An intelligent real-time prediction method for battery SOH

Abstract

The present invention provides an intelligent method for real-time prediction of battery SOH, belonging to the field of battery health status prediction. The content includes: obtaining historical data of battery health status; setting parameters required for the ISCO-GRU-OL dynamic model; optimizing the hyperparameters of the gated recurrent unit neural network based on the improved shark cooperation optimization algorithm and historical data of battery health status; establishing and training the ISCO-GRU model based on the optimized hyperparameters in combination with historical battery health status data; performing parameter fine-tuning on the ISCO-GRU model based on the online learning module according to the obtained latest battery health status data and online learning frequency, so as to update the ISCO-GRU model; predicting future battery health status data based on the updated ISCO-GRU model. The present invention can accurately perform real-time prediction of SOH, providing a basis for battery health management and prediction of remaining service life.

Description

Technical Field

[0001] The present invention relates to the field of battery health status prediction, and more particularly, to an intelligent real-time prediction method for battery SOH. Background Art

[0002] The State of Health (SOH) of a battery is a key indicator for evaluating battery performance and life, reflecting the ratio of the current available capacity of the battery to its rated capacity. Accurately evaluating SOH is crucial for the effective operation of the battery management system, battery life prediction, and estimation of the remaining useful life, which will directly affect the reliability and safety of the battery. There are many challenges in existing SOH prediction methods. On the one hand, there is a contradiction between model accuracy and real-time performance. Many high-precision SOH prediction models have high computational complexity and are difficult to meet the requirements of real-time prediction. On the other hand, although machine learning-based data-driven methods can fit complex non-linear relationships, their generalization ability and robustness to noisy data need to be further improved. Overfitting and insufficient model training data are common problems.

[0003] Therefore, how to establish a dynamic real-time prediction model suitable for battery SOH prediction, while improving model accuracy, real-time performance, and robustness, and minimizing the computational cost to the greatest extent, is crucial for high-precision and high-efficiency battery SOH prediction. Summary of the Invention

[0004] The present invention proposes an intelligent real-time prediction method for battery SOH, which combines an improved shark swarm optimization algorithm (ISCO), a gated recurrent unit (GRU) neural network, and an online learning (OL) strategy. Through the strong local convergence ability and global search ability of the ISCO algorithm, the efficient processing ability of the GRU network for time series data, and the ability of the online learning strategy to continuously update model parameters, the battery health status (SOH) is dynamically and real-time predicted. The purpose of the present invention is to improve the accuracy, real-time performance, and robustness of battery SOH prediction, and provide reliable technical support for dynamic real-time prediction of battery SOH.

[0005] The present invention is implemented as follows:

[0006] The technical solution for achieving the object of the present invention is: an intelligent real-time prediction method for battery SOH, comprising the following steps:

[0007] Step 1: Obtain historical data of the battery health status (SOH);

[0008] Step 2: Set the parameters required for the ISCO-GRU-OL dynamic model;

[0009] The ISCO-GRU-OL dynamic model consists of an ISCO-GRU model and an online learning module; the required parameters include the parameters required for the improved shark cooperation optimization algorithm, the parameters required for building the ISCO-GRU model, and the parameters required for the online learning module;

[0010] Step 3: Based on the improved shark cooperation optimization algorithm and the historical data of the battery health state, adopt an optimization strategy based on sliding window transfer training, and use the root mean square error of the predicted value of the gated recurrent unit neural network under sliding window transfer training as the fitness function to optimize the hyperparameters of the gated recurrent unit neural network;

[0011] Step 4: Based on the optimized hyperparameters, establish and train an ISCO-GRU model in combination with the historical battery health state data;

[0012] Step 5: According to the obtained latest battery health state data and the online learning frequency F OL , based on the online learning module, perform parameter fine-tuning on the ISCO-GRU model to obtain the ISCO-GRU model after parameter fine-tuning;

[0013] Step 6: Based on the latest ISCO-GRU model and the latest battery health state data, predict the future battery health state data.

[0014] Further, in Step 2, the parameters required for the improved shark cooperation optimization algorithm include the total number of individuals N, the proportion of whitetip reef shark individuals in the total individuals P1 , the proportion of shark leaders in the total individuals P2 , the maximum number of iterations T ; the parameters required for the ISCO-GRU model include the training window size of the ISCO-GRU model W 1 , the training iteration number of the ISCO-GRU model EPO 1 , the number of sliding window transfer training times WN , the upper limit of the hyperparameter optimization range ub and the lower limit lb ; the parameters required for the online learning module include the online learning frequency F OL , the online learning window size W 2 .

