CN114578234A - A Degradation and Capacity Prediction Model for Li-ion Batteries Considering Causal Features - Google Patents

Abstract

The present invention proposes a lithium-ion battery degradation and capacity prediction model considering causal features, extracts and filters the causal features, and uses the causal features as the input of the model to predict the available capacity of the battery, so as to enhance the model's performance in terms of prediction and features. Interpretability, improving the prediction accuracy of capacity. Specifically: (1) In view of the insufficient explanation of the causal relationship between input parameters and output by other neural network models, it is proposed to consider causal features to predict capacity, which improves the interpretability of the model. (2) The impulse response analysis method is proposed to analyze the influence of the features on the capacity fading, and combined with the analysis of the battery aging mechanism, the exact influence and mechanism of the selected features on the capacity fading are clarified, and the interpretability of the features is improved. (3) Taking causal features as input, a capacity prediction model based on long short-term memory LSTM network is constructed, which achieves a more accurate prediction of capacity compared with the prediction results using the original features and the prediction results of other methods.

Description

A Degradation and Capacity Prediction Model for Li-ion Batteries Considering Causal Features

technical field

The invention provides a lithium ion battery degradation and capacity prediction model considering causal characteristics, and belongs to the technical field of new energy batteries.

Background technique

Lithium-ion batteries have become a new energy source that is widely used in many fields such as electric vehicles, aerospace, and computers due to their advantages of large storage capacity, high recharge and discharge times, recyclability, and low pollution. However, due to the influence of the working environment and working conditions, a series of complex changes will occur inside the battery, causing various side reactions, and the capacity will gradually decrease, thus causing some safety problems. Therefore, accurate lifetime estimation is crucial for the safety and reliability of Li-ion batteries, and battery lifetime prediction is an important research direction in this field. With the rise of artificial intelligence and machine learning, data-driven lithium-ion battery life prediction methods are becoming a mainstream and effective method. The data-driven prediction method extracts features from the collected battery data as input variables, uses battery life as the target variable, and learns relevant knowledge through the model to predict battery life, such as SOC (State Of Charge), SOH (State Of Health) ), RUL (Remaining Useful Life) and capacity, etc. Although the existing methods have achieved good prediction results, the causal relationship between the input and output of the model cannot be well explained, so how to enhance the interpretability of the model is an urgent problem to be solved.

Data-driven machine learning prediction methods, such as Support Vector Machine (SVM), Extreme Learning Machine (ELM), Gaussian process regression (GPR), Deep Neural Network, DNN), Broad Learning System (BLS), etc., have achieved good prediction results. There are also some methods for the setting and selection of model input features, such as setting the variance and kurtosis of the original data, Shannon entropy and Hausdorff distance of the charging curve, etc., and random segments of the charging curve as the input features of the model. In feature selection, features such as Pearson and Spearman correlation coefficients, Random Forest (RF), Grey Relational Analysis (GRA) or recursive feature elimination with cross-validation are used to screen features.

Although the selection of input features through existing feature settings and extraction methods has shown promising results in battery life prediction, whether there is a causal relationship between the input features and the predicted target variable is not well explained or confirmed, i.e. Whether there is a causal relationship between the "input and output" of the model is not explained, and there is a problem of insufficient model interpretability.

SUMMARY OF THE INVENTION

Aiming at the problem that the existing prediction model can not explain enough, the present invention considers extracting and screening causal features that have an impact on battery capacity degradation, and uses the causal features as the input of the model to predict the available capacity of the battery, so as to enhance the model's ability to predict and feature Interpretability of angles and improved prediction accuracy of capacity. The core technical points of the present invention are:

(1) Aiming at the problem that other neural network models do not adequately explain the causal relationship between input parameters (or features) and outputs, the present invention proposes to consider causal features to predict capacity, which improves the interpretability of the model.

(2) The impulse response analysis method is proposed to analyze the influence of the features on the capacity fading, and combined with the analysis of the battery aging mechanism, the exact influence and mechanism of the selected features on the capacity fading are clarified, and the interpretability of the features is improved.

(3) Taking causal features as input, a capacity prediction model based on Long Short-Term Memory (LSTM) network is constructed. more accurate predictions.

