A probabilistic prediction method for the remaining life of lithium-ion batteries

Through the probability prediction method based on quantile regression random forests, the uncertainty and complexity problems in lithium-ion battery life prediction are solved, and accurate battery life interval prediction is provided, which improves battery usage safety and user experience.

CN115436814BInactive Publication Date: 2024-11-08JILIN UNIVERSITY
View PDF 0 Cites 2 Cited by

Patent Information

Application Number
CN202211268274.X
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-10-17
Publication Date
2024-11-08
Estimated Expiration
Not applicable · inactive patent

AI Technical Summary

Technical Problem

The existing technology is difficult to accurately predict the remaining effective life of lithium-ion batteries, especially when considering uncertainty factors and complexity in the actual application environment, resulting in large differences between the prediction results and the actual situation, and the inability to provide effective maintenance suggestions, which affects the impact of Battery safety.

Method used

Using a probability prediction method based on quantile regression random forest (QRRF), by obtaining the aging test data of lithium-ion batteries, selecting appropriate feature subsets and model hyperparameters, constructing a probability prediction model for the remaining life of the battery, and outputting different confidence probabilities The interval distribution below provides more accurate battery life prediction.

Benefits of technology

This method can effectively utilize actual data, consider the unmodeled reaction mechanism and unperceived impact, provide a more accurate battery life prediction interval, improve battery usage safety and user experience, simplify the model structure, and only need one training Obtain prediction results for all target quantiles.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN115436814B_ABST
    Figure CN115436814B_ABST
Patent Text Reader

Abstract

The invention discloses a probability prediction method for the remaining life of a lithium-ion battery, which relates to the technical field of lithium-ion batteries and comprises the following steps: acquiring aging test data of the lithium-ion battery and preprocessing the test data; extracting features from the preprocessed data and performing feature selection using a pinballloss loss function; selecting model hyperparameters using a method of minimizing the pinballloss loss function or minimizing a coverage probability deviation index, establishing a quantile regression random forest model QRRF based on the model hyperparameters, and training the model; inputting the test data into a battery remaining life probability prediction model, outputting a target quantile prediction value of the battery remaining life, and obtaining a probability prediction interval of the battery remaining life.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of lithium ion batteries, and in particular to a probability prediction method for the remaining life of a lithium ion battery. Background Art

[0002] As one of the core components of pure electric vehicles, the safety of power batteries has always been criticized by users. Most electric vehicle power batteries use lithium-ion batteries. In the safety research of electric vehicles, the remaining useful life (RUL) is one of the research focuses. RUL refers to the number of cycles that the battery undergoes before the battery capacity decays to 80% of the rated capacity during the charge and discharge cycle.

[0003] Most studies on RUL use deterministic methods to predict their expected values, that is, to predict the remaining useful life of lithium-ion batteries with a certain value through algorithms. However, due to the measurement errors of historical data and the influence of unperceived / unpredictable factors in the environment and working conditions, there will inevitably be deviations between the actual remaining life of the battery and the predicted value. Only through the evaluation indicators of deterministic estimation, such as root mean square error, mean absolute percentage error, etc., it is impossible to fully reflect the characteristics of the role of this uncertainty on individuals, which makes it difficult for users and vehicle manufacturers to effectively use deterministic methods to specify clear battery maintenance or replacement recommendations to improve the safety of battery use. Some studies have considered introducing probabilistic prediction models into the prediction of remaining effective life, but in order to simplify the algorithm, they directly assume that the probability distribution of the remaining effective life is a normal distribution, which is inconsistent with the actual situation. A large number of experiments have shown that the probability distribution of the remaining effective life of the battery is an asymmetric distribution.

