Method for diagnosing battery state of grid-forming type energy storage power station based on peak regulation stage

Abstract

The invention discloses a peak shaving stage-based battery state diagnosis method for a network-constructed energy storage power station, and the method mainly comprises the steps: 1, collecting the voltage and temperature data of a lithium ion battery in a charging / discharging process under a working condition environment, and constructing an initial data set; 2, on the acquired voltage and temperature time sequence, the sampling frequency reaches 30 seconds based on an average interpolation method, and feature sequence alignment is achieved; 3, adding noise to part of normal sample data by using a Gaussian white noise and baseline drift method to generate an abnormal sample; step 4, normalizing the feature sequence of each sample by using a normalization method; step 5, fusing XGBoost and GAN classification models to complete battery state diagnosis; step 6, according to the model classification error, recalculating the fusion weight of the XGBoost model and the GAN model; and step 7, finishing increment fine adjustment of the model based on voltage and temperature data in the battery charging / discharging process of nearly one month.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of energy storage power station fault diagnosis, and particularly relates to a battery state diagnosis method for grid-connected energy storage power station based on a peak regulation stage. BACKGROUND

[0002] In the field of energy storage, lithium ion batteries are widely used as a clean energy storage device, especially in energy storage power stations, for the storage of new energy and the support of power grids. In practical applications, as the number of overall charging and discharging of the energy storage power station increases, lithium ion batteries will inevitably be damaged. Severe damage will affect the normal operation of the energy storage power station and cause safety accidents. At the same time, due to the differences in materials and manufacturing processes during the manufacturing process of lithium batteries, the chemical reactions of each single battery under different operating conditions also differ. These problems will cause inconsistencies in the same battery cluster or the same battery pack, which will exacerbate the battery degradation rate and increase the risk of thermal runaway in the power station. Therefore, it is necessary to find abnormal batteries in the battery cluster / pack in a timely and accurate manner.

[0003] The present application collects voltage and temperature data of the lithium ion battery from the start of charging / discharging to the end of charging / discharging and then static for 5 minutes, and diagnoses whether the battery is abnormal based on XGBoost and GAN models through data cleaning, data noise adding, feature extraction and other steps. SUMMARY

[0004] In order to overcome the deficiencies in the prior art, the present application provides a battery state diagnosis method for grid-connected energy storage power station based on a peak regulation stage, which realizes real-time monitoring and diagnosis of energy storage lithium batteries in the energy storage system.

[0005] In order to achieve the above application purposes and solve the technical problems, the technical scheme adopted is as follows: A battery state diagnosis method for grid-connected energy storage power station based on a peak regulation stage, comprising the following steps: Step 1: In the actual working condition environment, collect voltage and temperature data of the lithium ion battery from the start of charging / discharging to the end of charging / discharging and then static for 5 minutes, and construct an initial data set; Step 2: On the collected voltage and temperature time series data, use the average interpolation method for interpolation to make the existing time series data collection frequency reach 30 seconds, and align the two time series data; Step 3: Use Gaussian white noise and baseline drift method to add noise to the voltage and temperature time series of part of the normal samples to make them become abnormal samples; Step 4: Use the standard normalization method to normalize the voltage and temperature time series of each sample respectively; Step 5: The state diagnosis of the battery is completed by fusing the classification model XGBoost and the GAN model; Step 6: According to the classification error of the model, the fusion weight of XGBoost and GAN model is recalculated 、 ; Step 7: Based on the voltage and temperature data of the battery charging / discharging process in the past month, the XGBoost and GAN models are respectively completed for incremental fine-tuning.

[0006] Further, in step 1, the voltage and temperature time series data of the lithium ion battery in the energy storage power station from the beginning of charging / discharging to the end of charging / discharging and static for 5 minutes are collected, and the data set composed of the time series of the above two characteristics is used to judge the battery state. The battery state is divided into normal and fault.

[0007] Further, in step 2, the voltage and temperature time series of the sample are interpolated by the average interpolation method, so that the two time series not only meet the 30-second sampling frequency, but also meet the time alignment and equal length. The interpolation calculation method used is as follows: The value of each insertion point is equal to the average of the values of the two points closest to it, and the calculation formula of the ith insertion point is as follows: In the above formula, represents the value closest to the left, represents the value closest to the right.