In order to further clarify the technical scheme of the present invention, it is now described in detail as follows:

The present invention considers selecting causal features in the capacity degradation process to predict capacity. Specifically, the present invention proposes a lithium-ion battery degradation and capacity prediction model based on a vector autoregressive (Vector autoregressive, VAR) model and LSTM, which can select causal features in the process of lithium-ion battery capacity decay and analyze its The specific impact on the capacity fading enhances the interpretability of the features, and then uses the selected causal features to predict the battery capacity, which enhances the interpretability of the model in terms of prediction and improves the prediction accuracy of the capacity. The technical scheme of the proposed method of the present invention comprises the following steps:

S1. Extract initial features

First, the voltage, current, temperature and time data of the battery during each charge-discharge cycle obtained by the detection equipment are recorded as V, I, T, t. Then, preliminary data processing (data cleaning, dimensionality reduction, normalization, normalization) is performed on the initial data, and the initial characteristics of all cycles of the battery are constructed. Specifically, a total of 8 initial features are constructed by processing the four kinds of physical data in each cycle, denoted by F1-F8, so the total feature dimension is 8×N, where N represents the total number of cycles.

S2. Battery Feature Screening Model

Based on the VAR model and the Granger causality (GC) test, a feature screening system is constructed to screen all the initial features (F1-F8) of the battery constructed in S1, extract the battery capacity as the target variable, and compare the features with The capacities are denoted as y _i , where i=1, 2, . . . , N, and N=9. Specific steps are as follows:

S2.1, assuming that all initial features (F1-F8) and capacities are endogenous variables, construct a multivariate VAR(p) system for Granger causality test, where p represents the VAR model order, and the model is as follows:

Y _t =B _t +α ₁ Y _t-1 +α ₂ Y _t-1 +...+α _p Y _tp +e _t

Y_t-p则代表变量Y_t在t-p时刻的滞后值。具体而言，y_i,t分别代表t时刻(循环)的8个特征(F1-F8)与容量数据。而α、b、e则分别代表代表模型计算过程中权重参数、常数项参数以及误差项参数。模型通过使用容量的滞后值以及初始特征的滞后值来进行回归分析，因此可以用来分析内生变量与目标变量之间的影响关系。本发明也将利用此特性分析电池初始特征与电池容量之间的关系，对于特征进行初步筛选。in

Y _tp represents the lag value of variable Y _t at time tp. Specifically, y _{i, t} represent the eight features (F1-F8) and capacity data at time t (cycle), respectively. α, b, and e represent the weight parameters, constant term parameters, and error term parameters in the model calculation process, respectively. The model performs regression analysis by using the lagged values of the capacity and the lagged values of the initial features, so it can be used to analyze the influence relationship between the endogenous variable and the target variable. The present invention will also use this characteristic to analyze the relationship between the initial characteristics of the battery and the capacity of the battery, and conduct preliminary screening of the characteristics.

S2.2 takes capacity as the target variable, and conducts Granger causality test to analyze and verify the causal relationship between the hypothesized health factor and capacity. GC evaluates how past values of one variable lead to another. Taking feature F1 and capacity as an example, there is a Granger causality between feature F1 and capacity (lag order l, k) if and only if:

f(y9 _,t |y9 _,tk ,y1 _,tl )≠f(y9 _,t |y9 _,tk )

According to the above definition, in order to describe the screening principle more clearly, according to the model formula in S2.1, the null hypothesis that the feature F1 and the capacity have Granger causality is:

H0: _α19,1 = _α19,2 =…=α19 _,p =0

H1: At least one alpha _{19, i} is not 0

If the null hypothesis is not rejected, then y _{1, t} has no Granger causality for y _{9, t} , that is, the feature F1 has no predictive ability for battery capacity data; conversely, if the null hypothesis is rejected, then y _{1, t} for y _{9 , t} has Granger causality, that is, feature F1 has predictive power for battery capacity data. The above hypothesis test can be done with the chi-square test, and the test statistic asymptotically obeys χ ² (p).

S2.3 Change the order p, reconstruct the VAR model, repeat steps S2.1-2.2, and complete all the inspection processes of the single feature F1 of the single battery.

S2.4 uses other features of a single battery (F2-F8) as explanatory variables, and repeats steps S2.2-2.3 until all features of this battery are tested.

S2.5 Repeat steps S2.2-2.4 until all feature inspections of all batteries are completed.

S2.6 Comprehensive screening of all test results to obtain screening results of causal characteristics. The invention performs multiple GC inspections on all features of multiple batteries, thereby eliminating the randomness of single inspection results and single battery results, increasing the persuasiveness of the feature selection results, and making the feature selection results more objective.