[0004] The main methods of life prediction are model method and data-driven method. The model method involves physical and chemical reactions inside the battery, and the workload of parameter identification is large. It is difficult to consider the actual use environment, so it is rarely used in practice. The current focus of data-driven models for lithium-ion battery life prediction is mostly on the expected value of battery life, that is, deterministic prediction. However, due to the influence of uncertainty factors, there must be significant differences between the actual battery life and the prediction, which greatly limits the application of the remaining effective life prediction results. For example, the patent "A lithium-ion battery life estimation method based on relaxation time distribution" provides a lithium-ion battery life estimation method that associates the relaxation time distribution function of AC impedance with life, but this method can only obtain deterministic estimation results and cannot achieve probabilistic prediction. There are also individual patents that propose a probabilistic prediction method for the remaining effective life of the battery, such as: The patent "A method for probabilistic prediction of the remaining life of lithium-ion batteries based on a gray model" provides a lithium-ion battery life probability estimation method that combines a gray model with a Markov chain and a related vector machine, but the historical data of this method only uses the battery capacity decay sequence as a feature. Considering the high complexity of the uncertainty influencing factors of the battery decay process, the lack of its input features will affect the accuracy of the probabilistic prediction. In fact, probabilistic prediction has become a trend in technological progress in recent years. For example, the patent "Wind power probabilistic prediction method based on quantile regression forest and variable bandwidth evaluation" provides a wind power probabilistic prediction method based on quantile regression forest and variable bandwidth evaluation. However, this method is only applied to wind power prediction, and the correlation between wind power in different seasons, months and dates is weak. In the actual application of the remaining effective life of the battery, the historical data of the battery's attenuation process has a strong correlation with the prediction target. How to consider the correlation between the data throughout its life cycle and select reasonable data features are key issues that have not yet been resolved in the remaining effective life prediction technology of lithium-ion batteries. Summary of the invention

[0005] In view of the above problems, the present invention proposes a probability prediction method based on Quantile Regression Random Forest (QRRF), which predicts the interval distribution of RUL under different confidence probabilities when the battery runs to a specified number of cycles based on any historical data of the battery life cycle (greater than 100 cycles), providing more accurate and comprehensive information for battery maintenance or replacement, which is of great significance to improving the safety of electric vehicle batteries and improving the user experience.

[0006] The data-driven method used in the present invention can directly apply actual data, so as to better consider the reaction mechanisms not modeled in the model and the unperceived impacts in practical applications. Although it weakens the theoretical interpretability, it is more conducive to efficient engineering practice applications.

[0007] The present invention provides a method for predicting the probability of the remaining life of a lithium-ion battery, which specifically comprises the following steps:

[0008] Obtain aging test data of lithium-ion batteries;

[0009] Determine the candidate feature set consisting of the following features and their statistical analysis quantities: the discharge capacity difference ΔQ at the corresponding voltages of two cycles, the internal resistance difference of two cycles, the polynomial fitting parameters of the curve of discharge voltage changing with cycle number, and the discharge capacity of a specific cycle;

[0010] Use the pinball loss function to select a feature subset from the candidate feature set, and input the feature subset into the quantile regression random forest model, output the pinball loss of the quantile prediction result, and determine the feature subset that minimizes the pinball loss as the final selected feature set;

[0011] The hyperparameters of the quantile regression random forest model are selected by minimizing the pinball loss function or minimizing the coverage probability deviation index method, and the finally selected feature set is input into the quantile regression random forest model. The hyperparameters are changed to minimize the pinball loss or CPDI of the quantile prediction results output by the model, thereby determining the optimal model hyperparameters and constructing a battery remaining life probability prediction model based on quantile regression random forest.

[0012] The test data is input into the battery remaining life probability prediction model, the target quantile prediction value of the battery remaining life is output, and the probability prediction interval of the battery remaining life is obtained. The probability prediction result is expressed by a confidence interval symmetric about the median.

[0013] Furthermore, the obtaining of aging test data of the lithium-ion battery specifically includes:

[0014] The voltage, capacity, internal resistance and temperature data of each test cycle of the lithium-ion battery are obtained, where voltage, discharge capacity and internal resistance are required data, and charge capacity and temperature are optional data. The sampling frequency is f.

[0015] Furthermore, after obtaining the aging test data of the lithium-ion battery, the test data needs to be preprocessed, specifically including:

[0016] For all aging test data, the data with a division ratio of p is used as the training set, and the remaining data is used as the test set, where 0<p<1.

[0017] Furthermore, the preprocessing of the test data further includes:

[0018] Sampling is performed with a sampling interval of c cycles to achieve oversampling of the test data.

