A two-dimensional support domain direct inference health status prediction method for lithium battery with few cycles

Abstract

The present invention discloses a two-dimensional support domain direct-induction health status prediction method for lithium batteries with few cycles. This method takes into account the batch characteristics of multiple batteries of the same model, and uses historical data and batch data to construct a two-dimensional support domain to expand the model information source, thereby providing a coarse range of selectable samples for model establishment, and uses direct-induction modeling to comprehensively consider the information of offline and online sample feature spaces, and carefully divide each sample, and then select and model each sample according to the different importance of each sample, so as to solve the problem of inaccurate health status modeling prediction when the historical charge and discharge cycle data of lithium batteries is small. This method innovatively introduces two-dimensional support domain and direct-induction modeling to expand and filter data, which effectively solves the problem of health status prediction in the case of a small amount of cycle data.

Description

Technical Field

[0001] The present invention belongs to the field of lithium battery health status prediction, and in particular relates to a two-dimensional support domain modeling prediction method based on a small number of cycles. Background Art

[0002] In recent years, lithium-ion batteries have become widely used due to their high energy density, low self-discharge rate, high voltage, long life cycle, and enhanced reliability. However, the safety issues they present cannot be underestimated. As one of the primary means of ensuring safety, lithium battery state of health (SOH) management has received increasing attention. SOH, as an indicator of battery aging, reflects the ratio of the battery's actual capacity to its nominal capacity. A brand new battery has an SOH of 100%. With continued use, the SOH gradually decreases. Typically, when the battery's capacity drops to 80% of its initial value (SOH of 80%), it is considered end-of-life (EOL) and should be replaced. Generally, accurate SOH can only be directly measured under laboratory conditions; in practice, it can only be estimated using other variables, such as voltage and current. However, due to the nonlinearity and uncertainty of battery aging, SOH prediction is particularly difficult.

[0003] Due to the development of machine learning theory and the improvement of computing power in recent years, data-driven methods have received increasing attention in academia and industry. Data-driven battery SOH prediction methods have been widely used in the field of lithium battery SOH prediction. They establish data-based SOH prediction models without relying on any complex domain knowledge, and therefore have stronger generalization capabilities.

[0004] While data-driven approaches have garnered widespread attention in SOH assessment, their effectiveness depends heavily on the size and quality of historical datasets. In real-world scenarios, a single charge and discharge cycle for a lithium-ion battery can often take hours or even days. When historical charge and discharge cycle data is scarce, collecting sufficient battery aging data for modeling is inherently time-consuming. Using accelerated aging experiments to shorten the aging period is likely to alter the battery's aging profile, rendering the collected data unsuitable for modeling. Insufficient historical data for modeling can lead to overfitting. Furthermore, lithium batteries undergo capacity regeneration during charge and discharge cycles, a characteristic feature of lithium-ion battery degradation. If a battery is left idle longer than normal during aging, this regeneration process can occur, increasing the available capacity for the next cycle and rendering the model's predictions unreliable. Furthermore, for the current analysis target, model development and subsequent predictions often require a sufficient amount of data to be collected after a certain period of charge and discharge. This delays the understanding of the lithium battery's health status. Based on the above analysis, we can see that real-world SOH prediction for lithium batteries is typically hampered by the lack of data. How to cleverly design modeling and analysis strategies with limited charge-discharge cycle data to ensure model prediction accuracy, and how to fully leverage the information from training samples to effectively model the samples to be predicted, are pressing issues of both theoretical and practical significance. Summary of the Invention

[0005] The purpose of the present invention is to address the problem of difficulty in modeling and prediction when a small amount of historical cycle data is accumulated, and to provide a two-dimensional support domain direct modeling method for lithium batteries with a small number of cycles. On the one hand, the present invention uses similar characteristics of other batteries in the same batch as data expansion, constructs a support domain in two dimensions using batch data and historical data, and roughly divides the data, providing sufficient samples to choose from. On the other hand, a direct learning framework is proposed, which takes into account the similarity information of the feature space of offline samples and online samples, realizes the refined and full utilization of samples under a small amount of cycle data, and further realizes the accurate prediction of SOH under a small number of charge and discharge cycles.

