Energy storage battery capacity attenuation short-term prediction method and system based on multi-model fusion

Abstract

The invention discloses an energy storage battery capacity attenuation short-term prediction method and system based on multi-model fusion, relates to the technical field of battery management, and aims to solve the problem that short-term capacity attenuation prediction of an energy storage battery is difficult in the prior art. Taking the battery parameter variable data of the to-be-tested time period as input, and obtaining a plurality of machine learning model prediction results based on the output of the plurality of machine learning models; averaging and splicing a plurality of machine learning model prediction results to obtain a preliminary battery energy storage capacity attenuation prediction result; and taking the preliminary battery energy storage capacity attenuation prediction result as input, and obtaining a battery energy storage capacity attenuation short-term prediction result based on linear regression model output. According to the method, the instantaneous characteristics in the operation of the battery are used as input, short-term prediction of capacity fading can be reasonably and accurately carried out, and a new basis is provided for real-time control and strategy adjustment of energy storage.

Description

Technical Field

[0001] The present invention relates to a method and system for short-term prediction of energy storage battery capacity attenuation based on multi-model fusion, belonging to the technical field of battery management. Background Art

[0002] In various systems, battery capacity degradation not only affects the energy supply of the system, but also increases the system's operation and maintenance costs and the frequency of equipment replacement. In particular, in application scenarios that rely on high-efficiency batteries, such as electric vehicles, energy storage systems, and smart phones, the impact of capacity degradation on user experience and economic benefits is more obvious. For some key application scenarios, such as emergency power supplies, smart grids, and medical equipment, the reliability of battery capacity is more important, and the problems caused by degradation may affect the stability and even safety of the system.

[0003] In order to effectively manage the battery life and improve system efficiency, it is critical to accurately predict the trend of battery capacity decay. Especially when the system operation mode changes greatly, accurate prediction of battery short-term capacity decay can help adjust the control strategy in time to avoid unexpected downtime or insufficient energy supply.

[0004] At present, the prediction methods of battery capacity decay are mainly divided into three categories: methods based on physical models, methods based on empirical models, and methods based on data-driven methods. Physical models usually require an in-depth understanding of the complex chemical reactions inside the battery, so they are difficult to build and have limited prediction accuracy; empirical models rely on experimental data, but have poor adaptability to battery decay under specific working conditions. Another method for predicting the life of lithium-ion battery systems uses a simulation model to simulate the battery aging process and realize capacity decay prediction, but because the simulation model parameters are difficult to obtain accurately, the short-term prediction effect is not good. In contrast, data-driven prediction methods use a large amount of historical data, combined with machine learning and data analysis algorithms, can significantly improve prediction accuracy and are applicable to various types of batteries and various usage scenarios.

[0005] These methods have certain advantages in long-term prediction, but are difficult to apply to real-time energy storage control strategy adjustment, especially when the energy storage control operation mode is changeable, the accurate prediction of short-term capacity decay is still difficult. Therefore, there is an urgent need for a data-driven short-term capacity decay prediction method that can combine real-time data to provide more accurate predictions for short-term capacity decay and help the system operate efficiently. Summary of the invention

[0006] The purpose of the present invention is to overcome the deficiencies in the prior art and to provide a method and system for short-term prediction of energy storage battery capacity attenuation based on multi-model fusion. The method and system take the instantaneous characteristics of the battery operation as input, and can reasonably and accurately perform short-term prediction of capacity attenuation, thus providing a new basis for real-time control and strategy adjustment of energy storage.

[0007] To achieve the above object, the present invention is implemented by adopting the following technical solutions:

[0008] On the one hand, the present invention provides a short-term prediction method for energy storage battery capacity attenuation based on multi-model fusion, comprising:

[0009] Obtaining battery parameter variable data for the time period to be tested;

[0010] Taking the battery parameter variable data of the time period to be tested as input, and obtaining multiple machine learning model prediction results based on the outputs of multiple machine learning models;

[0011] The prediction results of multiple machine learning models are averaged and concatenated to obtain the preliminary prediction results of battery energy storage capacity attenuation;

[0012] The preliminary battery energy storage capacity attenuation prediction results are taken as input, and the short-term prediction results of battery energy storage capacity attenuation are obtained based on the linear regression model output.