[0019] Furthermore, the statistical analysis quantities include mean, maximum, variance, skewness, and kurtosis;

[0020] The total number of features in the candidate feature set is denoted as k.

[0021] Furthermore, the selecting a feature subset from the candidate feature set specifically includes:

[0022] Select a feature subset containing m to n features from the candidate feature set, where 1<m<n<k, and get feature subset.

[0023] Furthermore, the pinball loss function is defined as Where q is the quantile, is the predicted result of the qth quantile of y.

[0024] Furthermore, the construction of a battery remaining life probability prediction model based on quantile regression random forest specifically includes:

[0025] By setting appropriate hyperparameters, such as the number of trees in the forest and the maximum depth of the trees, a battery remaining life probability prediction model based on quantile regression random forest QRRF is established.

[0026] Furthermore, the hyperparameter selection methods include minimizing the pinball loss function method or minimizing the coverage probability deviation index method;

[0027] The coverage probability deviation index is defined as: CPDI = |PICP-Confidence Level|, where PICP is the proportion of observed values ​​of the target variable that fall within the prediction interval: Where N is the number of samples, U i ,L i are the upper and lower limits of the prediction interval for sample i, y i is the observed value of the target variable for sample i.

[0028] Compared with the prior art, the method for probabilistically predicting the remaining life of a lithium-ion battery provided by the present invention has the following beneficial effects:

[0029] The present invention uses the pinball loss function for feature selection, applies the quantile regression random forest to obtain the quantile prediction value of the remaining effective life of the battery, and estimates the remaining effective life interval of the battery under different confidence probabilities based on the quantile prediction value. This can not only effectively utilize the data and obtain a feature subset that is more suitable for the actual data; but also because the model structure is simple, all target quantiles can be obtained by training once (compared with the common quantile regression algorithm), and finally the accurate remaining effective life interval of the battery under different confidence probabilities can be obtained. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 A flowchart of a method for probabilistically predicting the remaining life of a lithium-ion battery provided by the present invention;

[0031] Figure 2 Schematic diagram of the probability prediction result of battery 1 in Example 2 of the present invention;

[0032] Figure 3 Schematic diagram of the probability prediction result of battery 2 in embodiment 2 of the present invention;

[0033] Figure 4 It is a schematic diagram of the probability prediction result of battery 3 in embodiment 2 of the present invention. DETAILED DESCRIPTION

[0034] The following is combined with Figures 1 to 4 The following examples 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.

[0035] Example 1: Figure 1 As shown, the present invention proposes a probability prediction method for the remaining life of a lithium-ion battery, comprising the following steps:

[0036] 1. Perform lithium-ion battery aging test to obtain voltage, capacity, internal resistance, temperature and other data for each test cycle, where voltage, discharge capacity and internal resistance are required data, and charge capacity and temperature are optional data. The sampling frequency is f;

[0037] 2. Preprocess the test data: ① Divide the data with a ratio of p into the training set and the data with a ratio of (1-p) into the test set, where 0<p<1; ② Sampling is performed with a sampling interval of c cycles to achieve oversampling of the test data;

[0038] 3. Perform feature extraction and selection in the preprocessed data:

[0039] ① Extract features and determine the candidate feature set consisting of the following features and their statistical analysis quantities (such as mean, maximum, variance, skewness, kurtosis, etc.): the discharge capacity difference ΔQ(V) under the corresponding voltage of the two cycles, the internal resistance difference of the two cycles, the polynomial fitting parameters of the curve of the discharge voltage changing with the number of cycles, and the discharge capacity of a specific cycle. The total number of features in the feature set is recorded as k;

[0040] ② Use the pinball loss function to perform feature selection, select a feature subset containing m to n features from the candidate feature set described in the above steps, where 1<m<n<k, and get the common feature subsets, input the data contained in these feature subsets into the QRRF model in turn, calculate the pinball loss of the quantile prediction results output by the model respectively, and determine the feature subset with the smallest pinball loss as the final selected feature set as the input of the model;

[0041] The pinball loss function is defined as

[0042]