[0008] Further, in step 3, the Gaussian white noise and baseline drift method are used to add noise to 20% of the normal samples in the charging phase or the normal samples in the discharging phase, so that they become abnormal samples; reduce the problem that the number of abnormal samples is small under actual working conditions, so that the number of abnormal samples and normal samples in the model training is balanced, and the noise adding process is as follows: First, the normal sample time series is added with Gaussian white noise to obtain time series A, and then the baseline drift method is used to add noise to time series A to obtain the final abnormal sample time series; In the Gaussian white noise noise adding method, noise is sampled from a normal distribution with a mean of 0 and a standard deviation of , and then added to the time series of the normal sample, where the standard deviation is generally taken as 1, and the voltage and temperature time series are respectively added with noise. The Gaussian noise calculation formula is as follows: ​​After adding Gaussian white noise, a second noise addition is performed using the baseline drift method. In the baseline drift method, a slowly changing signal is generated and added to the original signal. A low-frequency sine wave signal is selected, and the noise addition process is as follows: In the above formula, sine represents a low-frequency sine wave signal, thus completing the normal sample time series. Transforming time series samples into anomalous samples by adding noise .

[0009] Furthermore, in step 4, the classification model is used to process the resulting two-dimensional matrix. Before further analysis, the matrix needs to be normalized by feature to ensure that the time series values ​​of each feature follow a standard normal distribution, i.e., a mean of 0 and a variance of 1, to prevent overfitting of the model. The standard normalization method is used to normalize each feature dimension of each sample, and the specific calculation method is as follows: In the above formula, n represents the sample length. This represents the mean of the data. The standard deviation of the data represents the value that needs to be standardized, and z represents the standardized value.

[0010] Furthermore, in step 5, the XGBoost classification model and the GAN model are fused to complete the battery state diagnosis. By integrating the output of the integrated model using the fusion weights of the two models, the final state of the battery sample to be tested, i.e., normal or faulty, is obtained. The integration process is as follows: (1) Obtain the classification result of the XGBoost algorithm. The probability that the sample to be detected is a positive sample is: The probability that the sample to be tested is a negative sample is ; (2) Obtain the classification results of the GAN model. The probability that the sample to be detected is a positive sample is... The probability that the sample to be tested is a negative sample is ; (3) The probability that the sample to be tested will eventually be classified as a positive sample or a negative sample can be calculated by the following formulas: Positive sample probability: Negative sample probability: In the above formula, For the fusion weights of the XGBoost model, These are the fusion weights for the GAN model; (4) When When the sample to be tested is positive, it means there is no fault; otherwise, it is negative and there is a fault.

[0011] Furthermore, in step 6, the fusion weights of the XGBoost and GAN models are recalculated based on the model's classification error. , ,include: The fusion weights of the two models are updated according to their error rates within a certain period, so that the model with the smaller error receives a higher weight in the fusion. The threshold for updating the weights is set to 10. After the model completes 10 discriminations, the weights are recalculated according to the update strategy. The corresponding weight calculation method is as follows: Diagnostic error of XGBoost model in 10 iterations for: The diagnostic error of the GAN model in 10 iterations for: In the above formula, Let be the label output by the model in the t-th diagnosis. The true label of the sample in the t-th diagnosis; Based on the error statistics of the two models, the initial weights are calculated using the exponential error-weighted Softmax function, as follows: In the above formula, and These represent the fusion weights of the XGBoost model and the GAN model, respectively.