[0006] The purpose of the present invention is achieved through the following technical solutions:

[0007] A two-dimensional support domain direct-inference health status prediction method for lithium batteries with a small number of cycles includes the following steps:

[0008] (1) Collecting raw data, including the current charge and discharge cycle n, historical charge and discharge cycles, and charge and discharge cycle samples of other batches of batteries of the battery to be predicted, each charge and discharge cycle sample including at least a voltage curve; the other batches of batteries are batteries with the same model and charge and discharge cycle test conditions as the battery to be predicted;

[0009] (2) constructing a two-dimensional support domain, wherein the first dimension of the two-dimensional support domain includes all historical charge-discharge cycle samples of the battery to be predicted, and the second dimension includes the first (l-1) / 2 charge-discharge cycle samples and the last (l-1) / 2 charge-discharge cycle samples of the charge-discharge cycle n of other batches of batteries, where l is an odd number greater than 1;

[0010] (3) extracting features from each sample data in the current charge and discharge cycle n of the battery to be predicted and the two-dimensional support domain to construct a training set; the features are closely related to the battery capacity;

[0011] (4) Using the features of the samples in the training set as input and the predicted health status as output, direct modeling is performed based on feature similarity to obtain a well-established model.

[0012] (5) The characteristics of the current charge and discharge cycle samples of the battery to be predicted are input into the model to obtain the predicted value of the battery health status.

[0013] Furthermore, the step (3) is specifically as follows:

[0014] The training set of the two-dimensional support domain construction is

[0015]

[0016] Among them, the superscript j represents the number of the battery to be predicted, and the superscript r represents the number of other batches of batteries, r≠j, represents the characteristics of the current charge and discharge cycle n of the battery to be predicted, represents the label set of the historical charge and discharge cycle samples of the battery to be predicted, and k represents the total number of historical charge and discharge cycles accumulated by the battery to be predicted; Represents the feature set of the battery history charge and discharge cycle samples to be predicted, Represents the feature set of other batches of battery charge and discharge cycle samples added, Indicates the label set of other batches of battery charge and discharge cycle samples.

[0017] Furthermore, in step (3), feature extraction is performed on the current charge and discharge cycle n of the battery to be predicted and each sample data in the two-dimensional support domain, specifically:

[0018] Convert the voltage curve of each charge-discharge cycle sample into an incremental capacity curve;

[0019] The vertical coordinate of the incremental capacity curve is dQ / dV, that is, the incremental capacity value, and the horizontal coordinate is the voltage value;

[0020] Smoothing the incremental capacity curve using a filtering method, wherein the filtering method is a Gaussian filtering method, a sliding average method, an SG filtering method, or the like;

[0021] The features closely related to the battery capacity are extracted from the smoothed incremental capacity curve, specifically six feature points, which are the voltage values ​​and dQ / dV values ​​of the two peaks and one trough on the incremental capacity curve.

[0022] Furthermore, in step (4), the least squares method is used for direct inference modeling, and the established model is obtained by minimizing the loss function; the loss function is the error between the true value and the predicted value, wherein each sample is assigned an independent weight, which is expressed as follows:

[0023]

[0024] in represents the characteristics of the qth sample in the training set, τ is a hyperparameter, w q Represents the weight of the qth sample, where the smaller the weight, the smaller the impact on modeling. The loss function is specifically expressed as follows:

[0025]

[0026] in represents the true label of the qth sample in the training set, represents the predicted value of the sample, and k+3l is the total number of samples in the training set.

[0027] Compared with the prior art, the present invention has the following beneficial effects:

[0028] A SOH prediction method for lithium batteries with a small number of cycles based on two-dimensional support domain and direct modeling is proposed. By constructing a two-dimensional support domain and integrating batch data and historical data, sufficient selectable samples are provided for the model. Furthermore, through direct modeling, the similarity between offline and online samples is considered, and the refined utilization of samples under a small amount of cycle data and the accurate prediction of health status are achieved. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 It is a flow chart of the method of the present invention.

[0030] Figure 2 is the construction diagram of the two-dimensional support domain.

[0031] Figure 3The results of using the present invention to predict the SOH value of battery No.18 are shown in Figure 2. DETAILED DESCRIPTION

[0032] The present invention will be further described below with reference to the accompanying drawings and specific examples.