[0043] Where q is the quantile, is the predicted result of the qth quantile of y;

[0044] 4. Establish a battery remaining life probability prediction model based on quantile regression random forest (QRRF) and complete the modeling by setting appropriate model hyperparameters, such as the number of trees in the forest and the maximum depth of the trees;

[0045] There are two ways to select hyperparameters: ① minimize the pinball loss function, or ② minimize the coverage probability difference index (CPDI): preset a set of model hyperparameter values, input the selected feature set described in step 3② into the model, change the hyperparameters to minimize the pinball loss or CPDI of the quantile prediction results output by the model, thereby determining the optimal model hyperparameters and completing the modeling;

[0046] CPDI is defined as

[0047] CPDI=|PICP-Confidence Level|#(2)

[0048] Among them, PICP (prediction interval coverage probability) is the proportion of observed values ​​of the target variable falling within the prediction interval:

[0049]

[0050]

[0051] Where N is the number of samples, U i ,L i are the upper and lower limits of the prediction interval for sample i, y i is the observed value of the target variable for sample i.

[0052] 5. Train the model and predict RUL. Input the training set into the QRRF model for training. After the training is completed, input the test set into the model and output the target quantile prediction value of RUL to obtain the probability prediction interval of RUL. The probability prediction result is expressed as a confidence interval symmetric about the median. Compare the prediction result with the label value of RUL in the test set to evaluate the accuracy of the model.

[0053] Example 2: Figure 2-4 As shown, according to the steps in Example 1, the battery test data set published by the Massachusetts Institute of Technology (MIT) is used to verify the feasibility and effectiveness of the method proposed in the present invention. The data set contains test data of 124 battery cells and a total of about 96,700 test cycles. The detailed information of the data set is shown in https: / / www.nature.com / articles / s41560-019-0356-8 .

[0054] Data preprocessing: In step 2, p=0.75 is taken, and the batteries in the data set are divided into a training set and a test set in a ratio of 3:1. Each cell is sampled with a sampling interval of c=100 cycles to achieve oversampling of the data set; therefore, the training set contains 707 cycles of data from 83 battery cells, and the test set contains 242 cycles of data from 21 batteries.

[0055] Feature extraction and selection: Based on the description in step 3, 12 features with a strong correlation with battery life were extracted, that is, k = 12; m = 5, n = 7 were taken, and the obtained feature subsets were traversed. Finally, the 6 features that minimized the pinball loss function were determined as the input of the model.

[0056] Model building: Based on method ① in step 4, the hyperparameters of the model are determined by minimizing the pinball loss function to complete the modeling.

[0057] Model training and prediction: In this example, q = 1, 2, ..., 98, 99 is taken, and the training set data is input into the model for training. After the training is completed, the test set is input into the model for prediction, and the predicted value is compared with the label value in the test set to evaluate the accuracy of the model. In this example, the probability prediction result is represented by a 90% confidence interval, which is the interval determined by the 5th quantile and the 95th quantile.

[0058] The schematic diagram of the probability prediction results of three battery cells in the test set is as follows Figure 2-4 . It can be seen that the distribution of RUL is asymmetric and does not conform to the normal distribution assumption. The significance of the probability interval predicted by the present invention is that, assuming that the deterministic prediction result of the RUL of a battery is 500 cycles with an error of 100 cycles, but due to the asymmetric distribution of RUL, it is difficult for users to arrange battery maintenance based on this information, and the 90% confidence prediction interval given by the present invention is 480-580 cycles. Users can know from this information that it is safer to perform maintenance at 450 cycles.

[0059] In summary, the present invention uses the pinball loss function for feature selection, applies the quantile regression random forest to obtain the quantile prediction value of the remaining useful life of the battery, and estimates the remaining useful life interval of the battery under different confidence probabilities based on the quantile prediction value. This can not only effectively utilize the data and obtain a feature subset that is more suitable for the actual data; but also because the model structure is simple, all target quantiles can be obtained by training once (compared with the common quantile regression algorithm), and finally the accurate remaining useful life interval of the battery under different confidence probabilities can be obtained.

