A lithium battery SOH estimation model training method and system based on improved CPO optimized Bi-LSTM

Abstract

The present invention discloses a lithium battery SOH estimation model training method based on improved CPO optimized Bi-LSTM, including: obtaining multiple initial feature sequences of lithium batteries, normalizing each of the obtained initial feature sequences and the SOH value sequence composed of multiple SOH values ​​of the lithium battery to obtain normalized feature sequences and normalized SOH value sequences, using an association analysis method to calculate the correlation between each feature sequence in a feature sequence set and the normalized SOH value sequence, and filtering feature sequences with a correlation lower than a preset threshold from the feature sequence set, combining multiple updated feature sequences and the normalized SOH value sequence into a data set, and dividing the data set into a training set, a validation set and a test set according to a predetermined ratio. The present invention can solve the technical problem that the existing lithium battery SOH estimation method based on a data-driven model is difficult to accurately and comprehensively reflect the complex degradation trend of lithium batteries, thereby reducing the accuracy of SOH estimation.

Description

Technical Field

[0001] The present invention belongs to the technical field of deep learning and lithium battery state prediction, and more specifically, relates to a training method and system for a lithium battery health state estimation model based on CPO optimization. Background Art

[0002] In recent years, lithium batteries have been widely used in consumer electronics, electric vehicles, energy storage devices and other fields due to their advantages such as high operating voltage, high energy density, low self-discharge rate, no memory, environmental friendliness, and flexible size. They have even become the preferred lithium battery for electric vehicles. The state of health (SOH) is a key indicator in the lithium battery system that reflects the battery performance, service life and degradation state. The percentage of the actual battery capacity to the initial capacity is generally used to represent the battery health state. During the continuous charge and discharge cycle, the active substances inside the battery decrease, the side reaction products increase, and the SEI film on the positive electrode surface thickens, which eventually leads to an increase in the internal resistance of the battery and capacity decay. It is generally believed that after the lithium battery capacity decays to 80% of the initial capacity, it no longer meets the performance requirements of electric vehicles. In order to ensure the smooth operation of electric vehicles, it is necessary to estimate the health state of lithium batteries.

[0003] The existing SOH estimation method for lithium batteries is mainly based on the SOH prediction method of data-driven model, which extracts battery aging health characteristics such as voltage, current, temperature, etc. from the lithium battery charge and discharge data, and uses machine learning or deep learning algorithms to fit the nonlinear mapping relationship between health characteristics and SOH to achieve SOH estimation. This SOH prediction method based on data-driven model has strong flexibility and scalability because it does not consider the internal mechanism and complex model of lithium batteries.

[0004] However, the above-mentioned lithium battery SOH estimation method based on the data-driven model has some defects that cannot be ignored: First, in terms of feature extraction, this method often only extracts a single feature or multiple features of the same type as a feature, which is difficult to accurately and comprehensively reflect the complex degradation trend of lithium batteries, thereby reducing the accuracy of SOH estimation; Second, the internal structure and parameters of the data-driven model cannot be flexibly set according to the characteristic data of the lithium battery (for example, when the data-driven model is an LSTM model, the hyperparameters of the model such as the learning rate and the number of hidden layer neurons are usually manually set based on experience), which will cause the generalization ability of the data-driven model to deteriorate, thereby reducing the accuracy and efficiency of SOH estimation; Third, when the data-driven model is introduced into the algorithm to optimize the model parameters, the initial population of the algorithm is prone to uneven distribution, resulting in a slow global search speed or falling into a local optimum; Fourth, when the data-driven model is introduced into the algorithm to optimize the model parameters, when the position of the current optimal individual is not the global optimum, as the number of iterations increases, the individuals in the population will mistakenly gather near the local optimal area, resulting in premature convergence of the algorithm, thereby reducing the accuracy of the algorithm optimization. Summary of the invention

[0005] In view of the above defects or improvement needs of the prior art, the present invention provides a lithium battery SOH estimation model training method based on an improved CPO optimized Bi-LSTM, the purpose of which is to solve the technical problem that the existing lithium battery SOH estimation method based on the data-driven model often only extracts a single feature or multiple features of the same type as a feature, which is difficult to accurately and comprehensively reflect the complex degradation trend of the lithium battery, thereby reducing the accuracy of the SOH estimation, and the internal structure and parameters of the data-driven model cannot be flexibly set according to the characteristic data of the lithium battery, resulting in poor generalization ability of the data-driven model, thereby reducing the accuracy and efficiency of the SOH estimation, and the data-driven model is based on The technical problem that when introducing an algorithm to optimize the model parameters, the initial population of the algorithm is prone to uneven distribution, resulting in a slow global search speed or falling into a local optimum, and when introducing an algorithm to optimize the model parameters, when the position of the current optimal individual is not the global optimum, as the number of iterations increases, the individuals in the population will mistakenly gather near the local optimal area, resulting in premature convergence of the algorithm, thereby reducing the algorithm's optimization accuracy.

[0006] To achieve the above object, according to one aspect of the present invention, a lithium battery SOH estimation model training method based on improved CPO optimized Bi-LSTM is provided, comprising the following steps:

[0007] (1) obtaining multiple initial feature sequences of the lithium battery, and normalizing each of the obtained initial feature sequences and the SOH value sequence composed of multiple SOH values ​​of the lithium battery to obtain a normalized feature sequence and a normalized SOH value sequence, wherein all normalized feature sequences constitute a feature sequence set, wherein the length of each initial feature sequence is the number of charge and discharge cycles N of the lithium battery;

[0008] (2) using an association analysis method to calculate the degree of association between each feature sequence in the feature sequence set obtained in step (1) and the normalized SOH value sequence, and filtering out feature sequences with a degree of association lower than a preset threshold from the feature sequence set, thereby obtaining a plurality of updated feature sequences;

[0009] (3) combining the multiple updated feature sequences obtained in step (2) and the normalized SOH value sequence obtained in step (1) into a data set, and dividing the data set into a training set, a validation set, and a test set according to a predetermined ratio;

[0010] (4) Use all feature sequences and normalized SOH value sequences in the training set to train the Bi-LSTM model to obtain a preliminarily trained Bi-LSTM model;

[0011] (5) All feature sequences in the validation set are input into the Bi-LSTM model preliminarily trained in step (4) to obtain a predicted SOH value sequence consisting of multiple predicted SOH values, the length of which is equal to the length of each feature sequence in the validation set;

