Lithium ion battery health state prediction method for optimizing CNN-BiLSTM-AM model based on triangular topological polymerization optimization algorithm

Abstract

The invention relates to a lithium ion battery health state prediction method for optimizing a CNN-BiLSTM-AM model based on a triangular topology polymerization optimization algorithm, and belongs to the field of lithium ion battery health state estimation. Acquiring a data set, extracting health factors from the data set, establishing feature sample data after correlation analysis, and dividing the feature sample data into a training set and a test set; a CNN-BiLSTM-AM model used for predicting the state of health of the lithium ion battery is established; screening initial key parameters of the model, and optimizing the initial key parameters by adopting a triangular topology aggregation optimization algorithm; training and testing the model by combining the training set and the test set according to the optimized initial key parameters; and acquiring data of the lithium ion battery in real time, and predicting the health state of the lithium ion battery in real time by using the trained model. According to the method, the initial optimal parameter group of the model is searched by using the TTAO algorithm, so that the training speed and the learning ability of the model are improved, and the health state of the lithium battery can be predicted more quickly and more accurately.

Description

Technical Field

[0001] The present invention belongs to the field of lithium-ion battery health state estimation, and relates to a lithium-ion battery health state prediction method based on a CNN-BiLSTM-AM model optimized by a triangle topology aggregation optimization algorithm. Background Art

[0002] In the technical field of lithium-ion battery state of health (SOH) prediction, the application of CNN-BiLSTM-AM (convolutional neural network-bidirectional long short-term memory network-attention mechanism) model is gradually becoming a research hotspot. With the continuous development of deep learning technology, its advantages in processing complex time series data are becoming increasingly prominent, which provides strong support for the accurate prediction of lithium-ion battery SOH.

[0003] As an important device for modern energy storage, the health status of lithium-ion batteries is directly related to the safety and reliability of the equipment. Therefore, accurate prediction of the SOH of lithium-ion batteries is of great significance for optimizing battery use, maintenance and management. However, traditional prediction methods often fail to achieve ideal prediction results when dealing with high-dimensional, nonlinear, and time-varying lithium-ion battery performance data.

[0004] The CNN-BiLSTM-AM model combines the advantages of convolutional neural networks (CNN), bidirectional long short-term memory networks (BiLSTM), and attention mechanisms (AM). CNN can effectively extract spatial features in battery performance data, such as the correlation pattern between different parameters; BiLSTM can capture long-term dependencies and contextual information in time series data, thereby more comprehensively understanding the changing trend of battery performance data; and the attention mechanism can further highlight key information and improve the prediction accuracy and robustness of the model.

[0005] Through the CNN-BiLSTM-AM model, the historical data and current status of lithium-ion batteries can be deeply analyzed, and the potential laws in the data can be mined to achieve accurate prediction of lithium-ion battery SOH. This not only helps to improve the safety and reliability of equipment, but also provides strong support for enterprises to reasonably arrange battery replacement plans and reduce operating costs.

[0006] Since the use of data-driven models to predict the health status of lithium batteries is affected by the accuracy of the network model used by the data-driven model, the initial parameter values of the neural network, especially the initial weights and thresholds, determine the performance of the network, which directly affects the convergence speed, convergence performance and final learning effect of the network. Therefore, maximizing the optimization of the random initial value of the network is the key issue in using machine learning methods to improve the accuracy of lithium battery health status estimation.

[0007] A large number of efforts have been made in the health state prediction models of lithium batteries, including data-driven models, circuit models, and electrochemical models. The data-driven models are related to the selection and optimization algorithms. The optimization algorithms are used to find the optimal key parameters and improve the overall performance of the models. However, in conventional optimization algorithms, there are problems such as being prone to falling into local optimal solutions and having low convergence accuracy, resulting in poor exploration ability in the solution space, imbalance between local and global optimization, and affecting the overall estimation of the health of lithium batteries by the subsequent network models. Summary of the Invention

[0008] In view of this, the purpose of the present invention is to provide a method for predicting the health state of lithium-ion batteries by optimizing the CNN-BiLSTM-AM model based on the triangular topology aggregation optimization algorithm. By using the TTAO algorithm, it is committed to finding the initial optimal parameter group of the network to improve the network's ability to learn typical features of large-scale data inputs and learning efficiency, improve the training speed and learning ability of the CNN-BiLSTM-AM network, and enable it to predict the health state of lithium batteries more quickly and accurately.

[0009] To achieve the above purpose, the present invention provides the following technical solutions:

[0010] A method for predicting the health state of lithium-ion batteries by optimizing the CNN-BiLSTM-AM model based on the triangular topology aggregation optimization algorithm, the method comprising the following steps:

[0011] Obtain a data set, extract the health factor HI from the data set, establish feature sample data after correlation analysis, and divide it into a training set and a test set;

[0012] Establish a CNN-BiLSTM-AM model for predicting the health state of lithium-ion batteries;

[0013] Screen the initial key parameters of the CNN-BiLSTM-AM model, and optimize the initial key parameters using the triangular topology aggregation optimization algorithm;

[0014] According to the optimized initial key parameters, train and test the CNN-BiLSTM-AM model in combination with the training set and the test set;

[0015] Obtain lithium-ion battery data in real time, and use the trained TTAO-CNN-BiLSTM-AM model to predict the health state of lithium-ion batteries in real time.