[0060] The embodiments described above are only preferred specific implementation modes of the present invention, and the protection scope of the present invention is not limited thereto. Any simple changes or equivalent replacements of the technical solutions that can be obviously obtained by any technician familiar with the field within the technical scope disclosed in the present invention belong to the protection scope of the present invention.

Claims

1. A method for predicting the probability of the remaining life of a lithium-ion battery, characterized in that: The following steps are involved: Obtain aging test data of lithium-ion batteries; Determine the candidate feature set consisting of the following features and their statistical analysis quantities: the discharge capacity difference ΔQ at the corresponding voltages of two cycles, the internal resistance difference of two cycles, the polynomial fitting parameters of the curve of discharge voltage changing with cycle number, and the discharge capacity of a specific cycle; Use the pinballloss loss function to select a feature subset from the candidate feature set, input the feature subset into the quantile regression random forest model, output the pinballloss of the quantile prediction result, and determine the feature subset that minimizes the pinballloss as the final selected feature set; The hyperparameters of the quantile regression random forest model are selected by minimizing the pinballloss loss function or minimizing the coverage probability deviation index method, and the finally selected feature set is input into the quantile regression random forest model. The hyperparameters are changed to minimize the pinballloss or CPDI of the quantile prediction results output by the model, thereby determining the optimal model hyperparameters and constructing a battery remaining life probability prediction model based on quantile regression random forest. Input the test data into the battery remaining life probability prediction model, output the target quantile prediction value of the battery remaining life, obtain the probability prediction interval of the battery remaining life, and the probability prediction result is expressed by a confidence interval symmetric about the median; The coverage probability deviation index is defined as: CPDI=PICP-ConfidenceLevel Among them, PICP is the proportion of observed values ​​of the target variable falling within the prediction interval: Where N is the number of samples, U i ,L i are the upper and lower limits of the prediction interval for sample i, y i is the observed value of the target variable for sample i.

2. A method for predicting the probability of the remaining life of a lithium-ion battery as claimed in claim 1, characterized in that: The obtaining of aging test data of the lithium-ion battery specifically includes: The voltage, capacity, internal resistance and temperature data of each test cycle of the lithium-ion battery are obtained, where voltage, discharge capacity and internal resistance are required data, and charge capacity and temperature are optional data. The sampling frequency is f.

3. A method for predicting the probability of the remaining life of a lithium-ion battery as claimed in claim 1, characterized in that: After obtaining the aging test data of the lithium-ion battery, the test data needs to be preprocessed, which specifically includes: For all aging test data, the data with a division ratio of p is used as the training set, and the remaining data is used as the test set, where 0<p<1.

4. A method for predicting the probability of the remaining life of a lithium-ion battery as claimed in claim 3, characterized in that: The preprocessing of the test data further includes: Sampling is performed with a sampling interval of c cycles to achieve oversampling of the test data.

5. The method for predicting the remaining life of a lithium-ion battery according to claim 1, wherein: The statistical analysis quantities include mean, maximum, variance, skewness, and kurtosis; The total number of features in the candidate feature set is denoted as k.

6. A method for predicting the probability of the remaining life of a lithium-ion battery as claimed in claim 5, characterized in that: The selecting of a feature subset from the candidate feature set specifically includes: Select a feature subset containing m to n features from the candidate feature set, where 1<m<n<k, and get feature subset.

7. The method for predicting the remaining life of a lithium-ion battery according to claim 1, wherein: The pinball loss function is defined as Where q is the quantile, is the predicted result of the qth quantile of y.

8. A method for predicting the probability of the remaining life of a lithium-ion battery as claimed in claim 1, characterized in that: The construction of a battery remaining life probability prediction model based on quantile regression random forest specifically includes: By setting appropriate hyperparameters, such as the number of trees in the forest and the maximum depth of the trees, a battery remaining life probability prediction model based on quantile regression random forest QRRF is established.

Citation Information

Cited By

  • RELIABILITY PREDICTION OF AGING INFORMATION FOR A RECHARGEABLE BATTERY IN A SYSTEM

    FR3171286A1

  • Reliability prediction for ageing information relating to a rechargeable battery of a system

    WO2026150173A1