[0013] Furthermore, the machine learning model includes PGBM, XGBoost and LightGBM, and the model output of PGBM is the probability distribution of the short-term prediction value of battery energy storage capacity decay ,in, represents the predicted mean, Represents the prediction variance, and its expressions are: ; ;

[0014] in, represents the initial forecast mean, represents the mean learning rate, Indicates The fitted mean of the iterations, represents the initial prediction variance, represents the variance learning rate, Indicates The fitted variance of the iteration, Indicates the maximum number of iterations;

[0015] The model output of XGBoost and LightGBM is the short-term prediction value of battery energy storage capacity attenuation;

[0016] The linear regression model is expressed as: ; ;

[0017] in, represents the model output, represents the model input, represents the weight, represents the bias term, express The output probability of represents the output of a machine learning model other than PGBM, Represents machine learning models other than PGBM.

[0018] Furthermore, it also includes pre-training the machine learning model and the linear regression model, and the pre-training method includes: a. Obtain a long-term capacity attenuation data set of batteries of the same type as the battery to be tested; b. Based on the long-term capacity attenuation dataset of the same type of batteries, a time series dataset of energy storage battery capacity attenuation is constructed through feature engineering; c. Divide the energy storage battery capacity decay time series data set into multiple subsets; d. Randomly select a subset as the validation set, and the remaining subsets as the training set; e. Use the training set data to train each machine learning model, calculate the loss function, adjust the machine learning model parameters according to the loss function, and repeat the iterations d to e until the loss function range is within the preset range or the preset number of iterations is reached, thus obtaining a pre-trained machine learning model; f. Take the validation set data as input and obtain multiple prediction results through the output of each machine learning model; g. Take the mean of multiple prediction results and concatenate them as input, train the linear regression model, calculate the loss function, adjust the linear regression model according to the loss function, and repeat f~g continuously to obtain the pre-trained linear regression model.

[0019] Furthermore, the long-term capacity attenuation dataset of the same type of batteries is used to construct a time series dataset of energy storage battery capacity attenuation through feature engineering, including:

[0020] The long-term capacity attenuation data set of the same type of batteries is normalized, and the expression is: ;

[0021] in, Indicates The capacity decay data of the same type of battery, Represents the minimum value in the capacity attenuation data set of the same type of battery, Represents the maximum value in the capacity attenuation data set of the same type of battery, Represents the normalized Capacity decay data of batteries of the same type;

[0022] The outlier data is judged and deleted by the box curve method, and its expression is: UpperLimit=Q3+1.5IQR; LowerLimit=Q1-1.5IQR;

[0023] Among them, UpperLimit represents the upper limit value, Q1 represents the 25th percentile data after all data in the normalized long-term capacity attenuation data set are arranged from small to large, Q3 represents the 75th percentile data after all data in the normalized long-term capacity attenuation data set are arranged from small to large, and IQR represents the difference between Q1 and Q3;

[0024] By linear interpolation, the missing data is filled to obtain the battery parameter variable data within the set time period, and its expression is: Y = ( (X - X1) (Y2 - Y1) / (X2 - X1) ) + Y1;

[0025] Where X represents the missing point index, X1 and X2 represent the indexes of two adjacent data points of X, Y1 represents the value of the X1 data point index, Y2 represents the value of the X2 data point index, and Y represents the value of the X data point index;

[0026] The battery parameter variable data within a set time period constitutes a time series data set of energy storage battery capacity decay, and the battery parameters include current, voltage, temperature, current SOH and current SOC.