[0015] Further, in Step 3, the improved shark cooperation optimization algorithm includes the cooperation of two different types of individuals, whitetip reef sharks and grey reef sharks, and its search steps include four stages: (1) initialization; (2) team hunting; (3) individual tracking; (4) local reinforcement and global exploration;

[0016] (1)In the initialization stage, randomly generate N 1 whitetip reef sharks, N 1 The expression is:

[0017] N 1 = round( N × P1 )

[0018] where N represents the total number of individuals, P1 represents the proportion of whitetip reef shark individuals in the total individuals, and the round() function represents rounding the value inside the parentheses to an integer;

[0019] The initial position of each whitetip reef shark is calculated as follows:

[0020] M i =lb + ( ub - lb ) × rand(0, 1)

[0021] where, M i represents the position of the i th whitetip reef shark individual, i = 1, 2, … N1, lb is the lower limit of the search domain, ub is the upper limit of the search domain, M i , lb , ub are all D dimensional vectors, where D represents the dimension of the problem to be solved, and rand(0, 1) represents a D dimensional vector composed of random numbers between 0 and 1;

[0022] (2)In the team hunting stage, it includes three sub - stages: (i) electing a leader; (ii) generating grey reef sharks and forming a team; (iii) team hunting;

[0023] (i)In the sub - stage of electing a leader, calculate the fitness values of N 1 whitetip reef sharks, and select the P whitetip reef sharks with the best fitness values as the leader:

[0024] P = round( N × P2 )

[0025] where, P2Represents the proportion of shark leaders in the total individuals;

[0026] (ii) Generate grey reef sharks and form a team sub-phase. Around each leader, S i grey reef sharks are generated. The leader and the grey reef sharks generated around the leader form a leader team. The number of grey reef sharks depends on the fitness value of the leader and the total number of grey reef sharks. When the i th whitetip reef shark is the leader, the number of grey reef sharks generated S i is:

[0027] S i = round((1 / F(M i ) / F total × N ×(1 - P1 ))

[0028] where F total is the reciprocal sum of the fitness values of all leader whitetip reef sharks, as follows: , when i belongs to the leader;

[0029] where F () is the fitness function;

[0030] After all leaders generate grey reef sharks, judge whether the actual total number of generated grey reef sharks ∑ S i is consistent with the expected total number of generated grey reef sharks 1 - P1 . If:

[0031] ∑ S i > 1 - P1 , then increase the number of grey reef sharks generated by the leader with the optimal fitness. The increased number is ∑ S i - (1 - P1 );

[0032] ∑ S i = 1 - P1 , then the number of grey reef sharks remains unchanged;

[0033] ∑ S i < 1 - P1 , then reduce the number of grey reef sharks generated by the leader with the worst fitness. The reduced number is (1 -P1 ) - ∑ S i , if the number of grey reef sharks generated by the leader with the worst fitness is less than (1 - P1 ) - ∑ S i , then continue to reduce the number of grey reef sharks generated by the leader with the second-worst fitness until ∑ S i = 1 - P1 ;

[0034] The position of the generated grey reef shark is calculated as follows:

[0035] M ij = M i + 1 / T ×( ub - lb )×(1 - t / T ) j = 1, 2…, S i

[0036] where M ij represents the position of the i th j grey reef shark generated by the leader; T is the maximum number of iterations, t is the current iteration number; if the position of the grey reef shark exceeds the search space boundary in any dimension, the value of that dimension will be adjusted to the closest boundary value;

[0037] (iii) Team hunting sub-phase, calculate the fitness value of each grey reef shark and update the position of the whitetip reef shark to the position of the individual with the optimal fitness value in the team where the whitetip reef shark is located;

[0038] (3) In the individual tracking phase, it includes three sub-phases: (i) Select the target leader; (ii) Calculate the expected update point; (iii) Position update;

[0039] (i) Select the target leader sub-phase, each whitetip reef shark will select a specific leader to follow, and the probability

[0040] R ( i ) is expressed as: , when i belongs to the leader;

[0041] where, e is the base of the natural logarithm;

[0042] (ii) Calculating the expected update point sub - stage, the expected update point of the whitetip reef shark is determined by three factors: the individual following term, the global following term, and the inertia term. If the whitetip reef shark j chooses to target the leader i , then its expected update point M pj is:

[0043] M pj =M j +AG j × ( dk1 ) j +(1 - AG j ) × ( dk2 ) j + ( dk3 ) j

[0044] Wherein, M j represents the position of the whitetip reef shark j , (dk1) j is the target tracking term, (dk2) j is the global tracking term, and (dk3) j is the inertia term, AG j is the iteration factor, and the expressions are respectively:

