Lithium ion battery health state estimation method based on GJO optimization multi-characteristic quantity
By using a GJO-optimized multi-feature estimation method, selecting the optimal features and constructing a CNN-BiGRU-AM neural network model, the accuracy problem of lithium-ion battery health state estimation is solved, and the predictive ability and applicability are improved.
Patent Information
- Application Number
- CN202411089234.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-09
- Publication Date
- 2024-11-22
AI Technical Summary
Existing methods for estimating the health status of lithium-ion batteries neglect the optimization and adjustment of feature quantities, resulting in low prediction accuracy in complex environments and failing to effectively utilize the intrinsic mechanisms and statistical properties of multiple feature quantities.
A multi-feature estimation method based on GJO optimization is adopted. Through feature data extraction, preprocessing, evaluation index calculation and neural network model construction, the optimal features are selected and a CNN-BiGRU-AM neural network model is established for health status estimation.
It improves the accuracy and applicability of lithium-ion battery health status estimation, enhances the predictive ability of the model, simplifies the calculation process, and is suitable for engineering practice.
Smart Images

Figure CN119001459A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of new energy batteries, and is specifically a method for estimating the health status of lithium-ion batteries based on multiple feature quantities optimized by GJO. The method selects appropriate optimal feature quantities for health status prediction by optimizing the statistical characteristics and internal mechanisms of feature quantities. The method is applicable to related technical fields such as life prediction, degradation analysis, and reliability analysis of new energy batteries. Background Art
[0002] With the huge consumption of fossil energy, how to efficiently utilize clean energy has become a topic of concern in academia and industry. Lithium-ion batteries are considered to be one of the most promising energy storage devices due to their long service life, low self-discharge rate, and high specific energy. However, during use, lithium-ion batteries have the problem of life decay. This type of capacity decay problem will cause irreversible degradation of lithium-ion batteries. If this degradation is not paid attention to in time, it may cause power system failures, and some may even lead to serious accidents. Therefore, efficient and accurate prediction of the health status of lithium-ion batteries is of great significance to the safe operation of power systems.
[0003] The external manifestations of lithium-ion battery aging include capacity degradation, increased internal resistance and decreased power. The lithium inventory loss, active material loss and conductivity loss inside the lithium-ion battery can reveal the deeper reasons for battery aging. Lithium inventory loss is mainly caused by solid electrolyte interface layer growth, lithium plating, particle cracking and other phenomena. Active material loss and conductivity loss are caused by electrode particle cracking, loss of electrical contact layer, decomposition of binder and other reasons. These aging behaviors will gradually deplete the lithium ions that can migrate during the battery charging and discharging process, and ultimately lead to a decrease in the available capacity of the lithium-ion battery.
[0004] The estimation level of the health status of lithium-ion batteries mainly depends on the extracted features and the algorithm of the model. Although the existing features ensure a high correlation with the health status of the battery, they ignore the evaluation of the alternative feature quantities from a mathematical perspective, which reduces the accuracy of the health status estimation. In addition, various deep learning models excel in different fields. Most of the data-driven methods currently used only consider the prediction level of the neural network and ignore the optimization and adjustment level of the parameters. Therefore, they do not have high prediction accuracy in complex working conditions. Summary of the invention
[0005] In view of the shortcomings of the existing technology, a health state estimation method for lithium-ion batteries based on GJO optimization of multiple feature quantities is proposed. This method comprehensively considers various alternative feature quantities and selects the optimal feature quantities and parameter quantities for lithium-ion battery health state estimation.
[0006] The technical solution adopted by the present invention to solve the technical problem is to provide a method for estimating the health state of a lithium-ion battery based on GJO optimization of multiple feature quantities, the method comprising the following steps:
[0007] Step 1: Extraction and preprocessing of feature data
[0008] Data related to voltage and capacity during the charging process of a lithium-ion battery are obtained as original feature quantities, wherein several original feature quantities are independent of each other and are in a series relationship; the deviation value of each original feature quantity data is eliminated according to the criterion that the correlation characteristic coefficient is higher than 0.9; then, new feature quantity data is added through accumulation processing, extreme value taking or root mean square value method, and the new feature quantity is used as an alternative feature quantity.
