Energy storage lithium ion battery health state prediction method

Abstract

The invention relates to the technical field of lithium ion batteries, in particular to an energy storage lithium ion battery health state prediction method, which comprises the following steps that sparse test point data are acquired through a battery aging experiment, and each test point comprises a cycle number, internal resistance, an increment health state characteristic and a health state label; systematic slicing and resampling are carried out on a time axis by utilizing an aging mining method, training samples containing different circulation history paths are generated, and a mapping relation between circulation history and health state attenuation is learned by adopting a deep learning model, so that high-precision health state prediction is realized; the method has the beneficial effects that the problem of scarcity of battery aging data is solved, and the health state prediction precision is remarkably improved; the method has good generalization ability and has flexibility and expandability.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of lithium ion batteries, in particular to a method for predicting the state of health of energy storage lithium ion batteries. BACKGROUND

[0002] As the core component of new electrochemical energy storage systems, the health state decay prediction of lithium ion batteries is of great significance for battery management, safety monitoring and maintenance decision. Accurate health state prediction can not only provide early warning of battery performance decline, but also optimize charging and discharging strategies, prolong battery life and ensure user safety. However, the health state decay prediction of batteries faces a fundamental dilemma: there is an irreconcilable contradiction between the time scale of aging experiments and the data requirements of machine learning models. A complete battery aging experiment usually requires hundreds to thousands of charge and discharge cycles, and takes several months or even years, while deep learning models often require a large amount of training data to obtain good generalization performance.

[0003] Currently, the battery state of health prediction methods based on data-driven mainly include support vector machine, random forest, neural network, long short-term memory network and other machine learning algorithms. Although these methods have achieved certain results, they generally face problems such as large demand for training data, weak early prediction ability, limited generalization ability, etc. The authors of document [1] (Zhang Y, Wu X, Tian Z, Gao W, Peng H, Yang K. Comparison of random forest, support vector regression, and long short term memory for performance prediction and optimization of a cryogenic organic rankine cycle[J]. Energy Conversion and Management, 2023) compared the prediction performance of SVR, RF and LSTM under different scales and noisy data. LSTM performance declines when there are more features but limited data; RF is sensitive to outliers; SVR generalization declines in high-dimensional data. The authors of document [2] (Yang N, Hofmann H, Sun J. Remaining useful life prediction of lithium-ion batteries with limited degradation history using random forest[J]. IEEE Transactions on Energy Conversion, 2023) pointed out that LSTM has large error under limited historical data, revealing the problems of early prediction difficulty and strong dependence on training data.

[0004] From the data level, battery aging data has typical sparsity characteristics. Limited by experimental cost and time period, researchers can usually only perform complete tests at limited cycle nodes, which makes the number of data points available for training severely insufficient. Traditional data enhancement methods such as interpolation, noise injection, etc. can expand the data scale to a certain extent, but often cannot effectively mine the path dependence and dynamic preference in the battery aging process, and the expansion effect is limited.

[0005] Therefore, how to extract the rich information contained in the battery degradation process through reasonable data mining and expansion methods based on limited experimental data has become the key to improving the prediction accuracy of the state of health. SUMMARY

[0006] The present application aims to provide a method for predicting the state of health of a lithium-ion battery for energy storage, which systematically samples and reconstructs the degradation history path of the battery at different time points, extracts rich path-dependent information from sparse aging test data, and greatly expands the training data. Unlike traditional data enhancement methods, the aging mining method is based on the physical nature of battery aging, i.e. the state of health of the battery at a certain cycle number depends not only on the current state, but also on the complete cycle history from the initial state to the current state, to build a complete data mining framework to solve the problems raised in the background art.

[0007] To achieve the above-mentioned purpose, the present application provides the following technical scheme: a method for predicting the state of health of a lithium-ion battery for energy storage, comprising the following steps: Sparse test point data is collected through battery aging experiments, each test point containing cycle number, internal resistance, incremental state of health feature, and state of health label; Using the aging mining method, systematic slicing and resampling are performed on the time axis to generate training samples containing different cycle history paths; A deep learning model is used to learn the mapping relationship between cycle history and state of health decay, achieving high-precision state of health prediction.