[0045] ( dk1 ) j = S max × F ij × AG j × ( M i -M j )

[0046] ( dk2 ) j = S max × F jbest × AG j ) × ( M i -M j )

[0047] ( dk3 ) j =0.2×( AG j ×( dk1 ) j +( 1 - AG j ) ×( dk2 ) j ) × AG j

[0048] AG j =e (-t / T)

[0049] where F ij is the target tracking influence term, F jbest is the global tracking influence term, S max is the maximum step size, and the expressions are respectively:

[0050] F ij = 1 - ( F(M j ) - F(M i ) ) / ( F(M) max -F(M) min )

[0051] F ibest = 1 - ( F(M j ) - F(M) min ) / ( F(M) max -F(M) min )

[0052] S max = ( ub - lb ) / ( N × T ) (1 / 2)

[0053] where F(M) max andF(M) min respectively represent the optimal fitness value and the worst fitness value among all whitetip reef sharks;

[0054] If the expected update point exceeds the search space boundary, the value of this dimension will be adjusted to the closest boundary value;

[0055] (iii) Position update sub-phase. First, evaluate the fitness value of the expected update point of each whitetip reef shark. Second, for non-leader whitetip reef sharks, directly update their positions to the expected update points. Third, for leader whitetip reef sharks, their positions will only be updated to the expected update points when the fitness value of the expected update point position is better than the fitness value of the current position;

[0056] (4) In the local intensification and global exploration phase, for the whitetip reef shark with the worst fitness, its position is updated to the position of the second-best individual in the team where the leader whitetip reef shark with the best fitness is located; for the whitetip reef shark with the second-worst fitness, a new position is regenerated within the search domain according to the method in the initialization phase;

[0057] After each iteration, record the optimal position and fitness value, delete all grey reef shark individuals, and evaluate the termination criterion. If the criterion is met, the calculation terminates, and the optimal parameters and optimal fitness value are output. Otherwise, repeat calculation phases (2) to (4).

[0058] Further, in step three, the gated recurrent unit neural network includes an input layer, a GRU layer with L2 regularization, a dropout layer, a Dense fully connected layer, and a regression layer using the Adam optimizer; the hyperparameters include the number of units in the gated recurrent unit model, the learning rate, the Dropout rate, and the L2 regularization strength.

[0059] Further, in step three, the optimization strategy of the sliding window transfer training means that when optimizing the hyperparameters, for each given set of hyperparameters, training and prediction are performed in the sliding window transfer manner, and WN times of training and prediction are carried out, and after WN times of sliding window transfer training are completed, the root mean square error is obtained from the WN times of predicted values and the true values, and the root mean square error is used as the fitness function required for optimizing the hyperparameters, where WN is the number of sliding window transfer training times.

[0060] Further, in step five, the online learning module means that for every F OL latest battery state data obtained, using the W 2The previous data is the training input, and the latest data is the training output. Further training the ISCO-GRU model realizes parameter fine-tuning, where W 2 is the online learning window size.

[0061] The beneficial effects of the present invention are as follows: The present invention provides an intelligent method for real-time prediction of battery SOH. Among them, the improved shark cooperative optimization algorithm (ISCO) has fewer parameters, stronger local search ability, and faster convergence speed than the original SCO algorithm; the ISCO-GRU model can fully learn the historical state information of battery SOH to obtain an accurate prediction model; online learning (OL) can fine-tune the ISCO-GRU model according to the latest battery SOH data and correct the prediction accuracy. The present invention can provide a basis for accurately predicting SOH in real time and for predicting battery health management and remaining service life. Description of the Drawings

[0062] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the drawings required for the embodiments. It should be understood that the following drawings only show some embodiments of the present invention, so they should not be regarded as limiting the scope. For those of ordinary skill in the art, without creative efforts, other related drawings can also be obtained based on these drawings.

[0063] Figure 1 is a flowchart of an intelligent method for real-time prediction of battery SOH provided by an embodiment of the present invention;

[0064] Figure 2 is a calculation flowchart of the improved shark cooperative optimization algorithm (ISCO) provided by an embodiment of the present invention;

[0065] Figure 3 is a comparison chart of the prediction results of the ISCO-GRU-OL method and the comparative method provided by an embodiment of the present invention; Detailed Embodiments

[0066] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the following will clearly and completely describe the technical solutions in the embodiments of the present invention in combination with the implementation cases and drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments in the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention. Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the claimed present invention, but merely represents the selected embodiments of the present invention. Based on the embodiments in the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention.

