Battery system operation data generation method based on generative pre-training architecture

Abstract

The invention relates to the technical field of battery systems, in particular to a battery system operation data generation method based on a generative pre-training architecture, which comprises the following steps: collecting a small amount of real battery system operation data; cleaning the collected small amount of real battery system operation data based on a preset data preprocessing strategy, and removing abnormal values and noise data; according to the cleaned real battery system operation data, based on a preset feature extraction strategy, key operation features related to the battery system operation state and performance are extracted; sampling noise vectors are randomly acquired from the noise distribution vectors, and model parameters corresponding to a generator and a discriminator are optimized based on the generator and the discriminator of a pre-constructed generative adversarial network model and a preset adversarial training optimization strategy; and outputting a large amount of battery system operation data based on the noise distribution vector according to the optimized generator.

Description

Technical Field

[0001] The present invention relates to the technical field of battery systems, and particularly to a method for generating battery system operation data based on a generative pre-training architecture. Background Art

[0002] In the field of operation monitoring and analysis of battery systems, a large amount of high-quality battery operation data plays a decisive role in comprehensively understanding the operation state of the battery system and achieving accurate performance evaluation and fault prediction. However, at present, the acquisition of battery operation data faces many challenges, which seriously restricts the development of this field, mainly reflected in the following three aspects.

[0003] Traditional battery operation data collection relies on deploying sensor devices in the actually operating battery system. This not only requires purchasing a large number of high-precision sensors, increasing the hardware cost, but also requires regular calibration and maintenance of the sensors, consuming a large amount of human and material resources, resulting in high data collection costs.

[0004] Limited by the actual application scenario, the existing collected data is difficult to cover the operation states of the battery system under various complex conditions. The lack of such data diversity leads to the fact that the models built based on these data cannot comprehensively learn the operation rules of the battery system, reducing the accuracy and reliability of the models in analyzing battery performance and predicting faults under different working conditions, resulting in insufficient data collection diversity.

[0005] The operation data of the battery system usually needs to be gradually accumulated during a long-term operation process. Taking the research and development of new battery technologies as an example, in order to obtain enough data to evaluate the long-term performance and lifespan of the battery, it may take several months or even years for actual tests.

[0006] In summary, the existing methods for obtaining battery operation data have problems such as high cost, insufficient diversity, and long acquisition cycle, and are difficult to meet the data requirements for the operation monitoring and analysis of battery systems.

[0007] Based on this, there is an urgent need for a method for generating battery system operation data based on a generative pre-training architecture, which can solve the problems of high cost, insufficient diversity, and long acquisition cycle in the existing acquisition of battery operation data, so as to achieve low cost, diversity, and high efficiency in the acquisition of battery operation data. Summary of the Invention

[0008] One of the purposes of the present invention is to provide a method for generating battery system operation data based on a generative pre-training architecture, which can solve the problems of high cost, insufficient diversity, and long acquisition cycle in the existing acquisition of battery operation data, so as to achieve low cost, diversity, and high efficiency in the acquisition of battery operation data.

[0009] To achieve the above object, a method for generating battery system operation data based on a generative pre-training architecture is provided, including the following steps:

[0010] S1. Collect a small amount of real battery system operation data;

[0011] S2. Based on a preset data preprocessing strategy, clean the collected small amount of real battery system operation data to remove outliers and noise data;

[0012] S3. According to the cleaned real battery system operation data, based on a preset feature extraction strategy, extract key operation features related to the operation state and performance of the battery system;

[0013] S4. Randomly obtain a sampling noise vector from the noise distribution vector, and based on the generator and discriminator of a pre-constructed generative adversarial network model and a preset adversarial training optimization strategy, optimize the model parameters corresponding to the generator and discriminator, and output the optimized generator and discriminator;

[0014] S5. According to the optimized generator, based on the noise distribution vector, output a large amount of battery system operation data.