[0008] Preferably, it further comprises feature variable selection: The feature variables reflect the aging mechanism and state of health of the battery; the feature variables do not directly contain state of health information; the features include ohmic resistance R0, charge transfer resistance Rct, main peak position V_peak and peak height IC_max in incremental state of health analysis, or other available electrochemical features.

[0009] Preferably, the core mechanism of the aging mining method is that the state of health of the battery at a certain cycle number depends not only on the current state, but also on the complete cycle history from the initial state to the current state; by systematically constructing different cycle history paths, rich training samples are extracted from limited test points.

[0010] Preferably, the specific steps of the aging mining method include: Select a series of slicing points on the time axis, set the initial cycle number as n0, the maximum cycle number as n max , and the slicing interval as Δn; for each slicing point, construct a complete history sequence from the initial state to the point, and interpolate the features on the history sequence to obtain a complete input matrix; generate an expanded data set, each sample corresponding to a unique degradation path from the initial state to a certain cycle number.

[0011] Preferably, the deep learning model adopts a temporal model, including but not limited to: Long Short-Term Memory Network (LSTM); Gated Recurrent Unit (GRU); Temporal Convolutional Network (TCN); Transformer model.

[0012] Preferably, when using the LSTM model, its network architecture includes: The input layer receives a cyclic history sequence matrix of size [L×d]; the first LSTM layer contains 128 hidden units and outputs the complete sequence, followed by a dropout layer; the second LSTM layer contains 64 hidden units and outputs only the hidden state at the last time step, followed by a dropout layer; the fully connected layer performs feature transformation, and the output layer is a single neuron that provides a health status prediction.

[0013] Preferably, the method employs a leave-one-battery cross-validation strategy to evaluate the model's generalization ability, and the specific steps include: In the multi-battery dataset, one battery is selected as the test set each time, and the extended data of the remaining batteries is used as the training set; cyclic validation ensures the model's predictive ability for unseen batteries.

[0014] Preferably, it also includes a data preprocessing step: Calculate the mean μ and standard deviation σ of the training set for each feature dimension; apply the transformation X'=(X-μ) / σ to standardize the data.

[0015] Preferably, the method adjusts hyperparameters such as slice interval Δn and sequence length L to adapt to raw test data of different densities, wherein: Smaller Δn can generate more samples but may introduce redundancy; larger Δn has the opposite effect; the sequence length L needs to capture enough temporal dependencies without causing excessively long sequences to increase the difficulty of training.

[0016] Compared with the prior art, the beneficial effects of the present invention are: The energy storage lithium ion battery health state prediction method provided by the application realizes exponential expansion of data quantity by systematically constructing different cycle history paths and extracting rich training samples from sparse test points. The expansion is not simple data replication, but deep mining based on physical mechanism, and each expanded sample carries unique path dependence information. The method can expand 20 original test points into hundreds of training samples, effectively alleviating the deep learning model's demand for large data, and greatly reducing the time cost and economic cost of battery aging experiments. On this basis, the model trained based on the expanded data shows higher accuracy and stability in health state prediction, and can more fully learn the complex patterns and path dependence characteristics in the battery degradation process, providing more reliable prediction results for battery health state evaluation and management; solve the problem of battery aging data scarcity, and significantly improve the health state prediction accuracy.

