of classifier model training times does not reach the preset first training threshold, the fractional order parameters are updated to determine new fractional order parameters;

[0176] S002, using the intermediate energy storage system performance evaluation classifier model as a new initial energy storage system performance evaluation classifier model;

[0177] S003. Based on the new fractional order parameters, a new initial energy storage system performance evaluation classifier model is used to determine the model learning rate and the second model gradient according to the dimension reduction data of each feature to be trained;

[0178] S004, using the model learning rate and the second model gradient to update the model parameters of the new initial energy storage system performance evaluation classifier model, determine the intermediate energy storage system performance evaluation classifier model, and count the number of classifier model training times in real time until the number of classifier model training times reaches a preset first training threshold;

[0179] S005. Using the intermediate energy storage system performance evaluation classifier model determined when the number of classifier model training times reaches a preset first training threshold as the trained target energy storage system performance evaluation classifier model.

[0180] It should be noted that the updating process of fractional-order parameters can be expressed as:

[0181] ;

[0182] in, is the fractional order parameter of the tth iteration, i.e., the new fractional order parameter; is the initial fractional order parameter; is the adjustment coefficient, the value is 3; is the mean square error of the tth iteration.

[0183] Furthermore, after determining the new fractional-order parameters, the intermediate energy storage system performance evaluation classifier model is used as the new initial energy storage system performance evaluation classifier model, and the process jumps to step S41 until the number of classifier model training times reaches a preset first training threshold, and the intermediate energy storage system performance evaluation classifier model determined when the number of classifier model training times reaches the preset first training threshold is used as the trained target energy storage system performance evaluation classifier model.

[0184] Optionally, when a plurality of feature dimensionality reduction data to be detected are received, each feature dimensionality reduction data to be detected is input into a target energy storage system performance evaluation classifier model for classification, and a target classification evaluation result is output.

[0185] It should be noted that the trained model is used to process new samples to achieve the performance evaluation of the energy storage system of new energy power facilities. Among them, the data for the performance evaluation of the energy storage system of new energy power facilities, that is, the collected original data (property data of the energy storage system to be detected), is processed by data expansion, and then input into the trained preset feature extraction model and preset feature dimensionality reduction model in sequence for feature processing. Further, the processed features are input into the initial energy storage system performance evaluation classifier model for classifier training. After the model training is completed, when the feature dimensionality reduction data to be detected is received, it is processed by the trained classifier model to obtain the target classification evaluation result. The classification categories include: normal operation, performance degradation, maintenance required and fault state, a total of 4 categories. Finally, different operations are performed according to the evaluation results to achieve maintenance of the energy storage system.

[0186] In this embodiment, based on the preset distributed robust optimization strategy, the dimensionality reduction data of each feature to be trained is used to train the initial energy storage system performance evaluation classifier model to determine the target energy storage system performance evaluation classifier model.

[0187] As a comparison of technical effects, it can be combined with existing technologies for reference. Traditional data expansion methods may not be able to effectively generate diverse data points when processing high-dimensional data, resulting in data redundancy or anomalies in the data set, affecting the effect of model training and generalization ability; existing feature extraction models may encounter gradient vanishing or explosion problems when the network structure is complex, affecting the stability of training and the performance of the model. In addition, traditional feature dimensionality reduction methods may lose key features of the data during the compression process, reducing the effectiveness of the model; existing technologies lack adaptive processing of nonlinear and complex data patterns in classifier model design, and are not adaptable enough to data deviations and outliers that may be encountered in actual operations, affecting the accuracy of classification and the robustness of the model.

[0188] In view of the above problems, the present invention adopts the SMOTE algorithm based on quantum projection embedding, and accurately expands high-dimensional data through state simulation of quantum bits and quantum gate operations. The adjustment of the quantum bit state reflects the data characteristics, allowing data points to be generated more efficiently in high-dimensional space. The data points are selected for interpolation through the measurement results of the quantum bits, ensuring the diversity and applicability of the generated data. In addition, adaptive multiple interpolation adjusts the interpolation process according to the local density and attribute differences of the data points to improve the data quality; in the feature extraction stage, a six-layer fully connected neural network is used, and the exploratory gradient descent algorithm is used to optimize the network parameters. This method adjusts the learning step size by predicting the trend of gradient changes, effectively avoiding the problem of gradient disappearance or explosion; feature dimensionality reduction uses an autoencoder based on inverse reconstruction. This strategy maintains the key features of the data by minimizing the encoding and decoding processes between dependencies; the classifier model adopts a fractional-order neural network based on fractional robust optimization, and uses fractional-order derivatives to enhance the adaptability of the model to nonlinear and complex data patterns, and optimizes the performance of the model under data deviation.