[0015] The technical principle and effect of this solution: In this solution, data is the basis of the entire method. Real battery system operation data can truthfully reflect the state and performance of the battery during actual operation. This data can be obtained from the battery management system (BMS), sensors, or historical records. However, in actual scenarios, obtaining a large amount of real data may face problems such as high cost and long time, so this solution only collects a small amount of real data.

[0016] Then, the real collected data often contains outliers and noise data, which will have an adverse impact on subsequent feature extraction and model training. Therefore, it is necessary to clean the data according to the preset data preprocessing strategy.

[0017] The cleaned data contains a large amount of information, but not all information is valuable for the analysis of the operation state and performance of the battery system. Through the preset feature extraction strategy, key operation features closely related to the operation state and performance of the battery system can be extracted. For example, features such as voltage change rate and average voltage are extracted from voltage data; features such as charge and discharge current peaks and average current are extracted from current data. These key features can more accurately describe the operation state of the battery and provide more effective inputs for subsequent model training.

[0018] After that, the generative adversarial network (GAN) consists of a generator and a discriminator, which oppose and promote each other. The role of the generator is to randomly obtain a sampled noise vector from the noise distribution vector and then generate simulated battery system operation data; the discriminator is responsible for judging whether the input data is real or generated. During the adversarial training process, the generator and the discriminator continuously optimize their model parameters. Specifically, the goal of the generator is to generate simulated data that can deceive the discriminator, while the goal of the discriminator is to accurately distinguish between real data and generated data. The preset adversarial training optimization strategy (such as the gradient descent method) updates their model parameters according to the loss functions of the generator and the discriminator, and finally outputs the optimized generator and discriminator.

[0019] Finally, the generator optimized through adversarial training can learn the distribution characteristics of the real battery system operation data. Based on the noise distribution vector, the generator can output a large amount of simulated battery system operation data. These generated data are similar to the real data in distribution and can be used for subsequent battery system research, simulation, and testing, etc.

[0020] Traditional methods rely on deploying a large number of sensors on the battery system to collect data, while this solution only needs to collect a small amount of real battery system operation data. Compared with traditional technologies that need to deploy sensors at multiple key nodes, this solution significantly reduces the number of sensors used. At the same time, there is no need to maintain and calibrate these sensors for a long time, reducing the human and material costs.

[0021] Traditional methods not only consume a large amount of manpower for data collection, but also need to perform complex sorting on the collected data. This solution generates a large amount of data through the generative adversarial network, eliminating the need for continuous on-site data collection and cumbersome data sorting work, further saving human and time costs.

[0022] Due to the limitations of the actual application scenario, it is difficult for traditional data acquisition methods to comprehensively cover the operating states of the battery system under various complex working conditions and environmental conditions. In this solution, the generative adversarial network can learn the distribution characteristics of real data during the training process and generate a large amount of diverse simulated data based on the noise distribution vector. These data can simulate the operating states of the battery system under different conditions such as temperature, humidity, charge and discharge rate, etc., effectively making up for the problem of insufficient diversity of traditional data.

[0023] Traditional data acquisition methods take a long time to accumulate a certain amount of data, while this solution can quickly generate a large amount of simulated data by collecting a small amount of real data and using the generative adversarial network, greatly shortening the data acquisition cycle. That is, it can solve the problems of high cost, insufficient diversity, and long acquisition cycle in obtaining existing battery operation data, thus achieving low cost, diversity, and high efficiency in obtaining battery operation data.

[0024] Furthermore, the preset data preprocessing strategy is as follows:

[0025] According to the parameter values corresponding to the various parameters in the real battery system operation data, based on the preset calculation formula for the kernel density estimate value, calculate the kernel density estimate values corresponding to the various parameters;

[0026] The preset calculation formula for the kernel density estimate value is:

[0027]

[0028] In the formula, f(x) is the independent variable value for which the density is to be estimated corresponding to a certain parameter, n is the total number of a certain parameter, h is the bandwidth parameter, K is the kernel function, which is a non - negative function with an integral equal to 1, and x i is the i - th parameter value in a certain parameter;

[0029] According to the kernel density estimate values corresponding to the various parameters, based on the kernel density thresholds corresponding to the respective parameters, determine whether the corresponding kernel density estimate value is lower than the corresponding kernel density threshold. If so, the corresponding parameter value is an outlier; otherwise, it is a normal value.