[0012] Furthermore, in step 7, incremental fine-tuning is performed on the XGBoost and GAN models based on voltage and temperature data from the battery charging / discharging process over the past month, including: The XGBoost and GAN models are optimized incrementally based on the latest sampled data, and the two models are incrementally trained each month based on the latest data. The XGBoost model, while retaining prior knowledge of the original tree structure, adds new trees through gradient boosting to adapt to new data changes. Based on the feature splitting patterns of the historical tree model, it accelerates the fitting of new data. It employs forward step-by-step training, and the method for calculating the classification label of the nth tree is as follows: In the above formula, This represents the cumulative classification values ​​of the first m-1 trees. For the newly generated regression tree, Let be the learning rate, where the optimal new tree is calculated by minimizing the diagnostic error, as shown in the following formula: In the above formula, For the monthly data sample size, Let m be the input features of the m-th regression tree. For the first m-1 tree pairs of samples The predicted value; The GAN model selects the battery charging / discharging data from the past month as its training set each month, and updates the parameters of the GAN model using an incremental learning approach. The objective function of the model is as follows: In the above formula, To combat the losses, Given the total number of samples in the training set, during the training of the GAN model, parameters from the previous month's GAN model are retained as initialization, based on... Train the model and update the parameters in the model.

[0013] Furthermore, step 8 involves training two battery state diagnostic models based on data collected during the charging and discharging phases, respectively: a battery state diagnostic model for the charging phase and a battery state diagnostic model for the discharging phase. Different models are selected according to the phase to which the battery sample data belongs, thus completing the diagnosis.

[0014] By employing the above technical solutions, this invention has the following advantages and positive effects compared with the prior art: This invention discloses a battery state diagnosis method for energy storage power stations based on a peak-shaving phase and dynamic weight fusion model. It collects voltage and temperature data from the start of charging / discharging to a 5-minute rest period after the battery's charge / discharge ends. Interpolation is used to align the data of the two features in the time dimension while maintaining a 30-second sampling frequency. Simultaneously, Gaussian white noise and baseline drift are used to add noise to some normal samples in the time series data, transforming them into anomalous samples. This balances the number of normal and anomalous battery samples in the training samples, enhancing the model's robustness. The final diagnostic result for the battery sample to be tested is obtained by integrating the probabilities of positive and negative samples output by the XGBoost and GAN models using fusion weights. By combining the two algorithms, the false detection rate of the model is reduced. Furthermore, within a specified period, the fusion weights of the models are updated according to the error rates of the two models, reducing the weight of the model with the larger error in the integration. Incremental fine-tuning is used to periodically update the parameters in the XGBoost and GAN models, improving the model's robustness to data and ensuring the accuracy of the model's diagnosis. This invention collects battery data during the charging and discharging phases, and trains battery state diagnostic models for the charging and discharging phases using the same steps, making battery state diagnosis more convenient and faster. Attached Figure Description

[0015] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In the drawings: Figure 1 This is a schematic diagram of the framework of a battery status diagnosis method for energy storage power stations based on a peak shaving stage and dynamic weight fusion model according to the present invention. Detailed Implementation

[0016] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] like Figure 1 As shown in the figure, this embodiment discloses a battery status diagnosis method for grid-connected energy storage power stations based on the peak shaving phase, including the following steps: Step 1: Under actual working conditions, collect voltage and temperature data of the lithium-ion battery from the start of charging / discharging to 5 minutes of rest after charging / discharging, and construct the initial dataset.

[0018] The voltage of a lithium-ion battery gradually increases during charging and immediately drops back to a steady state when charging ends. Similarly, the voltage gradually decreases during discharging and immediately rises again when discharging ends, eventually reaching a steady state. This steady-state voltage is also known as the resting voltage. During both charging and discharging processes, the battery temperature gradually rises until these two states are completed, at which point the temperature slowly returns to a relatively stable level.

[0019] When a lithium battery malfunctions or experiences significant capacity decay, both voltage and temperature will change to some extent during charging and discharging, becoming more pronounced at the end of the charge / discharge cycle. Therefore, this embodiment selects voltage and temperature time series from the start of charging / discharging to a 5-minute rest period after the charge / discharge cycle to diagnose whether the battery cluster has malfunctioned. The voltage and temperature sampling frequency is set to 30 seconds whenever possible. Because voltage changes are opposite during charging and discharging, this embodiment trains separate charging and discharging data models for each of the two scenarios to diagnose the battery status and determine whether it is normal or faulty.

[0020] Step 2: On the collected voltage and temperature time series data, use the average interpolation method to interpolate, so that the existing time series data acquisition frequency reaches 30 seconds, and align the two time series data.