[0009] Step 2: Calculate the evaluation index of each candidate feature quantity, which includes correlation index, reliability index, time series index, trend index and stability index, among which the correlation index is the maximum eigenvalue λ of the reliable judgment matrix that meets the consistency requirements. max The absolute value of the corresponding eigenvector is used as the internal correlation measure of each candidate feature quantity; the Hurst index is used as the reliability index, Corr is used as the time series index, Mon is used as the trend index, and Rob is used as the stability index;
[0010] Step 3: normalize the values of each evaluation index, define the weighted sum of the normalized evaluation index values as the comprehensive evaluation index; use the maximum membership principle to determine the best weight combination, and use the best weight combination to obtain the comprehensive evaluation index value of each candidate feature quantity;
[0011] Step 4: Arrange the candidate feature quantities in descending order according to their comprehensive evaluation index values. The comprehensive evaluation index value of the n candidate feature quantities after arrangement is S i , all candidate feature quantities are selected based on feature engineering algorithm to obtain the final optimal feature quantity;
[0012] Step 5: Establish a GJO-optimized CNN-BiGRU-AM neural network model for health status estimation
[0013] A CNN-BiGRU-AM neural network model is constructed. The CNN-BiGRU-AM neural network model includes a CNN convolutional neural network composed of a convolutional layer, a pooling layer and a fully connected layer. The BiGRU structure and the attention mechanism AM are connected after the pooling layer. The input of the fully connected layer is the output of the attention mechanism AM. After being processed by the fully connected layer, the health status SOH is output.
[0014] The golden jackal algorithm GJO is used to optimize the learning rate, hidden layer nodes, and regularization coefficient of the CNN-BiGRU-AM neural network model; the optimized CNN-BiGRU-AM neural network model is used to predict the health status SOH.
[0015] The formula for battery health status expressed in terms of capacity is as follows:
[0016]
[0017] Among them, C1 is the predicted current battery capacity, C0 is the initial capacity, and the health status of the lithium-ion battery is estimated by the SOH value. Each set of optimal feature quantities in the data set corresponds to a health status SOH data for use in training models.
[0018] Furthermore, the CNN-BiGRU-AM neural network model includes a CNN convolutional neural network composed of a convolutional layer, a pooling layer and a fully connected layer. The BiGRU structure and the attention mechanism AM are connected after the pooling layer. The input of the fully connected layer is the output of the attention mechanism AM, and the health status SOH is output after processing by the fully connected layer.
[0019] Furthermore, the battery dataset from the Center for Advanced Lifecycle Engineering (CALCE) or the University of Oxford is used to train the optimized CNN-BiGRU-AM neural network model for lithium-ion battery health state estimation.
[0020] Furthermore, the golden jackal algorithm GJO is initialized using forward and reverse synchronous learning, and the specific process is:
[0021] Let X = (x1, x2, ..., x n ) is a set of random solutions in n-dimensional space. After reverse learning, the corresponding reverse solution X' is:
[0022]
[0023] Let Y = (y1, y2, ..., y n ) is a set of random solutions in the reverse solution of n-dimensional space. The result of the reverse solution is iterated forward to obtain the corresponding forward solution Y':
[0024]
[0025] Where: rand(0,1) represents a random number between 0 and 1, i∈[1,n], x i is an element of X, y' i is the element in the forward solution Y' obtained by forward learning the directional solution; x max is the maximum value of the random solution, x min is the minimum value of the random solution.
[0026] Furthermore, the expression of the prey energy consumption function E1 in the golden jackal algorithm GJO is
[0027]
[0028] Among them, T is the maximum number of iterations; c1 is a constant value, set to 1.5; t is the current number of iterations.
[0029] Furthermore, the number of convolutional layers is 2 to 4; the number of filters of the CNN convolutional neural network is 16 to 256; the number of layers of the BiGRU structure is 1 to 3; and the number of attentions in the attention mechanism AM is 8.
[0030] Compared with the prior art, the present invention has the following beneficial effects:
[0031] 1) The present invention comprehensively characterizes the evaluation indexes of the advantages and disadvantages of various candidate characteristic quantities of lithium-ion batteries, and selects the optimal characteristics that can characterize battery degradation. The optimal characteristics are more in line with the degradation trend, and solves the problem of estimating the battery health status of multiple characteristic quantities.
[0032] 2) The present invention takes into account the actual needs of lithium-ion battery health status estimation, the actual characteristics of multiple characteristic indicators, and the statistical characteristics and internal mechanisms when optimizing characteristic parameters. The GJO optimization fusion neural network model accelerates its early convergence speed and increases its optimization accuracy, thereby improving the accuracy of model prediction.