[0067] The following takes a specific case of real-time prediction of landslide displacement as an example to illustrate the method of the present invention.

[0068] Such as Figure 1 , an intelligent real-time prediction method for battery SOH, comprising the following steps:

[0069] Step 1: Obtain historical data of battery state of health (SOH);

[0070] Step 2: Set the parameters required for the ISCO-GRU-OL dynamic model;

[0071] The ISCO-GRU-OL dynamic model is composed of an ISCO-GRU model and an online learning module; the required parameters include the parameters required for the improved shark cooperation optimization algorithm, the parameters required for constructing the ISCO-GRU model, and the parameters required for the online learning module;

[0072] Step 3: Based on the improved shark cooperation optimization algorithm and the historical data of battery state of health, adopt an optimization strategy based on sliding window rotation training, and use the root mean square error of the prediction value of the gated recurrent unit neural network under sliding window rotation training as the fitness function to optimize the hyperparameters of the gated recurrent unit neural network;

[0073] Step 4: Based on the optimized hyperparameters, establish and train an ISCO-GRU model in combination with the historical battery state of health data;

[0074] Step 5: According to the obtained latest battery state of health data and the online learning frequency F OL , based on the online learning module, perform parameter fine-tuning on the ISCO-GRU model to obtain an ISCO-GRU model with fine-tuned parameters;

[0075] Step 6: Based on the latest ISCO-GRU model and the latest battery state of health data, predict the future battery state of health data.

[0076] Further, in step 2, the parameters required for the improved shark cooperation optimization algorithm include the total number of individuals N, the proportion of whitetip reef shark individuals in the total individuals P1 , the proportion of shark leaders in the total individuals P2 , the maximum number of iterations T ; the parameters required for the ISCO-GRU model include the training window size of the ISCO-GRU model W 1 , the number of training iterations of the ISCO-GRU model EPO 1 , the number of times of sliding window transfer training WN , the upper limit of the hyperparameter optimization range ub and the lower limit lb ; the parameters required for the online learning module include the online learning frequency F OL , the online learning window size W 2 .

[0077] Further, in step 3, the improved shark cooperation optimization algorithm involves the cooperation of two different types of individuals, whitetip reef sharks and grey reef sharks. Its search steps include four stages: (1) initialization; (2) team hunting; (3) individual tracking; (4) local intensification and global exploration;

[0078] (1) In the initialization stage, randomly generate N 1 whitetip reef sharks, N 1 The expression is:

[0079] N 1 = round( N × P1 )

[0080] where N represents the total number of individuals, P1 represents the proportion of whitetip reef shark individuals in the total individuals, and the round() function represents rounding the value inside the parentheses to an integer;

[0081] The initial position of each whitetip reef shark is calculated as follows:

[0082] M i =lb + ( ub - lb ) × rand(0,1)

[0083] where, M i represents thei The positions of i N1 white-tip reef sharks, lb where ub is the lower limit of the search domain, M i , lb , ub are all D d-dimensional vectors, where D represents the dimension of the problem to be solved, and rand(0,1) represents a D d-dimensional vector composed of random numbers between 0 and 1;

[0084] (2) In the team hunting stage, it includes three sub-stages: (i) electing a leader; (ii) generating grey reef sharks and forming a team; (iii) team hunting;

[0085] (i) In the sub-stage of electing a leader, calculate the N 1 fitness values of P N1 white-tip reef sharks, and select the

[0086] P with the best fitness value as the leader: N × P2 )

[0087] where P2 represents the proportion of shark leaders in the total individuals;

[0088] (ii) In the sub-stage of generating grey reef sharks and forming a team, the number of grey reef sharks generated around each leader is S i , and a leader team is formed by the leader and the grey reef sharks generated around the leader. The number of grey reef sharks depends on the fitness value of the leader and the total number of grey reef sharks. When the i th white-tip reef shark is the leader, the number of grey reef sharks S i generated is:

[0089] S i = round((1 / F(M i ) / F total × N ×(1- P1 ))

[0090] where F total is the sum of the reciprocals of the fitness values of all leader white-tip reef sharks, as follows: , when iWhen it belongs to the leader;

[0091] Where F () is the fitness function;

[0092] After all leaders generate grey reef sharks, judge the actual total number of grey reef sharks generated ∑ S i And the expected total number of grey reef sharks generated 1 - P1 Whether they are consistent. If:

[0093] ∑ S i > 1 - P1 , then increase the number of grey reef sharks generated by the leader with the best fitness. The increased quantity is ∑ S i -(1 - P1 );