S3. Battery characteristic causality analysis model

According to the causality feature screening result obtained from S2, the present invention constructs a battery feature causality analysis system based on the VAR system and the system impulse response function (Impulse Response Functions, IRF) in this step, and analyzes the impact of the screened features on the battery capacity, That is, the specific influence and mechanism of these features in the process of battery capacity degradation, to further verify the causal features of the selected battery capacity degradation. Proceed as follows:

S3.1 According to the preliminary screening results of the features and combined with the test results in S2, multiple multivariate VAR systems are reconstructed, and each system is used to analyze the influence of battery causal features on battery capacity degradation.

S3.2 Before analyzing the characteristic causality of the impulse response, the stability test of the constructed VAR model is required. The inspection method is as follows:

It can be obtained by the multivariate VAR(p) model formula in S2.1. If p=t, t-1, ..., t-k+1, k can be obtained (t-k+1 time to t time) Multivariate VAR model. If it is represented in the form of a block matrix, that is, if:

则可将k个VAR(p)模型通过友矩阵变化写为VAR(1)，也即：

Then the k VAR(p) models can be written as VAR(1) through the change of the friend matrix, that is:

Z _t =C+φZ _t-1 +θ _t

This model is also the characteristic causality analysis model of the battery. Specifically, Y _t represents 8 features (F1-F8) and capacity data at time t (cycle), Y _tp represents the lag value of variable Y _t at time tp, and B _t and e _t represent the VAR at time t The constant term parameters and error term parameters of the model, α _j represents the weight parameter matrix of the jth VAR(p) model, j=1, 2,..., k, E represents the identity matrix. Then, in the constructed VAR(1), C and θ _t represent the constant term parameter matrix and error term parameter matrix of the analytical model, respectively, and Z _t and Z _t-1 represent the selected battery causal characteristics and capacity construction matrix current and hysteresis values.

的全部特征值全在单位圆以内，也即其特征方程

的全部根在单位圆之内。如果不稳定，则改变模型阶数，重构VAR系统。若模型稳定则可进行脉冲响应分析，分析因果特征对于电池容量衰减的具体影响。For the integrated multivariate VAR model currently constructed, the condition for model stability is the total weight parameter matrix

All eigenvalues of are all within the unit circle, that is, its characteristic equation

All roots of are within the unit circle. If it is unstable, change the model order and reconstruct the VAR system. If the model is stable, impulse response analysis can be performed to analyze the specific impact of causal features on battery capacity decay.

S3.3 Through the impulse response function of the battery feature causality analysis system VAR(1) constructed in S3.2, the specific impact of each causal feature in the model on the capacity can be analyzed. The impulse response function describes the impact on the current and future values of endogenous variables after applying a standard deviation shock to the random error term, so it can be used to analyze the impact of a time series variable on another in time series. The influence of time series variables. First, we can get from VAR(1):

可得：make

Available:

Z _t+r = θ _t+r +ψ ₁ θ _t+r-1 +ψ ₂ θ _t+r-2 +…+ψ _r θ _t +…

Then there are:

代表模型变换和计算过程中的参数矩阵。Ψ_r代表整个电池特征因果性分析系统模型的脉冲响应函数，其中第i行j列的元素看为滞后期数r的函数，此元素表示令其他误差项在任何时期都不变的条件下，当第j个变量y_j,t的误差项e_t在t时刻收到一个单位的冲击后，对于第i个内生变量y_i,t在t+r期造成的影响。在本发明中，以电池容量为目标变量，通过电池因果性分析系统的IRF获取容量对于所选因果性特征的脉冲响应结果。Among them, Z _t and Z _t+r represent the causal characteristics of the battery and the value of the capacity matrix at time t and t+r, and θ represents the error term parameter matrix of the model. E stands for the identity matrix, L and

Represents the parameter matrix during model transformation and computation. Ψ _r represents the impulse response function of the entire battery characteristic causality analysis system model, in which the element in the i-th row and j column is regarded as a function of the lag period r. When the error term e _t of the j-th variable y _j,t receives a unit shock at time t, the impact of the i-th endogenous variable y _i,t in period t+r. In the present invention, taking the battery capacity as the target variable, the impulse response result of the capacity to the selected causality feature is obtained through the IRF of the battery causality analysis system.