[0030] Beneficial effects: This strategy uses the kernel density estimate value to judge the outliers in the data, and can more accurately capture the data points with a large difference from the overall data distribution. Kernel density estimation can estimate the probability density of each data point according to the actual distribution of the data. When the kernel density estimate value of a certain parameter value is lower than the corresponding kernel density threshold, it indicates that the probability of this data point appearing in the overall data distribution is relatively low and is very likely to be an outlier. Compared with some simple statistical methods (such as the mean or standard deviation method), this method can better adapt to the complexity of the data distribution and improve the accuracy of outlier identification. This means that while removing outliers, it can better retain the true characteristics and distribution form of the data. For the battery system operation data, retaining these characteristics is very important for accurately analyzing the performance and state of the battery subsequently.

[0031] Furthermore, the preset adversarial training optimization strategy is as follows:

[0032] S40. When training the generator and the discriminator, initialize the generator parameters and discriminator parameters corresponding to the generator and the discriminator in the pre - constructed generative adversarial network model;

[0033] S41. Fix the generator parameters corresponding to the initialized generator, input the obtained sampled noise vector into the fixed generator, output the corresponding first generated battery system operation data, and input the first generated battery system operation data and the real battery system operation data into the discriminator to output the corresponding first judgment result; Based on the first generated battery system operation data, the real battery system operation data, and the first judgment result, calculate the first loss value of the corresponding first generated battery system operation data and the real battery system operation data according to the preset generator loss function calculation formula;

[0034] S42. According to the calculated first loss value of the first generated battery system operation data and the real battery system operation data, update the discriminator parameters corresponding to the discriminator based on the preset parameter update rule, and re - execute S41 until the iteration requirement is met, and output the discriminator corresponding to the final discriminator parameters;

[0035] S43. Fix the discriminator corresponding to the output final discriminator parameters, input the obtained sampled noise vector into the initialized generator, output the corresponding second generated battery system operation data, and input the second battery system operation data and the real battery system operation data into the discriminator corresponding to the final discriminator parameters to output the corresponding second judgment result; Based on the second battery system operation data, the real battery system operation data, and the second judgment result, calculate the second loss value of the corresponding second battery system operation data and the real battery system operation data according to the preset generator loss function calculation formula;

[0036] S44. According to the calculated second loss value of the second generated battery system operation data and the real battery system operation data, update the discriminator parameters corresponding to the generator based on the preset parameter update rule, and re - execute S43 until the iteration requirement is met, and output the generator corresponding to the final generator parameters. At this time, the training of the generator and the discriminator is completed.

[0037] Beneficial effects: Through the clear steps (S40 to S44), the training process of the generator and the discriminator is carefully divided. Starting from the initial parameters, first fix the generator to train the discriminator, and then fix the discriminator to train the generator, alternating. This orderly training method makes the training process easy to understand and implement, can effectively guide the training of the generative adversarial network model, reduces the complexity and error probability of model training, and helps developers better control the training process.

[0038] During the training process, by alternately fixing one party and updating the parameters of the other party, the generator and the discriminator can be gradually optimized in the process of competing with each other. When the discriminator is trained with the generator fixed, the discriminator can focus on improving its ability to distinguish between real data and generated data; when the generator is trained with the discriminator fixed, the generator can improve the quality of the generated data according to the feedback of the discriminator. This adversarial training mechanism helps the model converge to a better state faster, and at the same time improves the stability of model training, avoiding training failures or instabilities caused by one party being too strong or too weak. For example, in the generation of battery system operation data, it can make the data generated by the generator closer to the real data distribution, and the discriminator can also more accurately judge the authenticity of the data.