[0012] (6) Calculating the loss value using the loss function based on the predicted SOH sequence obtained in step (5) and the normalized SOH value sequence in the validation set obtained in step (3);

[0013] (7) Setting the range of the number of hidden neurons, the range of the number of iterations, and the range of the learning rate of the Bi-LSTM model, setting the internal parameters of the crown porcupine optimization algorithm CPO, and generating an initial population using the Chebyshev chaos mapping technology according to the set range of the number of hidden neurons, the range of the number of iterations, the range of the learning rate, and the internal parameters of the CPO algorithm;

[0014] (8) The number of hidden neurons, the number of iterations and the learning rate in the initial population obtained in step (7) are used as the search starting point of the crown porcupine optimization algorithm CPO, and the loss function obtained in step (6) is used as the objective function of the CPO algorithm to perform iterative processing of the CPO algorithm. After each iteration of the CPO algorithm, a random difference mutation strategy is used to mutate the population in the CPO algorithm, and finally the hidden neurons, the maximum number of iterations and the learning rate that minimize the loss value are obtained as the optimal solution, and the Bi-LSTM model corresponding to the optimal solution is used as the initially trained Bi-LSTM model.

[0015] Preferably, in step (1), a plurality of initial feature sequences of the lithium battery are obtained, and all the initial feature sequences are normalized to obtain a feature sequence set. This process includes the following sub-steps:

[0016] (1-1) Obtaining the time interval for the voltage of the lithium battery to change from 3.8V to 4.0V during the constant current charging stage of the first charge and discharge cycle, the time interval for the voltage to change from 3.8V to 4.0V during the constant current charging stage of the second charge and discharge cycle, ..., the time interval for the voltage to change from 3.8V to 4.0V during the constant current charging stage of the Nth charge and discharge cycle, all of which constitute a first characteristic sequence;

[0017] (1-2) obtaining the time interval for the voltage of the lithium battery to change from 4.0 V to 4.2 V during the constant current charging stage of the first charge and discharge cycle, the time interval for the voltage to change from 4.0 V to 4.2 V during the constant current charging stage of the second charge and discharge cycle, ..., the time interval for the voltage to change from 4.0 V to 4.2 V during the constant current charging stage of the Nth charge and discharge cycle, and all time intervals constitute a second characteristic sequence;

[0018] (1-3) obtaining the time interval when the current of the lithium battery changes from 1.0 A to 0.5 A during the constant voltage charging stage of the first charge and discharge cycle, the time interval when the current changes from 1.0 A to 0.5 A during the constant voltage charging stage of the second charge and discharge cycle, ..., the time interval when the current changes from 1.0 A to 0.5 A during the constant voltage charging stage of the Nth charge and discharge cycle, and all time intervals constitute a third characteristic sequence;

[0019] (1-4) obtaining the time interval for the voltage of the lithium battery to change from 4.0 V to 3.6 V during the constant current discharge stage of the first charge and discharge cycle, the time interval for the voltage to change from 4.0 V to 3.6 V during the constant current discharge stage of the second charge and discharge cycle, ..., the time interval for the voltage to change from 4.0 V to 3.6 V during the constant current discharge stage of the Nth charge and discharge cycle, and all time intervals constitute a fourth characteristic sequence;

[0020] (1-5) obtaining the time interval for the voltage of the lithium battery to change from 3.6 V to 3.2 V during the constant current discharge stage of the first charge and discharge cycle, the time interval for the voltage to change from 3.6 V to 3.2 V during the constant current discharge stage of the second charge and discharge cycle, ..., the time interval for the voltage to change from 3.6 V to 3.2 V during the constant current discharge stage of the Nth charge and discharge cycle, and all time intervals constitute a fifth characteristic sequence;

[0021] (1-6) Obtain the total time spent by the lithium battery in the constant current discharge stage of the first charge and discharge cycle from the start of constant current discharge to the time when the temperature of the lithium battery reaches the highest, the total time spent in the constant current discharge stage of the second charge and discharge cycle from the start of constant current discharge to the time when the temperature of the lithium battery reaches the highest, ..., the total time spent in the constant current discharge stage of the Nth charge and discharge cycle from the start of constant current discharge to the time when the temperature of the lithium battery reaches the highest, and all total times constitute the sixth characteristic sequence;

[0022] (1-7) Obtain the capacity increment IC curve of the lithium battery in the constant current discharge stage of the first charge and discharge cycle, and obtain the IC curve The peak value and the corresponding discharge voltage value are obtained to obtain the IC curve of the lithium battery in the constant current discharge stage of the second charge and discharge cycle, and the The peak value and the corresponding discharge voltage value, ..., obtain the IC curve of the lithium battery in the constant current discharge stage of the Nth charge and discharge cycle, and obtain the IC curve The peak values ​​and their corresponding discharge voltage values, all peak values ​​constitute the seventh characteristic sequence, and all voltage values ​​constitute the eighth characteristic sequence, where the abscissa of the IC curve is the discharge voltage value and the ordinate is That is, the differential of the discharge capacity of the lithium battery to the discharge voltage;

[0023] (1-8) Normalize the first to eighth feature sequences obtained in steps (1-1) to (1-7), and combine the normalized first to eighth feature sequences into a feature sequence set.

[0024] Preferably, the calculation formula of the IC curve in step (1-7) is: Among them, Q, V, I, and T are the discharge capacity, discharge voltage, discharge current, and discharge time of the lithium battery respectively.

[0025] Preferably, the normalization process in step (1-8) adopts the following formula:

[0026]

[0027] Where x represents any element in the feature sequence, x' represents the normalized value of element x, and x min Represents the minimum value of all elements in the characteristic sequence where element x is located, xmax Represents the maximum value of all elements in the feature sequence where element x is located.

[0028] Preferably, the association analysis method used in step (2) is the Spearman correlation analysis method;

[0029] The absolute value of the preset threshold in step (2) ranges from 0.8 to 1.0.

[0030] Preferably, the loss function used in step (6) is a root mean square error (RMSE) function.

