CN-121831563-B - Two-stage lithium ion battery capacity fading track prediction method based on residual life and knee point

CN121831563BCN 121831563 BCN121831563 BCN 121831563BCN-121831563-B

Abstract

The invention discloses a two-stage lithium ion battery capacity decay track prediction method based on residual life and knee points, and belongs to the technical field of lithium ion battery health state assessment and life prediction. The method comprises the steps of constructing a feature set by utilizing capacity-cycle data obtained by a cycle test, realizing joint prediction of RUL and knee points based on a mixed network of a convolution-gating cycle unit, dividing a capacity curve into a slow fading stage and an accelerated fading stage by taking the knee points as boundaries, respectively modeling by adopting a LOESS smooth fitting and a double-exponential model, realizing continuous conductive fusion of two sections of curves in the neighborhood of the knee points by a splicing function with smooth coefficients, solving global parameters by taking the minimum life-span error as a criterion, and adaptively determining network super-parameters and stage model parameters by a dung beetle optimizing method. The method can simultaneously capture the differential degradation mechanism before and after the knee point, ensure the smoothness and consistency of the full-life capacity track, and remarkably improve the prediction precision of the later accelerated degradation stage.

Inventors

SHEN JIANGWEI
ZHOU YI
CHEN ZHENG
SHEN SHIQUAN
Wei fuxing
XIA XUELEI

Assignees

昆明理工大学

Dates

Publication Date: 20260508
Application Date: 20260313

Claims (5)