[0027] Furthermore, the training process of each iteration of the linear regression model includes:

[0028] The multiple prediction results are averaged and concatenated as input to train the linear regression model;

[0029] The loss function is calculated during the training process. The loss function uses the mean square error, and its expression is: ;

[0030] in, represents the mean square error, represents the number of data samples, represents the true value of the i-th data sample, Represents the predicted value of the i-th data sample;

[0031] According to the loss function, the weights and bias terms in the linear regression model are updated to obtain the linear regression model after this iteration, which is expressed as: ; ;

[0032] in, represents the updated weight, represents the learning rate, Represents the updated bias term.

[0033] Furthermore, when the machine learning model is PGBM, step e specifically includes:

[0034] S1. Initialize the prediction mean and prediction variance, where the prediction mean is initialized to the mean of the training data set and the prediction variance is initialized to a constant;

[0035] S2. Use the training set data to train PGBM and calculate the gradient of the loss function;

[0036] S3, fitting the residuals of the predicted mean and the predicted variance through a weak learner to obtain a fitted mean and a fitted variance;

[0037] S4. Update the predicted mean and predicted variance according to the fitted mean and the fitted variance respectively to obtain updated predicted mean and predicted variance;

[0038] S5. Repeat S2 to S5 until the loss function no longer decreases significantly or reaches the preset number of iterations, and obtain the pre-trained PGBM.

[0039] Furthermore, the loss function adopts negative log-likelihood, which is expressed as: ;

[0040] in, represents the negative log-likelihood loss function, is the variance penalty term, is the prediction error penalty term, represents the true value of the i-th data sample, Indicates i The predicted mean of the data samples, Indicates i The prediction variance of the data samples;

[0041] The gradient of the loss function includes the predicted mean gradient and the predicted variance gradient. The calculation expression of the predicted mean gradient is: ;

[0042] in, Indicates i The data sample is in The predicted mean gradient of the iteration, Indicates i The data sample is in The mean prediction in iterations, Indicates i The data sample is in The prediction variance in iterations;

[0043] The calculation expression of the prediction variance gradient is: ;

[0044] in, Indicates i The data sample is in The predicted variance gradient of the iteration;

[0045] The calculation expression of the fitting mean is: ;

[0046] in, Indicates The fitted mean of the iterations, represents the number of training set data samples, Indicates i The data sample is in The fitted mean of the iterations;

[0047] The calculation expression of the fitting variance is: ;

[0048] in, Indicates The fitted variance of the iteration, Indicates i The data sample is in The fitted variance for iteration .

[0049] The calculation expression of the updated predicted mean is: ;

[0050] in, is the predicted mean after the tth iteration, that is, the updated predicted mean, represents the mean learning rate, For the t -The predicted mean after 1 iteration;

[0051] The calculation expression of the updated prediction variance is: ;

[0052] in, is the prediction variance after the tth iteration, that is, the updated prediction variance, represents the variance learning rate, Indicates t - The prediction variance after 1 iteration.

[0053] Furthermore, before averaging and splicing the prediction results of multiple machine learning models, correction processing is required. The expression of the correction processing is: ; ;

[0054] in, It represents the prediction result after correction. express The deviation correction value at the moment, Represents the formula parameters, express The residual random value at time, represents the residual random value at time t, .

[0055] Furthermore, the calculation expression for the preliminary battery energy storage capacity attenuation prediction result obtained by averaging and concatenating the prediction results of multiple machine learning models is: ;

[0056] Among them, y represents the preliminary battery energy storage capacity attenuation prediction result, np.mean represents the mean operation, w1, w2, and w3 are all weights, Represents the PGBM model The output result of this time is Indicates the XGBoost model The output result of this time is Indicates the LightGBM model Output results of times; ，K Indicates the number of subsets.

[0057] On the other hand, the present invention also provides a short-term prediction system for energy storage battery capacity attenuation based on multi-model fusion, comprising:

[0058] A memory for storing instructions;

[0059] A processor is used to execute the instructions so that the device performs operations to implement the short-term prediction method for energy storage capacity decay based on multi-model fusion as described in any one of the above items.