[0039] Furthermore, the preset adversarial training optimization strategy further includes:

[0040] During the training process of the generator and the discriminator, after each time the generator and the discriminator output corresponding output data, based on the preset dynamic training coefficient adjustment strategy, adjust the dynamic training coefficients α and β corresponding to the generator and the discriminator at this time;

[0041] The preset dynamic training coefficient adjustment strategy is:

[0042] Judge the magnitude relationship between the iteration number C corresponding to this iteration and the first preset iteration number A and the second preset iteration number B;

[0043] When C ≤ A, the dynamic training coefficients α and β corresponding to the generator and the discriminator at this time are the first fixed value and the second fixed value respectively;

[0044] When A < C ≤ B, obtain the loss value of the generator and the loss value of the discriminator corresponding to this iteration in real time, and extract the feature vectors of the generated battery system operation data and the real battery system operation data corresponding to this iteration in the intermediate layer of the discriminator;

[0045] According to the two obtained feature vectors, calculate the corresponding feature matching degree based on the preset feature matching degree calculation formula;

[0046] The preset feature matching degree calculation formula is:

[0047]

[0048] In the formula, FM is the feature matching degree, D mid (G(z)) is the feature vector of the generated battery system operation data in the intermediate layer of the discriminator; D mid (x) is the feature vector of the real battery system operation data in the intermediate layer of the discriminator;

[0049] Based on the loss value of the generator, the loss value of the discriminator, and the feature matching degree corresponding to this iteration, calculate the dynamic training coefficients α and β corresponding to the generator and discriminator in this iteration based on the preset dynamic training coefficient calculation formula;

[0050] Preset dynamic training coefficient calculation formula:

[0051]

[0052] In the formula, α t is the dynamic training coefficient of the generator corresponding to the iteration number t, and β t is the dynamic training coefficient of the discriminator corresponding to the iteration number t, σ is the Sigmoid activation function, W α , W β is the corresponding parameter matrix, b α , b β is the corresponding bias term, the parameter matrix W α , W β and the bias term b α , b β are obtained through offline pre-training; is the corresponding loss value of the generator, is the loss value of the discriminator;

[0053] When C > B, the adjustment ranges of the dynamic training coefficients α and β corresponding to the generator and discriminator at this time are within ±0.1.

[0054] Beneficial effects: In the initial stage of training, when the iteration number C ≤ A, set the dynamic training coefficients α and β of the generator and discriminator to the first fixed value and the second fixed value respectively. This method provides a stable training direction for the model at the beginning of training. Because in the initial stage of training, the model has limited understanding of the data distribution and features. Fixed training coefficients can avoid excessive fluctuations in parameter adjustment of the model, enabling the generator and discriminator to start learning the basic features of the data in a relatively stable environment and laying a solid foundation for subsequent training. The setting of fixed coefficients can enable the model to quickly converge to a relatively reasonable state in the initial stage of training.

[0055] During the training of population A < C ≤ B, the loss value of the generator, the loss value of the discriminator, and the feature matching degree between the generated data and the real data in the intermediate layer of the discriminator are obtained in real time, and the training coefficients α and β are dynamically adjusted according to these factors. This multi-factor real-time adjustment mechanism enables the model to be flexibly optimized according to the real-time state of training. One of the training difficulties of the generative adversarial network is to maintain the balance between the generator and the discriminator. By dynamically adjusting the training coefficients, this balance can be effectively maintained. When the discriminator is too powerful and can easily distinguish between real data and generated data, the loss value of the generator will increase. At this time, α can be increased to enhance the ability of the generator; conversely, when the data generated by the generator is too realistic and the discriminator has difficulty distinguishing, β can be increased to enhance the discrimination ability of the discriminator. This adjustment of dynamic balance helps to improve the training effect of the model and makes the data generated by the generator closer to the real data.