1. The method for predicting the capacity fading track of the two-stage lithium ion battery based on the residual service life and the knee point is characterized by comprising the following steps of: s1, cyclic data acquisition and capacity fading sequence construction; Carrying out multi-cycle charge and discharge test on different batteries of the same model under the constant temperature condition until the discharge capacity is attenuated to 80% of rated capacity, adopting different charge strategies in the test process, wherein the test process comprises constant current-constant voltage and multi-stage constant current-constant voltage modes, and when the SOC reaches 80%, carrying out multi-stage constant current cut-off, standing the battery for 10s, setting the charge state as constant voltage charge; S2, performing data preprocessing on the data obtained in the step S1, and performing feature construction to obtain a multi-channel input matrix; S3, constructing a CNN-GRU residual life and knee point joint prediction model based on the multi-channel input matrix; In the CNN-GRU residual life and knee point joint prediction model, performing super-parameter self-adaptive optimizing operation through a dung beetle optimizing algorithm DBO; s4, predicting the residual life RUL and knee points; The operation of S4 includes: S4.1, defining knee points as turning points of transition from a slow decay phase to an acceleration decay phase in a curve of capacity variation along with cycle times, and calculating a capacity curve corresponding to a position at which the capacity variation rate is obviously increased Is expressed based on a smooth turning model, and has the following expression: ; in the formula, To smooth the model first, second and third parameters, , wherein, , The local linear slopes before and after the knee point, As an intermediate parameter, a parameter which is a function of the parameter, Is a turning center parameter; Is a transition smoothing coefficient; Is the number of cycles; based on the local linear slope before and after the knee point, a knee point is obtained by calculating the slope of the angular bisector, and the expression is as follows: ; in the formula, Is the slope of the angular bisector; by calculating the knee point cycle number, the knee point determination is completed, and the expression is as follows: ; in the formula, Representing the adjustment parameters; The number of knee point cycles; S4.2, training and outputting a model, taking 20% of input sequence in front of a target battery as input, dividing a sample into a training set and a test set according to the proportion of 8:2, taking target battery data into the test set, training by using a CNN-GRU model configured by an optimal super parameter set determined by a DBO algorithm, and inputting the test set into the model to obtain a prediction And ; S5, modeling based on a two-stage capacity fading curve of the residual life RUL and the knee point; the modeling of the two-stage capacity fading curve comprises the steps of adopting local weighted regression (LOESS) smooth fitting before the knee point and adopting a double-exponential function model after the knee point; s6, outputting a capacity fading track prediction result; And selecting a fitting curve with the minimum integral of the Euclidean distance and the capacity deviation as a final prediction curve.
2. The method for predicting capacity fade trajectories of two-stage lithium ion batteries based on remaining life and knee points of claim 1, wherein the step of S2 comprises: performing wavelet noise reduction and smoothing on voltage and capacity data acquired by the lithium ion battery in a multi-cycle charge-discharge process, and eliminating abnormal cycle points; Extracting voltage characteristic parameters of charge-discharge stage to form voltage characteristic And combining capacity fading characteristic matrix Forming a comprehensive input matrix: ; Using belts The linear regression model of regularization term performs sparse constraint on the feature importance of the comprehensive input matrix, and the optimization objective function is as follows: ; in the formula, Is an intercept term; is a regression coefficient vector; Is a regularization coefficient; Is a target variable; for the total number of battery cycle period samples, An index representing the total number of battery cycle samples; Representing the total number of features, An index representing the total number of features; Represent the first The first battery cycle period sample A characteristic value; Represent the first Regression coefficient absolute values of the individual features; Determination of optimum by cross-validation And reserving a key voltage and capacity feature subset with non-zero coefficients, and defining a feature importance index according to the correlation of the feature variance and the target variable: ; in the formula, Represent the first A composite importance score for each feature; Represent the first The variance of the features is such that, For the feature and the target variable for the feature Correlation coefficients between; For weighting coefficients according to The first 4 related features are arranged in descending order and selected to form a final input matrix For a pair of Performing normalization and constructing a multi-channel timing feature: ; in the formula, Representing first order difference, i.e. representing the first Cycle and the first The amount of variation between cycles; Representing an identification of the type of data, , A sequence of capacities is indicated and, Representing a sequence of voltages; Representing different data types A value of each cycle; Representing different data types A value of each cycle; representing a first order difference, i.e., a change in the amount of change, to capture speed information; Representing different data types A value of each cycle; representing a running average; Representing a cyclic index; representing a sliding window length; Is a lag index within the window; sequencing capacity First order difference Second order difference With a sliding average Aligned and spliced according to time steps to form a capacity sequence multichannel input matrix : ; Wherein, the Representing a time step; Sequencing the voltages First order difference Second order difference With a sliding average Aligned and spliced according to time steps to form a voltage sequence multichannel input matrix : ; Splicing the two types of time sequence blocks in the channel dimension to obtain a multi-channel input matrix 。