[0033] 3) The health status estimation method proposed in the present invention is simple to calculate, convenient to optimize and adjust parameters, and has better applicability and adjustability. BRIEF DESCRIPTION OF THE DRAWINGS
[0034] Figure 1 It is a practical flow chart of a method for estimating the health status of a lithium-ion battery based on multiple characteristic quantities optimized by GJO according to the present invention.
[0035] Figure 2 It is a correlation diagram between the normalized multiple feature quantities extracted by the present invention and the health status of the lithium-ion battery.
[0036] Figure 3 It is a comparison curve between the lithium-ion battery health status estimation method based on GJO optimization of multiple feature quantities of the present invention and the actual value; wherein (a) is the comparison result of the predicted value and the actual value of Cell3 predicted by using Oxford data sets Cell1 and Cell2, and (b) is the comparison result of the predicted value and the actual value of Cell5 predicted by using Oxford data sets Cell3 and Cell4.
[0037] Figure 4It is the error curve of comparing the health status estimation method of lithium-ion battery based on GJO optimization of multiple features of the present invention with other neural network models; wherein (a) is the error result of comparing the health status estimation method of lithium-ion battery based on GJO optimization of multiple features of the present invention with other neural networks; (b) is the error of the health status estimation method of lithium-ion battery based on GJO optimization of multiple features of the present invention alone; (c) is the error result of predicting Oxford data set Cell1 using Maryland data sets CS2_35, CS2_36 and CS2_37; (d) is the error result of predicting Maryland data set CS2_35 using Oxford data sets Cell1, Cell2 and Cell3.
[0038] Figure 5 It is a structural diagram of the CNN-BiGRU-AM neural network model. DETAILED DESCRIPTION
[0039] The technical solution of the present invention is further described below in conjunction with specific embodiments and drawings.
[0040] The present invention provides a method for estimating the health state of a lithium-ion battery based on GJO optimization of multiple characteristic quantities, the method comprising the following steps:
[0041] Step 1: Extraction and preprocessing of feature data
[0042] The data related to the voltage and capacity during the charging process of the lithium-ion battery are obtained as the original feature quantities. Several original feature quantities are independent of each other and are in series relationship. The deviation value of each original feature quantity data is eliminated according to the criterion that the correlation characteristic coefficient is higher than 0.9. After that, the feature quantity data is added by accumulating, taking extreme values or calculating the root mean square value. The newly added feature quantities are all candidate feature quantities. Table 1 shows the extracted original feature quantities.
[0043] Table 1 Extracted original features
[0044]
[0045] In this embodiment, after the original feature quantities are accumulated, candidate feature quantities are obtained as shown in Table 2, where F1-F9 represent candidate feature quantities respectively, and the correlation relationship data between the candidate feature quantities and the battery health status are as shown in Table 2. Figure 2 shown.
[0046] Table 2 Extracted candidate feature quantities
[0047]
[0048] Step 2: Calculate the evaluation index of each candidate feature quantity
[0049] 1) Use the angle analysis method to determine the contribution of each feature to battery degradation, classify the candidate feature quantities, and determine the impact of the candidate feature quantity i on product performance θ i , i = 1, 2, ..., m; n is the total number of candidate feature quantities; the degree of influence is usually divided into 1 to 9 levels, corresponding to values 1 to 9, the higher the value, the greater the degree of correlation;
[0050] 2) Using the obtained influence value θ i Construct the degree level information judgment matrix A according to the following formula:
[0051]
[0052] 3) Judge the consistency of the constructed judgment matrix A:
[0053] Calculate the maximum eigenvalue λ of the judgment matrix A max , and then calculate the consistency index CI=(λ max -n) / (n-1) and consistency ratio index CR=CI / RI;
[0054] Among them, RI is the random consistency index, which can be obtained by looking up Table 3;
[0055] When CR<1, it is considered to have passed the consistency test; otherwise, the impact value needs to be adjusted to change the judgment matrix until it passes the test and obtains a reliable judgment matrix. Table 3 gives the reference values of random consistency indicators.