[0016] Furthermore, the way to obtain the data set includes experimental acquisition and using public data sets;

[0017] According to the measured charge-discharge curve, the health index HI is extracted. The specific extraction method is as follows: The constant-current charging time of the battery is extracted as HI1; the equal-voltage discharge time is extracted as HI2;

[0018] In the data preprocessing stage, for the outliers in the extracted health index HI, the mean value of adjacent data points is used for filling;

[0019] Then, for the extracted health index HI data, Pearson correlation coefficient and Spearman coefficient are used for correlation analysis. Through correlation analysis, the correlation degree between the health index and SOH is quantitatively analyzed;

[0020] For the data after correlation analysis, data normalization processing is carried out. The data is adjusted through linear transformation, and the minimum-maximum normalization method is adopted to scale the data to the range of 0 to 1; the normalized data is used as the feature sample data composed of the health index HI, and it is divided into a training set and a test set according to a certain proportion.

[0021] Furthermore, the established CNN-BiLSTM-AM model for predicting the health state of lithium-ion batteries includes an input layer Input, a convolutional neural network layer CNN, a bidirectional long short-term memory network layer BiLSTM, an attention mechanism layer AM, and a fully connected output layer FC;

[0022] The input layer Input receives the input data of the battery;

[0023] The convolutional neural network layer CNN includes a convolutional layer Conv and a pooling layer pooling. The convolutional layer Conv extracts local features in the input data through convolutional kernels; the pooling layer pooling performs pooling operations on the output of the convolutional layer;

[0024] The bidirectional long short-term memory network layer BiLSTM includes several long short-term memory networks LSTM with bidirectional structures. The long short-term memory network LSTM is used to process sequence data; the bidirectional structure helps BiLSTM process sequence information in both forward and backward directions;

[0025] In the attention mechanism layer AM, the Softmax function is used to normalize the output of BiLSTM to obtain the attention weights at each moment; then the attention weights are used to perform weighted summation on the output of BiLSTM to obtain a context vector containing important information;

[0026] The fully connected output layer FC comprehensively analyzes the output of the previous layers and obtains the prediction result, that is, the health state SoH of the lithium-ion battery, through a fully connected manner.

[0027] Furthermore, the key parameter of the obtained CNN-BiLSTM-AM model is the number of hidden units hd , the initial learning rate l r and the dropout rate D u , and the triangular topology aggregation optimization algorithm optimizes the number of hidden units h d , the initial learning rate l r and the dropout rate D u The specific process is as follows:

[0028] Population initialization: Initialize the key parameters, the number of hidden units h d , the initial learning rate l r and the dropout rate D u for each individual in the population, then determine the population size and the variable dimension, divide the population into multiple triangular topology units, and obtain the initial solution set X1;

[0029] Formation of triangular topology units: According to each first vertex X in the solution set X1 i,1 , determine the second vertex X of the triangular topology unit i,2 , the third vertex X i,3 and the fourth vertex X inside the triangular topology unit i,4 ;

[0030] General aggregation: Simulate the gene crossover mechanism for general aggregation, and information interaction occurs between the best individual in each triangular topology unit and the best individual randomly selected from other units to update the optimal individual;

[0031] Local aggregation: The updated optimal or sub-optimal individual forms a temporary triangular topology unit with the other two vertices, and local aggregation is performed on the temporary triangular topology unit to update the optimal individual again;

[0032] Judge whether the iteration ends. If it does not end, return to step S33 to continue the iteration. If it ends, output the optimal individual.

[0033] Furthermore, during the population initialization process, the parameters h to be optimized d , l r , D u are encoded as a multi-dimensional vector x = [h d , l r , D u , and each multi-dimensional vector is an individual of the population. Among them, the number of hidden units h d is a discrete parameter, and its range is expressed as (h dmin , h dmax ), the initial learning rate l r and the dropout rate D u are continuous parameters, and their ranges are expressed as (l rmin , l rmax ) and (D umin , D umax), from which the population size \(N\) and the variable dimension \(D\) of the initial population are obtained, and then an initial solution set \(X_1\) of \(N / 3\) solutions is generated to provide an initial solution space for the subsequent global search and local optimization phases; where each search individual in the initial solution set \(X_1\) is represented as:

[0034] X i,1 =r o ×(UB - LB)+LB

[0035] In the formula, \(X\) i,1 is the first search individual of the \(i\)-th triangular topological unit; \(LB\) and \(UB\) are the lower and upper bounds of the search space respectively; \(r_0\) is a random value from 0 to 1.