[0060] Compared with the prior art, the present invention has the following beneficial effects:

[0061] The present invention can combine real-time data, take the instantaneous characteristics of battery operation as input, integrate the prediction results of multiple machine learning models, and obtain the final short-term prediction results of battery capacity through a linear regression model, so that the prediction results are more intuitive and reliable, the operation is simple, and the short-term prediction of capacity attenuation can be reasonably and accurately performed, providing a new basis for real-time control and strategy adjustment of energy storage;

[0062] The PGBM used in the present invention can not only predict the target value, but also quantify the uncertainty of the prediction, providing a reference for the results of model fusion;

[0063] The present invention adopts the ECM method to correct the prediction results, solves the problem of incoherent prediction of time series data, and improves the prediction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 It is a schematic flow chart of a method for short-term prediction of energy storage battery capacity attenuation based on multi-model fusion in an embodiment of the present invention;

[0065] Figure 2 It is a schematic diagram of training and testing of model fusion of the short-term prediction method of energy storage battery capacity attenuation based on multi-model fusion in the present invention. DETAILED DESCRIPTION

[0066] The present invention will be further described below in conjunction with the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and cannot be used to limit the protection scope of the present invention.

[0067] Embodiment 1:

[0068] like Figure 1 As shown, the present invention provides a short-term prediction method for energy storage battery capacity attenuation based on multi-model fusion, including a training phase and a testing phase.

[0069] During the training phase:

[0070] Construct multiple machine learning models and linear regression models. In this embodiment, the machine learning models include PGBM, XGBoost and LightGBM.

[0071] Pre-train multiple machine learning models and linear regression models. The pre-training methods include:

[0072] a. Obtain the capacity decay data set of the same type of battery as the battery to be tested;

[0073] b. Based on the long-term capacity attenuation dataset of the same type of batteries, a time series dataset of energy storage battery capacity attenuation is constructed through feature engineering.

[0074] Since the existing capacity decay data sets of the same type of batteries usually only have long-term capacity decay data, it is necessary to construct a capacity decay time series data set of energy storage batteries using a public data set through a data processing method. In this embodiment, the data processing method is outlier detection and linear interpolation. There are a large number of abnormal data that exceed the data range in the battery data, and there are abnormal fluctuations within the normal range. The normalized box curve method is required to detect and delete the abnormal data, which specifically includes:

[0075] The long-term capacity attenuation data set of the same type of batteries is normalized, and the expression is: ;

[0076] in, Indicates The capacity decay data of the same type of battery, Represents the minimum value in the capacity attenuation data set of the same type of battery, Represents the maximum value in the capacity attenuation data set of the same type of battery, Represents the normalized The capacity decay data of the same type of batteries.

[0077] The outlier data is judged and deleted by the box curve method, and its expression is: UpperLimit=Q3+1.5IQR; LowerLimit=Q1-1.5IQR;

[0078] Among them, UpperLimit represents the upper limit value, Q1 represents the 25th percentile data after all data in the normalized long-term capacity attenuation data set are arranged from small to large, Q3 represents the 75th percentile data after all data in the normalized long-term capacity attenuation data set are arranged from small to large, and IQR represents the difference between Q1 and Q3.

[0079] By linear interpolation, the missing data is filled to obtain the battery parameter variable data within the set time period. The battery parameters reflect the characteristic data of the battery itself and the battery operation parameter data. In this embodiment, they include current, voltage, temperature, current SOH and current SOC, etc., and their expressions are: Y = ( (X - X1) (Y2 - Y1) / (X2 - X1) ) + Y1;

[0080] Where X represents the missing point index, X1 and X2 represent the indexes of two adjacent data points of X, Y1 represents the value of the X1 data point index, Y2 represents the value of the X2 data point index, and Y represents the value of the X data point index.

[0081] The battery parameter variable data within the set time period is the energy storage battery capacity decay time series data set.

[0082] c. In this embodiment, the training adopts the K-fold cross-validation method, that is, the energy storage battery capacity decay time series data set is divided into K subsets of similar size, and the data of each subset is as consistent as possible with the capacity decay data set of the same type of battery, such as Figure 2 As shown, the K value in this embodiment is set to 5.