[0021] In actual operating conditions, the sampling frequency for voltage and temperature data may not reach 30 seconds, and the timestamps for voltage and current may not align, hindering subsequent anomaly detection. Therefore, this embodiment employs the average value interpolation method within linear interpolation to achieve a sampling frequency of 30 seconds for existing voltage and temperature data.

[0022] In real-world data, there may be situations where multiple points need to be interpolated between two points. Interpolation is performed by calculating the average value between the two points. The specific calculation method is as follows: Time series Assumption and These are points collected from actual working conditions. If these are the points that need interpolation, then the calculation method for the corresponding values ​​of these three points is as follows: From the above formula, we can also deduce that the value of each insertion point is equal to the average of the values ​​of its two nearest neighbors. Therefore, the value of the i-th insertion point... The calculation formula is as follows: In the above formula, Indicates and The nearest left value, Indicates and The value closest to the right.

[0023] Based on the above method, the battery voltage and temperature data are supplemented, ultimately forming a two-dimensional matrix with row representation features and columns representing the voltage / temperature sampling values ​​per minute. Where m represents the feature dimension and n represents the number of minutes included in the total sampling time, in this embodiment m=2.

[0024] Step 3: Add noise to the voltage and temperature time series of some normal samples using Gaussian white noise and baseline drift methods to turn them into abnormal samples.

[0025] In real-world operating conditions, collecting anomalous data samples is difficult, while collecting normal data samples is relatively easy. Therefore, in this embodiment, Gaussian white noise and baseline drift are used to add noise to 20% of the normal samples during the charging or discharging phases, turning them into anomalous samples. This reduces the problem of a small number of anomalous samples under actual operating conditions, ensuring a balance between the number of anomalous and normal samples during model training. The noise addition process is as follows: First, Gaussian white noise is used to add noise to the normal sample time series to obtain time series A. Then, the baseline drift method is used to add noise to time series A to obtain the final abnormal sample time series.

[0026] In the Gaussian white noise addition method, the mean is 0 and the standard deviation is... Noise is sampled from a normal distribution and then added to the time series of normal samples, where the standard deviation is... Generally, the value is 1. Noise is added separately for the voltage and temperature time series. The Gaussian noise addition calculation formula is as follows: After adding Gaussian white noise, a second noise addition is performed using the baseline drift method. In the baseline drift method, a slowly changing signal is generated and added to the original signal. In this embodiment, a low-frequency sine wave signal is selected, and the noise addition process is as follows: In the above formula, sine represents a low-frequency sine wave signal, thus completing the normal sample time series. Transforming time series samples into anomalous samples by adding noise .

[0027] Step 4: Normalize the voltage and temperature time series of each sample using the standard normalization method.

[0028] Using the classification model to form a two-dimensional matrix Before further analysis, the matrix needs to be normalized by feature to ensure that the time series values ​​of each feature follow a standard normal distribution, i.e., a mean of 0 and a variance of 1, to prevent overfitting of the model. In this embodiment, the standard normalization method is used to normalize each feature dimension of each sample. The specific calculation method is as follows: In the above formula, n represents the sample length. This represents the mean of the data. The standard deviation of the data represents the value that needs to be standardized, and z represents the standardized value.

[0029] Step 5: Combine the classification model XGBoost and the GAN model to complete the battery state diagnosis.

[0030] To improve fault diagnosis accuracy and robustness, based on the ensemble learning approach, XGBoost and GAN models are used to classify the data after feature extraction. The probabilities of positive and negative samples from each algorithm are then ensembled to obtain the final classification result. Both algorithms possess strong modeling capabilities for complex data structures, can handle missing values ​​and nonlinear relationships, and exhibit high accuracy. In this embodiment, the XGBoost and GAN classification models are fused to complete battery status diagnosis. The final status of the battery sample (normal or faulty) is obtained by integrating the output of the ensemble model using the respective fusion weights of the two models. The ensemble process is as follows: (1) Obtain the classification result of the XGBoost algorithm. The probability that the sample to be detected is a positive sample is: The probability that the sample to be tested is a negative sample is ; (2) Obtain the classification results of the GAN model. The probability that the sample to be detected is a positive sample is... The probability that the sample to be tested is a negative sample is ; Generative Adversarial Networks (GANs) are networks that learn data distributions through adversarial training. They consist of a generator G and a discriminator D. The generator G maps random noise z to simulated data G(z), while the discriminator distinguishes between real data x and generated data G(z). In anomaly detection, GANs model a normal data distribution and consider data deviating from this distribution as anomalies.