[0056] Table 3 Random consistency index reference values
[0057]
[0058] 4) Calculate five evaluation indicators for each candidate feature quantity:
[0059] The correlation evaluation index is: the maximum eigenvalue λ of the reliable judgment matrix obtained above max The absolute value of the corresponding eigenvector is used as a measure of the internal correlation of each candidate feature quantity;
[0060] Reliability index, time series index, trend index and stability index are important indicators that characterize the statistical characteristics of feature quantities, among which reliability index: evaluates the strong correlation of feature groups;
[0061] Time series index: measures the correlation between candidate feature data and time;
[0062] Trend indicator: measures the increasing or decreasing trend of the candidate feature quantity data;
[0063] Stability index: measures the tolerance of candidate features to abnormal data;
[0064] 1) Reliability index calculation: The Hurst index is used as the reliability index. The index is obtained by the range analysis method. The process is as follows:
[0065] ①Given a length of m i The candidate feature quantity I i , divide this candidate feature into L adjacent sub-intervals of length l. If it cannot be evenly divided, the redundant data will be discarded; any sub-interval is represented by Q α , in Q α The elements in are represented by N k,α , k=1,2,…,l;α=1,2,…,L;N k,α The mean of is:
[0066]
[0067] ②Subinterval Q α The cumulative intercept of the mean is defined as:
[0068]
[0069] ③The range is defined as:
[0070]
[0071] ④Subinterval Q α The standard deviation of is defined as:
[0072]
[0073] ⑤(R / S) l Defined as:
[0074]
[0075] ⑥l starts from 3, repeat steps ① to ⑤ until l = m i , get the sequence (R / S) l , l=3,…,m i . Taking log(l) as the independent variable, log[(R / S) l ] is used as the dependent variable for linear regression:
[0076] log[(R / S) l ]=H·log(l)+log(c)
[0077] The slope H in the above formula is the candidate feature quantity I i Hurst index, c is a constant;
[0078] 2) Alternative feature quantity I iThe calculation formulas for the timing Corr, trend Mon and stability Rob indicators are as follows:
[0079]
[0080] Among them, I i,j,T and I i,j,R To select the candidate feature quantity I i,j The trend term and random term obtained by empirical mode decomposition are combined with the degradation characteristics of lithium-ion batteries. The trend term is determined by the degradation mechanism of lithium-ion batteries; the random term is determined by uncertainty factors; and I i,j =I i,j,T +I i,j,R ;I i,j',T To select the candidate feature quantity I i,j' The trend term obtained by empirical mode decomposition, I i,j' is the i-th candidate feature quantity I i The j'th monitoring value of i -1; δ(*) is the unit step function, and its expression is as follows:
[0081]
[0082] In this example, in order to obtain the evaluation results of each performance evaluation index, the influence degree of the candidate feature quantity is scored respectively. The scores are shown in Table 4. This score represents the influence degree value θ i .
[0083] Table 4 Alternative feature importance score table
[0084]
[0085]
[0086] Substitute the scores in Table 4 into the degree level information judgment matrix A, and obtain the reliable judgment matrix after consistency judgment. Then calculate the maximum eigenvalue λ of the reliable judgment matrix max The absolute value of the corresponding feature vector is used as the internal correlation measure of each feature quantity, and the score of the internal correlation of each candidate feature quantity can be obtained. Then, the scores of reliability, time series, trend and stability are obtained according to the above formulas, and finally the evaluation index values of each candidate feature quantity can be obtained, as shown in Table 5.
[0087] Table 5 Evaluation index values of each candidate feature quantity
[0088]
[0089] Step 3: Calculate the comprehensive evaluation index of alternative feature quantities
[0090] First, the following formula is used to normalize the evaluation index values obtained in step 2:
[0091]
[0092] Where q is the evaluation index number, q = 1, 2, 3, 4, 5, corresponding to the five evaluation indicators in step 2, namely, correlation, stability, time series, trend, and reliability index; I i represents the i-th candidate feature quantity, i = 1, 2, ..., n, n is the total number of candidate feature quantities, then, Z q (I i ) is the candidate feature quantity I i Evaluation index value in the qth aspect; Z' q (I i ) is the normalized candidate feature quantity I i The evaluation index value in the qth aspect; are the minimum and maximum values of the evaluation index values of all candidate feature quantities in the qth aspect respectively; furthermore, the candidate feature quantity I i The comprehensive evaluation index Z(I i ) is expressed as follows:
[0093] Z(I i )=w1Z′1(I i )+w2Z′2(I i )+w3Z′3(I i )+w4Z′4(i)+w5Z′5(i)
[0094] Among them, Z'1(I i ),Z'2(I i ),Z'3(I i ),Z'4(I i ),Z'5(I i ) are normalized data of the evaluation index values of five aspects: internal correlation, stability, time series, trend and reliability; w1, w2, w3, w4, w5 are the index weights, and there are:
[0095] w1+w2+w3+w4+w5=1
[0096] Define the membership degree DM under the e-th weight combination e for:
[0097]
[0098] Where e is the weight combination number, e = 1, 2, ..., N W , that is, the total weight is N W weighted combinations; It represents the candidate feature quantity with the highest comprehensive evaluation index under the e-th weight combination. Its calculation formula is as follows:
[0099]
[0100] Membership Degree DM e It represents the minimum normalized value of the normalized values of the five evaluation index values of the candidate feature quantities with the highest comprehensive evaluation index under the e-th weight combination;
[0101] Represents the candidate feature quantity The normalized result of the normalized value of the qth aspect evaluation index value under the eth weight combination is calculated as follows:
[0102]
[0103] Among them, Z' q,max and Z' q,min They represent the normalized values of the maximum and minimum values of the evaluation index of all candidate feature quantities in the qth aspect, respectively. The calculation formula is as follows:
[0104]
[0105] Weight combination W e =(w1,w2,w3,w4,w5), all its combinations are as follows:
[0106]
[0107] Where W = 5 × Q, Q is a positive integer; c 1,e 、c 2,e 、c 3,e 、c 4,e 、c 5,e For any 1,e +c 2,e +c 3,e +c 4,e +c 5,e ) / W=1;
[0108] Different weight combinations correspond to different membership values, and the group with the largest membership value indicates that this weight combination is the best combination; however, the largest membership value may be distributed in different weight combinations. Therefore, the selection of weight combinations in different situations is set according to the maximum membership principle, which has the following characteristics:
[0109] 1) If the maximum membership degree corresponds to a set of weight combinations, then the weight combination is selected as the final result;
[0110] 2) If the maximum membership degree corresponds to multiple weight combinations, the weight combination that maximizes the comprehensive evaluation index of the candidate feature quantity is selected as the final result.
[0111] Using the selected optimal weight combination, the comprehensive evaluation index value of each candidate feature quantity is obtained;
[0112] Table 6 Evaluation index values of candidate feature quantities
[0113]
[0114] Step 4: Optimization of candidate feature quantities
[0115] The feature engineering algorithm is used to select all candidate feature quantities. The specific process is as follows:
[0116] 1) Arrange the candidate feature quantities in descending order according to their comprehensive evaluation index values, and record the comprehensive evaluation index value of the n candidate feature quantities after arrangement as S i ;
[0117] 2) Select 1 to i in the order of comprehensive evaluation index values from large to small b The corresponding candidate feature quantity is the effective feature quantity to characterize the degradation process, and the remaining candidate feature quantities are discarded. b ≤n, determine the value based on actual conditions.
[0118] The five effective feature quantities selected in this example are:
[0119] 1) Equal current drop rate;
[0120] 2) Valley voltage;
[0121] 3) valley slope;
[0122] 4) Valley temperature;
[0123] 5) Left peak factor.
[0124] Step 5: Establish the GJO optimized CNN-BiGRU-AM neural network model
[0125] According to the characteristics of the neural network structure, the network is integrated to predict the performance of multiple feature quantities. The basic process is as follows:
[0126] 1) Establish a GJO optimization model;
[0127] The golden jackal optimization algorithm (GJO) is an optimization algorithm inspired by the foraging behavior of golden jackals in nature; this algorithm focuses on multiple local optimal solutions in the search space to improve the comprehensiveness of the search. Adding this algorithm can more easily turn the problem of characteristic quantity into a problem of local optimal solution, thereby accelerating the parameter adjustment process of the health status characteristics of lithium-ion batteries.
[0128] GJO is a population-based method where the initial solutions are uniformly distributed in the search space as first trials:
[0129] Y0=Y min +rand(Y max -Y min )
[0130] where Y max and Y min are the upper and lower bounds of the variable, and rand is a uniform random vector in the range 0 to 1. Initialization creates the initial matrix Prey, where the first and second adaptations are jackal pairs. The prey is displayed as follows:
[0131]
[0132] Y i,h represents the jth dimension of the i-th prey. There are n preys and d variables in total. The fitness (objective) function is used to estimate the fitness value of each prey during the optimization process, and the following matrix collects the fitness values of all preys:
[0133]
[0134] where F OA is the matrix that stores the fitness of each prey, and f is the objective function. The fittest one is called the male jackal, and the second fittest one is called the female jackal. A pair of jackals obtain the corresponding prey positions.