[0031] In step (7), the number of hidden neurons ranges from 50 to 150, the number of iterations ranges from 50 to 150, and the learning rate ranges from 0.001 to 0.01;

[0032] The internal parameters of the CPO algorithm in step (7) are set as follows: population size N = 100, maximum number of iterations = 100, optimization dimension = 3, minimum number of individuals in the population Nmin = 80, number of population reduction cycles T = 2, and convergence speed factor α = 0.1.

[0033] Preferably, the initial population X obtained in step (7) is:

[0034]

[0035] Where X N represents the Nth candidate solution in the initial population X, x i,j represents the jth position of the i-th candidate solution, and j∈[1,d], d represents the size of the optimization dimension, which is equal to 3.

[0036] The i-th candidate solution X in the initial population X i equal:

[0037]

[0038] Where L represents the number of hidden neurons, the number of iterations, and the lower limit of the learning rate, U represents the number of hidden neurons, the number of iterations, and the upper limit of the learning rate, r is a random number between 0 and 1, and:

[0039] x k+1 =cos(kcos -1 (x k ))

[0040] Where k is the order, its value is 4, and x0 is a random number between -1 and 1.

[0041] Preferably, the formula of the random difference mutation strategy is as follows:

[0042] Y(t+1)=r1(Y* -Y(t))+r2(Y′(t)-Y(t))

[0043] Among them, Y(t+1) represents the individual position obtained at the t+1th iteration, r1 and r2 are random numbers between 0 and 1, t is the current iteration number, Y* is the optimal individual position obtained from the 1st iteration to the tth iteration, and Y′(t) is the individual randomly selected from the population at the tth iteration.

[0044] According to another aspect of the present invention, a lithium battery SOH estimation model training system based on improved CPO optimized Bi-LSTM is provided, comprising:

[0045] The first module is used to obtain multiple initial feature sequences of the lithium battery, and normalize each of the obtained initial feature sequences and the SOH value sequence composed of multiple SOH values ​​of the lithium battery to obtain a normalized feature sequence and a normalized SOH value sequence. All normalized feature sequences constitute a feature sequence set, wherein the length of each initial feature sequence is the number of charge and discharge cycles N of the lithium battery;

[0046] The second module is used to calculate the correlation between each feature sequence in the feature sequence set obtained by the first module and the normalized SOH value sequence by using the correlation analysis method, and to filter out the feature sequences with correlation lower than a preset threshold from the feature sequence set, thereby obtaining multiple updated feature sequences;

[0047] The third module is used to combine the multiple updated feature sequences obtained by the second module and the normalized SOH value sequence obtained by the first module into a data set, and divide the data set into a training set, a validation set and a test set according to a predetermined ratio;

[0048] The fourth module is used to train the Bi-LSTM model using all feature sequences and normalized SOH value sequences in the training set to obtain a preliminarily trained Bi-LSTM model;

[0049] The fifth module is used to input all feature sequences in the validation set into the Bi-LSTM model preliminarily trained in the fourth module to obtain a predicted SOH value sequence consisting of multiple predicted SOH values, whose length is equal to the length of each feature sequence in the validation set;

[0050] The sixth module is used to calculate the loss value using the loss function according to the predicted SOH sequence obtained in the fifth module and the normalized SOH value sequence in the validation set obtained in the third module;

[0051] The seventh module is used to set the range of the number of hidden neurons, the range of the number of iterations, and the range of the learning rate of the Bi-LSTM model, set the internal parameters of the crown porcupine optimization algorithm CPO, and generate the initial population using the Chebyshev chaos mapping technology according to the set range of the number of hidden neurons, the range of the number of iterations, the range of the learning rate, and the internal parameters of the CPO algorithm;

[0052] The eighth module is used to use the number of hidden neurons, the number of iterations and the learning rate in the initial population obtained in the seventh module as the search starting point of the crown porcupine optimization algorithm CPO, and use the loss function obtained in the sixth module as the objective function of the CPO algorithm to iterate the CPO algorithm. After each iteration of the CPO algorithm, a random difference mutation strategy is used to mutate the population in the CPO algorithm, and finally the hidden neurons, the maximum number of iterations and the learning rate that minimize the loss value are obtained as the optimal solution, and the Bi-LSTM model corresponding to the optimal solution is used as the initially trained Bi-LSTM model.

[0053] In general, the above technical solutions conceived by the present invention can achieve the following beneficial effects compared with the prior art:

[0054] (1) The present invention adopts step (1) and step (2), thereby extracting a total of 8 feature sequences with a high correlation with SOH in multiple charge and discharge cycles of lithium batteries, which can comprehensively reflect the degradation trend of lithium batteries. Therefore, it can solve the technical problem that the existing lithium battery SOH estimation method based on data-driven model often only extracts a single feature or multiple features of the same type as a feature, which is difficult to accurately and comprehensively reflect the complex degradation trend of lithium batteries, thereby reducing the accuracy of SOH estimation;

[0055] (2) The present invention adopts step (4) and step (5), and uses the CPO optimization algorithm to optimize the internal structure and parameters of the Bi-LSTM, thereby improving the accuracy of the Bi-LSTM model in predicting SOH. Therefore, it can solve the technical problem that the existing lithium battery SOH estimation method based on the data-driven model has poor generalization ability due to the inability to flexibly set the internal structure and parameters of the data-driven model according to the characteristic data of the lithium battery, thereby reducing the accuracy and efficiency of the SOH estimation.

[0056] (3) The present invention adopts step (7) and step (8), and the initial population distribution of the CPO improved by using the Chebyshev chaos mapping technology is more uniform, so it can solve the technical problem that when the data-driven model is introduced into the algorithm to optimize the model parameters, the initial population of the algorithm is easily unevenly distributed, resulting in a slow global search speed or falling into a local optimum;

[0057] (4) The present invention adopts steps (7) and (8) and uses the improved CPO using the random difference mutation strategy to increase the diversity of the population and can jump out of the local optimum. Therefore, it can solve the technical problem that when a data-driven model introduces an algorithm to optimize model parameters, when the position of the current optimal individual is not the global optimal, as the number of iterations increases, the individuals in the population will mistakenly gather near the local optimal area, resulting in premature convergence of the algorithm, thereby reducing the algorithm's optimization accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 It is a comparison curve between the predicted value and the true value of SOH by the existing LSTM model;

[0059] Figure 2 It is a comparison curve between the predicted value and the true value of SOH by the existing Bi-LSTM model;