[0094] ∑ S i = 1 - P1 , then the number of grey reef sharks remains unchanged;

[0095] ∑ S i < 1 - P1 , then reduce the number of grey reef sharks generated by the leader with the worst fitness. The reduced quantity is (1 - P1 ) - ∑ S i . If the number of grey reef sharks generated by the leader with the worst fitness is less than (1 - P1 ) - ∑ S i , then continue to reduce the number of grey reef sharks generated by the leader with the second-worst fitness until ∑ S i = 1 - P1 ;

[0096] The position for generating grey reef sharks is calculated as follows:

[0097] M ij = M i + 1 / T ×( ub - lb )×(1 - t / T ) j = 1, 2…, S i

[0098] Where let M ijIndicates the leader i The j position of the th grey reef shark; T is the maximum number of iterations, t is the current iteration number; if the position of the grey reef shark exceeds the search space boundary in any dimension, the value of that dimension will be adjusted to the closest boundary value;

[0099] (iii) Team hunting sub-phase, calculate the fitness value of each grey reef shark, and update the position of the whitetip reef shark to the position of the individual with the optimal fitness value in the team where the whitetip reef shark is located;

[0100] (3) In the individual tracking phase, it includes three sub-phases: (i) Select the target leader; (ii) Calculate the expected update point; (iii) Update the position;

[0101] (i) Select the target leader sub-phase, each whitetip reef shark will select a specific leader to follow, and the probability that leader i is selected

[0102] R ( i ) is expressed as: , when i belongs to the leader;

[0103] Among them, e is the base of the natural logarithm;

[0104] (ii) Calculate the expected update point sub-phase, the expected update point of the whitetip reef shark is determined by three factors: the individual following term, the global following term, and the inertia term. If the whitetip reef shark j chooses to target leader i , then its expected update point M pj is:

[0105] M pj =M j +AG j ×( dk1 ) j +(1 - AG j ) ×( dk2 ) j + ( dk3 ) j

[0106] Among them, M j represents the position of the whitetip reef shark j , (dk1) jis the target tracking item, (dk2) j is the global tracking item, and (dk3) j is the inertial item, AG j is the iteration factor, and the expressions are respectively:

[0107] ( dk1 ) j = S max × F ij × AG j ×( M i -M j )

[0108] ( dk2 ) j = S max × F jbest × AG j )×( M i -M j )

[0109] ( dk3 ) j =0.2×( AG j ×( dk1 ) j +( 1 - AG j ) ×( dk2 ) j ) × AG j

[0110] AG j =e (-t / T)

[0111] where F ij is the target tracking influence item, F jbest is the global tracking influence item, S max is the maximum step size, and the expressions are respectively:

[0112] F ij =1 - ( F(M j ) - F(M i ) ) / ( F(M) max -F(M) min )

[0113] F ibest = 1 - ( F(M j ) - F(M) min ) / ( F(M) max -F(M) min )

[0114] S max = ( ub - lb ) / ( N × T ) (1 / 2)

[0115] where F(M) max and F(M) min represent the optimal fitness value and the worst fitness value among all whitetip reef sharks, respectively;

[0116] If the expected update point exceeds the search space boundary, the value of this dimension will be adjusted to the closest boundary value;

[0117] (iii) Position update sub - phase. First, evaluate the fitness value of the expected update point of each whitetip reef shark. Second, for non - leader whitetip reef sharks, directly update their positions to the expected update points. Third, for leader whitetip reef sharks, their positions will only be updated to the expected update points when the fitness value of the expected update point position is better than the fitness value of the current position;

[0118] In the local intensification and global exploration phase, for the whitetip reef shark with the worst fitness, its position is updated to the position of the second - best individual in the team where the leader whitetip reef shark with the best fitness is located; for the whitetip reef shark with the second - worst fitness, a new position is regenerated within the search domain according to the method in the initialization phase.

[0119] After each iteration, record the optimal position and fitness value, delete all grey reef shark individuals, and evaluate the termination criterion. If the criterion is met, the calculation terminates, and the optimal parameters and optimal fitness value are output; otherwise, repeat calculation phases (2) - (4).

[0120] Further, in step three, the gated recurrent unit neural network includes an input layer, a GRU layer with L2 regularization, a dropout layer, a Dense fully connected layer, and a regression layer using the Adam optimizer; the hyperparameters include the number of units in the gated recurrent unit model, the learning rate, the Dropout rate, and the L2 regularization strength.