[0135] The search process in GJO starts with the creation of an arbitrary prey population. In an iterative process, the possible locations of prey are estimated by male and female jackal hunting pairs. Each candidate in the population has its relative distance to the jackal updated.
[0136] 2) Building a CNN model;
[0137] Convolutional neural network (CNN) is a typical artificial intelligence network structure, which is widely used in processing spatial data. CNN has three layers, namely convolution layer, pooling layer and fully connected layer. The convolution layer extracts local features in two-dimensional data and learns feature information from the input data; the pooling layer applies convolution operation to replace matrix operation and extracts complex features; the fully connected layer classifies the features extracted by the previous two layers.
[0138] 3) Establish the BiGRU model;
[0139] On the basis of GRU, forward and reverse hidden layer structures are added to form the BiGRU structure, which avoids the problem of gradient discreteness and makes the calculation more efficient. The expression of the model is as follows:
[0140]
[0141] h t is the current hidden layer state. is the forward hidden layer state, is the inverted hidden layer state. x is the current input, t is the time, is the forward hidden state at time t-1. is the state of the last hidden layer at time t+1. i ,i∈(1,2,…,6) is the weight of each layer.
[0142] 4) Establishing AM model;
[0143] Adding BiGRU layer and attention mechanism after the pooling layer of CNN to increase the weight of the main features enables the model to adjust automatically as a whole, thereby optimizing the dependency between each feature and time step, thereby improving the model's resistance to abnormal capacity mutations. In addition, adding the attention mechanism AM to BiGRU can strengthen key information and weaken useless information. The expression formula of the attention mechanism is as follows:
[0144] e t =tanh(ω i *h t +b t )
[0145]
[0146] s t =a t h t
[0147] ω i represents the weight, b represents the bias term. s represents the weight of each attribute. After paying attention to the weight a t The normalized result is output as the final result.
[0148] 5) Establish the GJO optimized CNN-BiGRU-AM model;
[0149] During exploration, the jackal's relative position to each prey item in the population is as follows:
[0150]
[0151] Where t represents the current iteration number. Indicates the position of the i-th individual in the current iteration, They represent the male and female jackals in the current iteration, that is, the best and second-best individuals in the population. Therefore, there is a pair of position vectors corresponding to each of them. E is the escape ability, which is calculated as follows:
[0152] E=E1·E0
[0153] Among them, E1 is used to reduce the prey's escape energy, and E0 is the initial escape energy. The calculation is as follows:
[0154] E0=2·r-1
[0155]
[0156] Among them, r is a random number between [0,1]; T is the maximum number of iterations; c1 is a constant value, set to 1.5. Therefore, E1 decreases from 1.5 to 0 throughout the iteration, changing the prey energy consumption function E1 into a nonlinear decreasing function.
[0157] rl represents a D-dimensional random vector based on Levy distribution, and D represents the problem dimension. The calculation is as follows:
[0158]
[0159]
[0160] Γ(y)=(y-1)!
[0161] Among them, μ and γ are random values between (0,1), σ is a constant set to 1.5; β is the default constant, set to 1.5, and Γ is the standard Gamma function integration operation.
[0162] Step 5: Use the improved golden jackal algorithm to optimize the parameters of the CNN-BiGRU-AM neural network model
[0163] The Golden Jackal Algorithm (GJO) is a newly proposed metaheuristic algorithm, which is inspired by the changes in a series of behaviors of the golden jackal during the foraging process. It can be divided into the initialization stage, the exploration stage and the development stage. Compared with other metaheuristic algorithms, the Golden Jackal Algorithm has the advantages of fewer adjustment parameters, strong optimization ability and easy implementation. The original algorithm uses random initialization, which may narrow the initial search range of the population and fail to achieve the effect of searching the entire feasible domain. The initial solution is interactively learned in both positive and negative directions to improve the distribution ability of the initial solution in the search domain.
[0164] 1) Forward and reverse synchronous learning initialization
[0165] Let X = (x1, x2, ..., x n ) is a set of random solutions in n-dimensional space. After reverse learning, the corresponding reverse solution X' is:
[0166]
[0167] Let Y = (y1, y2, ..., y n ) is a set of random solutions in the reverse solution of n-dimensional space. The result of the reverse solution is iterated forward to obtain the corresponding forward solution Y':
[0168]
[0169] Where: rand(0,1) represents a random number between 0 and 1, i∈[1,n], x i is an element of X, y' i is the element in the forward solution Y' obtained by forward learning the directional solution; x max is the maximum value of the random solution, x min is the minimum value of the random solution.