3. The method for predicting capacity fade trajectories of two-stage lithium ion batteries based on remaining life and knee points of claim 1, wherein the step of S3 comprises: s3.1, aligning input serialization with a label; The multi-channel input matrix obtained in the step S2 Time series samples as models, each sample consisting of a succession of A time step composition, each time step comprising The number of channels is 8, a sample is obtained by sliding along a time axis with a step length of 1, and a starting anchor point is remembered as Terminate as The total number of samples is The label is the number of remaining life cycles Cycle number with knee point Stacking all samples in time sequence to form input tensor samples of a model and corresponding supervision labels of the input tensor samples and center anchor points of time sequence samples Aligning to obtain an input sequence; S3.2, a convolution feature extraction layer CNN, wherein the input sequence firstly extracts time local features through a one-dimensional convolution layer, and a convolution calculation formula is as follows: ; in the formula, Representing a convolution operation along a time dimension; Representing a convolution kernel weight matrix; Representing the bias vector; Representing an input sequence; Is a nonlinear activation function ReLU; the layer is used for automatically extracting local trend information in capacity fading characteristics, and the extracted result is transmitted into a circulating layer after batch normalization and Dropout operation; s3.3 gating cyclic unit layer GRU, convolution layer output sequence Is sent to the gate control loop unit GRU layer, GRU at time step The status update process of (1) is: ; in the formula, Expressed in time steps A feature vector input to the GRU unit; activating a function for Sigmoid; To update the activation value of the gate; an activation value for a reset gate; Representing candidate hidden states; Is a time step Is output hidden state of (a); representing per-element products; , , respectively inputting corresponding weight matrixes; , , weight matrixes which are respectively connected in a recursion way in a hidden state; , , bias items of each gate and candidate state respectively; Is a hyperbolic tangent activation function; s3.4 output mapping layer and joint prediction, namely final hiding state of GRU output Mapping to a prediction space through a full connection layer, wherein the calculation formula is as follows: ; in the formula, As a matrix of weights, the weight matrix, As a result of the offset vector, Activating the function for ReLU and outputting Corresponding to Or (b) ; S3.5, super-parameter self-adaptive optimizing based on a dung beetle optimizing algorithm DBO, namely introducing the dung beetle optimizing algorithm to carry out the self-adaptive optimizing on key super-parameters of the network structure before model training; The hyper-parameter set to be optimized is: ; in the formula, Is the number of convolution kernels; Is the convolution kernel time length; A Dropout proportion; The size of the micro-Batch; The number of GRU units; for the residual life and knee point of the joint output, adopting a mean square error MSE as a training target: ; in the formula, As a total number of samples, Is a true RUL value or knee point value; to predict RUL values or knee points values; the root mean square error RMSE of the validation set is defined as an fitness function of the DBO as: ; in the formula, To verify the total number of samples; To verify the true RUL value or knee point value in the set; to verify and intensively predict RUL value or knee point value, fixing the ultrasonic parameters after the DBO ultrasonic parameter search is completed, and passing Updating weights in training set and verifying set Early-stop and model selection are implemented.
4. The method for predicting capacity fade trajectories of two-stage lithium ion batteries based on remaining life and knee points of claim 1, wherein the step of S5 comprises: s5.1, carrying out stage division and data input according to Number of end cycles corresponding to remaining life Dividing the battery capacity fading curve into a slow fading stage and an acceleration fading stage, wherein the input data comprises a capacity sequence The abscissa and ordinate of knee points Cycle transverse and longitudinal termination points ; S5.2, for the slow decay phase before the knee point, carrying out smooth fitting on the capacity data by adopting a local weighted regression LOESS method, wherein the fitting function is as follows: ; in the formula, Is a slow decay phase fitting curve; for the total number of samples of the locally weighted regression method, An index representing the sample; To at the first Capacity values of the individual local sample points; As the weight coefficient of the light-emitting diode, Representing the first sample of the local weighted regression method The number of cycles of the local sample points; S5.3, modeling in an accelerated degradation stage, normalizing the cycle number and the capacity in the accelerated degradation stage after the knee point, and unifying the scales of different life intervals: ; ; in the formula, In order to normalize the number of cycles, Normalized capacity; Is the actual cycle number; Is the actual capacity value; migration learning is introduced based on the similarity of the battery aging mechanism: The migration function is specifically formed as follows: ; in the formula, A capacity amplitude adjustment coefficient representing a migration function; a time scale factor representing a migration function; a lateral translation term representing a migration function; representing the longitudinal offset of the migration function; normalized cycle number; wherein the migration parameter is determined by minimizing the following equation: ; in the formula, Representing an objective function of the transfer learning; to accelerate the number of samples in the decay phase; Represent the first Migration prediction capacity of individual samples; Capacity data for accelerating the decay phase; Represent the first Number of cycles of the individual samples; S5.4, double-index enhancement modeling, namely carrying out enhancement fitting on capacity fading in an accelerated fading stage by adopting an improved double-index function form after transfer learning correction, wherein the expression is as follows: ; in the formula, And Amplitude coefficients of slow and accelerated decay components, respectively; And Is a slow and accelerated decay rate parameter; For normalizing the cycle number, transfer learning the obtained parameters The method comprises the steps of correcting initial values of an amplitude coefficient and a decay rate parameter; And S5.5, curve fusion and parameter optimization, wherein fitting functions of the first stage and the second stage are spliced to establish an integral capacity fading prediction function, and the expression is as follows: ; in the formula, Is a transition smoothing coefficient; The number of knee point cycles; representing a slow decay phase fitting function; Is the actual cycle number; representing an accelerated decay phase fitting function; by minimizing the objective function: , wherein, Represent the first Global model predictors for the individual sample loop points, Represent the first Combining the parameter constraint of transfer learning output to realize the global optimization of the integral curve; Solving by adopting a DBO optimization algorithm to obtain an optimal parameter combination And the full life capacity prediction function obtained after optimization is as follows: ; in the formula, Is a transition smoothing coefficient; The number of knee point cycles; Representing the optimized slow decay phase fitting function; Is the actual cycle number; And representing the optimized fitting function of the accelerated decay phase.
5. The method for predicting capacity fade trajectories of two-stage lithium ion batteries based on residual life and knee points according to claim 1, wherein the calculation mode of the Euclidean distance and capacity deviation integral in S6 is as follows: Integral fitting curve obtained in step S5 And measured capacity data Interpolatively at the same cycle number And comparing, namely calculating Euclidean distance of each sampling point, wherein the calculation formula is as follows: ; Wherein, in the formula, Representation of post-optimization item Fitting capacity values of the individual samples; Represent the first Experimental observations of individual sample circulation points; Is the first Cycle points of the individual samples; define the integral form of the capacity deviation: ; Wherein, the The number of end cycles corresponding to the remaining life.