[0121] Further, in step three, the optimization strategy of the sliding window transfer training means that when optimizing hyperparameters, for each given set of hyperparameters, training and prediction are performed in a sliding window transfer manner, WN times, and after WN times of sliding window transfer training are completed, the root mean square error is obtained from the WN times of predicted values and the true values, and the root mean square error is used as the fitness function required for optimizing hyperparameters, where WN is the number of sliding window transfer training times.

[0122] Further, in step five, the online learning module means that for every F OL latest battery state data obtained, using the W 2 data before the latest data as the training input and the latest data as the training output, the ISCO-GRU model is further trained to achieve parameter fine-tuning, where W 2 is the online learning window size.

[0123] The experimental data comes from the publicly available lithium battery SOH test data of NASA. 166 data are selected in the experiment, among which the first 75 are used as historical data to train the ISCO-GRU model, and the subsequent 91 data are gradually added to the historical data as the latest data, and the model is fine-tuned and predicted through online learning.

[0124] The specific implementation process is as follows:

[0125] As Figure 1 , an intelligent battery SOH real-time prediction method includes the following steps:

[0126] Step one: Obtain the historical data of the state of health (SOH) of the battery;

[0127] Step two: Set the parameters required for the ISCO-GRU-OL dynamic model;

[0128] The ISCO-GRU-OL dynamic model is composed of an ISCO-GRU model and an online learning module; the required parameters include the parameters required for the improved shark cooperation optimization algorithm, the parameters required for constructing the ISCO-GRU model, and the parameters required for the online learning module;

[0129] Among them, the total number of individuals N = 20, the proportion of whitetip reef sharks in the total individuals P1 = 0.7, the proportion of shark leaders in the total individuals P2 = 0.3, and the maximum number of iterations T = 20; the training window size of the ISCO-GRU model W1 = 70, the training iteration number of the ISCO-GRU model EPO1 = 500, the number of sliding window transfer trainings WN = 5, the upper limit of the hyperparameter optimization range ub = [200, 0.1, 0.5, 0.01], and the lower limit lb = [20, 1e-4, 0.0, 0.0]. The hyperparameters correspond to the number of units, learning rate, Dropout rate, and L2 regularization strength of the gated recurrent unit model respectively; the parameters required for the online learning module include the online learning frequency F OL = 1, and the online learning window size W2 = 50.

[0130] Step 3: Based on the improved shark cooperation optimization algorithm and the historical data of the battery health state, adopt an optimization strategy based on sliding window transfer training, and use the root mean square error of the predicted value of the gated recurrent unit neural network under sliding window transfer training as the fitness function to optimize the hyperparameters of the gated recurrent unit neural network;

[0131] The calculation process of the improved shark cooperation optimization algorithm is as Figure 2 shown.

[0132] Step 4: Based on the optimized hyperparameters, establish and train the ISCO-GRU model in combination with the historical battery health state data;

[0133] Step 5: According to the obtained latest battery health state data and the online learning frequency F OL , based on the online learning module, fine-tune the parameters of the ISCO-GRU model to obtain the ISCO-GRU model after parameter fine-tuning;

[0134] Step 6: Based on the latest ISCO-GRU model and the latest battery health state data, predict the future battery health state data.

[0135] To verify the advantages of the method of the present invention, the prediction results of the GRU model, the ISCO-GRU model, and the ISCO-GRU-OL model of the present invention were compared respectively. Among them, the number of units, learning rate, Dropout rate, and L2 regularization strength of the GRU model are 100, 0.01, 0.25, and 0.005 respectively, and other involved model parameters are the same.

[0136] Figure 3 The comparison of the prediction effects of the three models is shown. It can be seen from Figure 3 that the ISCO-GRU-OL model of the present invention has obvious advantages over the comparison models.

[0137] To further explore the performance of each model, four indicators, namely Mean Absolute Error (MAE), Mean Absolute Percentage Error (MAPE), Root Mean Square Error (RMSE), and coefficient of determination R 2 are used to quantitatively evaluate each model. Among them, the closer MAE, MAPE, and RMSE are to 0, the better the effect, and the closer R 2 is to 1, the more accurate the prediction. The evaluation results of each model are shown in Table 1:

[0138]

[0139] As shown in Table 1, the ISCO-GRU-OL model reached 0.005, 0.648, 0.008, and 0.989 respectively in the four indicators of MAE, MAPE, RMSE, and coefficient of determination R 2 , all significantly outperforming the comparative models, fully demonstrating the advantages of the ISCO-GRU-OL model in battery SOH prediction. The present invention can provide a basis for accurately predicting SOH in real time and for battery health management and remaining useful life prediction.