[0056] In the later stage of training, the adjustment range of the dynamic training coefficients of the generator and the discriminator is limited within ±0.1. This measure can prevent the model from having excessive parameter adjustments when approaching convergence, thus ensuring the stable convergence of the model. Description of the Drawings

[0057] Figure 1 It is a flowchart of the method for generating battery system operation data based on the generative pre-training architecture in the first embodiment of the present invention. Detailed Embodiments

[0058] The following is a further detailed description through specific embodiments:

[0059] Embodiment 1

[0060] A method for generating battery system operation data based on the generative pre-training architecture is basically as Figure 1 shown, and includes the following steps:

[0061] S1. Collect a small amount of real battery system operation data, where the real battery system operation data includes voltage, current, temperature, and charge and discharge state parameters;

[0062] S2. Based on a preset data preprocessing strategy, clean the collected small amount of real battery system operation data to remove outliers and noise data;

[0063] The preset data preprocessing strategy is:

[0064] According to the parameter values corresponding to the respective parameters in the real battery system operation data, calculate the kernel density estimation values corresponding to the respective parameters based on a preset kernel density estimation value calculation formula;

[0065] The preset kernel density estimation value calculation formula is:

[0066]

[0067] Wherein, f(x) is the independent variable value of the density to be estimated corresponding to a certain parameter, n is the total number of a certain parameter, h is the bandwidth parameter, K is the kernel function, which is a non - negative function with an integral equal to 1, and x i is the value of the i - th parameter among a certain parameter;

[0068] According to the kernel density estimates corresponding to each parameter, based on the kernel density thresholds corresponding to each parameter, it is judged whether the corresponding kernel density estimate is lower than the corresponding kernel density threshold. If so, the corresponding parameter value is an outlier, otherwise it is a normal value. In this embodiment, the noise data is removed by methods such as smoothing filtering to remove the random disturbance in the data.

[0069] S3. According to the cleaned real - time battery system operation data, based on a preset feature extraction strategy, extract the key operation features related to the operation state and performance of the battery system;

[0070] The feature extraction strategy is as follows:

[0071] According to the cleaned real - time battery system operation data, based on a preset similarity calculation formula, calculate the correlation between each data and the operation state and performance of the battery system;

[0072] The preset similarity calculation formula is:

[0073]

[0074] Wherein, R XY is the correlation degree between data X and the operation state performance index Y of the battery system; x(j) is the i - th data value of data X, and y(j) is the i - th data value of the operation state performance index Y of the battery system, are the corresponding sample means respectively, ε is the corresponding weight value, and R is the corresponding basic correlation degree;

[0075] According to the correlation between each data and the operation state and performance of the battery system, arrange them in descending order of correlation, and select the data corresponding to the preset ranking value as the corresponding key operation features.

[0076] S4. Randomly obtain a sampling noise vector from the noise distribution vector, and based on the generator and discriminator of the pre - constructed generative adversarial network model, and a preset adversarial training optimization strategy, optimize the model parameters corresponding to the generator and discriminator, and output the optimized generator and discriminator;

[0077] The preset adversarial training optimization strategy is as follows:

[0078] S40. When training the generator and the discriminator, initialize the generator parameters and discriminator parameters corresponding to the generator and the discriminator in the pre-constructed generative adversarial network model.

[0079] S41. Fix the generator parameters corresponding to the initialized generator, input the obtained sampled noise vector into the fixed generator, output the corresponding first generated battery system operation data, and input the first generated battery system operation data and the real battery system operation data into the discriminator to output the corresponding first judgment result; based on the first generated battery system operation data, the real battery system operation data, and the first judgment result, calculate the first loss value of the corresponding first generated battery system operation data and the real battery system operation data according to the preset generator loss function calculation formula.