[0083] d. Perform K cycles, randomly select a subset as the validation set in each cycle, and the remaining subsets as the training set;

[0084] e. Use the training set data to train each machine learning model and calculate the loss function. Adjust the machine learning model parameters according to the loss function, and repeat the iterations d~e until the loss function range is within the preset range or reaches the preset number of iterations to obtain the pre-trained machine learning model.

[0085] The following is a detailed description of PGBM:

[0086] The calculation process of PGBM is based on the traditional gradient boosting decision tree and introduces the idea of ​​probabilistic modeling, so that the model can not only predict the value of the target variable, but also quantify the uncertainty of the prediction results.

[0087] Assume that the data set consists of n samples and the input data is , the corresponding label is , Represents battery parameter variable data, Representation and Corresponding battery capacity attenuation data.

[0088] For each input , the output of PGBM includes not only the predicted value , including the uncertainty of the predicted value It is usually assumed that the model output follows a Gaussian distribution, which is expressed as: ;

[0089] in, represents the PGBM output, represents the predicted mean, represents the prediction variance.

[0090] In this way, PGBM can provide confidence intervals for prediction results.

[0091] In PGBM, the loss function not only needs to minimize the difference between the predicted value and the true value, but also needs to consider the prediction variance. In this embodiment, the loss function uses negative log-likelihood, which can be expressed as: ;

[0092] in, represents the negative log-likelihood loss function, is the variance penalty term, which is used to constrain the prediction uncertainty of the model; is the prediction error penalty term, which is used to measure the difference between the predicted value and the true value. represents the true value of the i-th data sample, Indicates i The predicted mean of the data samples, Indicates i The prediction variance for a data sample.

[0093] By optimizing the above loss function, the model is able to fit both the predicted value and the uncertainty of the prediction.

[0094] The pre-training process of PGBM specifically includes:

[0095] Initialize the predicted mean and predicted variance, where the predicted mean is initialized to the mean of the training data set and the predicted variance is initialized to a constant, such as 1, respectively. , .

[0096] In each round of iteration, the PGBM is trained using the training set data and the gradient of the loss function is calculated. The gradient of the loss function includes the predicted mean gradient and the predicted variance gradient. The predicted mean gradient and the predicted variance gradient are used to update the predictions of the mean and variance, respectively.

[0097] The calculation expression of the predicted mean gradient is: ;

[0098] in, Indicates i The data sample is in The predicted mean gradient of the iteration, Indicates i The data sample is in The mean prediction in iterations, Indicates i The data sample is in The prediction variance in iterations, , Indicates the maximum number of iterations;

[0099] The calculation expression of the predicted variance gradient is: ;

[0100] in, Indicates i The data sample is in The predicted variance gradient at iteration .

[0101] The residuals of the predicted mean and predicted variance are fitted by a weak learner (such as a decision tree) to obtain the fitted mean and fitted variance. The calculation expression of the fitted mean is: ;

[0102] in, represents the fitted mean of the t-th iteration, represents the number of training set data samples, Indicates i The data sample is in The fitted mean of the iterations;

[0103] The calculation expression of the fitted variance is: ;

[0104] in, represents the fitted variance of the t-th iteration, Indicates i The data sample is in The fitted variance for iteration .

[0105] The predicted mean and predicted variance are updated according to the fitted mean and fitted variance respectively to obtain the updated predicted mean and predicted variance. The calculation expression of the updated predicted mean is: ;

[0106] in, is the predicted mean after the tth iteration, that is, the updated predicted mean, represents the mean learning rate, For the t -The predicted mean after 1 iteration;

[0107] The calculation expression of the updated prediction variance is: ;

[0108] in, is the prediction variance after the tth iteration, that is, the updated prediction variance, represents the variance learning rate, Indicates t - The prediction variance after 1 iteration.

[0109] Repeat the iterations until the loss function value no longer decreases significantly or reaches the preset number of iterations, and the pre-trained PGBM is obtained.