[0140] The above is only the preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention can have various changes and modifications. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. An intelligent real-time prediction method for battery SOH, characterized in that, It includes the following steps: Step 1: Obtain historical data on battery health status; Step 2: Set the parameters required for the ISCO-GRU-OL dynamic model; The ISCO-GRU-OL dynamic model consists of an ISCO-GRU model and an online learning module; the required parameters include the parameters required for the improved shark cooperation optimization algorithm, the parameters required for constructing the ISCO-GRU model, and the parameters required for the online learning module; Step 3: Based on the improved shark cooperation optimization algorithm and the historical data on battery health status, adopt an optimization strategy based on sliding window rotation training, and use the root mean square error of the predicted values of the gated recurrent unit neural network under sliding window rotation training as the fitness function to optimize the hyperparameters of the gated recurrent unit neural network; The improved shark cooperation optimization algorithm involves the cooperation of two different types of individuals, the whitetip reef shark and the grey reef shark, and its search steps include four stages: (1) initialization; (2) team hunting; (3) individual tracking; (4) local intensification and global exploration; (1) In the initialization stage, randomly generate N1 whitetip reef sharks, and the expression of N1 is: N1 = round(N × P1) where N represents the total number of individuals, P1 represents the proportion of whitetip reef shark individuals in the total individuals, and the round() function represents rounding the value in the parentheses to an integer; The initial position of each whitetip reef shark is calculated as follows: M i = lb+(ub - lb)×rand(0,1) Among them, M i represents the position of the i-th individual of the whitetip reef shark, where i = 1, 2, …, N1, lb is the lower bound of the search domain, ub is the upper bound of the search domain, and M i , lb, and ub are all D-dimensional vectors, where D represents the dimension of the problem to be solved, and rand(0, 1) represents a D-dimensional vector composed of random numbers between 0 and 1; (2) In the team hunting stage, it includes three sub-stages: (i) electing a leader; (ii) generating grey reef sharks and forming a team; (iii) team hunting; (i) In the sub-stage of electing a leader, calculate the fitness values of N1 whitetip reef sharks, and select the P whitetip reef sharks with the best fitness values as the leaders: P = round(N × P2) where P2 represents the proportion of shark leaders in the total individuals; (ii) Generate grey reef sharks and form a team sub-stage. The number of grey reef sharks generated around each leader is S i , and a leader team is formed by the leader and the grey reef sharks generated around the leader. The number of grey reef sharks depends on the fitness value of the leader and the total number of grey reef sharks. When the i-th whitetip reef shark is the leader, the number of grey reef sharks S i is as follows: S i = round((1 / F(M i ) / F total × N × (1 - P1)) where F total is the sum of the reciprocals of the fitness values of all the leader grey reef sharks, as follows: When i belongs to the leader; where F() is the fitness function; After all leaders generate grey reef sharks, judge the actual total number of grey reef sharks generated ∑S i Is it consistent with the expected total number of grey reef sharks generated 1 - P1? If: ∑S i If it is > 1 - P1, then increase the number of grey reef sharks generated by the leader with the optimal fitness. The increased number is ∑S i - (1 - P1); ∑S i = 1 - P1, then the number of grey reef sharks remains unchanged; ∑S i <If it is < 1 - P1, then reduce the number of grey reef sharks generated by the leader with the worst fitness. The reduced number is (1 - P1) - ∑S i , if the number of grey reef sharks generated by the leader with the worst fitness is less than (1 - P1) - ∑S i , then continue to reduce the number of grey reef sharks generated by the leader with the second-worst fitness until ∑S i = 1 - P1; The position of the generated grey reef shark is calculated as follows: M ij = M i + 1 / T × (ub - lb) × (1 - t / T) j = 1, 2…, S i Let M ij denote the position of the j-th grey reef shark generated by leader i; T is the maximum number of iterations, and t is the current number of iterations; if the position of the grey reef shark exceeds the search space boundary in any dimension, the value of that dimension will be adjusted to the closest boundary value; (iii) In the sub-stage of team hunting, calculate the fitness value of each grey reef shark, and update the position of the whitetip reef shark to the position of the individual with the best fitness value in the team where the whitetip reef shark is located; (3) In the individual tracking stage, it includes three sub-stages: (i) selecting a target leader; (ii) calculating the expected update point; (iii) position update; (i) In the sub-stage of selecting a target leader, each whitetip reef shark will choose a specific leader to follow, and the probability R(i) of leader i being selected is expressed as: When i belongs to the leader; where e is the base of the natural logarithm; (ii) In the expected update point calculation sub-phase, the expected update point of the whitetip reef shark is determined by three factors: the individual following term, the global following term, and the inertia term. If whitetip reef shark j chooses to target leader i, then its expected update point M pj is as follows: M pj = M j + AG j × (dk1) j +(1 - AG j ) × (dk2) j +(dk3) j Among them, M j represents the position of the whitetip reef shark j, (dk1) j is the target tracking term, (dk2) j is the global tracking term, while (dk3) j is the inertial term, AG j is the iteration factor, and the expressions are respectively: (dk1) j = S max × F ij × AG j × (M i - M j ) (dk2) j = S max × F jbest × AG j ) × (M i - M j ) (dk3) j = 0.2 × (AG j × (dk1) j + (1 - AG j ) × (dk2) j ) × AG j AG j = e (-t / T) where F ij is the target tracking influence term, F jbest is the global tracking influence term, S max is the maximum step size, and the expressions are respectively: F ij = 1 - (F(M j ) - F(M i )) / (F(M) max - F(M) min ) F ibest = 1 - (F(M j ) - F(M) min ) / (F(M) max - F(M) min ) S max = (ub - lb) / (N × T) (1 / 2) where F(M) max and F(M) min represent the optimal fitness value and the worst fitness value among all whitetip reef sharks, respectively; If the expected update point exceeds the search space boundary, the value of this dimension will be adjusted to the closest boundary value; (iii) In the position update sub-stage, first, evaluate the fitness value of the expected update point of each whitetip reef shark; second, for non-leader whitetip reef sharks, directly update their positions to the expected update point; third, for leader whitetip reef sharks, only when the fitness value of the expected update point position is better than the fitness value of the current position, will their positions be updated to the expected update point; (4) In the local reinforcement and global exploration phase, for the whitetip reef shark with the worst fitness, its position is updated to the position of the individual with the second-best fitness in the team of the leader whitetip reef shark with the best fitness; for the whitetip reef shark with the second-worst fitness, a new position is regenerated within the search domain according to the method in the initialization phase; After each iteration, record the optimal position and fitness value, delete all grey reef shark individuals, and evaluate the termination criterion. If the criterion is met, the calculation terminates, and the optimal parameters and optimal fitness value are output. Otherwise, repeat calculation steps (2) to (4); Step Four: Based on the optimized hyperparameters, establish and train the ISCO-GRU model in combination with historical battery health status data; Step 5: Based on the obtained latest battery health status data and the online learning frequency F OL , fine-tune the parameters of the ISCO-GRU model based on the online learning module to obtain the ISCO-GRU model with fine-tuned parameters; Step Six: Based on the latest ISCO-GRU model and the latest battery health status data, predict the future battery health status data.