[0080] The preset generator loss function calculation formula is as follows:

[0081]

[0082] In the formula, x is the real data, p(x) is the distribution of the real data, D(x) is the judgment result of the discriminator on the real data, is the loss value of the real data, is the loss value corresponding to the generated data. G(z) is the data generated by the generator, D(G(z)) is the judgment result of the discriminator on the generated data, and p(z) is the distribution of the random noise z.

[0083] S42. According to the calculated first loss value of the first generated battery system operation data and the real battery system operation data, update the discriminator parameters corresponding to the discriminator based on the preset parameter update rule, and re-execute S41 until the iteration requirement is met, and output the discriminator corresponding to the final discriminator parameters.

[0084] In this embodiment, the preset parameter update rule is: update the discriminator parameter θ corresponding to the discriminator by the gradient descent method D :

[0085]

[0086] where η is the learning rate.

[0087] S43. Fix the discriminator corresponding to the final discriminator parameters, input the obtained sampled noise vector into the initialized generator, output the corresponding second generated battery system operation data, and input the second battery system operation data and the real battery system operation data into the discriminator corresponding to the final discriminator parameters to output the corresponding second judgment result; calculate the second loss value of the generator at this time based on the preset generator loss function calculation formula according to the second battery system operation data, the real battery system operation data, and the second judgment result.

[0088] The second loss value L of the generator at this time G is:

[0089] L G =-E z～p(z) [logD(G(z))]

[0090] S44. Update the discriminator parameters corresponding to the generator based on the calculated second loss value of the second generated battery system operation data and the real battery system operation data according to the preset parameter update rule, and re - execute S43 until the iteration requirement is met, and output the generator corresponding to the final generator parameters. At this time, the training of the generator and the discriminator is completed. The preset parameter update rule is to update the parameters θ of the generator by the gradient descent method G :

[0091]

[0092] The preset adversarial training optimization strategy further includes:

[0093] During the training of the generator and the discriminator, after the generator and the discriminator output the corresponding output data each time, adjust the dynamic training coefficients α and β corresponding to the generator and the discriminator at this time based on the preset dynamic training coefficient adjustment strategy.

[0094] The preset dynamic training coefficient adjustment strategy is:

[0095] Judge the size relationship between the iteration number C corresponding to this iteration and the first preset iteration number A and the second preset iteration number B;

[0096] When C ≤ A, the dynamic training coefficients α and β corresponding to the generator and the discriminator at this time are the first fixed value and the second fixed value respectively;

[0097] When A < C ≤ B, obtain the loss value of the generator and the loss value of the discriminator corresponding to this iteration in real - time, and extract the feature vectors of the generated battery system operation data and the real battery system operation data corresponding to this iteration in the middle layer of the discriminator;

[0098] According to the two acquired feature vectors, the corresponding feature matching degree is calculated based on a preset feature matching degree calculation formula;

[0099] The preset feature matching degree calculation formula is:

[0100]

[0101] In the formula, FM is the feature matching degree, D mid (G(z)) is the feature vector of the generated battery system operation data in the middle layer of the discriminator; D mid (x) is the feature vector of the real battery system operation data in the middle layer of the discriminator;

[0102] According to the loss value of the generator, the loss value of the discriminator and the feature matching degree corresponding to this iteration, the dynamic training coefficients α and β of the generator and discriminator corresponding to this iteration are calculated based on the preset dynamic training coefficient calculation formula;

[0103] The preset dynamic training coefficient calculation formula is:

[0104]

[0105] In the formula, α t is the dynamic training coefficient of the generator corresponding to the number of iterations t, β t is the dynamic training coefficient of the discriminator corresponding to the number of iterations t, σ is the Sigmoid activation function, W α , W β is the corresponding parameter matrix, b α , b β is the corresponding bias term, the parameter matrix W α , W β and the bias term b α , b β Obtained through offline pre-training; is the loss value of the corresponding generator, is the loss value of the discriminator; in this embodiment, when |L G -L D |>δ triggers compensation training.

[0106] When C>B, the adjustment range of the dynamic training coefficients α and β corresponding to the generator and discriminator is within ±0.1.