[0189] In summary, the present invention effectively solves the problem of generating redundant or abnormal data points that may be encountered by traditional data expansion methods when processing high-dimensional data through data expansion, thereby improving the efficiency and quality of data expansion; the feature extraction model realizes fast and accurate feature learning through exploratory gradient descent, thereby enhancing the generalization ability of the model; the feature dimensionality reduction model ensures that important features are retained during data compression, thereby improving the accuracy of data processing; the fractional-order neural network improves the model's adaptability to data deviations and outliers that may be encountered in actual operations through a distributed robust optimization strategy, thereby improving the classification accuracy.

[0190] In an embodiment of the present invention, the present invention provides a method for training a classifier model for energy storage system performance evaluation. First, multiple attribute data of energy storage systems to be detected are obtained, and based on an improved artificial minority class oversampling method, data expansion is performed on the attribute data of each energy storage system to be detected, and multiple target interpolation samples to be detected are determined; then, each target interpolation sample to be detected is input into a preset target feature extraction model for feature extraction, and multiple feature extraction data to be detected are output; a preset target feature dimensionality reduction model is used to perform feature dimensionality reduction on each feature extraction data to be detected, and multiple feature dimensionality reduction data to be trained are generated; finally, based on the preset classification robustness, the target interpolation sample to be detected is input into a preset target feature extraction model for feature extraction, and multiple feature extraction data to be detected are output; Optimization strategy, using the dimension reduction data of each feature to be trained to train the initial energy storage system performance evaluation classifier model, and determine the target energy storage system performance evaluation classifier model; based on the above scheme, through the preset distribution blue robust optimization strategy, the feature dimension reduction data to be trained obtained by data expansion, preset target feature extraction model, and preset target feature dimension reduction model processing is used to train the initial energy storage system performance evaluation classifier model, and the process of determining the target energy storage system performance evaluation classifier model can improve the model's adaptability to data deviations and outliers that may be encountered in actual operations, and further improve the classification accuracy of the model.

[0191] For better explanation, refer to Figure 2 , shows a flowchart of the steps of the model training process of the preset target feature extraction model provided in the second embodiment of the present invention, and the process may include the following steps:

[0192] Step 201: When a plurality of target interpolation samples to be trained are received, each target interpolation sample to be trained is input into an initial feature extraction model for feature extraction, and a plurality of feature extraction data are output.

[0193] The target interpolation samples to be trained are interpolation samples used for model training of the initial feature extraction model.

[0194] It should be noted that the expanded data is input into the initial feature extraction model for training the feature extraction model, that is, when multiple target interpolation samples to be trained are received, each target interpolation sample to be trained is input into the initial feature extraction model for feature extraction, and multiple feature extraction data are output.

[0195] Furthermore, the model parameters of the initial feature extraction model include the weights and biases of the fully connected neural network. The initialization method of the model parameters can be expressed as:

[0196] ;

[0197] in, is the weight of the i-th layer of the initial feature extraction model (neural network); is the scaling factor for the weights of the initial feature extraction model; is a normal distribution with a mean of 0 and a standard deviation of 1; is a tiny noise, which is represented by a noise vector whose elements are all 0.001; is the bias of the i-th layer of the initial feature extraction model; Scaling factor for the bias of the initial feature extraction model.

[0198] In this embodiment, when a plurality of target interpolation samples to be trained are received, each target interpolation sample to be trained is input into an initial feature extraction model for feature extraction, and a plurality of feature extraction data are output.

[0199] Step 202: extract data according to each feature and calculate the third model gradient.

[0200] It should be noted that, by adopting an exploratory gradient descent update strategy, before each update of the model parameters of the initial feature extraction model, possible gradient changes are predicted, and the step size and direction of the gradient update are adjusted according to the prediction results, so as to optimize the parameter adjustment during the training process and avoid falling into the local optimum too early; among them, the calculation of the third model gradient according to the model results (feature extraction data) output by the feature extraction model can refer to the calculation process of the cross entropy loss function in the prior art, and the present invention will not go into details.

[0201] In this embodiment, data is extracted based on each feature and the third model gradient is calculated.

[0202] Step 203: Substitute the third model gradient into the preset update amount calculation formula to calculate the model parameter update amount.

[0203] The third model gradient includes the gradient of the loss function with respect to the weight and the gradient of the loss function with respect to the bias.

[0204] The model parameter update includes the weight update and the bias update.

[0205] It should be noted that the calculation process of the model parameter update amount can be expressed as:

[0206] ;

[0207] in, is the update amount of the weight of the i-th layer of the neural network; is the learning rate of the neural network, and its value is 0.03; is the loss function with respect to weights The gradient of To explore the strength coefficient; is the loss function of the neural network, which is the cross entropy loss function; is the prediction function of the gradient; is the update amount of the bias of the i-th layer of the neural network; is the loss function with respect to the bias gradient.

[0208] Furthermore, the exploratory gradient prediction function uses historical gradient data to perform linear regression prediction. Taking the processing of the gradient of the loss function with respect to the weight as an example, its calculation method can be expressed as:

[0209] ;

[0210] in, is the prediction function of the gradient; To explore the intensity coefficient, the value is 2; is the attenuation factor, which is set to 0.98; Tgz is the total number of iterations, i.e., the preset second training threshold; is the loss function at the t-1th iteration with respect to the weight gradient.