2. The intelligent battery SOH real-time prediction method according to claim 1, wherein In Step 2, the parameters required for the improved shark cooperation optimization algorithm include the total number of individuals N, the proportion P1 of whitetip reef shark individuals in the total individuals, the proportion P2 of shark leaders in the total individuals, and the maximum number of iterations T; the parameters required for the ISCO-GRU model include the training window size W1 of the ISCO-GRU model, the training iteration number EPO1 of the ISCO-GRU model, the number of sliding window transfer training times WN, the upper limit ub and the lower limit lb of the hyperparameter optimization range; the parameters required for the online learning module include the online learning frequency F OL , and the online learning window size W2.

3. The intelligent battery SOH real-time prediction method according to claim 1, wherein In Step Three, the gated recurrent unit neural network includes an input layer, a GRU layer with L2 regularization, a dropout layer, a Dense fully connected layer, and a regression layer using the Adam optimizer; the hyperparameters include the number of units in the gated recurrent unit model, the learning rate, the Dropout rate, and the L2 regularization strength.

4. The intelligent battery SOH real-time prediction method according to claim 1, wherein In Step Three, the optimization strategy of sliding window rotation training means that when optimizing hyperparameters, for each given set of hyperparameters, WN times of training and prediction are carried out in the way of sliding window rotation. After WN times of sliding window rotation training are completed, the root mean square error is calculated from the WN times of predicted values and the true values, and the root mean square error is used as the fitness function required for optimizing hyperparameters, where WN is the number of sliding window rotation training times.

5. The intelligent battery SOH real-time prediction method according to claim 1, wherein In step five, the online learning module means that every time F OL latest battery state data are obtained, using the W2 data before the latest data as the training input and the latest data as the training output, to further train the ISCO-GRU model to achieve parameter fine-tuning, where W2 is the online learning window size.

Images