[0060] Figure 3 It is a comparison curve between the predicted value and the true value of SOH by the existing Bi-LSTM model based on PSO optimization;

[0061] Figure 4 It is a comparison curve between the predicted value and the true value of SOH by the existing Bi-LSTM model optimized based on CPO;

[0062] Figure 5 It is a comparison curve of the predicted value and the true value of SOH of the lithium battery SOH estimation model based on the improved CPO optimized Bi-LSTM model of the present invention;

[0063] Figure 6 It is a flow chart of the lithium battery SOH estimation model training method based on the improved CPO optimized Bi-LSTM of the present invention. DETAILED DESCRIPTION

[0064] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0065] like Figure 6 As shown, the present invention provides a lithium battery SOH estimation model training method based on improved CPO optimized Bi-LSTM, comprising the following steps:

[0066] (1) obtaining multiple initial feature sequences of the lithium battery (the length of each initial feature sequence is the number of charge and discharge cycles of the lithium battery N, where in this example, the value of N is 168), and normalizing each of the obtained initial feature sequences and the SOH value sequence composed of multiple SOH values ​​of the lithium battery (whose value is also equal to the number of charge and discharge cycles of the lithium battery, which is equal to 168 in this example) to obtain a normalized feature sequence and a normalized SOH value sequence, and all normalized feature sequences constitute a feature sequence set;

[0067] Specifically, in this step, multiple initial feature sequences of lithium batteries are obtained, and all initial feature sequences are normalized to obtain a feature sequence set. This process includes the following sub-steps:

[0068] (1-1) Obtaining the time interval for the voltage of the lithium battery to change from 3.8V to 4.0V during the constant current charging stage of the first charge and discharge cycle, the time interval for the voltage to change from 3.8V to 4.0V during the constant current charging stage of the second charge and discharge cycle, ..., the time interval for the voltage to change from 3.8V to 4.0V during the constant current charging stage of the Nth charge and discharge cycle, all of which constitute a first characteristic sequence;

[0069] In the constant current charging stage, as the number of charge and discharge cycles increases, the constant current charging time gradually decreases. This phenomenon is caused by the aging of lithium battery materials over time. In addition, since the process of voltage change from 3.8V-4.0V in the constant current charging stage is of great significance, and the time experienced in this process decreases with the increase in the number of charge and discharge cycles, the present invention takes the time interval when the voltage changes from 3.8V-4.0V in the constant current charging stage in all cycles as the first characteristic sequence.

[0070] (1-2) obtaining the time interval for the voltage of the lithium battery to change from 4.0 V to 4.2 V during the constant current charging stage of the first charge and discharge cycle, the time interval for the voltage to change from 4.0 V to 4.2 V during the constant current charging stage of the second charge and discharge cycle, ..., the time interval for the voltage to change from 4.0 V to 4.2 V during the constant current charging stage of the Nth charge and discharge cycle, and all time intervals constitute a second characteristic sequence;

[0071] Similar to the above step (1-1), in the constant current charging stage, the process in which the voltage changes from 4.0V-4.2V is of great significance, and the time taken for this process decreases with the increase in the number of charge and discharge cycles. Therefore, the present invention takes the time interval when the voltage changes from 4.0V-4.2V in the constant current charging stage in all cycles as the second characteristic sequence.

[0072] (1-3) obtaining the time interval when the current of the lithium battery changes from 1.0 A to 0.5 A during the constant voltage charging stage of the first charge and discharge cycle, the time interval when the current changes from 1.0 A to 0.5 A during the constant voltage charging stage of the second charge and discharge cycle, ..., the time interval when the current changes from 1.0 A to 0.5 A during the constant voltage charging stage of the Nth charge and discharge cycle, and all time intervals constitute a third characteristic sequence;

[0073] In the constant voltage charging stage, as the number of charge and discharge cycles increases, the constant voltage charging time gradually increases, and in the constant voltage charging stage, the process of current changing from 1.0A-0.5A is of great significance, and the time difference of this process increases with the increase in the number of charge and discharge cycles. Therefore, the present invention takes the time interval when the current changes from 1.0A-0.5A in the constant voltage charging stage in all cycles as the third characteristic sequence.

[0074] (1-4) obtaining the time interval for the voltage of the lithium battery to change from 4.0 V to 3.6 V during the constant current discharge stage of the first charge and discharge cycle, the time interval for the voltage to change from 4.0 V to 3.6 V during the constant current discharge stage of the second charge and discharge cycle, ..., the time interval for the voltage to change from 4.0 V to 3.6 V during the constant current discharge stage of the Nth charge and discharge cycle, and all time intervals constitute a fourth characteristic sequence;

[0075] (1-5) obtaining the time interval for the voltage of the lithium battery to change from 3.6 V to 3.2 V during the constant current discharge stage of the first charge and discharge cycle, the time interval for the voltage to change from 3.6 V to 3.2 V during the constant current discharge stage of the second charge and discharge cycle, ..., the time interval for the voltage to change from 3.6 V to 3.2 V during the constant current discharge stage of the Nth charge and discharge cycle, and all time intervals constitute a fifth characteristic sequence;

[0076] In the constant current discharge stage, as the number of charge and discharge cycles increases, the constant current discharge time gradually decreases. In addition, in practical applications, the battery usually operates in a voltage range of 4.0V to 3.2V during discharge to protect the battery, and the voltage changes stably. In the constant current discharge process, the time it takes for the voltage to change from 4.0V to 3.2V decreases as the number of charge and discharge cycles increases. Therefore, the present invention takes the time interval for the voltage to change from 4.0V to 3.6V and the time interval for the voltage to change from 3.6V to 3.2V in the constant current discharge stage in all cycles as the fourth characteristic sequence and the fifth characteristic sequence.

[0077] (1-6) Obtain the total time spent by the lithium battery in the constant current discharge stage of the first charge and discharge cycle from the start of constant current discharge to the time when the temperature of the lithium battery reaches the highest, the total time spent in the constant current discharge stage of the second charge and discharge cycle from the start of constant current discharge to the time when the temperature of the lithium battery reaches the highest, ..., the total time spent in the constant current discharge stage of the Nth charge and discharge cycle from the start of constant current discharge to the time when the temperature of the lithium battery reaches the highest, and all total times constitute the sixth characteristic sequence;

[0078] Temperature is a major factor affecting the reliability and safety of lithium batteries and has an important impact on the electrochemical reactions inside the battery. In the constant current discharge stage, the greater the number of cycles, the shorter the time it takes for the battery temperature to reach its peak value. Therefore, the present invention uses the time it takes for the battery temperature to reach its peak value during the constant current discharge process in all cycles as the sixth characteristic sequence.