S3.4 interprets the results of the impulse response analysis in conjunction with the mechanism of the internal capacity degradation of Li-ion batteries, further enhancing the interpretability of the features.

S4. Battery capacity estimation

S4.1 Build a capacity prediction model. The present invention constructs a stacked LSTM network for capacity prediction. The network structure has multiple layers: input layer, LSTM layer, Dropout layer, Dense layer and output layer. The network structure is shown in Figure 3. The input layer is responsible for receiving the input data and converting the data into a format acceptable to the LSTM layer. In this layer, a sliding window technique with a "many-to-one" structure is used to make the output of the network and the input related at multiple times to better to learn the temporal information of the sequence. The LSTM layer is responsible for learning the characteristic information of the data transmitted by the input layer, and the nonlinear relationship between this information and the predictor variable (capacity). The Dropout layer is responsible for randomly deactivating the neurons of the LSTM layer to prevent the model from overfitting and reduce the prediction effect. The Dense layer is responsible for dimensional transformation of the output data of the LSTM layer-Dropout layer. Finally, the output layer receives the transformed data of the Dense layer and outputs the prediction result. The output of the network is the predicted capacity of the corresponding life cycle of the battery.

S4.2 According to S3, using the determined causal features as input data, use the LSTM prediction model constructed in S4.1 to predict the battery capacity. Specifically, firstly, the input data is converted into a format suitable for the LSTM layer through the sliding window technique at the input layer; secondly, the feature information of the data transmitted by the input layer is learned at the LSTM layer, and the weight parameters of the LSTM layer are retained; , in the Dropout layer, the Dropout technology is used to discard the calculation results of part of the LSTM neurons; then, the Dense layer reduces the dimension of the output results of the LSTM layer processed by the Dropout, and transmits the transformation results to the output layer; finally, the output layer outputs the capacity forecast result

S4.3 obtains the capacity prediction result.

The advantages and beneficial effects of the present invention are: the present invention uses the causal features in the battery degradation process to predict the battery usable capacity, which not only achieves a more accurate prediction of the battery capacity, but also improves the interpretability of the model in terms of prediction and features. , which is also an important research area in machine learning today, that is, the problem of interpretability. In addition, the proposed method can also be extended to other types of batteries or life prediction fields, so the present invention not only has important academic significance, but also has potential engineering application value for improving the reliability and safety of new energy electronic devices.

Description of drawings

Figure 1 is a flow chart of the entire battery degradation and capacity prediction model.

Figure 2 is a schematic diagram of the principle of the model.

Figure 3 is a structural diagram of a capacity prediction model constructed based on LSTM.

FIG. 4 is a certain characteristic curve diagram extracted from the detection data of a certain battery in the NASA public data set used in the embodiment.

FIG. 5 is a graph of a battery capacity degradation curve in the NASA public data set used in the examples.

Figure 6 is an example diagram of the stability test results of a VAR system.

FIG. 7 is an example graph of characteristic impulse response analysis results.

Figure 8a is a graph of the capacity prediction result of a certain battery using causal features and original features as model input features.

Figure 8b is a prediction error graph of the capacity prediction result of a certain battery using causal features and raw features as model input features.

Detailed ways

The present invention will be further described below with reference to the accompanying drawings and embodiments. As shown in Figures 1-2, the present invention provides a lithium-ion battery degradation and capacity prediction model considering causal characteristics. The present invention uses the public battery degradation dataset of NASA Ames Center of Excellence as an example dataset. The dataset contains four 18650-type batteries B5, B6, B7, and B18, charged, discharged, and impedance with three different operations at room temperature, and data was collected. The experiment was stopped when the rated capacity of the battery dropped by 30% (from 2Ah to 1.4Ah). Specific steps are as follows:

S1. Extract initial features. First, based on the battery data obtained by the monitoring equipment, the initial characteristics of the battery for all cycles are constructed. The characteristic curve extracted from the B5 battery discharge voltage data in the data set is shown in Figure 4.

S2. Feature preliminary screening.

S2.1 First, extract the battery capacity data. The capacity degradation curve of the B5 battery is shown in Figure 5.

S2.2 takes capacity as the target variable and assumes that all eigenfactors and capacity are endogenous variables to construct a multivariate VAR(p) system for Granger causality test. The test results of a certain characteristic of B5 battery are shown in the following table:

Feature Chi-sq df Prob. Discharge current 33.1035 10 0.0003

Where Chi-sq represents the Chi-squared test value of the null hypothesis, df represents the VAR model order, and Prob. represents the probability of rejecting the null hypothesis. The present invention sets the test significance level as 0.05, which means that the probability that the selected variable and the target variable are endogenous variables is at least 95%, that is, there is a Granger causality. In this test, the discharge current was initially chosen as the causal feature.