[0036] Furthermore, during the formation of the triangular topological unit, the way to generate the second vertex \(X\) i,1 and the third vertex \(X\) i,2 is as follows: i,3

[0037] X i,2 =X i,1 +l×f(θ)

[0038]

[0039] In the formula, \(l\) is the size of the triangular topological unit, which gradually decreases as the number of iterations increases; \(f(θ)\) and are the direction vectors of two sides obtained using \(X_1\);

[0040] \(l\) is defined as:

[0041]

[0042] In the formula, \(t\) and \(T\) are the current iteration number and the maximum iteration number respectively;

[0043] The specific formulas for \(f(θ)\) and are:

[0044] \(f(θ)=[\cosθ_1,\cosθ_2,\cdots,\cosθ\) D

[0045]

[0046] In the formula, \(θ\) j , \(j = 1,2,\cdots,D\) are random numbers from 0 to \(π\);

[0047] Each group of triangular topological units with \(X_1\), \(X_2\), and \(X_3\) as units aggregate into a fourth vertex in a linear weighted manner:

[0048] X i,4 =r_1×X​​i,1 +r2×X i,2 +r3×X i,3

[0049] Wherein, r1, r2 and r3 are random numbers between 0 and 1, and the sum of the three is always equal to 1.

[0050] Furthermore, the general aggregation process is as follows:

[0051] Generate a new individual by linearly combining different weights between the best individual and the random individual:

[0052] X t+1 i,n1 = r4×X t i,b +(1 - r4)×X t r,b

[0053] Wherein, X t r,b is the individual randomly selected at the t-th iteration; X t i,b is the best individual; r4 is a random number between 0 and 1;

[0054] Compare the fitness value of the newly generated individual with the current best individual and the sub-optimal unit, and update the optimal individual according to the greedy strategy. The specific formula is:

[0055]

[0056] Wherein, represents the fitness value of the newly generated individual at the (t + 1)-th time, represents the fitness value of the optimal individual at the (t + 1)-th iteration, X t+1 i,b represents the optimal individual at the (t + 1)-th iteration, represents the fitness value of the newly generated individual of unit i at the (t + 1)-th time, represents the fitness value of the sub-optimal individual at the (t + 1)-th iteration, X t+1 i,sb represents the sub-optimal individual at the (t + 1)-th iteration.

[0057] Furthermore, the local aggregation process is as follows:

[0058] Re-search each temporary triangular topology unit within the local area, and the new vertex is calculated as follows:

[0059] X t+1 i,n2 = X t+1 i,b + a(X t+1i,b -X t+1 i,sb )

[0060] In the formula, a is the size of the aggregation range, expressed as:

[0061]

[0062] Compare the fitness values of adjacent vertices in the local search to determine whether to perform position update. If the newly generated individual is better than the original individual, perform position adjustment; otherwise, keep the original position unchanged:

[0063]

[0064] In the formula, represents the fitness value of the newly generated individual of the i-th cell at the (t + 1)-th time.

[0065] Furthermore, it is characterized in that: the fitness function in the triangular topology aggregation optimization algorithm is set as the root mean square error between the SOH prediction value and the actual value, and its calculation formula is:

[0066]

[0067] In the formula, z i represents the SOH prediction value of the battery for the i-th sample; z i ′ represents the true value of the battery SOH, and M is the total number of samples used for training.

[0068] The beneficial effects of the present invention are as follows:

[0069] The present invention has a significant improvement compared with the traditional prediction method. Its triangular topology aggregation optimization algorithm further optimizes the initial key parameters of the CNN-BiLSTM-AM model, thereby ensuring that the model can accurately capture key information in the complex and changeable lithium-ion battery performance data and improving the prediction accuracy. This technical effect enables the present method to provide more reliable prediction results for the health state of lithium-ion batteries and provides strong support for the optimal use, maintenance, and management of equipment.

[0070] The generalization ability of the model of the present invention is significantly improved. Since the triangular topology aggregation optimization algorithm can search globally and gradually focus on finding the optimal solution as the number of iterations increases, the model constructed by the present method can show good adaptability when facing different types of lithium-ion battery data. This technical effect not only improves the generality of the model but also provides the possibility for predicting the health state of lithium-ion batteries in different scenarios.

[0071] The present invention has also made remarkable progress in terms of model training efficiency. Traditional deep learning models often consume a large amount of time and computing resources during the training process. However, through the optimization of initial key parameters by the triangular topology aggregation optimization algorithm, the model can converge to the optimal solution faster during the training process, thus greatly improving the training efficiency. This technical effect not only shortens the time for model training but also reduces the demand for computing resources, providing more convenience for the practical application of lithium-ion battery state of health prediction.

[0072] Other advantages, objectives, and features of the present invention will be described to some extent in the subsequent specification, and to some extent, will be obvious to those skilled in the art based on the study of the following text, or can be taught from the practice of the present invention. The objectives and other advantages of the present invention can be achieved and obtained through the following specification. BRIEF DESCRIPTION OF THE DRAWINGS

[0073] In order to make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be described in detail preferably with reference to the accompanying drawings, where:

[0074] Figure 1 is a schematic diagram of the overall process of the lithium-ion battery state of health prediction method for optimizing the CNN-BiLSTM-AM model based on the triangular topology aggregation optimization algorithm under the embodiment of the present invention;

[0075] Figure 2 is a schematic diagram of the detailed process of the lithium-ion battery state of health prediction method for optimizing the CNN-BiLSTM-AM model based on the triangular topology aggregation optimization algorithm under the embodiment of the present invention;

[0076] Figure 3 is a schematic diagram of the structure of the CNN-BiLSTM-AM model under the embodiment of the present invention;