[0079] (1-7) Obtain the incremental capacity (IC) curve of the lithium battery in the constant current discharge stage of the first charge and discharge cycle (where the abscissa of the curve is the discharge voltage value and the ordinate is That is, the differential of the discharge capacity of the lithium battery to the discharge voltage), and obtain the IC curve The peak value and the corresponding discharge voltage value are obtained to obtain the capacity increment (IC) curve of the lithium battery in the constant current discharge stage of the second charge and discharge cycle, and the IC curve is obtained. The peak value and the corresponding discharge voltage value, ..., obtain the capacity increment (IC) curve of the lithium battery in the constant current discharge stage of the Nth charge and discharge cycle, and obtain the IC curve The peak values ​​and their corresponding discharge voltage values, all the peak values ​​constitute the seventh characteristic sequence, and all the voltage values ​​constitute the eighth characteristic sequence;

[0080] The present invention uses an incremental capacity analysis method to extract IC features, and converts the voltage into an IC curve with obvious features based on the battery voltage and current parameters. The IC curve can depict the lithium ion information participating in the electrochemical reaction during the embedding process of the lithium ion active material under different aging conditions, thereby showing the phase change characteristics of the battery. During the constant current discharge process of the lithium battery, the IC calculation formula can be expressed as: Among them, Q, V, I, and T are the discharge capacity, discharge voltage, discharge current, and discharge time of the lithium battery respectively. As the voltage changes, it can be found that as the number of charge and discharge cycles increases, The peak value of gradually decreases and moves to the left. The reason for this phenomenon is that the active materials inside the battery gradually lose the ability to embed lithium ions during the aging process. The peak value and its corresponding voltage are taken as the IC characteristic sequence, which are recorded as the seventh characteristic sequence and the eighth characteristic sequence respectively.

[0081] The advantages of the above sub-steps (1-1) to (1-7) are that the use of these eight feature sequences extracted from the lithium battery charge and discharge cycle data can not only effectively reflect the complex degradation trend of the lithium battery during the charge and discharge cycle, but also when these eight feature sequences are input into the estimation model of the present invention, the eight feature sequences have a high correlation with the lithium battery SOH sequence, thereby improving the estimation accuracy of the model.

[0082] (1-8) normalizing the first to eighth feature sequences obtained in steps (1-1) to (1-7), and combining the first to eighth feature sequences after the normalization process into a feature sequence set;

[0083] Specifically, the normalization process in this sub-step uses the following formula:

[0084]

[0085] Where x' represents the normalized value of any element x in the feature sequence, x min Represents the minimum value of all elements in the characteristic sequence where element x is located, x max Represents the maximum value of all elements in the feature sequence where element x is located.

[0086] After normalization, all elements in all feature sequences are normalized to the interval [0, 1].

[0087] The advantage of this sub-step is that normalizing the data can limit the range of the data to a smaller interval, which helps to speed up the convergence of the model and reduce the differences between data features, helping to enhance the robustness of the model.

[0088] (2) using an association analysis method to calculate the degree of association between each feature sequence in the feature sequence set obtained in step (1) and the normalized SOH value sequence, and filtering out feature sequences with a degree of association lower than a preset threshold from the feature sequence set, thereby obtaining a plurality of updated feature sequences;

[0089] Specifically, the correlation analysis method used in this step is the Spearman correlation analysis method, and the calculated correlation degree is the Spearman correlation coefficient.

[0090] The absolute value of the preset threshold in this step ranges from 0.8 to 1.0, preferably from 0.9 to 1.0.

[0091] More specifically, the Spearman correlation coefficient is calculated in this step using the following formula:

[0092]

[0093] where r s is the Spearman correlation coefficient between each feature sequence in the feature sequence set obtained in step (1) and the normalized SOH value sequence, n is the length of the feature sequence (which is exactly the same as the normalized SOH value sequence), d i For X i and Y i The level difference between i and Y i are the two numbers at the i-th position in the feature sequence and the normalized SOH value sequence, respectively, and i∈[1, the length of the feature sequence n]); the rank of a number is the position of the number after all the numbers in the sequence are sorted from small to large, so after a set of data is sorted, the ranks of the numbers in it are 1, 2, ..., n from small to large. Note that when the same values ​​appear when the sequence is sorted, the ranks of the numbers with the same values ​​are the arithmetic mean of their positions.

[0094] Specific examples are shown in the following table:

[0095] <![CDATA[X i ]]> <![CDATA[Y i ]]> <![CDATA[X i Level]]> <![CDATA[Y i Level]]> <![CDATA[Level difference (d i )]]> 4 6 2 2.5 -0.5 6 2 4 1 3 5 6 3 2.5 0.5 9 7 5 4 1 1 8 1 5 -4

[0096] When the Spearman correlation coefficient is greater than 0, it means that the feature sequence and the normalized SOH value sequence are positively correlated; when it is less than 0, it means that the feature sequence and the normalized SOH value sequence are negatively correlated; the closer it is to 1 and -1, the stronger the correlation between the feature sequence and the normalized SOH value sequence.

[0097] The advantage of this step is that using a feature sequence that is highly correlated with the SOH sequence as the input of the model can effectively improve the prediction effect of the model on SOH.

[0098] (3) combining the multiple updated feature sequences obtained in step (2) and the normalized SOH value sequence obtained in step (1) into a data set, and dividing the data set into a training set, a validation set, and a test set according to a predetermined ratio (e.g., 2:1:1);

[0099] The advantage of this step is that the model is trained using the training set and validation set, and then the model effect is tested using the test set. This can prevent the model from being exposed to the test set data in advance during the training of the model, thereby affecting the final prediction effect.

[0100] (4) Use all feature sequences and normalized SOH value sequences in the training set to train the Bi-LSTM model to obtain a preliminarily trained Bi-LSTM model;

[0101] (5) All feature sequences in the validation set are input into the Bi-LSTM model preliminarily trained in step (4) to obtain a predicted SOH value sequence consisting of multiple predicted SOH values, the length of which is equal to the length of each feature sequence in the validation set;

[0102] (6) Calculating the loss value using the loss function based on the predicted SOH sequence obtained in step (5) and the normalized SOH value sequence in the validation set obtained in step (3);

[0103] Specifically, the loss function used in this step is the root mean square error (RMSE).