[0211] In this embodiment, the third model gradient is substituted into the preset update amount calculation formula to calculate the model parameter update amount.

[0212] Step 204: update the model parameters of the initial feature extraction model using the model parameter update amount, determine the intermediate feature extraction model, and count the number of feature extraction model training times in real time.

[0213] It should be noted that the model parameters of the initial feature extraction model, namely the weights and bias parameters of the neural network, are updated by using the error back propagation method and the model parameter update amount.

[0214] In this embodiment, the model parameters of the initial feature extraction model are updated using the model parameter update amount and the third model gradient, the intermediate feature extraction model is determined, and the number of feature extraction model training times is counted in real time.

[0215] Step 205: Determine whether the number of feature extraction model training times reaches a preset second training threshold.

[0216] In this embodiment, it is determined whether the number of feature extraction model training times reaches a preset second training threshold.

[0217] Step 206: If reached, the intermediate feature extraction model is used as the trained preset target feature extraction model.

[0218] It should be noted that if the number of feature extraction model training times reaches the preset second training threshold, the intermediate feature extraction model is used as the trained preset target feature extraction model; if the number of feature extraction model training times does not reach the preset second training threshold, the intermediate feature extraction model is used as the new initial feature extraction model, and the process jumps to execute step 201 until the number of feature extraction model training times reaches the preset second training threshold, and the intermediate feature extraction model determined when the number of feature extraction model training times reaches the preset second training threshold is used as the trained preset target feature extraction model; wherein, the preset second training threshold can be set as needed, for example, it can be set to 1000 times, and the present invention is not limited to this.

[0219] In this embodiment, if it is achieved, the intermediate feature extraction model is used as the trained preset target feature extraction model.

[0220] For example, see Figure 3 First, initialize the model parameters of the feature extraction model (neural network), namely weights and biases. Then, input the training data (target interpolation samples to be trained) into the neural network, initialize each hidden layer in the feature extraction model for linear changes and nonlinear activation, that is, perform feature extraction operations, output feature extraction data (that is, calculate the output layer feature representation), and finally use exploratory gradient descent to update the neural network parameters, namely weights and biases, until the preset stop iteration condition is met, which means that the model training is completed.

[0221] In an embodiment of the present invention, the present invention adopts a neural network based on exploratory gradient descent as a preset feature extraction model, which not only considers the current gradient information, but also adopts a forward exploration mechanism to predict the possible change trend of the gradient before the gradient is updated, and adjust the learning step size and direction to achieve faster convergence speed and better local minimum, thereby enhancing the generalization ability of features. It can solve the problem in the prior art that when some schemes use neural networks for feature extraction, in certain neural network structures, they may encounter gradient disappearance, gradient explosion or falling into local optimal solutions, thereby further improving the stability of training and the performance of the model.

[0222] For better explanation, refer to Figure 4 , showing Figure 4 A flowchart of the steps of the model training process of the preset target feature dimensionality reduction model provided in the third embodiment of the present invention, the process may include the following steps:

[0223] Step 401: when a plurality of feature extraction data to be trained are received, each feature extraction data to be trained is input into an initial feature dimension reduction model for feature dimension reduction, and feature dimension reduction data corresponding to each feature extraction data to be trained is output.

[0224] The feature extraction data to be trained is feature extraction data used for model training of the initial feature dimensionality reduction model.

[0225] It should be noted that the model parameters of the initial feature dimensionality reduction model include the weights and bias parameters of the autoencoder, and their initialization method is expressed as

[0226] ;

[0227] in, is the initial weight of the autoencoder; is the number of neurons in the input layer; is the number of neurons in the output layer; is a random number uniformly distributed in the interval [-1,1]; is the initial bias of the autoencoder.

[0228] In this embodiment, when a plurality of feature extraction data to be trained is received, each feature extraction data to be trained is input into the initial feature dimension reduction model for feature dimension reduction, and feature dimension reduction data corresponding to each feature extraction data to be trained is output.

[0229] Step 402: Calculate multiple dimensional probability distributions based on the feature extraction data to be trained and the corresponding feature dimension reduction data.

[0230] In this embodiment, multiple dimensional probability distributions are calculated based on each feature extraction data to be trained and the corresponding feature dimension reduction data.

[0231] Step 403: Calculate information divergence based on the probability distribution of multiple dimensions.

[0232] It should be noted that KL divergence (information divergence) can measure the difference between the original data distribution and the reconstructed data distribution, and is used to ensure that the key features of the data are not lost during the dimensionality reduction process. The calculation method of information divergence is expressed as:

[0233] ;

[0234] in, is the information divergence; Mtr is the feature dimension before feature dimensionality reduction; is the probability distribution of the original data (data to be trained and feature extracted) in the i-th dimension; is the probability distribution of the reconstructed data (feature reduced dimension data) in the i-th dimension.

