An electric vehicle lithium battery residual life prediction method based on an XGBoost-LSTM optimization model

Abstract

The application relates to the technical field of electric vehicle lithium batteries, and discloses an electric vehicle lithium battery residual life prediction method based on an XGBoost-LSTM optimization model, which comprises an electric vehicle lithium battery information online collection technology.The electric vehicle lithium battery residual life prediction method based on the XGBoost-LSTM optimization model is used for experiments through actual electric vehicle charging data, can classify based on appropriate data volume, select different and appropriate training models, solves the RUL prediction problem of covering the whole attenuation process of the electric vehicle lithium battery, has remarkable significance for improving the long-term prediction performance of the battery RUL and improving prediction accuracy, and introduces the cumulative feature influence and the meta-reinforcement learning algorithm into the battery RUL prediction, sufficiently excavates the hidden battery health state information and change law in the cumulative features, and meanwhile, the meta-reinforcement learning algorithm has strong small sample learning ability, so that high-precision prediction can be realized in the whole life process of the battery.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of electric vehicle lithium batteries, in particular to a method for predicting the remaining life of an electric vehicle lithium battery based on an XGBoost-LSTM optimization model. BACKGROUND

[0002] An electric vehicle usually adopts a power battery as an energy storage device and source. The existing power battery is usually a chemical battery. After a period of charging and discharging, the maximum capacity of the battery decreases. Generally, the state of health (SOH) of the battery refers to the ratio of the capacity of the battery after being fully charged to the rated capacity. The SOH can reflect the health status of the current battery. Obtaining the SOH of the electric vehicle lithium battery can effectively help to calculate the remaining life of the electric vehicle. The remaining useful life (RUL) of the power battery refers to the number of cycle periods required for the maximum available capacity of the power battery to degrade to a specified failure threshold under a certain charging and discharging system. RUL prediction is a process of predicting and calculating the residual life of the power battery based on the historical data of the power battery by using a certain mathematical method. With the wide application of the power battery in various fields, the RUL prediction technology of the power battery has attracted widespread attention and research.

[0003] The prior art has the following defects and deficiencies:

[0004] 1. The algorithm for early life prediction based on a small amount of data is very important for preventing battery failure. However, the training data of most prediction algorithms accounts for 40%-70% of the total data. Therefore, developing a small sample training algorithm and extracting more effective battery health feature factors are the key and challenge to realize early prediction. The prediction accuracy of the battery life model obtained by fitting the capacity degradation data depends not only on the accuracy of the life model itself, but also on the size of the data amount. Too much data will increase the time cost and complexity of RUL prediction. Insufficient data will lead to inaccurate prediction models. In addition, most models can only have high accuracy and strong feasibility in short-term prediction when predicting the RUL of the lithium battery. The long-term prediction performance needs to be improved by adjusting parameters, data size or combining multiple models. Noise data may be generated during data collection. The detection standard can be improved and the noise data can be processed to improve the prediction accuracy. The accuracy of the prediction model is also due to various influencing factors that are complex and coupled with each other.

[0005] 2、More realistic battery data sets, existing experiments mostly use constant temperature and CC / CV condition cycle battery aging data to establish and verify the algorithm, however, in most battery energy storage applications, the battery will experience dynamic changes in working conditions, in addition, with the wide application of battery energy storage system, the life prediction technology of battery under the extreme condition of fast charging becomes increasingly important. SUMMARY

[0006] In view of the shortcomings of the prior art, the application provides a lithium battery residual life prediction method for electric vehicles based on an XGBoost-LSTM optimization model, which can solve the problem of the existing lithium battery residual life prediction method for electric vehicles based on an XGBoost-LSTM optimization model; the device can perform experiments on actual electric vehicle charging data, classify based on appropriate data volume, select different appropriate training models, solve the RUL prediction problem of electric vehicle lithium battery covering the whole attenuation process, and has significant significance for improving the long-term prediction performance of battery RUL and improving prediction accuracy,

[0007] To achieve the above-mentioned purpose of the lithium battery residual life prediction method for electric vehicles based on an XGBoost-LSTM optimization model, the application provides the following technical scheme: a lithium battery residual life prediction method for electric vehicles based on an XGBoost-LSTM optimization model, comprising the following steps:

[0008] (1) First, based on the electric vehicle lithium battery information online acquisition technology, the charging data of the electric vehicle lithium battery on the charging pile can bus is collected, if the collected electric vehicle lithium battery charging times are more than the estimated boundary times of the experiment, the charging data is sufficient, and the single charging data training set 1 is arranged, if the charging times are less than the boundary times, the relationship between the accumulated charging characteristics and the battery attenuation is not obvious, then the single charging data and the accumulated charging data are selected as the training set 2, the boundary times are the charging times when the weight of the accumulated charging characteristics reaches one order of magnitude of the weight of the single charging characteristics in the XGBoost model;

[0009] (2) Then the SOH health index sequence based on the XGBoost model is constructed, the decision tree adopts the classification and regression tree CART; a new CART is generated according to the training data segment to fit the residual generated by the last CART, and finally all the CARTs are integrated into the final XGBoost integrated model. Weight experiment is carried out to obtain the boundary frequency. The boundary frequency is the charging frequency when the weight of the cumulative charging feature in the constructed XGBoost model meets the order of magnitude of the single charging feature weight. If the charging frequency of the collected electric vehicle lithium battery is more than the estimated boundary frequency, the charging data is sufficient, and the training set 1 including the single charging data is arranged. If the charging frequency is less than the boundary frequency, the relationship between the cumulative charging feature and the battery degradation is not obvious, and the single charging data and the cumulative charging data are selected as the training set 2. When training the two cases of the training set 1 or the training set 2, the SOH sequence 1 or the SOH sequence 2 is obtained respectively.

[0010] (3) Finally, the SOH sequence 1 is substituted into the LSTM model optimized by the gradient descent method, the training parameters are optimized, the model is used to predict the degradation state of the battery capacity, the remaining life prediction result is obtained according to the failure threshold of the battery capacity, the residual life prediction technology of the electric vehicle lithium battery based on meta-reinforcement learning is proposed, the SOH sequence 2 uses the LSTM model to predict the battery degradation, calculates the gradient descent of the loss function, improves the efficiency of updating the parameters, and reduces the required training samples.

[0011] Preferably, in the step (1), the charging pile collects the battery charging current curve, the charging voltage curve, the initial SOC, the environmental temperature, the SOC cycle interval, the battery rated capacity signal data, and the charging frequency, the calendar days, the deep charging proportion, and the high temperature charging proportion record data through the can bus; according to the factors affecting the battery health, the data is classified and divided into single charging data and cumulative charging data.

[0012] Preferably, in the step (2), the XGBoost algorithm is used for feature extraction, the charging data features are converted into leaf nodes in the model according to the idea of gradient boosting decision tree, a decision tree model is established, the CART tree is cyclically increased, the optimal feature weight is calculated, the features are sorted according to the weight as the estimation basis of the battery SOH, the calculated derivative is set to 0, the optimal weight and the minimum loss of each leaf can be solved, the weight experiment is carried out in advance, and the boundary frequency is obtained. The boundary frequency is the charging frequency when the weight of the cumulative charging feature in the constructed XGBoost model meets the order of magnitude of the single charging feature weight.

[0013] Preferably, in the step (3); if it is SOH sequence 1, the general LSTM model training parameters are substituted into the optimal value, the model is used to predict the attenuation state of the battery capacity, and the prediction result of the remaining life is obtained according to the failure threshold of the battery capacity, if it is SOH sequence 2, the gradient descent method with parameters is used to optimize the model, the efficiency of updating the parameters can be improved, the required training samples of the LSTM model are reduced, the meta-learner determines the optimal value of the learning rate and the parameter of the forgetting gate through a series of updates, so that the optimization is quickly carried out, each iteration receives the gradient and loss information from the classifier based on the LSTM meta-learner to update, and the calculated value is returned to the meta-learner to train the meta-learner, and the parameters of the meta-learner are updated again, and the LSTM operation logic based on the gradient descent optimization is as follows:

[0014]

[0015] Wherein, θ is the parameter to be optimized, z f is the forgetting gate signal, z i is the selection gate signal, △ θ l is the gradient direction of the current position.