[0107] S5. According to the optimized generator, based on the noise distribution vector, a large amount of battery system operation data is output. In this embodiment, the generated data is verified to remove abnormal values that do not conform to actual physical laws to ensure the accuracy and reliability of the data.

[0108] Certainly, in this embodiment, during the training of the generator and the discriminator, various index data in the data quality evaluation index system corresponding to the generated data will also be continuously monitored;

[0109] When the following situations are detected, corresponding adjustments are triggered: when the fidelity index SSIM+≥0.85 and the diversity index IVR<0.7, start the diversity enhancement protocol (β* = 0.6, α+ = 0.3). When the stability index (σ)>0.15 lasts for 3 cycles, freeze the generator parameters and perform discriminator reinforcement training (β = 2.0, number of iterations × 3). When SSIM+∈[0.6,0.8) and IVR>0.75, enable the hybrid training mode (α = 1.0, β = 1.0, inject 5% historical generated samples). Among them, the fidelity index: adopt the improved structural similarity (SSIM+), and fuse the frequency-domain wavelet transform system; the diversity index: calculate the within-class variance ratio IVR of the generated samples in the Inception-v3 feature space; the stability index: monitor the sliding standard deviation of the L_G / L_D ratio in the last K iterations.

[0110] The above are only embodiments of the present invention. Specific structures and common knowledge such as characteristics that are well known in the art are not described in detail here. Those of ordinary skill in the art know all the common technical knowledge in the technical field to which the invention belongs before the application date or the priority date, can know all the prior arts in this field, and have the ability to apply the conventional experimental means before this date. Those of ordinary skill in the art can, under the inspiration given in this application, improve and implement this solution in combination with their own abilities. Some typical well-known structures or well-known methods should not be an obstacle for those of ordinary skill in the art to implement this application. It should be noted that for those skilled in the art, without departing from the structure of the present invention, several deformations and improvements can be made, and these should also be regarded as the protection scope of the present invention, and these will not affect the implementation effect of the present invention and the practicality of the patent. The protection scope required by this application should be based on the content of its claims, and the specific implementation manners described in the specification can be used to explain the content of the claims.

Claims

1. A method for generating battery system operation data based on a generative pre-training architecture, characterized in that: It includes the following steps: S1. Collect a small amount of real battery system operation data; S2. For the small amount of real battery system operation data collected, based on a preset data preprocessing strategy, clean the real battery system operation data to remove outliers and noise data; S3. According to the cleaned real battery system operation data, based on a preset feature extraction strategy, extract key operation features related to the operation state and performance of the battery system; S4. Randomly obtain a sampling noise vector from the noise distribution vector, and based on the generator and discriminator of a pre-constructed generative adversarial network model, and a preset adversarial training optimization strategy, optimize the model parameters corresponding to the generator and discriminator, and output the optimized generator and discriminator; S5. According to the optimized generator, based on the noise distribution vector, output a large amount of battery system operation data.

2. The method for generating battery system operation data based on a generative pre-training architecture according to claim 1, wherein: The preset data preprocessing strategy is: According to the parameter values corresponding to each parameter in the real battery system operation data, based on a preset kernel density estimation value calculation formula, calculate the kernel density estimation value corresponding to each parameter; The preset kernel density estimation value calculation formula is: Wherein, f(x) is the value of the independent variable of the density to be estimated corresponding to a certain parameter, n is the total number of a certain parameter, h is the bandwidth parameter, K is the kernel function, which is a non - negative function with an integral equal to 1, and x i is the value of the i - th parameter among a certain parameter; According to the kernel density estimation values corresponding to each parameter, based on the kernel density threshold corresponding to each parameter, judge whether the corresponding kernel density estimation value is lower than the corresponding kernel density threshold. If so, the corresponding parameter value is an outlier, otherwise it is a normal value.