[0077] Figure 4 is a schematic diagram of the SOH estimation results of different models for battery L001 under the embodiment of the present invention, where Figure 4 (a) is a curve graph of the estimation results of experimental battery L001, Figure 4 (b) is a curve graph of the relative error of experimental battery L001;

[0078] Figure 5 is a schematic diagram of the SOH estimation results of different models for battery L002 under the embodiment of the present invention, where Figure 5 (a) is a curve graph of the estimation results of experimental battery L002; Figure 5 (b) is a curve graph of the relative error of experimental battery L002;

[0079] Figure 6Schematic diagram of SOH estimation results of different models for L003 battery under the embodiments of the present invention, where Figure 6 (a) is the estimation result curve of experimental battery L003; Figure 6 (b) is the relative error curve of experimental battery L003;

[0080] Figure 7 Schematic diagram of SOH estimation results of different models for L004 battery under the embodiments of the present invention, where Figure 7 (a) is the estimation result curve of experimental battery L004; Figure 7 (b) is the relative error curve of experimental battery L004. Detailed implementation manners

[0081] The following specific examples illustrate the implementation manners of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific implementation manners. Various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the diagrams provided in the following embodiments only illustrate the basic concept of the present invention in a schematic manner. Without conflict, the following embodiments and the features in the embodiments can be combined with each other.

[0082] Among them, the drawings are only for illustrative purposes, showing only schematic diagrams, not physical diagrams, and should not be construed as a limitation to the present invention; for better illustration of the embodiments of the present invention, some components in the drawings will be omitted, enlarged or reduced, which do not represent the dimensions of actual products; for those skilled in the art, it is understandable that some well-known structures and their descriptions in the drawings may be omitted.

[0083] In the drawings of the embodiments of the present invention, the same or similar reference numerals correspond to the same or similar components; in the description of the present invention, it should be understood that if there are terms such as "upper", "lower", "left", "right", "front", "rear", etc. indicating the orientation or positional relationship, they are based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation. Therefore, the terms describing the positional relationship in the drawings are only for illustrative purposes and should not be construed as a limitation to the present invention. For those of ordinary skill in the art, the specific meanings of the above terms can be understood according to specific circumstances.

[0084] Please refer to Figures 1 to 7 , a lithium-ion battery health state prediction method based on a CNN-BiLSTM-AM model optimized by a triangular topology aggregation optimization algorithm.

[0085] Embodiment

[0086] This embodiment provides specific implementation steps of a method for predicting the health state of a lithium-ion battery by optimizing the CNN-BiLSTM-AM model based on a triangular topology aggregation optimization algorithm, as Figure 1 shown, which includes the following steps:

[0087] S1. Obtain a data set, extract the health indicator HI (Health Indicator, HI) from the data set, establish feature sample data after correlation analysis, and divide it into a training set and a test set;

[0088] S2. Establish a CNN-BiLSTM-AM model for predicting the health state of a lithium-ion battery;

[0089] S3. Screen the initial key parameters of the CNN-BiLSTM-AM model, and optimize the initial key parameters using a triangular topology aggregation optimization algorithm;

[0090] S4. According to the optimized initial key parameters, train and test the CNN-BiLSTM-AM model in combination with the training set and the test set;

[0091] S5. Obtain lithium-ion battery data in real time, and use the trained TTAO-CNN-BiLSTM-AM model to predict the health state of the lithium-ion battery in real time.

[0092] Figure 2 The following shows a schematic diagram of the overall detailed process of the method for predicting the health state of a lithium-ion battery by optimizing the CNN-BiLSTM-AM model based on a triangular topology aggregation optimization algorithm in this embodiment, where:

[0093] In step S1 of this embodiment, the ways to obtain the data set include experimental acquisition and using a public data set. The public data set can use the NASA battery data set or the CALCE battery data set. In this embodiment, the battery sample data is directly measured in the laboratory. According to the measured charge and discharge curves, the health indicator HI is extracted. The specific extraction method is as follows: Extract the constant current charging time of the battery as HI1; extract the equal voltage discharge time as HI2.

[0094] In the data preprocessing stage, the outliers in the extracted health indicator HI are filled with the mean value of adjacent data points.

[0095] Then, for the extracted health indicator HI data, Pearson correlation coefficient and Spear coefficient are used for correlation analysis. By performing correlation analysis, the correlation degree between the health indicator and the SOH is quantitatively analyzed.

[0096] Finally, for the data after correlation analysis, data normalization is performed. The data is adjusted through linear transformation while maintaining the original order of the data, which can improve the performance and training speed of the model. The min-max normalization method is adopted to scale the data to the range of 0 to 1;

[0097] The normalized data is used as the feature sample data composed of the health factor HI, and it is divided into a training set and a test set according to a ratio of 6:4.

[0098] In step S2 of this embodiment, the structural schematic diagram of the CNN-BiLSTM-AM model established for predicting the health state of lithium-ion batteries is as Figure 3 shown, which includes an input layer Input, a convolutional neural network layer CNN, a bidirectional long short-term memory network layer BiLSTM, an attention mechanism layer AM, and a fully connected output layer FC.

[0099] The input layer Input receives the input data of the battery, and these data can include operating parameters such as voltage, current, and temperature.