[0110] The prediction output of the trained PGBM is the probability distribution of the predicted value , where the predicted mean , prediction variance The expression is: ; ;

[0111] in, represents the initial forecast mean, represents the mean learning rate, Indicates The fitted mean of the iterations, represents the initial prediction variance, represents the variance learning rate, Indicates The fitted variance of the iteration, Indicates the maximum number of iterations.

[0112] This information can be used to construct confidence intervals. For example, a 95% confidence interval can be expressed as:

[0113] ;

[0114] In this way, PGBM not only gives a point forecast, but also quantifies the uncertainty of the forecast, thereby providing a more comprehensive forecast result.

[0115] f. Take the validation set data as input and obtain multiple prediction results through the output of each machine learning model;

[0116] g. Correct the multiple prediction results. The expression of the correction process is: ; ;

[0117] in, It represents the prediction result after correction. represents the deviation correction value at time t-1, Represents the formula parameters, represents the residual random value at time t-1, represents the residual random value at time t, .

[0118] The prediction results after correction are averaged and spliced. The expression of average splicing is: ;

[0119] Among them, y represents the preliminary battery energy storage capacity attenuation prediction result, np.mean represents the mean operation, w1, w2, and w3 are all weights, Represents the PGBM model The output result of this time is Indicates the XGBoost model The output result of this time is Indicates the LightGBM model Output results of times; ，K Represents the number of subsets. In this embodiment, K=5.

[0120] The prediction results after taking the mean concatenation are used as input to train the linear regression model and calculate the loss function. The linear regression model is adjusted according to the loss function, and f~g is repeated continuously until the preset number of iterations is reached to obtain the pre-trained linear regression model.

[0121] The training process of the linear regression model specifically includes:

[0122] Assume that there is a linear relationship between the input variable X and the output variable y, which is expressed as: ; ;

[0123] in, represents the model output, represents the model input, represents the weight, represents the bias term, represents the PGBM model output, represents the probability output of the PGBM model, represents the output of a machine learning model other than PGBM, Represents machine learning models other than PGBM.

[0124] The goal of the training process is to find the right , The value enables the model to minimize the error between the predicted value and the true value. The mean square error (MSE) is used as the loss function to calculate the error between the true value and the predicted value. The expression is: ;

[0125] in, represents the mean square error, represents the number of data samples, represents the true value of the i-th data sample, Represents the predicted value of the i-th data sample.

[0126] The weights and bias terms in the linear regression model are updated according to the loss function, and its expression is: ; ;

[0127] in, represents the updated weight, represents the learning rate, Represents the updated bias term.

[0128] Testing phase:

[0129] Obtain the battery parameter variable data for the time period to be tested.

[0130] The battery parameter variable data of the time period to be tested is used as input, and multiple machine learning model prediction results are obtained based on the outputs of multiple machine learning models.

[0131] The prediction results of multiple machine learning models are averaged and concatenated to obtain the preliminary battery energy storage capacity attenuation prediction results.

[0132] The preliminary battery energy storage capacity attenuation prediction results are taken as input, and the short-term prediction results of battery energy storage capacity attenuation are obtained based on the linear regression model output.

[0133] Embodiment 2:

[0134] Based on Example 1, this embodiment provides a short-term prediction system for energy storage battery capacity decay based on multi-model fusion, including:

[0135] Memory, used to store instructions.

[0136] A processor is used to execute the instructions so that the device performs operations to implement the short-term prediction method for energy storage capacity decay based on multi-model fusion as described in Example 1.

[0137] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A short-term prediction method for energy storage battery capacity attenuation based on multi-model fusion, characterized in that: include: Obtaining battery parameter variable data for the time period to be tested; Taking the battery parameter variable data of the time period to be tested as input, and obtaining multiple machine learning model prediction results based on the outputs of multiple machine learning models; The prediction results of multiple machine learning models are averaged and concatenated to obtain the preliminary prediction results of battery energy storage capacity attenuation; The preliminary battery energy storage capacity attenuation prediction results are taken as input, and the short-term prediction results of battery energy storage capacity attenuation are obtained based on the linear regression model output.