[0033] The present invention provides a two-dimensional support domain direct-pushing health status prediction method for lithium batteries with a small number of cycles, comprising:

[0034] (1) Collecting raw data, including the current charge and discharge cycle n, historical charge and discharge cycles, and charge and discharge cycle samples of other batches of batteries of the battery to be predicted, each charge and discharge cycle sample including at least a voltage curve V; the other batches of batteries are batteries with the same model and charge and discharge cycle test conditions as the battery to be predicted;

[0035] (2) constructing a two-dimensional support domain, using the data of charge and discharge cycle samples of other batches of batteries, and then integrating the original accumulated historical charge and discharge cycle sample data of the battery to be predicted, to construct the support domain from two dimensions, wherein the first dimension of the two-dimensional support domain includes all historical charge and discharge cycle samples of the battery to be predicted, and the second dimension includes the first (l-1) / 2 charge and discharge cycle samples and the last (l-1) / 2 charge and discharge cycle samples of the charge and discharge cycle n of other batches of batteries;

[0036] Specifically, the sample data of the second dimension can be selected using a sliding window around the current cycle n from the charge and discharge cycle data of other batches of batteries. The data window includes not only the data before the current cycle, but also the data after the current cycle. The sliding window length is represented by l. Since the sliding window selects data from both sides at the same time, (l-1) / 2 samples are selected on each side, so l should be an odd number.

[0037] (3) extracting features from each sample data in the current charge and discharge cycle n of the battery to be predicted and the two-dimensional support domain to construct a training set; the features are closely related to the battery capacity;

[0038] Among them, the features closely related to battery capacity include the duration of constant current charging, the duration of constant voltage charging, the slope and area of ​​the voltage curve before constant voltage charging begins, etc. Taking the extraction of features closely related to battery capacity based on the incremental capacity curve as an example, the details are as follows:

[0039] The original voltage curve of the current charge and discharge cycle n of the battery to be predicted and each sample in the two-dimensional support domain is converted into an incremental capacity curve, and the curve needs to be smoothed. The sliding average method, SG filtering method or Gaussian filtering method can be used for denoising.

[0040] Specifically, for the voltage, current, and temperature curves generated by each charge and discharge cycle, represented by V, I, T, the voltage curve is converted into an incremental capacity curve using the following formula:

[0041]

[0042] The vertical axis of the curve is dQ / dV, that is, the incremental capacity value, and the horizontal axis is the voltage value. The incremental capacity curve can indicate the degree of battery aging. However, since it involves differential calculations, there are many noise signals in the curve. It is necessary to use a filtering method to smooth the curve. Here, the Gaussian filter method is used for denoising, which can be expressed as:

[0043]

[0044] Where μ is the mean, σ is the standard deviation, and z represents a data point on the incremental capacity curve. After filtering with a Gaussian filter, features closely related to battery capacity are extracted from the smoothed incremental capacity curve. In this example, six feature points are extracted from the curve: the voltage values ​​and dQ / dV values ​​of two peaks and one trough. These six features are used to predict SOH.

[0045] Furthermore, the constructed training set is like Figure 2 As shown, it means as follows:

[0046]

[0047] Wherein, the superscript j represents the number of the battery to be predicted, and the superscript r represents the number of other batches of batteries. In this embodiment, r∈{1,2,3,4}, r≠j, represents the characteristics of the current charge and discharge cycle n of the battery to be predicted, represents the label set of the historical charge and discharge cycle samples of the battery to be predicted, and k represents the total number of historical charge and discharge cycles accumulated by the battery to be predicted;

[0048] Represents the feature set of the battery history charge and discharge cycle samples to be predicted, Represents the feature set of other batches of battery charge and discharge cycle samples added, Represents the label set of other batches of battery charge and discharge cycle samples added, Indicates the number of other batches of batteries with number r Characteristics of charge and discharge cycles, For other batches of batteries numbered r Labels for charge and discharge cycles.

[0049] (4) Using the features of the samples in the training set as input and the predicted health status as output, direct modeling is performed based on feature similarity to obtain a well-established model.