Description

Two-stage lithium ion battery capacity fading track prediction method based on residual life and knee point Technical Field The invention belongs to the technical field of lithium ion battery health state assessment and life prediction, and particularly relates to a two-stage lithium ion battery capacity fading track prediction method based on residual life and knee points. Background Capacity fade in lithium ion batteries is a central manifestation of battery performance degradation. As the number of battery uses increases, the battery capacity typically exhibits a non-linear decay characteristic. In particular, the capacity drops more slowly in the early stages of decay, whereas after the knee point, the battery enters the stage of accelerated decay. Knee points refer to turning points in the battery capacity decay curve, marking a significant increase in decay rate. Accurate prediction of knee position not only provides early performance early warning, but also significantly improves the accuracy of RUL and capacity prediction, enabling the battery management system to more accurately assess the state of health of the battery. The battery state of health assessment method based on data driving mainly relies on historical data and real-time data, and the capacity fading process of the battery is analyzed through technologies such as machine learning, statistical analysis and the like, so that the state of health of the battery is assessed. However, most of the current battery state of health assessment methods focus on the prediction of RUL and the fitting of capacity fading track, neglect the accurate prediction of knee point position, so that the fitting result of the capacity fading track is not accurate enough and dynamic change of the battery state of health cannot be comprehensively reflected, although the prediction methods based on deep learning can provide more accurate results in some scenes, the methods generally rely on a large amount of data for training, and have higher calculation force requirements, and most of the mainstream prediction methods focus on the prediction of a single task, lack of joint multi-aspect prediction, so that the degradation characteristics of different stages in the battery capacity fading process cannot be fully captured, especially in the accelerated fading stages before and after the knee point. Currently, the research on the high-precision prediction fitting of the two-stage capacity fade locus aiming at the joint prediction RUL and knee points is deficient. The method which accords with the actual application scene is established, and the technology for precisely fitting the capacity fading track of the lithium battery on the basis of the RUL and knee point combined prediction is developed, so that the method has important significance in improving the accuracy of the battery health state evaluation. Disclosure of Invention In order to solve the technical problems, the invention provides a two-stage lithium ion battery capacity decay track prediction method based on residual service life and knee points. In order to realize the technical scheme, the method comprises the following steps: s1, cyclic data acquisition and capacity fading sequence construction; And (3) carrying out multi-cycle charge and discharge tests on different batteries of the same model under the constant temperature condition until the discharge capacity is reduced to 80% of the rated capacity, adopting different charge strategies including constant current-constant voltage (CC-CV) and multi-stage constant current-constant voltage (C1-C2-CV) modes in the test process, stopping the multi-stage constant current when the SOC reaches 80%, standing the battery for 10 seconds, converting the charge state from multi-stage constant current charge to constant voltage charge, and recording the charge voltage, the cycle number and the discharge capacity. S2, data preprocessing and feature construction; performing wavelet noise reduction and smoothing on voltage and capacity data acquired by the lithium ion battery in a multi-cycle charge-discharge process, and eliminating abnormal cycle points; Extracting voltage characteristic parameters of charge-discharge stage to form voltage characteristic And combining capacity fading characteristic matrixForming a comprehensive input matrix:; Using belts The linear regression model of regularization term performs sparse constraint on the feature importance of the comprehensive input matrix, and the optimization objective function is as follows: in the formula, Is an intercept term; is a regression coefficient vector; Is a regularization coefficient; Is a target variable; for the total number of battery cycle period samples, An index representing the total number of battery cycle samples; Representing the total number of features, An index representing the total number of features; Represent the first The first battery cycle period sampleA characteristic value