[0017] Has good generalization ability, flexibility and scalability: the method shows stable prediction ability for unseen batteries under leave-one-battery cross-validation conditions, which shows that the model can learn the degradation rules with universal significance, rather than relying on individual characteristics of a single battery, which is particularly critical for building a general model in actual battery management system. Aging mining extracts common degradation patterns and path dependence information, so that the model can generalize from a small amount of battery data to a larger battery group. At the same time, the method has high flexibility in feature selection and model structure, and can select different electrochemical characteristics according to data availability and adapt to various time series modeling frameworks, and its idea can be naturally extended to other health indicators and different types of degradation systems. In the engineering implementation level, aging mining mainly relies on a simple and efficient data construction process, the overall process is clear, the computational overhead is low, the model reasoning speed is fast, and it is easy to integrate into the existing battery management system, and has the feasibility of moving from research to practical application. BRIEF DESCRIPTION OF DRAWINGS

[0018] Figure 1 The method flowchart of the application is shown in the figure; Figure 2 The prediction and actual comparison diagram of DE-LSTM of the application for different batteries is shown in the figure; Figure 3 The comparison diagram of DE-LSTM and traditional LSTM of the application is shown in the figure. DETAILED DESCRIPTION

[0019] In order to make the purpose, technical scheme of the present application, and the advantages are clearer and more apparent, the embodiments of the present application are further described in detail below with reference to the drawings. It should be understood that the specific embodiments described herein are part of the embodiments of the present application, rather than all embodiments, and are used to explain the embodiments of the present application, and do not limit the embodiments of the present application. All other embodiments obtained by those skilled in the art without creative labor fall within the scope of the present application.

[0020] In one embodiment, the present application provides a technical solution: a method for predicting the state of health of an energy storage lithium ion battery, comprising the following steps: First, sparse test point data is collected through battery aging experiments, each test point containing cycle number, internal resistance, incremental state of health and other characteristics, and health state label. Among them, the physical meaning and selection principle of the characteristic variable: In order to avoid data leakage and fully characterize the battery degradation state, the characteristic variables used in this method need to meet two basic principles: one is to reflect the aging mechanism and health state of the battery, and the other is not to directly contain health state information. Based on these two principles, the following types of characteristics are selected as model inputs, but it should be emphasized that the specific selection of characteristics should be adjusted flexibly according to the availability of actual data.

[0021] Ohmic internal resistance R0 reflects the electronic and ionic transport impedance inside the battery, and its evolution is closely related to the aging mechanisms such as current collector corrosion, electrode contact degradation, and electrolyte conductivity decline. Charge transfer resistance Rct represents the charge transfer kinetics at the electrode-electrolyte interface, and its change reflects processes such as active material surface area loss and SEI film thickening.

[0022] Incremental capacity analysis (Incremental Capacity, IC) provides important information about the phase transition process of electrode materials. The main peak position V_peak of the IC curve corresponds to the phase transition potential of the electrode material, reflecting the change in lattice parameters; the peak height IC_max represents the reaction kinetics and accessibility of active materials. These characteristics describe the degradation state of the battery from different angles and constitute a complete representation of the health state decay. It should be noted that if certain characteristics (such as incremental health state) cannot be obtained under experimental conditions, other electrochemical characteristics (differential voltage, temperature response, etc.) can be used instead, or only a subset of available characteristics can be used. The flexibility of feature selection is an important advantage of this method.

[0023] The prediction target is the maximum available state of health Q, which is only used as a supervision label for model output and cannot be part of the input characteristics in any case, to ensure the practicality of the method and the credibility of the prediction results.

[0024] Then, the aging mining method is used to slice and resample the data systematically along the time axis, generating a large number of training samples containing different cyclic historical paths. The core mechanism of the aging mining data is: The theoretical basis of the aging mining method lies in the path dependence of the battery aging process. The state of health Q(n) of the battery at cycle number n depends not only on the state variables at the current time (such as internal resistance, incremental state of health features, etc.), but also on the complete evolution history of the battery from the initial state n0 to the current state n. This historical dependence is due to the cumulative effect of multiple irreversible aging processes, such as the growth of SEI film, which can be represented as δSEI(n)=δ o +∫ o n This integral form explicitly reveals the essential dependence of the aging state on the historical path.