S2.3 changes the order, reconstructs the VAR model, and repeats steps S2.1-2.2 to complete the single feature inspection process for a single battery.

S2.4 Repeat steps S2.1-2.3 until all feature inspections of this battery are completed.

S2.5 Repeat steps S2.2-2.4 until all feature inspections of all batteries are completed.

S2.6 Comprehensive screening of all test results to obtain preliminary screening results.

S3. Feature causality analysis.

S3.1 reconstructs multiple multivariate VAR(p) systems according to the preliminary screening of features and the results combined with the test results in S2, and analyzes the impact of the screened features on capacity degradation.

S3.2 Before performing the impulse response, the stability test of the constructed VAR model is required. The stability test result of a VAR system is shown in Figure 6. It can be seen that the system is stable and the impulse response analysis can be performed.

S3.3 Through the impulse response function of the system, combined with the impulse response results of multiple systems, the specific impact of the screening features in the model on the capacity is analyzed. The impulse response diagram of the capacity to the discharge voltage in a certain system is shown in Figure 7. According to this result, it can be seen that although the influence of discharge voltage on capacity degradation is relatively fluctuating, it is basically negative. Therefore, according to the results of impulse response analysis, discharge voltage is preliminarily selected as the characteristic that has influence on capacity degradation.

S3.4 Interpret the results of the impulse response analysis in conjunction with the mechanism of internal capacity degradation of Li-ion batteries. For example, according to the analysis of the impulse response results in S3.3, the discharge cut-off voltage of this battery is greater than the maximum discharge voltage (2.75V), and the battery is in an “overdischarged” state in each cycle. This operating state can cause some negative effects, such as corrosion of the battery collector and loss of active materials, which will lead to the degradation of the battery capacity. Therefore, combined with the impulse response results, the discharge voltage is considered to be one of the reasons for the degradation of battery capacity.

S3.5 Determine final causal characteristics. In this embodiment, the final charge-discharge temperature, discharge voltage, current, and time are preliminarily screened as features that have an impact on capacity degradation.

S4. Capacity estimation

S4.1 Build a capacity prediction model.

S4.2 processes the data and rewrites the data dimension. According to S3, the causality feature is used as the input feature of the prediction model to predict the battery capacity.

S4.3 obtains the capacity prediction results. The prediction results of a certain battery are shown in Figure 8(a) and Figure 8(b). Figure 8(a) and Figure 8(b) show the prediction using the original feature and the causal feature. results and the corresponding prediction errors. It can be seen that compared with the original features, the capacity prediction results using causal features are closer to the real degradation curve, and the prediction error is closer to 0, which verifies the effectiveness of the selection method and selected features. The comparison with other methods is shown in the table below:

为第i个循环时电池实际容量和预测容量，计算方式如下所示：Among them, MAE (mean absolute error) and RMSE (root mean square error) are commonly used evaluation indicators for regression prediction problems. m is the number of prediction cycles, Qi and

is the actual capacity and predicted capacity of the battery at the i-th cycle, and the calculation method is as follows:

It can be seen that the MAE and RMSE of the method of the present invention are both the smallest, which shows the effectiveness of the proposed method.

Claims (6)

1. A lithium ion battery degradation and capacity prediction model considering causality characteristics is characterized by comprising the following specific steps:

s1, extracting initial characteristics:

firstly, voltage, current, temperature and time data obtained by detection equipment in each charge-discharge cycle process of the battery are recorded as V, I, T, t; then, carrying out primary data processing on the initial data to construct initial characteristics of all cycles of the battery;

s2, battery characteristic screening model:

constructing a characteristic screening system based on a VAR model and a Glankey causal relationship GC test, and screening all initial characteristics of the battery constructed in S1;

s3, a battery characteristic causality analysis model:

obtaining a causal characteristic screening result according to S2, constructing a battery characteristic causal analysis system based on a VAR system and a system Impulse Response Function (IRF), analyzing the influence of the screened characteristics on the battery capacity, namely the specific influence and action mechanism of the characteristics in the battery capacity degradation process, and further verifying the selected causal characteristics of the battery capacity degradation;

and S4, estimating the capacity of the battery.