[0235] In this embodiment, information divergence is calculated based on probability distributions of multiple dimensions.

[0236] Step 404: Substitute the information divergence and each feature dimension reduction data into a preset inverse reconstruction loss function and derive the function to determine the fourth model gradient.

[0237] It should be noted that the calculation of the inverse reconstruction loss of the autoencoder defines the calculation method of the inverse reconstruction loss function, that is, the preset inverse reconstruction loss function can be expressed as:

[0238] ;

[0239] in, is the preset inverse reconstruction loss function; Ntr is the number of samples input in the current batch, that is, the number of feature extraction data to be trained in the current batch; Extract data for the i-th feature to be trained; Dimensionality reduction data for the i-th feature; is the regularization parameter; is information divergence; is the L2 norm.

[0240] In this embodiment, the information divergence and each feature dimension reduction data are substituted into a preset inverse reconstruction loss function and derived to determine the fourth model gradient.

[0241] Step 405: Use the fourth model gradient to update the model parameters of the initial feature dimensionality reduction model, determine the intermediate feature dimensionality reduction model, and count the number of feature dimensionality reduction model training times in real time.

[0242] The fourth model gradient includes the encoder weight gradient, the encoder bias gradient, the decoder weight gradient, and the decoder bias gradient.

[0243] The model parameters of the initial feature dimensionality reduction model include weights and bias parameters of the autoencoder, the weights of the autoencoder include encoder weights and decoder weights, and the bias parameters of the autoencoder include encoder bias and decoder bias.

[0244] It should be noted that, according to the fourth model gradient calculated by the above preset inverse reconstruction loss function, the weights and biases of the encoder and decoder are updated using the gradient descent algorithm, and the update method is expressed as:

[0245] ;

[0246] in, is the updated encoder weight, i.e., the encoder weight of the intermediate feature dimensionality reduction model; is the encoder weight before updating, that is, the encoder weight of the initial feature dimensionality reduction model; is the learning rate; is the encoder weight gradient; is the updated encoder bias, i.e., the encoder bias of the intermediate feature dimensionality reduction model; is the encoder bias before updating, that is, the encoder bias of the initial feature dimensionality reduction model; is the encoder bias gradient; is the updated decoder weight, i.e., the decoder weight of the intermediate feature dimensionality reduction model; is the decoder weight before updating, that is, the decoder weight of the initial feature dimensionality reduction model; is the decoder weight gradient; is the updated decoder bias, i.e., the decoder bias of the intermediate feature dimensionality reduction model; is the decoder bias before updating, that is, the decoder bias of the initial feature dimensionality reduction model; is the decoder bias gradient.

[0247] It is worth mentioning that the gradient For example, the calculation method can be expressed as:

[0248] ;

[0249] in, is the encoder weight gradient; is the partial derivative of the loss function with respect to the reconstructed output, that is, the partial derivative of the preset inverse reconstruction loss function with respect to the feature dimensionality reduction data; is the partial derivative in the decoding process; is the partial derivative in the encoding process.

[0250] In this embodiment, the fourth model gradient is used to update the model parameters of the initial feature dimensionality reduction model, determine the intermediate feature dimensionality reduction model, and count the number of feature dimensionality reduction model training times in real time.

[0251] Step 406: Determine whether the number of feature dimensionality reduction model training times reaches a preset third training threshold.

[0252] In this embodiment, it is determined whether the number of feature dimension reduction model training times reaches a preset third training threshold.

[0253] Step 407: If reached, the intermediate feature dimensionality reduction model is used as the trained preset target feature dimensionality reduction model.

[0254] It should be noted that if the number of feature dimensionality reduction model training times reaches the preset third training threshold, it means that the model training is completed, and the intermediate feature dimensionality reduction model is used as the trained preset target feature dimensionality reduction model; if the number of feature dimensionality reduction model training times does not reach the preset third training threshold, the intermediate feature dimensionality reduction model is used as the new initial feature dimensionality reduction model, and the process jumps to execute step 401 until the number of feature dimensionality reduction model training times reaches the preset third training threshold, and the intermediate feature dimensionality reduction model determined when the number of feature dimensionality reduction model training times reaches the preset third training threshold is used as the trained preset target feature dimensionality reduction model; wherein, the preset third training threshold can be set as needed, for example, it can be set to 1200, and the present invention is not limited to this.

[0255] In this embodiment, if it is achieved, the intermediate feature dimensionality reduction model is used as the trained preset target feature dimensionality reduction model.