[0100] The convolutional neural network layer CNN includes a convolutional layer Conv and a pooling layer pooling. The convolutional layer Conv extracts local features from the input data through convolutional kernels. Convolutional operations can effectively capture patterns and changes in time series data. The pooling layer pooling performs a pooling operation on the output of the convolutional layer, which is usually used to reduce the dimension of the data, reduce the amount of calculation, and at the same time retain important features.

[0101] The bidirectional long short-term memory network layer BiLSTM includes several long short-term memory networks LSTM with bidirectional structures. The long short-term memory network LSTM is used to process sequence data. LSTM effectively solves the problem of gradient disappearance during the training process of long sequences through a gating mechanism. The bidirectional structure helps BiLSTM process sequence information in both the forward and backward directions at the same time, enabling the model to utilize the context information before and after the sequence data and improving the prediction accuracy.

[0102] In the attention mechanism layer AM, the Softmax function is used to normalize the output of BiLSTM, so that each output value is between 0 and 1, and the sum of all values is 1, thereby obtaining the attention weights at each moment. The attention weight α refers to determining the importance of information at each moment by calculating the similarity between the query vector and the output of BiLSTM. Then, the attention weights are used to perform a weighted sum on the output of BiLSTM to obtain a context vector containing important information.

[0103] The fully connected output layer FC comprehensively analyzes the output of the previous layer and obtains the prediction result, that is, the state of health (SoH) of the lithium-ion battery, through a fully connected manner.

[0104] In step S3 of this embodiment, the key parameters of the CNN-BiLSTM-AM model obtained are the number of hidden units h d , the initial learning rate l r and the dropout rate D u . The specific process of optimizing the number of hidden units h d , the initial learning rate l r and the dropout rate D u by the triangular topology aggregation optimization algorithm is as follows:

[0105] S31. Population initialization: Initialize the key parameters, namely the number of hidden units h d , the initial learning rate l r and the dropout rate D u for each individual in the population. Then determine the population size and the variable dimension, divide the population into multiple triangular topology units, and obtain the initial solution set X1;

[0106] Specifically, encode the parameters h d , l r , D u to a multi-dimensional vector x = [h d , l r , D u . Each multi-dimensional vector serves as an individual in the population. Among them, the number of hidden units h d is a discrete parameter, and its range is expressed as (h dmin , h dmax ). The initial learning rate l r and the dropout rate D u are continuous parameters, and their ranges are expressed as (l rmin , l rmax ) and (D umin , D umax ) respectively. Thus, obtain the population size N and the variable dimension D of the initial population, and then generate an initial solution set X1 of N / 3 solutions to provide an initial solution space for the subsequent global search and local optimization stages; where each search individual in the initial solution set X1 is expressed as:

[0107] X i,1 = r o ×(UB - LB)+LB

[0108] In the formula, X i,1 is the first search individual of the i-th triangular topology unit; LB and UB are the lower and upper bounds of the search space respectively; r0 is a random value between 0 and 1.

[0109] In this embodiment, the population size N is set to 30, and the dimension D is 3; the value range of the number of hidden units h d is (16, 512); the initial learning rate lr ranges from (0.001, 0.01); the discard rate D u ranges from (0.01, 0.1); the size of the convolutional kernel is set to 3; the maximum number of iterations is 100.

[0110] S32. Formation of triangular topological units: According to each first vertex X in the solution set X1 i,1 , determine the second vertex X of the triangular topological unit i,2 , the third vertex X i,3 and the fourth vertex X inside the triangular topological unit i,4 ;

[0111] Specifically, to ensure that equilateral triangles are formed on each two-dimensional plane. Through the conversion between polar coordinates and the conventional coordinate system, the second and third vertices are generated using the following equations:

[0112] X i,2 = X i,1 + l × f(θ)

[0113]

[0114] In the formula, l is the size of the triangular topological unit, which gradually decreases as the number of iterations increases, ensuring that the algorithm conducts global searches in the initial stage and focuses on finding the optimal solution in the later stage; f(θ) and are the direction vectors of two sides obtained using X1.

[0115] l is defined as:

[0116]

[0117] In the formula, t and T are the current iteration number and the maximum iteration number respectively.

[0118] In addition, the specific formulas for f(θ) and are:

[0119] f(θ) = [cosθ1, cosθ2, …, cosθ D

[0120]

[0121] In the formula, θ j , j = 1, 2, …, D are random numbers from 0 to π, and X1, X2, and X3 of each group of unit triangular topological units aggregate into the fourth vertex inside in a linear weighted manner:

[0122] X i,4 = r1 × X i,1 + r2 × X i,2 + r3 × X​i,3

[0123] Wherein, r1, r2, and r3 are random numbers between 0 and 1, and the sum of the three is always equal to 1.

[0124] S33. General aggregation: Simulate the gene crossover mechanism for general aggregation. Information interaction occurs between the best individual in each triangular topology unit and the best individuals randomly selected from other units, and the optimal individual is updated. The process is as follows:

[0125] Generate a new individual through linear combination with different weights between the best individual and the random individual:

[0126] X t+1 i,n1 = r4 × X t i,b + (1 - r4) × X t r,b

[0127] Wherein, X t r,b is the individual randomly selected at the t-th iteration; X t i,b is the best individual. r4 is a random number between 0 and 1. The algorithm simulates the gene crossover mechanism in the genetic algorithm and applies it to general aggregation to demonstrate the interaction between information.