3. A method for generating battery system operation data based on a generative pre-training architecture according to claim 2, characterized in that: The preset adversarial training optimization strategy is: S40. When training the generator and discriminator, initialize the generator parameters and discriminator parameters corresponding to the generator and discriminator in the pre-constructed generative adversarial network model; S41. Fix the generator parameters corresponding to the initialized generator, input the obtained sampling noise vector into the fixed generator, output the corresponding first generated battery system operation data, and input the first generated battery system operation data and the real battery system operation data into the discriminator, and output the corresponding first judgment result; Based on the first generated battery system operation data, the real battery system operation data, and the first judgment result, based on a preset generator loss function calculation formula, calculate the first loss value of the corresponding first generated battery system operation data and the real battery system operation data at this time; S42. According to the calculated first loss value of the first generated battery system operation data and the real battery system operation data, based on a preset parameter update rule, update the discriminator parameters corresponding to the discriminator, and re-execute S41 until the iteration requirement is met, and output the discriminator corresponding to the final discriminator parameters; S43. Fix the discriminator corresponding to the output final discriminator parameters, input the obtained sampling noise vector into the initialized generator, output the corresponding second generated battery system operation data, and input the second battery system operation data and the real battery system operation data into the discriminator corresponding to the final discriminator parameters, and output the corresponding second judgment result; According to the second battery system operation data, the real battery system operation data, and the second judgment result, based on the preset calculation formula of the generator loss function, calculate the second loss value corresponding to the second battery system operation data and the real battery system operation data at this time; S44. Based on the calculated second loss value of the generated battery system operation data and the real battery system operation data, update the discriminator parameters corresponding to the generator according to the preset parameter update rule, and re-execute S43 until the iteration requirement is met, and output the generator corresponding to the final generator parameters. At this time, the training of the generator and the discriminator is completed.

4. A method for generating battery system operation data based on a generative pre-training architecture according to claim 3, characterized in that: The preset adversarial training optimization strategy further includes: During the training of the generator and the discriminator, after each generator and discriminator output corresponding output data, based on the preset dynamic training coefficient adjustment strategy, adjust the dynamic training coefficients α and β corresponding to the generator and the discriminator at this time; The preset dynamic training coefficient adjustment strategy is: Judge the magnitude relationship between the iteration number C corresponding to this iteration and the first preset iteration number A and the second preset iteration number B; When C ≤ A, the dynamic training coefficients α and β corresponding to the generator and the discriminator at this time are the first fixed value and the second fixed value respectively; When A < C ≤ B, obtain the loss value of the generator and the loss value of the discriminator corresponding to this iteration in real time, and extract the feature vectors of the generated battery system operation data and the real battery system operation data at the intermediate layer of the discriminator corresponding to this iteration; Based on the two obtained feature vectors, calculate the corresponding feature matching degree according to the preset feature matching degree calculation formula; The preset feature matching degree calculation formula is: where FM is the feature matching degree, D mid (G(z)) is the feature vector of the generated battery system operation data in the middle layer of the discriminator; D mid (x) is the feature vector of the real battery system operation data in the middle layer of the discriminator; Based on the loss value of the generator, the loss value of the discriminator, and the feature matching degree corresponding to this iteration, calculate the dynamic training coefficients α and β corresponding to the generator and the discriminator corresponding to this iteration according to the preset dynamic training coefficient calculation formula; The preset dynamic training coefficient calculation formula: where α t is the dynamic training coefficient of the generator corresponding to the iteration number t, β t is the dynamic training coefficient of the discriminator corresponding to the iteration number t, σ is the Sigmoid activation function, W α , W β are the corresponding parameter matrices, b α , b β are the corresponding bias terms, the parameter matrices W α , W β and the bias terms b α , b β are obtained through offline pre-training; is the loss value of the corresponding generator, is the loss value of the discriminator; When C > B, the adjustment range of the dynamic training coefficients α and β corresponding to the generator and the discriminator at this time is within ±0.1.

Images