[0256] For example, see Figure 5 First, initialize the autoencoder parameters (that is, initialize the model parameters of the initial feature dimensionality reduction model). Then, input the feature extracted data (feature extraction data to be trained) into the initial feature dimensionality reduction model. The data is converted into a low-dimensional space representation through the encoding in the initial feature dimensionality reduction model, and then the decoder is used to reconstruct the original dimensional data, thereby outputting the feature dimensionality reduction data. Based on the feature dimensionality reduction data, the inverse reconstruction loss and the corresponding gradient are calculated, and the weights and biases of the initial feature dimensionality reduction model are updated using gradient descent until the preset stop iteration condition is met, indicating that the model training is completed.

[0257] In an embodiment of the present invention, the present invention adopts an autoencoder algorithm based on inverse reconstruction as a feature dimensionality reduction model, which includes two parts: an encoder and a decoder. The encoder is responsible for compressing the input data into a low-dimensional representation, and the decoder reconstructs the data from the low-dimensional representation, thereby enhancing the feature dimensionality reduction adaptability of the model; the feature-extracted data is input into the feature dimensionality reduction model for training the feature dimensionality reduction model, and the inverse reconstruction strategy is used to process the input high-dimensional data after feature extraction, so as to minimize the dependency between different features, thereby improving the quality of encoding and the accuracy of decoding.

[0258] See also Figure 6 , Figure 6 This is a structural block diagram of a device for training a classifier model for energy storage system performance evaluation provided in Embodiment 4 of the present invention.

[0259] The present invention provides a device for training a classifier model for energy storage system performance evaluation, comprising:

[0260] The acquisition module 601 is used to acquire attribute data of multiple energy storage systems to be detected, and based on the improved artificial minority class oversampling method, perform data expansion on the attribute data of each energy storage system to be detected, and determine multiple interpolation samples of targets to be detected;

[0261] The feature extraction module 602 is used to input each interpolation sample of the target to be detected into a preset target feature extraction model for feature extraction, and output a plurality of feature extraction data to be detected;

[0262] The feature dimension reduction module 603 is used to perform feature dimension reduction on each feature extraction data to be detected by using a preset target feature dimension reduction model to generate a plurality of feature dimension reduction data to be trained;

[0263] The training module 604 is used to train the initial energy storage system performance evaluation classifier model based on the preset distributed robust optimization strategy using the dimension reduction data of each feature to be trained, and determine the target energy storage system performance evaluation classifier model.

[0264] Furthermore, the acquisition module 601 is specifically used for:

[0265] Perform feature calculations according to the attribute data of each energy storage system to be detected, and output feature parameters corresponding to the attribute data of each energy storage system to be detected;

[0266] Using each characteristic parameter to perform quantum mapping on the corresponding property data of the energy storage system to be detected, and generating initial quantum state data corresponding to the property data of each energy storage system to be detected;

[0267] Performing quantum projection operations on each initial quantum state data respectively to generate target quantum state data corresponding to each energy storage system attribute data to be detected;

[0268] Perform quantum state measurement on each target quantum state data, output the probability density distribution value corresponding to each energy storage system attribute data to be detected, and compare each probability density distribution value with the preset probability threshold value respectively;

[0269] The property data of the energy storage system to be detected corresponding to any probability density distribution value greater than a preset probability threshold is used as the property data of the target energy storage system to be detected;

[0270] Adaptively perform multiple interpolation on the neighborhood attribute data in the neighborhood of each target detection energy storage system attribute data to generate multiple initial interpolation samples to be detected at the current moment;

[0271] Determine whether the sum of the number of initial interpolation samples to be detected at the current moment and the number of initial interpolation samples to be detected at the historical moments reaches a preset sample number threshold;

[0272] If it is reached, the initial interpolation samples to be detected at the current moment and the initial interpolation samples to be detected at the historical moment are respectively used as the corresponding target interpolation samples to be detected.

[0273] In an optional embodiment, the device further comprises:

[0274] The first module is used to use a preset prediction model to perform quality assessment on each of the initial interpolation samples to be detected at the current moment, and output a sample quality assessment result if the sum of the sample quantities of the initial interpolation samples to be detected at the current moment and the initial interpolation samples to be detected at the historical moments does not reach a preset sample quantity threshold;

[0275] The second module is used to substitute the sample quality assessment result into the preset loss function and derive it to determine the first model gradient;

[0276] The third module is used to update each characteristic parameter using the first model gradient to determine a new characteristic parameter corresponding to each property data of the energy storage system to be detected;

[0277] The fourth module is used to use the initial interpolation sample to be detected at the current moment as the initial interpolation sample to be detected at the historical moment;

[0278] The fifth module is used to determine a plurality of initial interpolation samples to be detected at the current moment based on the new characteristic parameters and the corresponding attribute data of the energy storage system to be detected, until the sum of the number of initial interpolation samples to be detected at the current moment and the number of initial interpolation samples to be detected at the historical moment reaches a preset sample number threshold;

[0279] The sixth module is used to use the initial interpolation samples to be detected at historical moments, the initial interpolation samples to be detected at the current moment, and the initial interpolation samples to be detected at historical moments determined when the sum of the sample quantities reaches a preset sample quantity threshold as corresponding target interpolation samples to be detected.