[0016] Compared with the prior art, the application provides a lithium battery remaining life prediction method for electric vehicles based on an XGBoost-LSTM optimization model, which has the following beneficial effects:

[0017] 1、The lithium battery remaining life prediction method for electric vehicles based on the XGBoost-LSTM optimization model, by establishing the XGBoost-LSTM optimization model to predict the lithium battery remaining life of the electric vehicle, through the actual electric vehicle charging data experiment, and based on the appropriate data amount, different, suitable training model is selected, the RUL prediction problem of the lithium battery of the electric vehicle covering the whole attenuation process is solved, the long-term prediction performance of the battery RUL is improved, the prediction accuracy is improved, and the cumulative feature influence and meta-reinforcement learning algorithm are introduced into the battery RUL prediction, the hidden battery health state information and its change rule in the cumulative feature are fully mined, and the strong small sample learning ability of the meta-reinforcement learning algorithm is also exerted, and high-precision prediction can be realized in the whole life process of the battery. BRIEF DESCRIPTION OF DRAWINGS

[0018] Fig. 1 The flow chart of the lithium battery remaining life prediction of the electric vehicle of the application is shown in the figure;

[0019] Fig. 2 The system flow chart of the XGBoost model for estimating the SOH of the lithium battery of the application is shown in the figure;

[0020] Fig. 3The gradient descent method is used to optimize the LSTM model structure diagram. DETAILED DESCRIPTION

[0021] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work are within the protection scope of the present application.

[0022] Please refer to Figs. 1-3 A lithium battery residual life prediction method for electric vehicles based on an XGBoost-LSTM optimization model, comprising the following steps:

[0023] (1) First, based on the online collection technology of electric vehicle lithium battery information, the charging data of the electric vehicle lithium battery on the charging pile can bus is collected. If the collected charging times of the electric vehicle lithium battery are more than the estimated boundary times in the experiment, the charging data is sufficient, and the single charging data training set 1 is sorted. If the charging times are less than the boundary times, the relationship between the cumulative charging characteristics and the battery attenuation is not obvious, and the single charging data and the cumulative charging data are selected as the training set 2. The boundary times are the charging times when the weight of the cumulative charging characteristics reaches the same order of magnitude as the single charging characteristic weight in the XGBoost model.

[0024] (2) Then, the SOH health index sequence based on the XGBoost model is constructed, and the decision tree adopts the classification and regression tree CART. According to the training data segment, a new CART is generated to fit the residual error generated by the previous CART, and finally all the CARTs are integrated into the final XGBoost integrated model. To conduct a weight experiment, the boundary times are calculated. The boundary times are the charging times when the weight of the cumulative charging characteristics meets the order of magnitude of the single charging characteristic weight in the constructed XGBoost model. If the collected charging times of the electric vehicle lithium battery are more than the estimated boundary times in the experiment, the charging data is sufficient, and the single charging data training set 1 is sorted. If the charging times are less than the boundary times, the relationship between the cumulative charging characteristics and the battery attenuation is not obvious, and the single charging data and the cumulative charging data are selected as the training set 2. When training the training set 1 or the training set 2, the SOH sequence 1 or the SOH sequence 2 is calculated respectively.

[0025] (3) Finally, the SOH sequence 1 is substituted into the LSTM model optimized by the gradient descent method, the parameters are trained to the optimal value, the model is used to predict the attenuation state of the battery capacity, the prediction result of the remaining life is obtained according to the failure threshold of the battery capacity, the residual life prediction technology of the lithium battery of the electric vehicle based on the meta-reinforcement learning is proposed, the SOH sequence 2 is used to predict the battery attenuation by the LSTM model, the gradient descent of the loss function is calculated, and the efficiency of updating the parameters is improved, and the required training samples are reduced.