2. The short-term prediction method for energy storage battery capacity attenuation based on multi-model fusion according to claim 1 is characterized in that: The machine learning model includes PGBM, XGBoost and LightGBM. The model output of PGBM is the probability distribution of the short-term prediction value of battery energy storage capacity decay. ,in, represents the predicted mean, Represents the prediction variance, and its expressions are: ； ； in, represents the initial forecast mean, represents the mean learning rate, Indicates The fitted mean of the iterations, represents the initial prediction variance, represents the variance learning rate, Indicates The fitted variance of the iteration, Indicates the maximum number of iterations; The model output of XGBoost and LightGBM is the short-term prediction value of battery energy storage capacity attenuation; The linear regression model is expressed as: ； ； in, represents the model output, represents the model input, represents the weight, represents the bias term, express The output probability of represents the output of a machine learning model other than PGBM, Represents machine learning models other than PGBM.

3. The short-term prediction method for energy storage battery capacity attenuation based on multi-model fusion according to claim 1 is characterized in that: It also includes pre-training the machine learning model and the linear regression model, and the pre-training method includes: a. Obtain a long-term capacity attenuation data set of batteries of the same type as the battery to be tested; b. Based on the long-term capacity attenuation dataset of the same type of batteries, a time series dataset of energy storage battery capacity attenuation is constructed through feature engineering; c. Divide the energy storage battery capacity decay time series data set into multiple subsets; d. Randomly select a subset as the validation set, and the remaining subsets as the training set; e. Use the training set data to train each machine learning model, calculate the loss function, adjust the machine learning model parameters according to the loss function, and repeat the iterations d to e until the loss function range is within the preset range or the preset number of iterations is reached, thus obtaining a pre-trained machine learning model; f. Take the validation set data as input and obtain multiple prediction results through the output of each machine learning model; g. Take the mean of multiple prediction results and concatenate them as input, train the linear regression model, calculate the loss function, adjust the linear regression model according to the loss function, and repeat f~g continuously to obtain the pre-trained linear regression model.

4. The short-term prediction method for energy storage battery capacity attenuation based on multi-model fusion according to claim 3 is characterized in that: The long-term capacity attenuation dataset based on the same type of batteries is used to construct a time series dataset of energy storage battery capacity attenuation through feature engineering, including: The long-term capacity attenuation data set of the same type of batteries is normalized, and the expression is: ； in, Indicates The capacity decay data of the same type of battery, Represents the minimum value in the capacity attenuation data set of the same type of battery, Represents the maximum value in the capacity attenuation data set of the same type of battery, Represents the normalized Capacity decay data of batteries of the same type; The outlier data is judged and deleted by the box curve method, and its expression is: UpperLimit=Q3+1.5IQR; LowerLimit=Q1-1.5IQR; Among them, UpperLimit represents the upper limit value, Q1 represents the 25th percentile data after all data in the normalized long-term capacity attenuation data set are arranged from small to large, Q3 represents the 75th percentile data after all data in the normalized long-term capacity attenuation data set are arranged from small to large, and IQR represents the difference between Q1 and Q3; By linear interpolation, the missing data is filled to obtain the battery parameter variable data within the set time period, and its expression is: Y = ( (X - X1) (Y2 - Y1) / (X2 - X1) ) + Y1; Where X represents the missing point index, X1 and X2 represent the indexes of two adjacent data points of X, Y1 represents the value of the X1 data point index, Y2 represents the value of the X2 data point index, and Y represents the value of the X data point index; The battery parameter variable data within a set time period constitutes a time series data set of energy storage battery capacity decay, and the battery parameters include current, voltage, temperature, current SOH and current SOC.