[0128] Then, the algorithm compares the fitness value of the newly generated individual with the current best individual and the sub-optimal unit, and updates the optimal individual according to the greedy strategy. The specific formula is:

[0129]

[0130] Wherein, represents the fitness value of the newly generated individual at the (t + 1)-th time, represents the fitness value of the optimal individual at the (t + 1)-th iteration, X t+1 i,b represents the optimal individual at the (t + 1)-th iteration, represents the fitness value of the newly generated individual of unit i at the (t + 1)-th time, represents the fitness value of the sub-optimal individual at the (t + 1)-th iteration, X t+1 i,sb represents the sub-optimal individual at the (t + 1)-th iteration.

[0131] S34. Local aggregation: After general aggregation optimization, the updated optimal or sub-optimal individual and the other two vertices in the group of individuals with better fitness form a temporary triangular topology unit, and local aggregation is performed on the temporary triangular topology unit to update the optimal individual again.

[0132] Due to the difference in motion vectors between the optimal and sub-optimal individuals, the position of the optimal individual will be locally perturbed. Therefore, a re-search is performed for each group within the local area to fully utilize each topological triangular unit. The new vertex is calculated as follows:

[0133] X t+1 i,n2 = X t+1 i,b + a(X t+1 i,b - X t+1 i,sb )

[0134] In the formula, a is the size of the aggregation range, which can be expressed as:

[0135]

[0136] To ensure that the convergence develops in a more promising direction, it is necessary to compare the fitness values of adjacent vertices in the local search to determine whether to perform a position update. If the newly generated individual is better than the original individual, a position adjustment is made; otherwise, the original position remains unchanged.

[0137]

[0138] In the formula, represents the fitness value of the newly generated individual of unit i at the (t + 1)-th time.

[0139] S35. Determine whether the iteration ends. If it does not end, return to step S33 to continue the iteration. If it ends, output the optimal individual.

[0140] In the above process, the fitness function is set as the root mean square error between the SOH prediction value and the actual value, and its calculation formula is:

[0141]

[0142] In the formula, z i represents the battery SOH prediction value of the i-th sample; z i ' represents the true value of the battery SOH, and M is the total number of samples used for training.

[0143] In step S4 of this embodiment, according to the optimized initial key parameters, the CNN-BiLSTM-AM model is trained and tested in combination with the training set and the test set; the model after training is named the TTAO-CNN-BiLSTM-AM model.

[0144] In step S5 of this embodiment, lithium-ion battery data is obtained in real time, and the trained TTAO-CNN-BiLSTM-AM model is used to predict the health state of the lithium-ion battery in real time.

[0145] Embodiment 2

[0146] In this embodiment, the TTAO-CNN-BiLSTM-AM model is trained based on the method described in Embodiment 1, and it is compared with the existing CNN-BiLSTM model, CNN-BiLSTM-AM model, and actual values for effect demonstration.

[0147] In this embodiment, the batteries included in the obtained experimental battery dataset are batteries L001, L002, L003, and L004. The constant current charging time during the charging process and the equal voltage drop discharge time during the discharging process are used as the health factor HI. The training set and the test set are divided at a ratio of 6:4. Specifically, the SOH estimation results of different models for the L001 battery in the experimental battery dataset are as Figure 4 shown, where Figure 4 (a) is the curve graph of the estimation results of the experimental battery L001; Figure 4 (b) is the curve graph of the relative error of the experimental battery L001; the SOH estimation results of different models for the L002 battery in the experimental battery dataset are as Figure 5 shown, where Figure 5 (a) is the curve graph of the estimation results of the experimental battery L002; Figure 5 (b) is the curve graph of the relative error of the experimental battery L002; the SOH estimation results of different models for the L003 battery in the experimental battery dataset are as Figure 6 shown, where Figure 6 (a) is the curve graph of the estimation results of the experimental battery L003; Figure 6 (b) is the curve graph of the relative error of the experimental battery L003; the SOH estimation results of different models for the L004 battery in the experimental battery dataset are as Figure 7 shown, where Figure 7 (a) is the curve graph of the estimation results of the experimental battery L004; Figure 7(b) is the relative error curve graph of the experimental battery L004. It can be seen from the graph that the CNN-BiLSTM and CNN-BiLSTM-AM models are not ideal in capturing the local capacity growth information of the battery. Although CNN-BiLSTM can track the trend of battery SOH decline, there are large fluctuations in the estimation, and it is impossible to accurately estimate the battery SOH based on only a small number of features. The model after introducing AM has significantly improved fitting effect in the early stage, indicating that the attention mechanism can better process the key information in the battery sequence and give a higher status to the key part. However, the effect is poor in the later stage of prediction and it does not adapt to the capacity growth behavior during the battery decline process. The TTAO-CNN-BiLSTM-AM model can better track the local fluctuations of the battery, and can more accurately fit the actual change curve of the battery SOH both in the early stage and the later stage.