[0170] 2) Golden Jackal Algorithm Exploration Phase
[0171] At this stage, the way golden jackals search for prey is proposed. According to the nature and instinct of golden jackals to search and wait for prey, the hunt is led by males, and females follow males. Male and female golden jackals follow prey and update their own positions.
[0172] 3) Golden Jackal Algorithm Development Phase
[0173] At this stage, the method of golden jackals to capture prey is proposed. When the prey finds that it is being harassed by the golden jackals, it starts to escape and hide, and its energy gradually decreases during the escape and hide process, and the golden jackals will capture the prey.
[0174] The improved golden jackal algorithm adopts forward and reverse synchronous learning initialization and changes the prey energy consumption function to a nonlinear decreasing function. Other settings and operations are the same as the conventional ones.
[0175] The specific process of optimizing the CNN-BiGRU-AM neural network model using the improved golden jackal algorithm is as follows:
[0176] Step 1: Initialize parameters, set the number of golden jackals N, the maximum number of algorithm iterations, and solve the lower and upper bounds of the variable x min 、x max .
[0177] Step 2: Use forward and reverse synchronous learning to initialize the population and determine the variable initialization position.
[0178] Step 3: Calculate the population fitness value and record the current optimal fitness value.
[0179] Step 4: Update the golden jackal prey position and the corresponding golden jackal position.
[0180] Step 5: Use the improved prey energy consumption function E1 to calculate the escape energy E of the current prey. When the absolute value of the escape energy E is less than 1, the golden jackal uses the development phase to capture the prey. When the absolute value of the escape energy E is greater than 1, the exploration phase is used to search for prey and calculate the final position of the golden jackal.
[0181] Step 6: Add 1 to the number of iterations and determine whether the current number of iterations is greater than the set maximum iteration value. If not, jump to step 3 and continue iterative calculation; if yes, the iteration ends and the optimal fitness value and global optimal solution obtained after the last iterative calculation are output, thereby obtaining the optimal learning rate, hidden layer nodes, and regularization coefficient of the CNN-BiGRU-AM neural network model.
[0182] Example 1
[0183] In order to verify the effectiveness of the proposed GJO optimized multi-feature lithium-ion battery health state estimation method, twelve sets of battery data from two different types of data sets were used in the experiment, which came from the Center for Advanced Lifecycle Engineering (CALCE) and Oxford University respectively. Table 7 gives an overview of the two battery data sets.
[0184] The GJO algorithm parameter settings are as follows: the population size is 100 and the maximum number of iterations is 1000. The scheduling period is 1 day, which is divided into 96 time periods, that is, recording once every 15 minutes. Table 8 gives the specific parameter settings of the neural network.
[0185] Table 7 Overview of the two battery datasets
[0186]
[0187] Table 8 Specific parameters of each part of the GJO optimized CNN-BiGRU-AM neural network model
[0188]
[0189]
[0190] The present invention uses root mean square error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE) and determination coefficient (R 2 ) is used to measure the prediction performance. First, the effectiveness of the improved golden jackal algorithm (GJO) for solving the model is verified. GJO is used to optimize CNN-BiGRU-AM, CNN-SAM-LSTM, BiGRU-AM and BP neural network to solve the optimal solution. Under the same parameter settings and experimental conditions, 10 experiments are conducted to eliminate randomness, and the average values are taken for comparison. At the same time, the standard deviation of each algorithm is compared. The comparison results are shown in Table 9.
[0191] Table 9 Specific data of comparison algorithm optimization results
[0192]
[0193] From the comparison results in Table 9, it can be seen that the GJO optimization algorithm has the best optimization accuracy compared with other algorithms. The average value of the GJO algorithm optimization is 1.619%, and the root mean square error of the solution after the GJO algorithm optimization is 1.344%.
[0194] To verify the effectiveness of the neural network combination, the training set and test set are changed, and the Maryland and Oxford datasets are added for comparison. The results are shown in Table 10.