[0025] Based on this understanding, the aging mining method extracts rich training samples from limited test points by systematically constructing different cyclic historical paths. Assume that m cycle numbers n1, n2,..., n m are tested, and each test point i obtains a feature vector X i and a health state label Q i . Traditional methods can only use these m data points for training, while the aging mining method selects a series of slicing points along the time axis. Let the initial cycle number be n0 (usually 0) and the maximum cycle number be n max , and the slicing interval be Δn, then the slicing point sequence is n j =n0+(j-1)·Δn, where j=1, 2,..., k, and k is the maximum number of slicing points not exceeding n max . The slicing interval Δn is an important hyperparameter, a smaller Δn can generate more samples but may introduce redundancy, and a larger Δn is the opposite, which can be adjusted adaptively according to the density of the original data in practical applications.

[0026] For the jth slicing point n j , the method constructs a complete history sequence from n0 to n j . Specifically, L points are uniformly sampled in this history interval to form the sequence N_CE^[j]=linspace(n0,n j , L), where L is the sequence length hyperparameter (such as 10, 50 or 100), representing the historical depth that the time series model can trace back, and linspace represents from n0 to n jUniformly sample L points. Then, interpolate the values of each feature dimension at the original test points to obtain the values of the features at the sampled points. Take the ohmic resistance as an example, R0_CE^[j]=interp(n,R0,N_CE^[j]), where n and R0 are the cycle number vector and the internal resistance vector of the original test points, respectively, and interp represents the interpolation operation (which can be linear interpolation, spline interpolation, etc.). After repeating this operation for all features, the complete input matrix X_CE^[j]=[N_CE^[j],R0_CE^[j],Rct_CE^[j],V_peak_CE^[j],IC_max_CE^[j],...]∈ ^(L×d), where d is the number of feature dimensions, and Rct, V_peak and IC_max represent the charge transfer resistance, the peak position and the peak height of the IC curve, respectively. These are just some examples of features that can be used, and are not limited to these features. The health status label corresponding to this sample is Q_CE^[j]=interp(n,Q,n j ), which is the health status value at the slice point.

[0027] By traversing all slice points j=1,2,...,k, the extended dataset D_CE={(X_CE^[j],Q_CE^[j])}_(j=1)^k can be generated. This extension process achieves a significant increase in data size: if the original data has m=20 test points covering 1000 cycles, and Δn=20 is selected, about 50 training samples can be generated, with an expansion factor of about 2.5 times; if Δn=5, the expansion factor can reach 10 times. More importantly, these extended samples are not simply data repetition or random perturbation, but carry different cycle history information, and each sample corresponds to a unique degradation path from the initial state to a specific cycle number.

[0028] Finally, a deep learning model is used to learn the mapping relationship between the cycle history and the health status decay, achieving high-precision health status prediction; it also includes the architecture and training of the time series model: The expanded data has obvious time series characteristics, and each training sample is a multi-dimensional time series with a length of L. To effectively learn this time series dependency, the method uses a time series model in deep learning to predict the health status. It is important to note that the specific time series model used is not a limitation of the present application, but can be selected flexibly according to actual needs. Common time series models include long short-term memory networks (LSTM), gated recurrent units (GRU), time convolutional networks (TCN), and Transformers, all of which can effectively process sequence data and capture long-term dependencies.

[0029] The LSTM is used as an example to illustrate the model building process. LSTM can learn and remember long-term dependencies through its special gating mechanism, making it suitable for handling problems such as battery aging that have long-term evolution characteristics. A typical network architecture can be designed as follows: the input layer receives a cyclic history sequence matrix with size [L x d]; the first LSTM layer contains 128 hidden units, outputs the complete sequence for extracting multi-scale time series features, followed by a dropout layer (dropout rate 0.2) to prevent overfitting; the second LSTM layer contains 64 hidden units, only outputs the hidden state at the last time for sequence encoding, also with a dropout layer; then a fully connected layer (e.g., 32 neurons, ReLU activation) is used for feature transformation, and the output layer is a single neuron that gives the health state prediction value. Of course, this is just an example configuration, and in actual applications, the network depth, width, and structure can be adjusted according to factors such as data size and computing resources. Other time series models can also be used instead of LSTM, or an ensemble learning method can be used to combine the prediction results of multiple models.