[0026] In summary, in step (1), the charging pile collects the battery charging current curve, the charging voltage curve, the initial SOC, the environmental temperature, the SOC cycle interval and the battery rated capacity signal data and the charging frequency, the calendar days, the deep charging proportion, the high temperature charging proportion record data through the can bus;According to the factors affecting the battery health, the data is classified and divided into single charging data and cumulative charging data, in step (2), the XGBoost algorithm is used for feature extraction, the charging data features are converted into leaf nodes in the model by referring to the idea of gradient boosting decision tree, a decision tree model is established, the CART tree is increased in a loop, the optimal feature weight is calculated, and the features are sorted according to the weight as the estimation basis of the battery SOH, the derivative is calculated and set to 0, the optimal weight and the minimum loss of each leaf can be solved, the weight experiment is carried out in advance, the boundary frequency is calculated, the boundary frequency is the charging frequency when the weight of the cumulative charging feature in the constructed XGBoost model meets the order of magnitude of the single charging feature weight, in step (3);If SOH sequence 1, substitute into the general LSTM model to train the parameters to the optimal value, use the model to predict the attenuation state of the battery capacity, and obtain the prediction result of the remaining life according to the failure threshold of the battery capacity;If SOH sequence 2, the model is optimized by the gradient descent method with parameters, the efficiency of updating the parameters can be improved, and the required training samples of the LSTM model are reduced, the meta-learner determines the optimal value of the learning rate and the forgetting gate parameters through a series of updates, so that the optimization is quickly carried out, and each iteration receives the gradient and loss information from the classifier based on the LSTM meta-learner to update, and the calculated value is returned to the meta-learner to train the meta-learner, and the parameters of the meta-learner are updated again, and the operation logic of the LSTM based on the gradient descent optimization is as follows:

[0027]

[0028] Wherein θ is the parameter to be optimized, z f Is the forgetting gate signal, z i Is the selection gate signal, △ θ L is the gradient direction of the current position.

[0029] The work of the application uses the specific steps as follows:

[0030] (1) Study the online collection technology of electric vehicle lithium battery information, collect the charging data of electric vehicle lithium battery on the can bus of charging pile, including single charging data and cumulative charging data, and get the initial training set;

[0031] (2) Based on the Xgboost model, the SOH health index sequence is constructed, the weight experiment is carried out, and the boundary number is calculated. The boundary number is the charging number when the weight of the cumulative charging feature in the constructed XGBoost model is in the same order of magnitude as the weight of the single charging feature. If the charging number of the collected electric vehicle lithium battery is more than the estimated boundary number in the experiment, the charging data is sufficient, and the single charging data training set 1 is sorted out; If the charging number is less than the boundary number, the relationship between the cumulative charging feature and the battery attenuation is not obvious, and the single charging data and the cumulative charging data are selected as the training set 2. Train the training set 1 or the training set 2 to obtain SOH sequence 1 or SOH sequence 2;

[0032] (3) The training set 2 is directly substituted into the LSTM model, and the corresponding parameters are trained to the optimal value. The model is used to predict the attenuation state of the battery capacity, and the prediction result of the remaining life is obtained according to the failure threshold of the battery capacity;

[0033] (4) The residual life prediction technology of electric vehicle lithium battery based on meta-reinforcement learning is proposed. The gradient descent method is used to optimize the LSTM model, and the training set 1 is substituted into the cycle training to calculate the gradient descent of the loss function. Through a small amount of data, the attenuation state of the battery capacity is predicted, and the prediction result of the remaining life is calculated;

[0034] For step (1):

[0035] 1) The charging pile collects the battery charging current curve, charging voltage curve, initial SOC, environmental temperature, SOC cycle interval and battery rated capacity signal data, and charging number, calendar days, deep charging proportion, high temperature charging proportion record data through can bus;

[0036] 2) According to the factors affecting the battery health, the data is classified into single charging data and cumulative charging data, and the empty value, noise and invalid value abnormal data are deleted to form the initial data set;

[0037] 3) The single charging data and the cumulative charging data are normalized. The normalization method adopts the following calculation formula:

[0038]

[0039] In the formula, x i is the i th sample data, μ is the mean of all sample data under the feature, max and min represent the maximum and minimum of sample data under the feature respectively;

[0040] For step (2):

[0041] 1) The data classified and processed in the process (1) is obtained;

[0042] 2) XGBoost algorithm is used for feature extraction, referring to the idea of gradient boosting decision tree, a decision tree model is established, and a CART tree is added in a loop,

[0043] The objective function of XGBoost algorithm is:

[0044]

[0045] Where T represents the number of leaf nodes, w j represents the weight of each leaf node, and γ and λ are parameters in the model that control the proportion of different parts, G j is the sum of the first-order derivatives of the loss function, and H j is the sum of the second-order derivatives of the loss function; the minimum loss of the model is:

[0046]

[0047] 3) Use the greedy algorithm to build the tree, and by taking the derivative of w j and setting it equal to 0, the optimal weight of the leaf is calculated using the first and second order derivatives of the loss function, and the optimal weight formula is as follows:

[0048]

[0049] 4) Use the built tree to iteratively optimize the function space. The iteration formula is as follows:

[0050]

[0051] Where f t (x i ) is the expression of the tree, is the f t (x i ) function,

[0052] 5) Recalculate the first and second order derivatives of the loss function;

[0053] 6) Repeat step (1) until K trees are generated, and the current SOH of the battery is output as the training model. The specific process is shown in Fig. 2 ;

[0054] 7) According to the historical charging data, the weight experiment is carried out, and the boundary number is calculated in advance. The boundary number is the number of charges when the cumulative charging feature weight in the built XGBoost model meets the order of magnitude of the single charging feature weight.