[0195] Table 10 Comparative data of different neural network tests
[0196]
[0197]
[0198]
[0199] As shown in Table 10, the GJO optimized CNN-BiGRU-AM neural network model has excellent performance in predicting the health status of lithium-ion batteries, achieving the purpose of estimating the health status of lithium-ion batteries with multiple features, meeting the actual needs of engineering, solving the problem of estimating the health status of lithium-ion batteries with multiple features, taking into account the actual needs, ensuring that the analysis process is scientific, reasonable and close to the actual situation, and the health status estimation results are credible. At the same time, the proposed method is simple to calculate, convenient for engineering and technical personnel to use, and has strong application value.
[0200] Any matters not described in the present invention are applicable to the prior art.
Claims
1. A method for estimating the health status of a lithium-ion battery based on GJO optimization of multiple feature quantities, the method comprising the following steps: Step 1: Extraction and preprocessing of feature data The data related to the voltage and capacity during the charging process of the lithium-ion battery are obtained as the original feature quantities, wherein the several original feature quantities are independent of each other and are in a series relationship; the deviation value of each original feature quantity data is eliminated according to the criterion that the correlation characteristic coefficient is higher than 0.9; then, the feature quantity data is added by accumulation processing, extreme value taking or root mean square value method, and the added feature quantity is used as the candidate feature quantity; Step 2: Calculate the evaluation index of each candidate feature quantity, which includes correlation index, reliability index, time series index, trend index and stability index, among which the correlation index is the maximum eigenvalue λ of the reliable judgment matrix that meets the consistency requirements. max The absolute value of the corresponding eigenvector is used as the internal correlation measure of each candidate feature quantity; the Hurst index is used as the reliability index, Corr is used as the time series index, Mon is used as the trend index, and Rob is used as the stability index; Step 3: normalize the values of each evaluation index, define the weighted sum of the normalized evaluation index values as the comprehensive evaluation index; determine the best weight combination, and use the best weight combination to obtain the comprehensive evaluation index value of each candidate feature quantity; Step 4: Arrange the candidate feature quantities in descending order according to their comprehensive evaluation index values. The comprehensive evaluation index value of the n candidate feature quantities after arrangement is S i , further screen and select all candidate feature quantities to obtain the final optimal feature quantity; Step 5: Establish a GJO-optimized CNN-BiGRU-AM neural network model for health status estimation Construct a CNN-BiGRU-AM neural network model, and use the golden jackal algorithm GJO to optimize the learning rate, hidden layer nodes, and regularization coefficient of the CNN-BiGRU-AM neural network model; use the optimized CNN-BiGRU-AM neural network model to predict the health status SOH; the input of the CNN-BiGRU-AM neural network model is the optimal feature value determined in step 4.
2. The estimation method according to claim 1, It is characterized in that The CNN-BiGRU-AM neural network model includes a CNN convolutional neural network composed of a convolutional layer, a pooling layer and a fully connected layer. The BiGRU structure and the attention mechanism AM are connected after the pooling layer. The input of the fully connected layer is the output of the attention mechanism AM, and the health status SOH is output after being processed by the fully connected layer.
3. The estimation method according to claim 1, It is characterized in that The optimized CNN-BiGRU-AM neural network model is trained using battery datasets from the Center for Advanced Lifecycle Engineering (CALCE) or the University of Oxford for lithium-ion battery health state estimation.
4. The estimation method according to claim 1, It is characterized in that The golden jackal algorithm GJO uses forward and reverse synchronous learning for initialization. The specific process is: Let X = (x 1 ,x 2 ,...,x n ) is a set of random solutions in n-dimensional space. After reverse learning, the corresponding reverse solution X' is: Let T = (y 1 ,y 2 ,...,y n ) is a set of random solutions in the reverse solution of n-dimensional space. The result of the reverse solution is iterated forward to obtain the corresponding forward solution Y': Where: rand(0,1) represents a random number between 0 and 1, i∈[1,n], x i is an element of X, y' i is the element in the forward solution Y' obtained by forward learning the directional solution; x max is the maximum value of the random solution, x min is the minimum value of the random solution.
5. The estimation method according to claim 1, It is characterized in that Prey energy consumption function E in the golden jackal algorithm GJO 1 The expression is Where T is the maximum number of iterations; c 1 is a constant value, set to 1.5; t is the current iteration number.
6. The estimation method according to claim 1, It is characterized in that The number of convolutional layers is 2 to 4; the number of filters of the CNN convolutional neural network is 16 to 256; the number of layers of the BiGRU structure is 1 to 3; and the number of attentions in the attention mechanism AM is 8.
Citation Information
Cited By
Model training method, battery SOC estimation method, electronic equipment and storage medium
CN121522468A