[0148] The error curve of the CNN-BiLSTM model fluctuates the most, and the error of some points even reaches 6%. In contrast, the error curve of the model after introducing the attention mechanism is smoother, and the maximum relative error does not exceed 4%. The further optimized TTAO-CNN-BiLSTM-AM model performs the best, with the smallest fluctuation of the error curve and the overall relative error always remaining within 2%.

[0149] The root mean square error (RMSE), mean absolute error (MAE) and mean absolute percentage error (MAPE) are selected to evaluate the performance of the SOH estimation model. As shown in Table 1 below:

[0150] Table 1

[0151]

[0152] It can be seen from Table 1 that the TTAO-CNN-BiLSTM-AM model is significantly superior to the CNN-BiLSTM and CNN-BiLSTM-AM models in terms of the RMSE, MAE and MAPE indicators of SOH estimation. The RMSE of the four experimental batteries is lower than 0.0030, the MAE is lower than 0.0025, and the MAPE is lower than 0.4%. Compared with the unoptimized model, the TTAO-optimized CNN-BiLSTM-AM model has an average decrease of 72.99%, 72.45% and 75.67% in RMSE, MAE and MAPE respectively. It shows that the TTAO-CNN-BiLSTM-AM model performs the best in the SOH estimation of the four experimental batteries, significantly improving the stability and accuracy of SOH estimation.

[0153] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the present technical solution, and they should all be covered by the scope of the claims of the present invention.

Claims

1. A method for predicting the state of health of a lithium-ion battery by optimizing the CNN-BiLSTM-AM model based on a triangular topology aggregation optimization algorithm, characterized in that: The method includes the following steps: Obtain a dataset, extract the health factor HI from the dataset, establish feature sample data after correlation analysis, and divide it into a training set and a test set; Establish a CNN-BiLSTM-AM model for predicting the health state of lithium-ion batteries; Screen the initial key parameters of the CNN-BiLSTM-AM model, and optimize the initial key parameters using the triangular topology aggregation optimization algorithm; According to the optimized initial key parameters, train and test the CNN-BiLSTM-AM model by combining the training set and the test set; Obtain lithium-ion battery data in real time, and use the trained TTAO-CNN-BiLSTM-AM model to predict the health state of lithium-ion batteries in real time.

2. A method for predicting the state of health of a lithium-ion battery by optimizing the CNN-BiLSTM-AM model based on a triangular topology aggregation optimization algorithm according to claim 1, characterized in that: The ways to obtain the dataset include experimental acquisition and using public datasets; According to the measured charge and discharge curves, extract the health factor HI. The specific extraction method is: extract the constant current charging time of the battery as HI1; extract the equal voltage discharge time as HI2; In the data preprocessing stage, fill the outliers in the extracted health factor HI with the mean of adjacent data points; Then, for the extracted health factor HI data, perform correlation analysis using the Pearson correlation coefficient and the Spearman coefficient. By performing correlation analysis, quantitatively analyze the correlation degree between the health factor and SOH; For the data after correlation analysis, perform data normalization processing. Adjust the data through linear transformation, and use the minimum-maximum normalization method to scale the data to the range of 0 to 1; the normalized data is used as the feature sample data composed of the health factor HI, and it is divided into a training set and a test set according to a certain proportion.

3. A method for predicting the state of health of a lithium-ion battery by optimizing the CNN-BiLSTM-AM model based on a triangular topology aggregation optimization algorithm according to claim 1, characterized in that: The established CNN-BiLSTM-AM model for predicting the health state of lithium-ion batteries includes an input layer Input, a convolutional neural network layer CNN, a bidirectional long short-term memory network layer BiLSTM, an attention mechanism layer AM, and a fully connected output layer FC; The input layer Input receives the input data of the battery; The convolutional neural network layer CNN includes a convolutional layer Conv and a pooling layer pooling. The convolutional layer Conv extracts local features in the input data through a convolutional kernel; the pooling layer pooling performs a pooling operation on the output of the convolutional layer; The bidirectional long short-term memory network layer BiLSTM includes several long short-term memory networks LSTM arranged in a bidirectional structure. The long short-term memory network LSTM is used to process sequence data; the bidirectional structure helps the BiLSTM to process sequence information in both the forward and backward directions simultaneously; In the attention mechanism layer AM, use the Softmax function to normalize the output of the BiLSTM to obtain the attention weights at each moment; Then use the attention weights to perform weighted summation on the output of the BiLSTM to obtain a context vector containing important information; The fully connected output layer FC comprehensively analyzes the output of the previous layers and obtains the prediction result, that is, the health state SoH of the lithium-ion battery, through a fully connected manner.