5. The short-term prediction method for energy storage battery capacity attenuation based on multi-model fusion according to claim 3 is characterized in that: The training process of each iteration of the linear regression model includes: The multiple prediction results are averaged and concatenated as input to train the linear regression model; The loss function is calculated during the training process. The loss function uses the mean square error, and its expression is: ； in, represents the mean square error, represents the number of data samples, represents the true value of the i-th data sample, Represents the predicted value of the i-th data sample; According to the loss function, the weights and bias terms in the linear regression model are updated to obtain the linear regression model after this iteration, which is expressed as: ； ； in, represents the updated weight, represents the learning rate, Represents the updated bias term.

6. The short-term prediction method for energy storage battery capacity attenuation based on multi-model fusion according to claim 3 is characterized in that: When the machine learning model is PGBM, step e specifically includes: S1. Initialize the prediction mean and prediction variance, where the prediction mean is initialized to the mean of the training data set and the prediction variance is initialized to a constant; S2. Use the training set data to train PGBM and calculate the gradient of the loss function; S3, fitting the residuals of the predicted mean and the predicted variance through a weak learner to obtain a fitted mean and a fitted variance; S4. Update the predicted mean and predicted variance according to the fitted mean and the fitted variance respectively to obtain updated predicted mean and predicted variance; S5. Repeat S2 to S5 until the loss function no longer decreases significantly or reaches the preset number of iterations, and obtain the pre-trained PGBM.

7. The short-term prediction method for energy storage battery capacity attenuation based on multi-model fusion according to claim 6 is characterized in that: The loss function uses negative log-likelihood, which is expressed as: ； in, represents the negative log-likelihood loss function, is the variance penalty term, is the prediction error penalty term, represents the true value of the i-th data sample, Indicates i The predicted mean of the data samples, Indicates i The prediction variance of the data samples; The gradient of the loss function includes the predicted mean gradient and the predicted variance gradient. The calculation expression of the predicted mean gradient is: ； in, Indicates i The data sample is in The predicted mean gradient of the iteration, Indicates i The data sample is in The mean prediction in iterations, Indicates i The data sample is in The prediction variance in iterations; The calculation expression of the prediction variance gradient is: ； in, Indicates i The data sample is in The predicted variance gradient of the iteration; The calculation expression of the fitting mean is: ； in, Indicates The fitted mean of the iterations, represents the number of training set data samples, Indicates i The data sample is in The fitted mean of the iterations; The calculation expression of the fitting variance is: ； in, Indicates The fitted variance of the iteration, Indicates i The data sample is in The fitted variance of the iteration; The calculation expression of the updated predicted mean is: ； in, is the predicted mean after the tth iteration, that is, the updated predicted mean, represents the mean learning rate, For the t -The predicted mean after 1 iteration; The calculation expression of the updated prediction variance is: ； in, is the prediction variance after the tth iteration, that is, the updated prediction variance, represents the variance learning rate, Indicates t - The prediction variance after 1 iteration.

8. The short-term prediction method for energy storage battery capacity attenuation based on multi-model fusion according to claim 1 is characterized in that: Before averaging and concatenating the prediction results of multiple machine learning models, correction processing is required. The expression of the correction processing is: ； ； in, It represents the prediction result after correction. express The deviation correction value at the moment, Represents the formula parameters, express The residual random value at time, express The residual random value at time, .

9. The short-term prediction method for energy storage battery capacity attenuation based on multi-model fusion according to claim 1 is characterized in that: The calculation expression for the preliminary battery energy storage capacity attenuation prediction result obtained by averaging and splicing the prediction results of multiple machine learning models is: ； Among them, y represents the preliminary battery energy storage capacity attenuation prediction result, np.mean represents the mean operation, w1, w2, and w3 are all weights, Represents the PGBM model The output result of this time is Indicates the XGBoost model The output result of this time is Indicates the LightGBM model The output result of this time is , K Indicates the number of subsets.

10. A short-term prediction system for energy storage battery capacity decay based on multi-model fusion, characterized in that: include: A memory for storing instructions; A processor is used to execute the instruction so that the device performs the operation of implementing the short-term prediction method of energy storage battery capacity decay based on multi-model fusion as described in any one of claims 1 to 9.

Images