[0031] The training objective of GAN is a minimax game, as shown in the following equation. The generator attempts to minimize this objective to deceive the discriminator, while the discriminator maximizes the objective to accurately distinguish between real and fake data.

[0032] In the above formula, To combat the losses, The output of the discriminator to the real data x, The probability distribution of the real data. Sampling from real data, Let p be a random noise vector sampled from the noise distribution. The simulated data is mapped by the generator through noise z.

[0033] In this embodiment, the generator adopts a 4-layer fully connected network structure: the input layer has a dimension of 100 (random noise vector), hidden layer 1 has a dimension of 256, hidden layer 2 has a dimension of 512, and the output layer dimension is consistent with the length of the input battery voltage and temperature feature sequences (if the input dimension is 2*150, it corresponds to two features, voltage and temperature, each feature has 150 points, corresponding to 2 hours and 30 minutes of data. Since the charging and discharging time is different each time, the column dimension of the input data will change each time). The activation function of hidden layer 1 and hidden layer 2 is ReLU, and the output layer uses Tanh to map the generated data to a numerical range consistent with the real feature data.

[0034] The discriminator employs a two-layer fully connected network structure: the input layer has a dimension of 2*N (time series of voltage and temperature), the hidden layer has a dimension of 256, and the output layer outputs the discrimination probability with a dimension of 1. The hidden layer activation function is LeakyReLU to alleviate the gradient vanishing problem, and the output layer uses Sigmoid.

[0035] (3) The probability that the sample to be tested will eventually be classified as a positive sample or a negative sample can be calculated by the following formulas: Positive sample probability: Negative sample probability: In the above formula, For the fusion weights of the XGBoost model, These are the fusion weights for the GAN model.

[0036] (4) When When the sample to be tested is positive, it means there is no fault; otherwise, it is negative and there is a fault.

[0037] Step 6: Based on the classification error of the model, recalculate the fusion weights of the XGBoost and GAN models. , .

[0038] To improve diagnostic accuracy, the fusion weights of the two models are updated periodically based on their error rates, allowing the model with the smaller error to receive a higher weight in the fusion process. In this embodiment, the threshold for updating the weights is set to 10. After each 10 discriminations, the weights are recalculated according to the update strategy. The corresponding weight calculation method is as follows: Diagnostic error of XGBoost model in 10 iterations for: The diagnostic error of the GAN model in 10 iterations for: In the above formula, Let be the label output by the model in the t-th diagnosis. represents the true label of the sample in the t-th diagnosis.

[0039] Based on the error statistics of the two models, the initial weights are calculated using the exponential error-weighted Softmax function, as shown in the following method: In the above formula, and These represent the fusion weights of the XGBoost model and the GAN model, respectively.

[0040] Step 7: Based on the voltage and temperature data during the battery charging / discharging process over the past month, perform incremental fine-tuning on the XGBoost and GAN models respectively.

[0041] To continuously improve the robustness and accuracy of the models, this embodiment performs periodic incremental optimization of the XGBoost and GAN models based on the latest sampled data, and incrementally trains the two models each month based on the data of the new month.

[0042] The XGBoost model, while retaining prior knowledge of the original tree structure, adds new trees through gradient boosting to adapt to new data changes. This method accelerates the fitting of new data based on the feature splitting patterns of historical tree models and employs forward step-by-step training. The classification label of the nth tree is calculated as follows: In the above formula, This represents the cumulative classification values ​​of the first m-1 trees. For the newly generated regression tree, Let be the learning rate, where the optimal new tree is calculated by minimizing the diagnostic error, as shown in the following formula: In the above formula, For the monthly data sample size, Let m be the input features of the m-th regression tree. For the first m-1 tree pairs of samples The predicted value; The GAN model selects the battery charging / discharging data from the past month as its training set each month, and updates the parameters of the GAN model using an incremental learning approach. The objective function of the model is as follows: In the above formula, To combat the losses, This represents the total number of samples in the training set. During the training of the GAN model, the parameters from the previous month's GAN model are retained as initialization, based on... Train the model and update the parameters in the model.