[0104] (7) Setting the range of the number of hidden neurons, the range of the number of iterations, and the range of the learning rate of the Bi-LSTM model, setting the internal parameters of the Crested Porcupine Optimizer (CPO), and generating an initial population (including the initial number of hidden neurons, the initial number of iterations, and the initial learning rate of the Bi-LSTM model) using the Chebyshev chaos mapping technique according to the set range of the number of hidden neurons, the range of the number of iterations, the range of the learning rate, and the internal parameters of the CPO algorithm;

[0105] Specifically, the number of hidden neurons ranges from 50 to 150, the number of iterations ranges from 50 to 150, and the learning rate ranges from 0.001 to 0.01.

[0106] The internal parameters of the CPO algorithm in this step are set as follows:

[0107] Population N = 100;

[0108] Maximum number of iterations = 100;

[0109] Optimization dimension (dimension) = 3;

[0110] The minimum number of individuals in the population Nmin = 80;

[0111] The number of population reduction cycles is T = 2; the convergence speed factor α = 0.1.

[0112] The advantage of this step is that the randomness, uniformity and orderliness of the Chebyshev chaotic map make the CPO population evenly dispersed in each iteration, which can improve the global search speed of the CPO optimization algorithm in the early iteration, while avoiding falling into the local optimum in the later stage, and achieving a better balance between global search and local development. Initializing the parameters of the Bi-LSTM model is helpful for the subsequent optimization of the CPO algorithm and can improve the optimization efficiency of the algorithm.

[0113] The present invention introduces the Chebyshev chaos mapping technology to perturb the population initialization of CPO, making the spatial distribution of the population more uniform, thereby improving the search efficiency.

[0114] The initial population obtained in this step can be expressed as:

[0115]

[0116] where x i,j represents the jth position of the i-th candidate solution, and j∈[1,d], d represents the size of the optimization dimension, which is equal to 3.

[0117] The CPO population initialization calculation formula after the introduction of Chebyshev chaos mapping technology is as follows:

[0118]

[0119] Where X i represents the i-th candidate solution and i∈[1,N], L represents the number of hidden neurons, the number of iterations, and the lower limit of the learning rate, U represents the number of hidden neurons, the number of iterations, and the upper limit of the learning rate, r is a random number between 0 and 1, and:

[0120] X k+1 =cos(kcos -1 (x k ))

[0121] Where k is the order, which is 4, and x0 is a random number between -1 and 1.

[0122] (8) The number of hidden neurons, the number of iterations, and the learning rate in the initial population obtained in step (7) are used as the search starting point of the Crested Porcupine Optimizer (CPO), and the loss function obtained in step (6) is used as the objective function of the CPO algorithm to perform iterative processing of the CPO algorithm. After each iteration of the CPO algorithm, a random differential mutation strategy is used to mutate the population in the CPO algorithm, and finally the hidden neurons, the maximum number of iterations, and the learning rate that minimize the loss value are obtained as the optimal solution, and the Bi-LSTM model corresponding to the optimal solution is used as the initially trained Bi-LSTM model;

[0123] In the optimization process of CPO, when the position of the current optimal individual is not the global optimal, as the number of iterations increases, the individuals in the population will mistakenly gather near the local optimal area, which will lead to premature convergence of the algorithm, thereby reducing the optimization accuracy. Therefore, the present invention incorporates the random difference mutation strategy into the CPO algorithm to help increase population diversity and escape from the local optimum.

[0124] The formula for the random difference mutation strategy is as follows:

[0125] Y(t+1)=r1(Y * -Y(t))+r2(Y′(t)-Y(t))

[0126] Among them, Y(t+1) represents the individual position obtained at the t+1th iteration, r1 and r2 are random numbers between 0 and 1, t is the current iteration number, and Y * is the optimal individual position obtained from the 1st iteration to the tth iteration, and Y′(t) is the individual randomly selected from the population at the tth iteration.

[0127] The advantage of this step is that the use of the root mean square error as the objective function of the CPO optimization algorithm can more effectively optimize the Bi-LSTM model. When the objective function value is minimized, the optimal solution combination of hidden neurons, maximum number of iterations, and learning rate is found. In addition, in each iteration of CPO, the random difference mutation strategy is used to mutate the population, which can generate candidate individuals with greater potential, thereby increasing the diversity of the population and helping the algorithm to escape from the local optimum.

[0128] The advantage of the above steps (4) to (8) is that the training set data is the basis for model adjustment. By repeatedly iteratively learning and adjusting model parameters, the model can gradually improve its ability to predict input data. The validation set is used to evaluate the performance of the model under different combinations of hyperparameters (number of hidden neurons, number of iterations, and learning rate), so as to select the optimal model hyperparameter settings and optimize the generalization ability of the model. In addition, the introduction of the validation set before the test set can prevent the model from contacting the test set before training is completed, so that the test set can be simulated as unknown data encountered in real applications in the final prediction stage. In this way, the prediction accuracy of the model for new data can be objectively evaluated to verify the actual application effect of the model.

[0129] Test Results

[0130] In order to verify the advantages of the model of the present invention, the performance of the present invention is tested and compared with that of four existing mainstream models (i.e., the existing LSTM model, the existing Bi-LSTM model, the existing Bi-LSTM model based on PSO optimization, and the existing Bi-LSTM model based on CPO optimization), and the following comparison results are finally obtained:

[0131] 1. Comparison of main indicators:

[0132] After obtaining the trained Bi-LSTM model, the present invention subsequently inputs all feature sequences in the test set obtained in step (3) into the Bi-LSTM model trained in step (8) of the present invention, the existing LSTM model, the existing Bi-LSTM, the existing Bi-LSTM model based on PSO optimization, and the existing Bi-LSTM model based on CPO optimization to obtain a predicted SOH value sequence consisting of multiple predicted SOH values, and obtains the mean absolute error (MAE), mean absolute percentage error (MAPE), and root mean square error (RMSE) based on the obtained predicted SOH value sequence and the normalized SOH value sequence in the test set obtained in step (3). The performance comparison of the five models on the above three indicators is shown in Table 1 below:

[0133]

[0134] Table 1

[0135] It can be seen from Table 1 above that the model of the present invention is superior to the four existing mainstream models in terms of MAE, MAPE and RMSE.