[0280] Furthermore, the preset target feature dimension reduction model includes a target encoder and a target decoder; a feature dimension reduction module 603, specifically used for:

[0281] The target encoder is used to perform low-dimensional conversion on each feature extraction data to be detected, so as to generate low-dimensional attribute data corresponding to each feature extraction data to be detected;

[0282] The target decoder is used to reconstruct the low-dimensional attribute data and output multiple feature dimensionality reduction data to be trained.

[0283] Furthermore, the training module 604 is specifically used for:

[0284] Based on the fractional order parameters of the initial energy storage system performance evaluation classifier model, the initial energy storage system performance evaluation classifier model is used to classify the dimension reduction data of each feature to be trained, and generate the classification evaluation results to be trained;

[0285] Substitute the evaluation result of the classification to be trained into the preset optimization strategy loss function and derive it to determine the gradient of the second model;

[0286] Calculate the momentum term based on the second model gradient at the historical moment;

[0287] Substitute the momentum term into the preset learning rate adjustment formula to calculate the model learning rate;

[0288] The model learning rate and the second model gradient are used to update the model parameters of the initial energy storage system performance evaluation classifier model, determine the intermediate energy storage system performance evaluation classifier model, and count the number of classifier model training times in real time;

[0289] Determine whether the number of classifier model training times reaches a preset first training threshold;

[0290] If it is achieved, the intermediate energy storage system performance evaluation classifier model is used as the trained target energy storage system performance evaluation classifier model.

[0291] In an optional embodiment, the device further comprises:

[0292] The seventh module is used to update the fractional order parameters and determine new fractional order parameters if the number of classifier model training times does not reach the preset first training threshold;

[0293] An eighth module is used to use the intermediate energy storage system performance evaluation classifier model as a new initial energy storage system performance evaluation classifier model;

[0294] The ninth module is used to determine the model learning rate and the second model gradient based on the new fractional order parameters and the new initial energy storage system performance evaluation classifier model according to the dimension reduction data of each feature to be trained;

[0295] The tenth module is used to update the model parameters of the new initial energy storage system performance evaluation classifier model using the model learning rate and the second model gradient, determine the intermediate energy storage system performance evaluation classifier model, and count the number of classifier model training times in real time until the number of classifier model training times reaches a preset first training threshold;

[0296] The eleventh module is used to use the intermediate energy storage system performance evaluation classifier model determined when the number of classifier model training times reaches a preset first training threshold as the trained target energy storage system performance evaluation classifier model.

[0297] In an optional embodiment, the device further comprises:

[0298] The evaluation module is used to input each of the feature dimensionality reduction data to be detected into the target energy storage system performance evaluation classifier model for classification when receiving multiple feature dimensionality reduction data to be detected, and output the target classification evaluation result.

[0299] In an optional embodiment, the device further comprises:

[0300] The first feature extraction training module is used for, when receiving a plurality of target interpolation samples to be trained, inputting each target interpolation sample to be trained into an initial feature extraction model for feature extraction, and outputting a plurality of feature extraction data;

[0301] A feature extraction second training module is used to extract data according to each feature and calculate the third model gradient;

[0302] The feature extraction third training module is used to substitute the third model gradient into the preset update amount calculation formula to calculate the model parameter update amount;

[0303] The fourth feature extraction training module is used to update the model parameters of the initial feature extraction model using the model parameter update amount, determine the intermediate feature extraction model, and count the number of feature extraction model training times in real time;

[0304] A fifth feature extraction training module is used to determine whether the number of feature extraction model training times reaches a preset second training threshold;

[0305] The sixth feature extraction training module is used to use the intermediate feature extraction model as the trained preset target feature extraction model if it is achieved.

[0306] In an optional embodiment, the device further comprises:

[0307] The first feature dimension reduction training module is used for, when receiving a plurality of feature extraction data to be trained, inputting each feature extraction data to be trained into the initial feature dimension reduction model for feature dimension reduction, and outputting feature dimension reduction data corresponding to each feature extraction data to be trained;

[0308] The second feature dimension reduction training module is used to calculate the probability distribution of multiple dimensions according to the feature extraction data to be trained and the corresponding feature dimension reduction data;

[0309] The third training module of feature dimension reduction is used to calculate information divergence based on the probability distribution of multiple dimensions;

[0310] The fourth feature dimension reduction training module is used to substitute the information divergence and each feature dimension reduction data into a preset inverse reconstruction loss function and derive it to determine the fourth model gradient;

[0311] The fifth feature dimension reduction training module is used to update the model parameters of the initial feature dimension reduction model using the fourth model gradient, determine the intermediate feature dimension reduction model, and count the number of feature dimension reduction model training times in real time;

[0312] The sixth feature dimension reduction training module is used to determine whether the number of feature dimension reduction model training times reaches a preset third training threshold;

[0313] The seventh feature dimension reduction training module is used to use the intermediate feature dimension reduction model as the trained preset target feature dimension reduction model if it is achieved.