[0043] At this point, the incremental fine-tuning of the XGBoost and GAN models is complete.

[0044] Furthermore, step 8 involves training two battery state diagnostic models based on data collected during the charging and discharging phases, respectively: a battery state diagnostic model for the charging phase and a battery state diagnostic model for the discharging phase. Different models are selected according to the phase to which the battery sample data belongs, thus completing the diagnosis.

[0045] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for battery status diagnosis in grid-connected energy storage power stations based on peak-shaving phases, characterized in that, Includes the following steps: Step 1: Under actual working conditions, collect voltage and temperature data of the lithium-ion battery from the start of charging / discharging to 5 minutes of rest after charging / discharging, and construct the initial dataset; Step 2: On the collected voltage and temperature time series data, use the average interpolation method to interpolate, so that the existing time series data acquisition frequency reaches 30 seconds, and align the two time series data. Step 3: Add noise to the voltage and temperature time series of some normal samples using Gaussian white noise and baseline drift methods to turn them into abnormal samples; Step 4: Normalize the voltage and temperature time series of each sample using the standard normalization method; Step 5: Fuse the classification model XGBoost and the GAN model to complete the battery state diagnosis; Step 6: Based on the classification error of the model, recalculate the fusion weights of the XGBoost and GAN models. , ; Step 7: Based on the voltage and temperature data during the battery charging / discharging process over the past month, perform incremental fine-tuning on the XGBoost and GAN models respectively.

2. The battery status diagnosis method for a grid-type energy storage power station based on the peak-shaving stage according to claim 1, characterized in that, In step 1, voltage and temperature time series data of lithium-ion batteries in the energy storage power station are collected from the start of charging / discharging to 5 minutes of rest after charging / discharging. The battery status is determined based on the dataset composed of the time series of the above two features, and the battery status is divided into normal and fault.

3. The battery status diagnosis method for a grid-type energy storage power station based on the peak-shaving stage according to claim 1, characterized in that, In step 2, the voltage and temperature time series of the samples are interpolated using the average interpolation method to ensure that the two time series not only meet the 30-second sampling frequency but also time alignment and equal length. The interpolation calculation method used is as follows: The value of each insertion point is equal to the average of the values ​​of its two nearest neighbors. Therefore, the value of the i-th insertion point is... The calculation formula is as follows: In the above formula, Indicates and The nearest left value, Indicates and The value closest to the right.

4. The battery status diagnosis method for a grid-type energy storage power station based on the peak-shaving stage according to claim 1, characterized in that, In step 3, Gaussian white noise and baseline drift are used to add noise to 20% of the normal samples during the charging or discharging phase, turning them into anomalous samples. This reduces the problem of a small number of anomalous samples under actual working conditions, ensuring a balance between the number of anomalous and normal samples during model training. The noise addition process is as follows: First, Gaussian white noise is used to add noise to the normal sample time series to obtain time series A. Then, the baseline drift method is used to add noise to time series A to obtain the final abnormal sample time series. In the Gaussian white noise addition method, the mean is 0 and the standard deviation is... Noise is sampled from a normal distribution and then added to the time series of normal samples, where the standard deviation is... Generally, the value is 1. Noise is added separately for the voltage and temperature time series. The Gaussian noise addition calculation formula is as follows: After adding Gaussian white noise, a second noise addition is performed using the baseline drift method. In the baseline drift method, a slowly changing signal is generated and added to the original signal. A low-frequency sine wave signal is selected, and the noise addition process is as follows: In the above formula, sine represents a low-frequency sine wave signal, thus completing the normal sample time series. Transforming time series samples into anomalous samples by adding noise .