[0136] 2. Difference between true value and predicted value:

[0137] from Figures 1 to 5 It can be seen that the fitting effect (i.e., prediction accuracy) of the predicted value curve and the true value curve of the present invention is far superior to the existing LSTM model, Bi-LSTM model, Bi-LSTM model based on PSO optimization, and Bi-LSTM model based on the existing CPO optimization.

[0138] It will be easily understood by those skilled in the art that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A lithium battery SOH estimation model training method based on improved CPO optimized Bi-LSTM, characterized in that: The following steps are involved: (1) obtaining multiple initial feature sequences of the lithium battery, and normalizing each of the obtained initial feature sequences and the SOH value sequence composed of multiple SOH values ​​of the lithium battery to obtain a normalized feature sequence and a normalized SOH value sequence, wherein all normalized feature sequences constitute a feature sequence set, wherein the length of each initial feature sequence is the number of charge and discharge cycles N of the lithium battery; (2) using an association analysis method to calculate the degree of association between each feature sequence in the feature sequence set obtained in step (1) and the normalized SOH value sequence, and filtering out feature sequences with a degree of association lower than a preset threshold from the feature sequence set, thereby obtaining a plurality of updated feature sequences; (3) combining the multiple updated feature sequences obtained in step (2) and the normalized SOH value sequence obtained in step (1) into a data set, and dividing the data set into a training set, a validation set, and a test set according to a predetermined ratio; (4) Use all feature sequences and normalized SOH value sequences in the training set to train the Bi-LSTM model to obtain a preliminarily trained Bi-LSTM model; (5) All feature sequences in the validation set are input into the Bi-LSTM model preliminarily trained in step (4) to obtain a predicted SOH value sequence consisting of multiple predicted SOH values, the length of which is equal to the length of each feature sequence in the validation set; (6) Calculating the loss value using the loss function based on the predicted SOH sequence obtained in step (5) and the normalized SOH value sequence in the validation set obtained in step (3); (7) Setting the range of the number of hidden neurons, the range of the number of iterations, and the range of the learning rate of the Bi-LSTM model, setting the internal parameters of the crown porcupine optimization algorithm CPO, and generating an initial population using the Chebyshev chaos mapping technology according to the set range of the number of hidden neurons, the range of the number of iterations, the range of the learning rate, and the internal parameters of the CPO algorithm; (8) The number of hidden neurons, the number of iterations and the learning rate in the initial population obtained in step (7) are used as the search starting point of the crown porcupine optimization algorithm CPO, and the loss function obtained in step (6) is used as the objective function of the CPO algorithm to perform iterative processing of the CPO algorithm. After each iteration of the CPO algorithm, a random difference mutation strategy is used to mutate the population in the CPO algorithm, and finally the hidden neurons, the maximum number of iterations and the learning rate that minimize the loss value are obtained as the optimal solution, and the Bi-LSTM model corresponding to the optimal solution is used as the final trained Bi-LSTM model.

2. According to claim 1, the lithium battery SOH estimation model training method based on improved CPO optimized Bi-LSTM is characterized in that: In step (1), multiple initial feature sequences of lithium batteries are obtained, and all initial feature sequences are normalized to obtain a feature sequence set. This process includes the following sub-steps: (1-1) Obtaining the time interval for the voltage of the lithium battery to change from 3.8V to 4.0V during the constant current charging stage of the first charge and discharge cycle, the time interval for the voltage to change from 3.8V to 4.0V during the constant current charging stage of the second charge and discharge cycle, ..., the time interval for the voltage to change from 3.8V to 4.0V during the constant current charging stage of the Nth charge and discharge cycle, all of which constitute a first characteristic sequence; (1-2) obtaining the time interval for the voltage of the lithium battery to change from 4.0 V to 4.2 V during the constant current charging stage of the first charge and discharge cycle, the time interval for the voltage to change from 4.0 V to 4.2 V during the constant current charging stage of the second charge and discharge cycle, ..., the time interval for the voltage to change from 4.0 V to 4.2 V during the constant current charging stage of the Nth charge and discharge cycle, and all time intervals constitute a second characteristic sequence; (1-3) obtaining the time interval when the current of the lithium battery changes from 1.0 A to 0.5 A during the constant voltage charging stage of the first charge and discharge cycle, the time interval when the current changes from 1.0 A to 0.5 A during the constant voltage charging stage of the second charge and discharge cycle, ..., the time interval when the current changes from 1.0 A to 0.5 A during the constant voltage charging stage of the Nth charge and discharge cycle, and all time intervals constitute a third characteristic sequence; (1-4) obtaining the time interval for the voltage of the lithium battery to change from 4.0 V to 3.6 V during the constant current discharge stage of the first charge and discharge cycle, the time interval for the voltage to change from 4.0 V to 3.6 V during the constant current discharge stage of the second charge and discharge cycle, ..., the time interval for the voltage to change from 4.0 V to 3.6 V during the constant current discharge stage of the Nth charge and discharge cycle, and all time intervals constitute a fourth characteristic sequence; (1-5) obtaining the time interval for the voltage of the lithium battery to change from 3.6 V to 3.2 V during the constant current discharge stage of the first charge and discharge cycle, the time interval for the voltage to change from 3.6 V to 3.2 V during the constant current discharge stage of the second charge and discharge cycle, ..., the time interval for the voltage to change from 3.6 V to 3.2 V during the constant current discharge stage of the Nth charge and discharge cycle, and all time intervals constitute a fifth characteristic sequence; (1-6) Obtain the total time spent by the lithium battery in the constant current discharge stage of the first charge and discharge cycle from the start of constant current discharge to the time when the temperature of the lithium battery reaches the highest, the total time spent in the constant current discharge stage of the second charge and discharge cycle from the start of constant current discharge to the time when the temperature of the lithium battery reaches the highest, ..., the total time spent in the constant current discharge stage of the Nth charge and discharge cycle from the start of constant current discharge to the time when the temperature of the lithium battery reaches the highest, and all total times constitute the sixth characteristic sequence; (1-7) Obtain the capacity increment IC curve of the lithium battery in the constant current discharge stage of the first charge and discharge cycle, and obtain the IC curve The peak value and the corresponding discharge voltage value are obtained to obtain the IC curve of the lithium battery in the constant current discharge stage of the second charge and discharge cycle, and the The peak value and the corresponding discharge voltage value, ..., obtain the IC curve of the lithium battery in the constant current discharge stage of the Nth charge and discharge cycle, and obtain the IC curve The peak values ​​and their corresponding discharge voltage values, all peak values ​​constitute the seventh characteristic sequence, and all voltage values ​​constitute the eighth characteristic sequence, where the abscissa of the IC curve is the discharge voltage value and the ordinate is That is, the differential of the discharge capacity of the lithium battery to the discharge voltage; (1-8) Normalize the first to eighth feature sequences obtained in steps (1-1) to (1-7), and combine the normalized first to eighth feature sequences into a feature sequence set.