[0314] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described devices and modules can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0315] In the several embodiments provided in the present application, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0316] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0317] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features thereof may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for training a classifier model for energy storage system performance evaluation, characterized in that: include: Acquire multiple energy storage system attribute data to be detected, and based on the improved artificial minority class oversampling method, perform data expansion on each of the energy storage system attribute data to be detected, and determine multiple target interpolation samples to be detected; the energy storage system attribute data to be detected is monitoring data of energy storage facilities; Inputting each of the interpolated samples of the target to be detected into a preset target feature extraction model for feature extraction, and outputting a plurality of feature extraction data to be detected; Using a preset target feature dimension reduction model to perform feature dimension reduction on each of the feature extraction data to be detected, to generate a plurality of feature dimension reduction data to be trained; Based on the preset distributed robust optimization strategy, the initial energy storage system performance evaluation classifier model is trained using the feature dimensionality reduction data to be trained to determine the target energy storage system performance evaluation classifier model; The step of performing data expansion on the attribute data of each energy storage system to be detected based on the improved artificial minority class oversampling method to determine multiple interpolation samples of the target to be detected includes: Perform characteristic calculations respectively according to the attribute data of each energy storage system to be detected, and output characteristic parameters corresponding to the attribute data of each energy storage system to be detected; Using each of the characteristic parameters to perform quantum mapping on the corresponding property data of the energy storage system to be detected, and generating initial quantum state data corresponding to each of the property data of the energy storage system to be detected; Performing quantum projection operations on each of the initial quantum state data respectively to generate target quantum state data corresponding to each of the energy storage system attribute data to be detected; Performing quantum state measurement on each of the target quantum state data, outputting the probability density distribution value corresponding to each of the energy storage system attribute data to be detected, and comparing each of the probability density distribution values ​​with a preset probability threshold value; Taking any to-be-detected energy storage system attribute data corresponding to a probability density distribution value greater than the preset probability threshold as target detection energy storage system attribute data; Adaptively performing multiple interpolation on the neighborhood attribute data in the neighborhood of each target detection energy storage system attribute data to generate multiple initial interpolation samples to be detected at the current moment; Determine whether the sum of the number of initial interpolation samples to be detected at the current moment and the number of initial interpolation samples to be detected at the historical moment reaches a preset sample number threshold; If it is reached, the initial interpolation samples to be detected at the current moment and the initial interpolation samples to be detected at the historical moment are respectively used as the corresponding target interpolation samples to be detected.

2. The energy storage system performance evaluation classifier model training method according to claim 1, characterized in that: Also includes: If the sum of the sample quantities of the initial interpolation samples to be detected at the current moment and the initial interpolation samples to be detected at the historical moments does not reach the preset sample quantity threshold, a preset prediction model is used to perform quality assessment on each of the initial interpolation samples to be detected at the current moment, and a sample quality assessment result is output; Substituting the sample quality assessment result into a preset loss function and taking a derivative to determine a first model gradient; Using the first model gradient to update each of the characteristic parameters, and determining a new characteristic parameter corresponding to each of the energy storage system attribute data to be detected; The initial interpolation samples to be detected at the current moment are used as the initial interpolation samples to be detected at the historical moment; Based on the new characteristic parameters and the corresponding property data of the energy storage system to be detected, a plurality of initial interpolation samples to be detected at the current moment are determined, until the sum of the number of initial interpolation samples to be detected at the current moment and the number of initial interpolation samples to be detected at the historical moments reaches the preset sample number threshold; The initial interpolation samples to be detected at the historical moment, the initial interpolation samples to be detected at the current moment and the initial interpolation samples to be detected at the historical moment determined when the sum of the sample quantities reaches the preset sample quantity threshold are respectively used as the corresponding target interpolation samples to be detected.

3. The energy storage system performance evaluation classifier model training method according to claim 1, characterized in that: The preset target feature dimension reduction model includes a target encoder and a target decoder; the step of using the preset target feature dimension reduction model to perform feature dimension reduction on each of the feature extraction data to be detected to generate a plurality of feature dimension reduction data to be trained includes: Using a target encoder to perform low-dimensional conversion on each of the feature extraction data to be detected, so as to generate low-dimensional attribute data corresponding to each of the feature extraction data to be detected; The target decoder reconstructs each of the low-dimensional attribute data and outputs a plurality of feature dimensionality reduction data to be trained.