[0050] In this embodiment, the least squares method is used to perform direct inference modeling based on feature similarity, and the established model is obtained by minimizing the loss function; when modeling, an independent weight is assigned to each sample, the weight is measured according to the similarity of the feature space, and the loss function takes the error between the true value and the predicted value.

[0051] Specifically, the sample weight is calculated as follows:

[0052]

[0053] in represents the features of the qth sample in the training set, Represents the characteristics of the current charge and discharge cycle sample of the battery to be predicted, w q represents the weight of the sample, and τ is a hyperparameter.

[0054] The loss function can be mean square error, sum of squares of the two norms, etc. Taking the sum of squares of the two norms as an example, the loss function is expressed as:

[0055]

[0056] in represents the true label of the qth sample in the training set, represents the predicted value of the sample, and k+3l is the total number of samples in the training set.

[0057] Minimize the loss function When the model converges or the iteration exceeds the maximum number, the iteration stops and the model is obtained. It should be noted that since a charge and discharge cycle takes at least several hours or even longer, the interval between each sample point is very long, and the least squares model has a simple structure, and the modeling and prediction time is in seconds. Therefore, online modeling can fully guarantee the timeliness and accuracy of the prediction.

[0058] Furthermore, the support domain is not necessarily a regular shape, and can be modeled based on the weight w q The size of the sample participating in the direct induction model is selected, where the weight w q The smaller it is, the smaller the impact on modeling; specifically, when calculating the weights between samples in the training set and samples to be predicted, a threshold for measuring the size of the weight can be defined in advance. If the weight is lower than this threshold, it can be considered that it has almost no effect on modeling, then these data points will be ignored, and the range of the support domain will also change, becoming an irregular shape, thereby realizing the automatic selection of relevant data.

[0059] (5) The characteristics of the current charge and discharge cycle n of the battery to be predicted Input into the model to get the predicted health status value

[0060] In this example, a NASA open source dataset was selected for experimental verification, and the batteries used were numbered No. 5, No. 6, No. 7, and No. 18. This set of batteries was cyclically charged, discharged, and impedance tested at room temperature (24°C), with the batteries undergoing 168, 168, 168, and 132 charge-discharge cycles, respectively. During the charging process, the battery was first charged in constant current (CC) mode at a current of 1.5A until the voltage rose to 4.2V; then it was charged in constant voltage (CV) mode until the charging current dropped to 20mA. The discharge process included CC mode, discharging at a constant current of 2A until the voltage dropped to 2.7V, 2.5V, 2.2V, and 2.5V, respectively. To verify the effectiveness of the method of the present invention in predicting a small number of cycles, the first 20% of the battery's charge and discharge cycle data was selected as the training set, and the remaining 80% was selected as the test set. Each battery was subjected to one experiment, for a total of four times, with the batch division shown in Table 1.

[0061] Table 1. Classification of batteries in different batches

[0062]

[0063] The evaluation criteria used in the experiment are root mean square error (RMSE) and mean absolute error (MAE). RMSE can measure the deviation between the observed value and the true value, while MAE can better reflect the actual situation of the predicted value error. The two can be expressed as follows:

[0064]

[0065] where y i represents the true value, Represents the predicted value, and M represents the total number of predicted samples.

[0066] The prediction accuracy of SOH under different methods is shown in Table 2. Three regression methods, linear regression (LR), support vector regression (SVR), and random forest (RF), were selected for comparison. Here, the other regression methods use historical data plus all other batches of battery data for modeling, while this method only uses data in the support domain (l value is 11) for modeling each time. For four batteries, the prediction accuracy of the present invention has an optimal RMSE of 0.92%, a worst of 1.56%, and an average of 1.26%. The MAE is optimal 0.75%, a worst of 1.17%, and an average of 0.92%. Its prediction results are better than those of linear regression (averages of 1.57% and 1.24%), support vector regression (averages of 1.91% and 1.51%), and random forest (averages of 1.47% and 1.12%), fully demonstrating the effectiveness of the present invention. Figure 3 The effect diagram of using the proposed invention to predict the SOH of No.18 battery is shown. It can be seen from the figure that the predicted value of the present invention is very close to the actual value, which further proves the effectiveness of the present invention.