[0055] If the number of charging times of the collected electric vehicle lithium battery is more than the estimated boundary number, the charging data is sufficient, and the training set 1 including single charging data and cumulative charging data is arranged; if the number of charging times is less than the boundary number, the relationship between the cumulative charging characteristics and the battery attenuation is not obvious, and the single charging data is selected as the training set 2,

[0056] For step (3):

[0057] 1) The original LSTM model mainly has three stages:

[0058] 1. The forgetting stage, which mainly selectively forgets the input transmitted from the previous node, and the calculated z f is used as the forgetting gate to control the c t-1 of the previous state, to determine the information to be retained or forgotten;

[0059] 2. The selection memory stage, the input of which will be selectively "remembered", mainly for selective memory of the input x t , the current input content is represented by the previously calculated z i , and the selection gate signal is controlled by z

[0060] Through the above two stages, c t transmitted to the next state can be obtained, and the specific formula is as follows:

[0061] c t =z f ec t-1 +z i ez (10)

[0062] Where z f is the forgetting gate signal, and z i is the selection gate signal;

[0063] 3. The output stage, which controls the output of the current state through z 0 , and also scales c 0 obtained in the previous stage, which is changed through the tanh activation function, and the specific formula is as follows:

[0064] h t =z 0 e tanh(c t ) (11)

[0065] y t =σ(W′h t ) (12)

[0066] For step (4):

[0067] The gradient descent method with parameter theta is used to optimize the model, so that the meta-learner can determine the optimal value of the learning rate and the forgetting gate parameter through a series of updates, so that the optimization can be carried out quickly. The gradient and loss information from the classifier are received from the LSTM-based meta-learner through each iteration to update the calculated value, which is returned to the meta-learner to train the meta-learner and update the parameters of the meta-learner again;

[0068] The LSTM model is optimized by the gradient descent method, and the formula of gradient descent and the formula of LSTM are combined. The input c t of the LSTM is updated as follows: t

[0069]

[0070] Where theta is the parameter to be optimized, and θ l is the gradient direction of the current position.

[0071] The update of the parameter not only depends on the current gradient, but also depends on the result obtained by adding the gradient all the time. Therefore, the past gradient and the present gradient are integrated, and an additional LSTM is added to the original optimized LSTM, and the parameter is defined as m. The past gradient is stored through m, so that the past gradient can also participate in the decision of the present gradient. The specific model is shown in Fig. 3

[0072] According to the size of the battery charging data to be predicted, a suitable model is selected. The LSTM model based on meta-reinforcement learning is suitable for training data with less training data, which can ensure high RUL prediction accuracy. The model trained by the training set 2 is suitable for the case of more training data, which can improve the accuracy of the SOH health index sequence. The remaining life of the battery is represented by the number of remaining charges of the battery, and the specific formula is as follows:

[0073]

[0074] Where, c s represents the number of remaining charges of the battery, and c l represents the number of historical charges of the battery

[0075] ​​The electric vehicle lithium battery residual life prediction method based on the XGBoost-LSTM optimization model can perform experiments through actual electric vehicle charging data, can classify based on appropriate data volume, select different appropriate training models, solve the RUL prediction problem of the full attenuation process of the electric vehicle lithium battery, improve the long-term prediction performance of the battery RUL, and has significant significance for improving the prediction accuracy, and introduces the cumulative feature influence and meta-reinforcement learning algorithm into the battery RUL prediction, fully excavates the hidden battery health state information and its change rule in the cumulative feature, and simultaneously plays the strong small sample learning ability of the meta-reinforcement learning algorithm, and can realize high-precision prediction in the whole life process of the battery.

[0076] It should be noted that, in this document, the terms such as first and second are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations, and the term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed or inherent to such a process, method, article or device.