4. The energy storage system performance evaluation classifier model training method according to claim 1, characterized in that: The step of training the initial energy storage system performance evaluation classifier model based on the preset distribution robust optimization strategy using the feature dimension reduction data to be trained to determine the target energy storage system performance evaluation classifier model includes: Based on the fractional order parameters of the initial energy storage system performance evaluation classifier model, the initial energy storage system performance evaluation classifier model is used to classify each of the feature dimensionality reduction data to be trained to generate a classification evaluation result to be trained; Substituting the evaluation result of the classification to be trained into the preset optimization strategy loss function and deriving it to determine the second model gradient; Calculate the momentum term based on the second model gradient at the historical moment; Substituting the momentum term into a preset learning rate adjustment formula to calculate the model learning rate; The model learning rate and the second model gradient are used to update the model parameters of the initial energy storage system performance evaluation classifier model, determine the intermediate energy storage system performance evaluation classifier model, and count the number of classifier model training times in real time; Determine whether the number of training times of the classifier model reaches a preset first training threshold; If achieved, the intermediate energy storage system performance evaluation classifier model is used as the trained target energy storage system performance evaluation classifier model.

5. The energy storage system performance evaluation classifier model training method according to claim 4 is characterized in that: Also includes: If the number of classifier model training times does not reach the preset first training threshold, updating the fractional order parameters to determine new fractional order parameters; The intermediate energy storage system performance evaluation classifier model is used as a new initial energy storage system performance evaluation classifier model; Based on the new fractional-order parameters, the new initial energy storage system performance evaluation classifier model is used to determine the model learning rate and the second model gradient according to the dimension reduction data of each feature to be trained; The model parameters of the new initial energy storage system performance evaluation classifier model are updated using the model learning rate and the second model gradient, the intermediate energy storage system performance evaluation classifier model is determined, and the classifier model training times are counted in real time until the classifier model training times reaches the preset first training threshold; The intermediate energy storage system performance evaluation classifier model determined when the number of classifier model training times reaches the preset first training threshold is used as the trained target energy storage system performance evaluation classifier model.

6. The energy storage system performance evaluation classifier model training method according to claim 1, characterized in that: After the step of training the initial energy storage system performance evaluation classifier model based on the preset distribution robust optimization strategy using the feature dimension reduction data to be trained to determine the target energy storage system performance evaluation classifier model, the method includes: When a plurality of feature dimensionality reduction data to be detected are received, each of the feature dimensionality reduction data to be detected is input into the target energy storage system performance evaluation classifier model for classification, and a target classification evaluation result is output.

7. The energy storage system performance evaluation classifier model training method according to claim 1, characterized in that: The model training process of the preset target feature extraction model includes: When a plurality of target interpolation samples to be trained are received, each of the target interpolation samples to be trained is input into an initial feature extraction model for feature extraction, and a plurality of feature extraction data are output; Extracting data based on each of the features, calculating a third model gradient; Substituting the third model gradient into a preset update amount calculation formula to calculate the model parameter update amount; The model parameters of the initial feature extraction model are updated using the model parameter update amount, an intermediate feature extraction model is determined, and the number of feature extraction model training times is counted in real time; Determine whether the feature extraction model training times reaches a preset second training threshold; If achieved, the intermediate feature extraction model is used as the trained preset target feature extraction model.

8. The energy storage system performance evaluation classifier model training method according to claim 1, characterized in that: The model training process of the preset target feature dimensionality reduction model includes: When a plurality of feature extraction data to be trained is received, each of the feature extraction data to be trained is input into the initial feature dimension reduction model for feature dimension reduction, and the feature dimension reduction data corresponding to each of the feature extraction data to be trained is output; Calculate multiple dimensional probability distributions based on the feature extraction data to be trained and the corresponding feature dimension reduction data; Calculating information divergence according to the probability distribution of the multiple dimensions; Substituting the information divergence and each of the feature dimension reduction data into a preset inverse reconstruction loss function and taking a derivative to determine a fourth model gradient; The fourth model gradient is used to update the model parameters of the initial feature dimensionality reduction model, determine the intermediate feature dimensionality reduction model, and count the number of feature dimensionality reduction model training times in real time; Determine whether the number of training times of the feature dimensionality reduction model reaches a preset third training threshold; If it is achieved, the intermediate feature dimensionality reduction model is used as the trained preset target feature dimensionality reduction model.

9. A device for training a classifier model for energy storage system performance evaluation, characterized in that: The energy storage system performance evaluation classifier model training method applied to claim 1 comprises: An acquisition module is used to acquire multiple property data of energy storage systems to be detected, and based on an improved artificial minority class oversampling method, perform data expansion on the property data of each energy storage system to be detected, and determine multiple interpolation samples of targets to be detected; A feature extraction module, used for inputting each of the interpolation samples of the target to be detected into a preset target feature extraction model for feature extraction, and outputting a plurality of feature extraction data to be detected; A feature dimension reduction module is used to perform feature dimension reduction on each of the feature extraction data to be detected using a preset target feature dimension reduction model to generate a plurality of feature dimension reduction data to be trained; The training module is used to train the initial energy storage system performance evaluation classifier model based on the preset distributed robust optimization strategy using the dimension reduction data of each feature to be trained, and determine the target energy storage system performance evaluation classifier model.

Images