2. The model of claim 1, wherein the model is for predicting lithium ion battery degradation and capacity based on causal characteristics, and comprises: in step S1, a total of 8 initial signatures, denoted as F1-F8, are constructed by processing the four types of physical data for each cycle, so that the overall signature dimension is 8 × N, where N represents the total number of cycles.

3. The model of claim 1, wherein the model is for predicting lithium ion battery degradation and capacity based on causal characteristics, and comprises: in step S2, the battery capacity is extracted as a target variable, and the characteristics and the capacity are respectively denoted as y_iWherein i is 1, 2.

4. The model of claim 1,2 or 3 for predicting lithium ion battery degradation and capacity considering causal characteristics, wherein: in step S2, the method further comprises the following steps:

s2.1: setting all initial characteristics F1-F8 and capacity as endogenous variables, constructing a multivariate VAR (p) system for carrying out Glanker causal relationship test, wherein p represents the VAR model order, and the model is shown as the following formula:

Y_t＝B_t+α₁Y_t-1+α₂Y_t-1+…+α_pY_t-p+e_t

wherein

Y_t-pRepresents the variable Y_tA hysteresis value at time t-p; y is_i,tRespectively representing 8 characteristics F1-F8 and capacity data at the time t; and alpha, b, e are respectively substitutedThe table represents a weight parameter, a constant term parameter and an error term parameter in the calculation process of the model; the battery capacity is subjected to regression analysis by using the hysteresis value of the capacity and the hysteresis value of the initial characteristic, so that the battery capacity is used for analyzing the influence relationship between the interpretation variable and the target variable and primarily screening the characteristic;

s2.2: carrying out Glangel causal relationship test by taking the capacity as a target variable, and carrying out analysis and verification on the causal relationship between the assumed health factor and the capacity; the GC evaluates how past values of one variable result in another variable; characteristic F1 has a grand causal relationship with capacity if and only if:

f(y_9，t|y_9，t-k，y_1，t-l)≠f(y_9，t|y_9，t-k)

from the model formula in S2.1, the primary assumption that the feature F1 and capacity have grand causality is:

H0:α_19,1＝α_19,2＝…＝α_19,p＝0

h1 having at least one alpha_19,iIs not 0

If the original hypothesis is not rejected, y_1,tFor y_9,tThere is no granger causality, i.e., feature F1 has no predictive power for battery capacity data; otherwise if the original hypothesis is rejected, y_1,tFor y_9,tThe characteristics F1 have the causality of Glankey, namely the characteristics F1 have the prediction capability on the battery capacity data; the hypothesis test is completed by chi-square test, and the test statistic asymptotically obeys chi²(p)；

S2.3: changing the order p, reconstructing a VAR model, repeating the steps S2.1-2.2, and completing all the inspection processes of the single characteristic F1 of the single battery;

s2.4: repeating steps S2.2-2.3 using the other characteristics F2-F8 of the monoblock battery as explanatory variables until all characteristic checks of the battery are completed;

s2.5: repeating the step S2.2-2.4 until all the characteristics of all the batteries are checked;

s2.6: screening by integrating all detection results to obtain a screening result with causal characteristics; and performing GC (gas chromatography) inspection on all the characteristics of the plurality of batteries for multiple times, thereby eliminating the randomness of a single inspection result and a single battery result, increasing the persuasion of a characteristic selection result and enabling the characteristic selection result to have more objectivity.

5. The model of claim 4, wherein the model is used for predicting lithium ion battery degradation and capacity by considering causality characteristics, and comprises: in step S3, the method further comprises the following steps:

s3.1: reconstructing a plurality of multivariable VAR systems according to the preliminary screening results of the characteristics and combined with the detection results in S2, wherein each system is used for analyzing the influence relationship of the battery causal characteristics on the battery capacity degradation;

s3.2: before the characteristic causality of impulse response analysis is carried out, stability inspection needs to be carried out on the constructed VAR model; the test method comprises the following steps:

obtaining k multivariate VAR models by using a multivariate VAR (p) model formula in S2.1, wherein if p is t, t-1, …, t-k + 1; if the block matrix is expressed, let:

the k VAR (p) models are written as VAR (1) by friend matrix change, i.e.:

Z_t＝C+φZ_t-1+θ_t

this model is a characteristic causality analysis model of the battery; y is_t8 features F1-F8 representing time t and volume data, Y_t-pThen represents the variable Y_tHysteresis value at time t-p, and B_t、e_tConstant term parameter and error term parameter, α, of the representative time-t VAR model_jThen, a weight parameter matrix representing the jth var (p) model, where j is 1, 2.., k, E represents an identity matrix; then, in the constructed VAR (1), C and θ_tA constant term parameter matrix and an error term parameter matrix, Z, respectively representing the analytical model_tAnd Z_t-1Then the current and hysteresis values of the selected battery causal characteristic and capacity construction matrix are represented;

for the integrated multivariate VAR model currently constructed, the condition of model stability is the total weight parameter matrix

All the eigenvalues of (A) are all within the unit circle, i.e. the eigenequation

Is within the unit circle; if the model is unstable, changing the order of the model, and reconstructing a VAR system; if the model is stable, performing impulse response analysis, and analyzing the specific influence of causal characteristics on the battery capacity attenuation;

s3.3: analyzing the specific influence of each causal feature in the model on the capacity through an impulse response function of a battery feature causal analysis system VAR (1) constructed in S3.2; the impulse response function describes the influence on the current value and the future value of an endogenous variable after a standard deviation impact is applied to a random error term, so that the impulse response function is used for analyzing the influence relation of a certain time series variable on another time series variable in time sequence; first obtained from VAR (1):

order to

Obtaining:

Z_t+r＝θ_t+r+ψ₁θ_t+r-1+ψ₂θ_t+r-2+…+ψ_rθ_t+...

then there are:

wherein Z is_tAnd Z_t+rThen represents electricityValues of the pool causality characteristic and the capacity matrix at t and t + r moments, and theta represents an error item parameter matrix of the model; e represents an identity matrix, L and

representing a parameter matrix in the process of model transformation and calculation; psi_rAn impulse response function representing the causality analysis system model of the whole battery characteristic, wherein the element of the ith row and the jth column is regarded as a function of the lag period number r, and the element represents that other error terms are not changed in any period when the jth variable y is_j,tError term e of_tAfter receiving a unit of impact at time t, for the ith endogenous variable y_i,tThe effect caused during the t + r period; acquiring an impulse response result of the capacity to the selected causal characteristic through an IRF (interference rejection filter) of a battery causal analysis system by taking the battery capacity as a target variable;

s3.4: the interpretability of the characteristics is further enhanced by interpreting the results of the impulse response analysis in conjunction with the mechanism of internal capacity degradation in the lithium ion battery.

6. The model of claim 5, wherein the model is used for predicting lithium ion battery degradation and capacity by considering causality characteristics, and comprises: in step S4, the method further comprises the following steps:

s4.1: constructing a capacity prediction model; constructing a stacked LSTM network for capacity prediction; the network architecture has multiple levels: an input layer, an LSTM layer, a Dropout layer, a Dense layer and an output layer; the input layer is responsible for receiving input data and converting the data into a format acceptable by the LSTM layer, and a sliding window technology with a 'many-to-one' structure is used in the input layer, so that the output and the input of a network are related at a plurality of moments, and the time information of a sequence is learned; the LSTM layer is responsible for learning characteristic information of data transmitted by the input layer and the nonlinear relation between the information and a prediction variable; the Dropout layer is responsible for carrying out random inactivation treatment on the neurons of the LSTM layer, and the phenomenon that the model is over-fitted to reduce the prediction effect is prevented; the Dense layer is responsible for carrying out dimension transformation on output data of the LSTM layer-Dropout layer; finally, an output layer receives the transformed data of the Dense layer and outputs a prediction result, wherein the output of the network is the predicted capacity of the corresponding life cycle of the battery;

s4.2: according to step S3, the battery capacity is predicted using the LSTM prediction model constructed in S4.1, with the determined causal characteristics as input data; firstly, converting input data into a format which is suitable for an LSTM layer to be accepted by an input layer through a sliding window technology; secondly, learning characteristic information of data transmitted by an input layer on an LSTM layer, and keeping weight parameters of the LSTM layer; then, discarding a part of the calculation results of the LSTM neurons by using Dropout technology in a Dropout layer; then, carrying out dimension reduction on the output result of the LSTM layer subjected to Dropout processing in the Dense layer, and transmitting the conversion result to the output layer; finally, the output layer outputs the capacity prediction result

S4.3: and obtaining a capacity prediction result.

Images