[0067] Table 2. SOH prediction accuracy of different methods

[0068]

[0069] Obviously, the above embodiments are merely examples for clarity of explanation and are not intended to limit the implementation methods. Those skilled in the art will appreciate that other variations or modifications can be made based on the above description. It is not necessary and impossible to enumerate all implementation methods here. Obvious variations or modifications derived therefrom remain within the scope of protection of the present invention.

Claims

1. A two-dimensional support domain direct-pushing health status prediction method for lithium batteries with few cycles, characterized by: The following steps are involved: (1) Collecting raw data, including the current charge and discharge cycle n, historical charge and discharge cycles, and charge and discharge cycle samples of other batches of batteries of the battery to be predicted, each charge and discharge cycle sample including at least a voltage curve; the other batches of batteries are batteries with the same model and charge and discharge cycle test conditions as the battery to be predicted; (2) constructing a two-dimensional support domain, wherein the first dimension of the two-dimensional support domain includes all historical charge and discharge cycle samples of the battery to be predicted, and the second dimension includes the first (l-1) / 2 charge and discharge cycle samples and the last (l-1) / 2 charge and discharge cycle samples of the charge and discharge cycle n of other batches of batteries, where l is an odd number greater than 1; the sample data of the second dimension are selected by using a sliding window around the current cycle n in the charge and discharge cycle data of other batches of batteries, and the data window includes not only the data before the current cycle, but also the data after the current cycle. The sliding window length is represented by l, and the sliding window selects data from both sides at the same time. (l-1) / 2 samples are selected on each side to obtain the samples of the second dimension; (3) extracting features from each sample data in the current charge and discharge cycle n of the battery to be predicted and the two-dimensional support domain to construct a training set; the features are features related to the battery capacity; (4) Using the features of the samples in the training set as input and the predicted health status as output, a direct inference model is performed based on feature similarity to obtain a well-established model; (5) The characteristics of the current charge and discharge cycle samples of the battery to be predicted are input into the model to obtain the predicted value of the battery health status.

2. The method according to claim 1, wherein In step (3), the training set constructed by the two-dimensional support domain is Among them, the superscript j represents the number of the battery to be predicted, and the superscript r represents the number of other batches of batteries, r≠j, represents the characteristics of the charge and discharge cycle k of the battery to be predicted, Represents the label set of the battery history charge and discharge cycle samples to be predicted, represents the feature set of historical charge and discharge cycle samples of the battery to be predicted, and k represents the total number of historical charge and discharge cycles accumulated by the battery to be predicted; Represents the feature set of other batches of battery charge and discharge cycle samples added, Indicates the label set of other batches of battery charge and discharge cycle samples.

3. The method according to claim 2, wherein In step (4), the least squares method is used for direct inference modeling, and the established model is obtained by minimizing the loss function; the loss function is the error between the true value and the predicted value, wherein each sample is assigned an independent weight, which is expressed as follows: in Represents the features of the qth sample in the training set, w q represents the weight of the qth sample, τ is a hyperparameter, Represents the characteristics of the current charge and discharge cycle n of the battery to be predicted.

4. The method according to claim 3, wherein In step (4), the loss function is specifically expressed as follows: in represents the true label of the qth sample in the training set, represents the predicted value of the sample, and k+3l is the total number of samples in the training set.

5. The method according to claim 3, wherein Also includes the weight w q The size of the sample participating in the direct induction model is selected, where the weight w q The smaller it is, the less impact it has on modeling.

6. The method according to claim 1, wherein In step (3), feature extraction is performed on the current charge and discharge cycle n of the battery to be predicted and each sample data in the two-dimensional support domain, specifically: Convert the voltage curve of each charge-discharge cycle sample into an incremental capacity curve; The vertical coordinate of the incremental capacity curve is dQ / dV, that is, the incremental capacity value, and the horizontal coordinate is the voltage value; Smoothing the incremental capacity curve using a filtering method; Features related to battery capacity are extracted from the smoothed incremental capacity curve.

7. The method according to claim 6, wherein The filtering method is Gaussian filtering, sliding average filtering or SG filtering.

8. The method according to claim 6, wherein The features related to the battery capacity are extracted from the smoothed incremental capacity curve, specifically six feature points, which are the voltage values ​​and dQ / dV values ​​of the two peaks and one trough on the incremental capacity curve.

Images