[0030] During training, first, the data is standardized by calculating the mean μ and standard deviation σ for each feature dimension of the training set, and then applying the transformation X' = (X - μ) / σ. To evaluate the model's generalization ability, a leave-one-battery cross-validation strategy is used: in the multi-battery data set, each battery is selected as the test set, and the extended data of the remaining batteries is used as the training set. The model's prediction ability for unseen batteries is verified by cycling validation. The Adam optimizer is used for optimization, with an initial learning rate of 0.001 and a learning rate decay strategy (e.g., reducing to 0.5 times the original every 30 rounds) to improve convergence stability. The mean squared error (MSE) is chosen as the loss function, and the training is performed for a maximum of 100 rounds, with an early stopping mechanism set to terminate training if the validation set loss does not decrease for 10 consecutive rounds to avoid overfitting. The batch size can be set according to the number of samples, such as 32 or 64 samples per batch.

[0031] In Example Two, the aging data of four groups of lithium-ion batteries from the NASA dataset is used for verification. Each group of batteries has undergone a complete aging cycle process, from the initial state to the significant degradation of the health state. During the aging process, the batteries are periodically tested for complete features, including electrochemical impedance spectroscopy (EIS) to obtain ohmic resistance R0 and charge transfer resistance Rct, low-rate charge-discharge tests to obtain voltage-health state curves and calculate incremental health state curves, from which the main peak position V_peak and peak height IC_max are extracted, as well as standard health state tests to obtain the maximum available health state Q.

[0032] After data collection, each battery obtains a dataset containing m test points, where m is usually between 20 and 50, and each test point records all feature values and health state labels at the corresponding cycle number n_i.

[0033] After the data preprocessing, the aging mining method is applied to each battery to generate extended data. The key parameter settings in this embodiment are: the sequence length L is 10, that is, each training sample contains 10 time step history information, which can capture sufficient time dependence without increasing the training difficulty due to too long sequence; the slice interval Δn is initially set to 20 cycles, but considering that the test point density of different batteries may be different, an adaptive adjustment strategy is adopted - for batteries with fewer test points (such as less than 30 test points in total), Δn is automatically adjusted to 5 or 10 to generate more training samples; the interpolation method is selected as piecewise cubic Hermite interpolation, which can well preserve the local features of the original data while maintaining data smoothness and avoiding overshooting.

[0034] To comprehensively evaluate the generalization ability of the method, a leave-one-battery cross-validation strategy is adopted. The four batteries are sequentially used as the test set, and the extended data of the remaining three batteries is used as the training set, a total of four rounds of training and testing. Before each round of training, first standardize the training set: calculate the mean μ and standard deviation σ of all training samples for each feature dimension, then apply the transformation (X-μ) / σ to the training set and test set to ensure that different features are in similar numerical ranges.

[0035] This embodiment selects LSTM as the time series model, but as mentioned earlier, this is not the only choice, and GRU, TCN or other time series models can also be used. The network structure is designed as follows: the input layer receives 10x5 sequence data; the first LSTM layer has 128 hidden units, outputs the complete sequence, and is configured with a dropout layer (dropout rate 0.2); the second LSTM layer has 64 hidden units, only outputs the last time state, and is also configured with a dropout layer; then a fully connected layer (32 neurons, ReLU activation) and another dropout layer (dropout rate 0.1) are connected; finally, the output layer is a single neuron that gives the health state prediction value. The training uses the Adam optimizer with an initial learning rate of 0.001, and the learning rate is multiplied by 0.5 every 30 rounds to decay. The batch size is set to 32, the maximum training rounds is 100, but an early stopping mechanism is set: 15% of the training set is randomly selected as the validation set, and the validation set is evaluated every 30 rounds, if the validation loss does not decrease for 10 consecutive rounds, the training is stopped in advance. In actual training, in most cases, convergence is reached around 40-60 rounds.