5. The battery status diagnosis method for a grid-type energy storage power station based on the peak-shaving stage according to claim 1, characterized in that, In step 4, the classification model is used to process the resulting two-dimensional matrix. Before further analysis, the matrix needs to be normalized by feature to ensure that the time series values ​​of each feature follow a standard normal distribution, i.e., a mean of 0 and a variance of 1, to prevent overfitting of the model. The standard normalization method is used to normalize each feature dimension of each sample, and the specific calculation method is as follows: In the above formula, n represents the sample length. This represents the mean of the data. The standard deviation of the data represents the value that needs to be standardized, and z represents the standardized value.

6. The battery status diagnosis method for a grid-type energy storage power station based on the peak-shaving stage according to claim 1, characterized in that, In step 5, the XGBoost classification model and the GAN model are fused to complete the battery state diagnosis. By integrating the output of the integrated model using the fusion weights of the two models, the final state of the battery sample to be tested, i.e., normal or faulty, is obtained. The integration process is as follows: (1) Obtain the classification result of the XGBoost algorithm. The probability that the sample to be detected is a positive sample is: The probability that the sample to be tested is a negative sample is ; (2) Obtain the classification results of the GAN model. The probability that the sample to be detected is a positive sample is... The probability that the sample to be tested is a negative sample is ; (3) The probability that the sample to be tested will eventually be classified as a positive sample or a negative sample can be calculated by the following formulas: Positive sample probability: Negative sample probability: In the above formula, For the fusion weights of the XGBoost model, These are the fusion weights for the GAN model; (4) When When the sample to be tested is positive, it means there is no fault; otherwise, it is negative and there is a fault.

7. The battery status diagnosis method for a grid-type energy storage power station based on the peak-shaving stage according to claim 1, characterized in that, In step 6, the fusion weights of the XGBoost and GAN models are recalculated based on the model's classification error. , ,include: The fusion weights of the two models are updated according to their error rates within a certain period, so that the model with the smaller error receives a higher weight in the fusion. The threshold for updating the weights is set to 10. After the model completes 10 discriminations, the weights are recalculated according to the update strategy. The corresponding weight calculation method is as follows: Diagnostic error of XGBoost model in 10 iterations for: The diagnostic error of the GAN model in 10 iterations for: In the above formula, Let be the label output by the model in the t-th diagnosis. The true label of the sample in the t-th diagnosis; Based on the error statistics of the two models, the initial weights are calculated using the exponential error-weighted Softmax function, as follows: In the above formula, and These represent the fusion weights of the XGBoost model and the GAN model, respectively.

8. The battery status diagnosis method for a grid-type energy storage power station based on the peak-shaving stage according to claim 1, characterized in that, In step 7, incremental fine-tuning of the XGBoost and GAN models is performed based on voltage and temperature data from the battery charging / discharging process over the past month, including: The XGBoost and GAN models are optimized incrementally based on the latest sampled data, and the two models are incrementally trained each month based on the latest data. The XGBoost model, while retaining prior knowledge of the original tree structure, adds new trees through gradient boosting to adapt to new data changes. Based on the feature splitting patterns of the historical tree model, it accelerates the fitting of new data. It employs forward step-by-step training, and the method for calculating the classification label of the nth tree is as follows: In the above formula, This represents the cumulative classification values ​​of the first m-1 trees. For the newly generated regression tree, Let be the learning rate, where the optimal new tree is calculated by minimizing the diagnostic error, as shown in the following formula: In the above formula, For the monthly data sample size, Let m be the input features of the m-th regression tree. For the first m-1 tree pairs of samples The predicted value; The GAN model selects the battery charging / discharging data from the past month as its training set each month, and updates the parameters of the GAN model using an incremental learning approach. The objective function of the model is as follows: In the above formula, To combat the losses, Given the total number of samples in the training set, during the training of the GAN model, parameters from the previous month's GAN model are retained as initialization, based on... Train the model and update the parameters in the model.

9. The battery status diagnosis method for a grid-type energy storage power station based on the peak-shaving stage according to claim 1, characterized in that, The process also includes step 8, which trains two battery state diagnostic models based on data collected during the charging and discharging phases, respectively. These models are the charging phase battery state diagnostic model and the discharging phase battery state diagnostic model. The appropriate model is selected based on the phase to which the battery sample data belongs, thus completing the diagnosis.

Images