3. The lithium battery SOH estimation model training method based on improved CPO optimized Bi-LSTM according to claim 2 is characterized in that: The calculation formula of IC curve in step (1-7) is: Among them, Q, V, I, and T are the discharge capacity, discharge voltage, discharge current, and discharge time of the lithium battery respectively.

4. The lithium battery SOH estimation model training method based on improved CPO optimized Bi-LSTM according to claim 3 is characterized in that: The normalization process in steps (1-8) is performed using the following formula: Where x represents any element in the feature sequence, x' represents the normalized value of element x, and x min Represents the minimum value of all elements in the characteristic sequence where element x is located, x max Represents the maximum value of all elements in the feature sequence where element x is located.

5. The lithium battery SOH estimation model training method based on improved CPO optimized Bi-LSTM according to claim 4 is characterized in that: The association analysis method used in step (2) is the Spearman correlation analysis method; The absolute value of the preset threshold in step (2) ranges from 0.8 to 1.

0.

6. The lithium battery SOH estimation model training method based on improved CPO optimized Bi-LSTM according to claim 5 is characterized in that: The loss function used in step (6) is the root mean square error RMSE function; In step (7), the number of hidden neurons ranges from 50 to 150, the number of iterations ranges from 50 to 150, and the learning rate ranges from 0.001 to 0.01; The internal parameters of the CPO algorithm in step (7) are set as follows: population size N = 100, maximum number of iterations = 100, optimization dimension = 3, minimum number of individuals in the population Nmin = 80, number of population reduction cycles T = 2, and convergence speed factor α = 0.

1.

7. The lithium battery SOH estimation model training method based on improved CPO optimized Bi-LSTM according to claim 6 is characterized in that: The initial population X obtained in step (7) is: Where X N represents the Nth candidate solution in the initial population X, x i,j represents the jth position of the i-th candidate solution, and j∈[1,d], d represents the size of the optimization dimension, which is equal to 3; The i-th candidate solution X in the initial population X i equal: Where L represents the number of hidden neurons, the number of iterations, and the lower limit of the learning rate, U represents the number of hidden neurons, the number of iterations, and the upper limit of the learning rate, r is a random number between 0 and 1, and: x k+1 =cos(kcos -1 (x k )) Where k is the order, its value is 4, and x0 is a random number between -1 and 1.

8. The lithium battery SOH estimation model training method based on improved CPO optimized Bi-LSTM according to claim 7 is characterized in that: The formula for the random difference mutation strategy is as follows: Y(t+1)=r1(Y * -Y(t))+r2(Y′(t)-Y(t)) Among them, Y(t+1) represents the individual position obtained at the t+1th iteration, r1 and r2 are random numbers between 0 and 1, t is the current iteration number, and Y * is the optimal individual position obtained from the 1st iteration to the tth iteration, and Y′(t) is the individual randomly selected from the population at the tth iteration.

9. A lithium battery SOH estimation model training system based on improved CPO optimized Bi-LSTM, characterized in that: include: The first module is used to obtain multiple initial feature sequences of the lithium battery, and normalize each of the obtained initial feature sequences and the SOH value sequence composed of multiple SOH values ​​of the lithium battery to obtain a normalized feature sequence and a normalized SOH value sequence. All normalized feature sequences constitute a feature sequence set, wherein the length of each initial feature sequence is the number of charge and discharge cycles N of the lithium battery; The second module is used to calculate the correlation between each feature sequence in the feature sequence set obtained by the first module and the normalized SOH value sequence by using the correlation analysis method, and to filter out the feature sequences with correlation lower than a preset threshold from the feature sequence set, thereby obtaining multiple updated feature sequences; The third module is used to combine the multiple updated feature sequences obtained by the second module and the normalized SOH value sequence obtained by the first module into a data set, and divide the data set into a training set, a validation set and a test set according to a predetermined ratio; The fourth module is used to train the Bi-LSTM model using all feature sequences and normalized SOH value sequences in the training set to obtain a preliminarily trained Bi-LSTM model; The fifth module is used to input all feature sequences in the validation set into the Bi-LSTM model preliminarily trained in the fourth module to obtain a predicted SOH value sequence consisting of multiple predicted SOH values, whose length is equal to the length of each feature sequence in the validation set; The sixth module is used to calculate the loss value using the loss function according to the predicted SOH sequence obtained in the fifth module and the normalized SOH value sequence in the validation set obtained in the third module; The seventh module is used to set the range of the number of hidden neurons, the range of the number of iterations, and the range of the learning rate of the Bi-LSTM model, set the internal parameters of the crown porcupine optimization algorithm CPO, and generate the initial population using the Chebyshev chaos mapping technology according to the set range of the number of hidden neurons, the range of the number of iterations, the range of the learning rate, and the internal parameters of the CPO algorithm; The eighth module is used to use the number of hidden neurons, the number of iterations and the learning rate in the initial population obtained in the seventh module as the search starting point of the crown porcupine optimization algorithm CPO, and use the loss function obtained in the sixth module as the objective function of the CPO algorithm to iterate the CPO algorithm. After each iteration of the CPO algorithm, a random difference mutation strategy is used to mutate the population in the CPO algorithm, and finally the hidden neurons, the maximum number of iterations and the learning rate that minimize the loss value are obtained as the optimal solution, and the Bi-LSTM model corresponding to the optimal solution is used as the final trained Bi-LSTM model.

Images