[0036] The comparison results of DE-LSTM and traditional LSTM are shown in Table 1.

[0037] Table 1 Comparison of DE-LSTM and traditional LSTM

[0038] While embodiments of the application have been shown and described, it is to be understood that the embodiments described are merely exemplary of the principles and application of the present application. Numerous modifications and adaptions can be made without departing from the spirit and scope of the present application, which is defined by the following claims and their equivalents.

Claims

1. A method for predicting the health status of energy storage lithium-ion batteries, characterized in that: Includes the following steps: Sparse test point data were collected through battery aging experiments. Each test point included cycle number, internal resistance, incremental health status characteristics, and health status label. Using the aging mining method, systematic slicing and resampling are performed on the timeline to generate training samples containing different cyclic history paths; A deep learning model is used to learn the mapping relationship between cyclic history and health status decay, so as to achieve high-precision health status prediction.

2. The method for predicting the health status of an energy storage lithium-ion battery according to claim 1, characterized in that: It also includes feature variable selection: Characteristic variables reflect the aging mechanism and health status of the battery; characteristic variables do not directly contain health status information; characteristics include ohmic internal resistance R0, charge transfer internal resistance Rct, main peak position V_peak and peak height IC_max in incremental health status analysis, or other available electrochemical characteristics.

3. The method for predicting the health status of an energy storage lithium-ion battery according to claim 1, characterized in that: The core mechanism of the aging mining method is that the health status of a battery at a certain number of cycles depends not only on the current state, but also on the complete cycle history from the initial state to the current state; by systematically constructing different cycle history paths, rich training samples are extracted from a limited number of test points.

4. The method for predicting the health status of an energy storage lithium-ion battery according to claim 3, characterized in that: The specific steps of the aging excavation method include: Select a series of slice points on the timeline, and let the initial number of loops be n0 and the maximum number of loops be n. max The slice interval is Δn; for each slice point, a complete history sequence from the initial state to that point is constructed, and each feature is interpolated on the history sequence to obtain a complete input matrix; an extended dataset is generated, where each sample corresponds to a unique degradation path from the initial state to a specific number of cycles.

5. The method for predicting the health status of an energy storage lithium-ion battery according to claim 1, characterized in that: The deep learning model employs a temporal model, including but not limited to: Long Short-Term Memory Network (LSTM); Gated Recurrent Unit (GRU); Temporal Convolutional Network (TCN); and Transformer model.

6. The method for predicting the health status of an energy storage lithium-ion battery according to claim 5, characterized in that: When using the LSTM model, its network architecture includes: The input layer receives a cyclic history sequence matrix of size [L×d]; the first LSTM layer contains 128 hidden units and outputs the complete sequence, followed by a dropout layer; the second LSTM layer contains 64 hidden units and outputs only the hidden state at the last time step, followed by a dropout layer; the fully connected layer performs feature transformation, and the output layer is a single neuron that provides a health status prediction.

7. The method for predicting the health status of an energy storage lithium-ion battery according to claim 1, characterized in that: The method employs a leave-one-battery cross-validation strategy to evaluate the model's generalization ability. The specific steps include: In the multi-battery dataset, one battery is selected as the test set each time, and the extended data of the remaining batteries is used as the training set; cyclic validation ensures the model's predictive ability for unseen batteries.

8. The method for predicting the health status of an energy storage lithium-ion battery according to claim 1, characterized in that: It also includes data preprocessing steps: Calculate the mean μ and standard deviation σ of the training set for each feature dimension; apply the transformation X'=(X-μ) / σ to standardize the data.

9. The method for predicting the health status of an energy storage lithium-ion battery according to claim 1, characterized in that: The method adjusts hyperparameters such as slice interval Δn and sequence length L to adapt to raw test data of different densities, wherein: Smaller Δn can generate more samples but may introduce redundancy; larger Δn has the opposite effect; the sequence length L needs to capture enough temporal dependencies without causing excessively long sequences to increase the difficulty of training.

Images