4. A method for predicting the state of health of a lithium-ion battery by optimizing the CNN-BiLSTM-AM model based on a triangular topology aggregation optimization algorithm according to claim 1, characterized in that: The key parameters of the obtained CNN-BiLSTM-AM model are the number of hidden units h d , the initial learning rate l r and the dropout rate D u , and the specific process of optimizing the number of hidden units h d , the initial learning rate l r and the dropout rate D u by the triangular topology aggregation optimization algorithm is as follows: Population initialization: Hide the number of key parameters, the number of hidden units h d , the initial learning rate l r and the dropout rate D u Initialize each individual in the population, then determine the population size and the variable dimension, divide the population into multiple triangular topological units, and obtain the initial solution set X1; Triangle topology unit formation: Based on each first vertex X in the solution set X1 i,1 , determine the second vertex X of the triangle topology unit i,2 , the third vertex X i,3 and the fourth vertex X inside the triangle topology unit i,4 ; General aggregation: General aggregation is performed by simulating the gene crossover mechanism. Information interaction occurs between the best individual in each triangular topology unit and the best individuals randomly selected from other units, and the optimal individual is updated; Local aggregation: The updated optimal or sub-optimal individual forms a temporary triangular topology unit with the other two vertices. Local aggregation is performed for the temporary triangular topology unit, and the optimal individual is updated again; Judge whether the iteration ends. If it does not end, return to step S33 to continue the iteration. If it ends, output the optimal individual.

5. A method for predicting the state of health of a lithium-ion battery by optimizing the CNN-BiLSTM-AM model based on a triangular topology aggregation optimization algorithm according to claim 4, characterized in that: During the population initialization process, the parameter h to be optimized d , l r , D u is encoded as a multi-dimensional vector x = [h d , l r , D u . Each multi-dimensional vector serves as an individual in the population. Among them, the number of hidden units h d is a discrete parameter, and its range is expressed as (h dmin , h dmax ). The initial learning rate l r and the dropout rate D u are continuous parameters, and their ranges are respectively expressed as (l rmin , l rmax ) and (D umin , D umax ). Thus, the population size N and the variable dimension D of the initial population are obtained, and then an initial solution set X1 of N / 3 solutions is generated to provide an initial solution space for the subsequent global search and local optimization phases; where each search individual in the initial solution set X1 is represented as: X i,1 = r o × (UB - LB)+LB where X i,1 is the first search individual of the i-th triangular topology unit; LB and UB are the lower and upper bounds of the search space respectively; r0 is a random value between 0 and 1.

6. A method for predicting the health state of a lithium-ion battery by optimizing the CNN-BiLSTM-AM model based on a triangular topology aggregation optimization algorithm, characterized in that: During the formation of the triangular topology unit, according to the first vertex X i,1 Generate the second vertex X i,2 , the third vertex X i,3 The method is as follows: X i,2 = X i,1 + l × f(θ) where l is the size of the triangular topological unit, which gradually decreases as the number of iterations increases; f(θ) and are the direction vectors of the two sides obtained using X1; l is defined as: In the formula, t and T are the current iteration number and the maximum iteration number respectively; f(θ) and The specific formula for f(θ) = [cosθ1, cosθ2, ···, cosθ D ​ where θ j , j = 1, 2, ..., D are random numbers from 0 to π; X1, X2, and X3 of each group of triangular topology units in the unit are aggregated into the fourth vertex internally in a linear weighted manner: X i,4 = r1 × X i,1 + r2 × X i,2 + r3 × X i,3 In the formula, r1, r2, and r3 are random numbers between 0 and 1, and the sum of the three is always equal to 1.

7. A method for predicting the state of health of a lithium-ion battery by optimizing a CNN-BiLSTM-AM model based on a triangular topology aggregation optimization algorithm according to claim 6, characterized in that: The general aggregation process is: Generate new individuals by performing linear combinations with different weights between the best individual and the random individual: X t+1 i,n1 = r4 × X t i,b + (1 - r4) × X t r,b where X t r,b is an individual randomly selected at the t-th iteration; X t i,b is the best individual; r4 is a random number between 0 and 1; Compare the fitness value of the newly generated individual with the current best individual and the sub-optimal unit, and update the optimal individual according to the greedy strategy. The specific formula is: In the formula, represents the fitness value of the newly generated individual at the (t + 1)-th time, represents the fitness value of the optimal individual in the (t + 1)-th iteration, X t+1 i,b represents the optimal individual in the (t + 1)-th iteration, represents the fitness value of the newly generated individual of unit i at the (t + 1)-th time, represents the fitness value of the sub-optimal individual in the (t + 1)-th iteration, X t+1 i,sb represents the sub-optimal individual in the (t + 1)-th iteration.

8. A method for predicting the state of health of a lithium-ion battery by optimizing a CNN-BiLSTM-AM model based on a triangular topology aggregation optimization algorithm, characterized in that: The local aggregation process is: Re-search each temporary triangular topology unit in the local area, and the new vertex is calculated as follows: X t+1 i,n2 = X t+1 i,b + a(X t+1 i,b - X t+1 i,sb ) In the formula, a is the size of the aggregation range, which is expressed as: Compare the fitness values of adjacent vertices in the local search to judge whether to perform position update. If the newly generated individual is better than the original individual, perform position adjustment; otherwise, keep the original position unchanged: In the formula, represents the fitness value of the newly generated individual of unit i at the (t + 1)-th time.

9. A method for predicting the health state of a lithium-ion battery by optimizing the CNN-BiLSTM-AM model based on a triangular topology aggregation optimization algorithm according to claim 8, characterized in that: The fitness function in the triangular topology aggregation optimization algorithm is set as the root mean square error between the SOH prediction value and the actual value, and its calculation formula is: where z i represents the predicted value of the battery SOH for the i-th sample; z i ' represents the true value of the battery SOH, and M is the total number of samples used for training.

Images