[0077] Although the embodiments of the present application have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present application, and the scope of the present application is defined by the appended claims and their equivalents.

Claims

1. A method for predicting the remaining life of a lithium battery of an electric vehicle based on an XGBoost-LSTM optimization model, characterized by, Comprise the following steps: (1) First, based on the electric vehicle lithium battery information online collection technology, collect the charging data of the electric vehicle lithium battery on the charging pile can bus, if the collected electric vehicle lithium battery charging times is more than the experimental estimated boundary times, the charging data is enough, sort out the single charging data training set 1; if the charging times is less than the boundary times, the relationship between the cumulative charging characteristics and the battery attenuation is not obvious, then select the single charging data and the cumulative charging data as the training set 2, the boundary times is the charging times when the weight of the cumulative charging characteristics reaches the same order of magnitude as the single charging characteristic weight in the XGBoost model; (2) Then build the SOH health index sequence based on the XGBoost model, the decision tree adopts the classification and regression tree CART; according to the training data segment, generate a new CART to fit the residual error generated by the last CART, and finally integrate all the CARTs to obtain the final XGBoost integrated model, and the weight experiment is carried out to obtain the boundary times, the boundary times is the charging times when the weight of the cumulative charging characteristics meets the order of magnitude of the single charging characteristic weight in the constructed XGBoost model, if the collected electric vehicle lithium battery charging times is more than the experimental estimated boundary times, the charging data is enough, sort out the single charging data training set 1; if the charging times is less than the boundary times, the relationship between the cumulative charging characteristics and the battery attenuation is not obvious, then select the single charging data and the cumulative charging data as the training set 2; when training the training set 1 or the training set 2, the SOH sequence 1 or the SOH sequence 2 is obtained respectively; (3) Finally, the SOH sequence 1 is substituted into the LSTM model optimized by gradient descent method, the training parameters are optimized to the optimal value, the model is used to predict the attenuation state of the battery capacity, the prediction result of the remaining life is obtained according to the failure threshold of the battery capacity, the electric vehicle lithium battery remaining life prediction technology based on meta reinforcement learning is proposed, the SOH sequence 2 uses the LSTM model to predict the battery attenuation, calculates the gradient descent of the loss function, improves the efficiency of updating parameters, and reduces the required training samples.

2. The method of claim 1, wherein the method is based on an XGBoost-LSTM optimized model. In step (1), the charging pile collects the battery charging current curve, charging voltage curve, initial SOC, environmental temperature, SOC cycle interval and battery rated capacity signal data and charging times, calendar days, deep charging proportion, high temperature charging proportion record data through the can bus; according to the factors affecting the battery health, the data is classified and divided into single charging data and cumulative charging data.

3. The method of claim 1, wherein the method is based on an XGBoost-LSTM optimized model. In the step (2), the XGBoost algorithm is used for feature extraction, referring to the idea of gradient boosting decision tree, the charging data features are converted into leaf nodes in the model, a decision tree model is established, the CART tree is increased in a loop, the optimal feature weight is calculated, the features are sorted according to the weight as the basis for estimating the battery SOH, and the derivative is set to 0, so that the optimal weight and minimum loss of each leaf can be solved. The weight experiment is carried out in advance, and the boundary frequency is calculated. The boundary frequency is the charging frequency when the cumulative charging feature weight in the constructed XGBoost model meets the order of magnitude of the single charging feature weight.

4. The method of claim 1, wherein the method is based on an XGBoost-LSTM optimized model. In the step (3), if the SOH sequence is 1, the general LSTM model training parameters are substituted into the optimal value, the model is used to predict the attenuation state of the battery capacity, and the prediction result of the remaining life is obtained according to the failure threshold of the battery capacity; if the SOH sequence is 2, the gradient descent method with parameters is used to optimize the model, which can improve the efficiency of updating parameters and reduce the required training samples of the LSTM model. The meta-learner determines the optimal value of the learning rate and the forgetting gate parameters through a series of updates, so that the optimization is quickly performed. Each iteration receives the gradient and loss information from the classifier based on the LSTM meta-learner to update, and the calculated value is returned to the meta-learner to train the meta-learner, and the parameters of the meta-learner are updated again. The LSTM operation logic based on gradient descent optimization is as follows: where θ is the parameter to be optimized, z f is the forget gate signal, z i is the select gate signal, is the gradient direction of the current position.

Images